Selected papers on the evaluation of healthcare costs of prematurity and necrotizing enterocolitis using large retrospective databases

PDF

Download

Open document

Flip pages

Summarize

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content SELECTED PAPERS ON THE EVALUATION OF HEALTHCARE COSTS OF
PREMATURITY AND NECROTIZING ENTEROCOLITIS USING LARGE
RETROSPECTIVE DATABASES

By
VAIDYANATHAN GANAPATHY

A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALFIORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(PHARMACEUTICAL ECONOMICS AND POLICY)

MAY 2015

IN LOVING MEMORY OF

My beloved teacher Smt. Radha Krishna (Maami), whose memory is the guiding light in all my
endeavors.

My dearest grandparents Shri. K.G. Harihara Iyer, Smt. T.V. Rajamma and Smt. Saraswathi and
benevolent father-in-law Shri. T.S.Easwaran, who have showered bountiful love and blessings
for my academic pursuit and success.

2

DEDICATED TO
The enlightened,
In whose vision all are equal,
To whom no task is trivial,
And to whom limitations of mind and body cease to exist in the pursuit of their quest.

My loving, parents Smt. Lalitha Ganapathy and Shri. H. Ganapathy, wife Priya and mother-in-law, Smt.
Ramani Easwaran, who were great sources of strength and support during my research and studies.

SPECIAL DEDICATION TO
The many children around the world who are affected by extreme prematurity at birth and its
consequences.

3

ACKNOWLEDGEMENTS
The research works contained in this dissertation would not have been possible without the
intellectual support provided by the distinguished faculty members on my Dissertation and Proposal
Committees.
I owe my deepest respect and gratitude for my dissertation advisor, Professor Joel W. Hay, Ph.D.
(USC School of Pharmacy / Schaeffer Center for Health Policy & Economics) whose insightful
mentorship has greatly helped me to plan and conduct my doctoral research with academic rigor and
to explore novel methodological ideas in my dissertation.
I am very grateful to Professors, John A. Romley, Ph.D. (USC Sol Price School of Public Policy /
Schaeffer Center for Health Policy & Economics) and Neeraj Sood, Ph.D. (USC School of Pharmacy
/ Schaeffer Center for Health Policy & Economics) for providing critical feedback and suggestions
on the econometric methods described in Chapter 4 of my dissertation. I am extremely thankful to
Professor Jae H. Kim, M.D., Ph.D. (Department of Pediatrics, University of California, San Diego)
for providing his insights on the study designs from a clinician’s perspective, and for his thoughtful
feedback, to ensure that policy relevant questions were addressed in my dissertation. I sincerely thank
Professors, Sue Ingles, Ph.D. (Division of Epidemiology, USC Keck School of Medicine) and
Kathleen Johnson, Ph.D. (USC School of Pharmacy) for their support and feedback on my research.
I am highly grateful to Professor Jeffrey S. McCombs, Ph.D. (USC School of Pharmacy / Schaeffer
Center for Health Policy & Economics) for his mentorship and motivation throughout my graduate
studies.
I am extremely thankful to research scholar Fernando Rios Avila, Ph.D. (Levy Economics Institute at
Bard College) who provided practical guidance on implementing the Blinder-Oaxaca decomposition
methodology and my researcher friend Arvind Iyer, Ph.D. (Dept. of Biomedical Engineering, Visual
4

Neuroscience, USC) for reviewing and providing thoughtful feedback and comments on my
dissertation defense. I wholeheartedly thank Susan Ruffalo, Pharm.D., for her support with
formatting the final version of this dissertation.
Academic life, perhaps, would have been wearisome, without the vibrant and multi-faceted culture
that USC offered, and the reassuring company and support of near and dear ones. Passing difficulties
would have seemed insurmountable, if not for Priya’s (wife) positive and cheerful spirit, prayers of
my parents and the support of my brothers and extended family members. I thank my friends Raj,
Jenny, Arvind, Jai, Quang, Jaejin, Anand, Adam, Karthik, Ning, Sara and Flavia, for their great
company and encouragement.

5

Table of Contents
1. INTRODUCTION ...................................................................................................................... 16
2. INITIAL HOSPITALIZATION COSTS OF NECROTIZING ENTEROCOLITIS,
AND COST-EFFECTIVENESS OF EXCLUSIVELY HUMAN-MILK BASED
PRODUCTS IN FEEDING EXTREMELY PREMATURE INFANTS
,
............................... 17
2.1 ABSTRACT ................................................................................................................................ 17
2.2 BACKGROUND .......................................................................................................................... 19
2.3 OBJECTIVES .............................................................................................................................. 21
2.4 SUBJECTS AND METHODS ........................................................................................................... 21
2.4.1 Analysis of NEC Hospitalization Costs ........................................................................... 23
2.4.2 Sensitivity Analyses ....................................................................................................... 24
2.5 RESULTS .................................................................................................................................. 25
2.5.1 Costs of NEC .................................................................................................................. 25
2.5.2 Cost-Effectiveness of HUM Arm Versus BOV Arm Feeding Strategies .......................... 31
2.6 DISCUSSION ............................................................................................................................. 33
2.6.1 Economic Value of NEC Risk Reduction ........................................................................ 35
2.7 LIMITATIONS ............................................................................................................................ 36
2.8 CONCLUSIONS .......................................................................................................................... 37
2.9 REFERENCES ............................................................................................................................. 37
6

3. LONG TERM HEALTHCARE COSTS OF INFANTS WHO SURVIVED
NEONATAL NECROTIZING ENTEROCOLITIS: A RETROSPECTIVE
LONGITUDINAL STUDY AMONG INFANTS ENROLLED IN TEXAS
MEDICAID
,
................................................................................................................................ 41
3.1 ABSTRACT ................................................................................................................................ 41
3.2 BACKGROUND .......................................................................................................................... 43
3.3 OBJECTIVES .............................................................................................................................. 44
3.4 STUDY DESIGN AND METHODS .................................................................................................... 44
3.4.1 Sample Preparation and Propensity Score Matching ................................................... 45
3.5 OUTCOMES & STATISTICAL ANALYSES .......................................................................................... 48
3.6 RESULTS .................................................................................................................................. 49
3.6.1 Baseline Characteristics ................................................................................................ 49
3.6.2 Prevalence of chronic conditions in the medical and surgical NEC groups vs.
controls ...................................................................................................................................... 53
3.6.3 Univariate distribution of healthcare utilization and costs .......................................... 53
3.6.4 Multivariate generalized linear modeling results ......................................................... 57
3.7 DISCUSSION ............................................................................................................................. 61
3.8 LIMITATIONS ............................................................................................................................ 66
3.9 CONCLUSIONS .......................................................................................................................... 67
7

3.10 REFERENCES ............................................................................................................................. 67
4. APPLICATION OF COUNTERFACTUAL DECOMPOSITION METHODS WITH
FIXED EFFECTS TO EVALUATE THE IMPACT OF PREMATURITY ON
LONG-TERM HEALTHCARE COSTS OF CHILDREN BELONGING TO LOW-
INCOME FAMILIES ................................................................................................................ 73
4.1 ABSTRACT ................................................................................................................................ 73
4.2 INTRODUCTION ......................................................................................................................... 75
4.2.1 Background & Brief Review of Literature ..................................................................... 76
4.3 STUDY OBJECTIVES .................................................................................................................... 79
4.4 STUDY DESIGN & METHODS ....................................................................................................... 80
4.4.1 Description of data ....................................................................................................... 80
4.4.2 Identification and classification of preterm infants ...................................................... 81
4.4.3 Exclusions ...................................................................................................................... 81
4.4.4 Criteria for determining attrition .................................................................................. 82
4.4.5 Database set-up ............................................................................................................ 82
4.4.6 Outcomes ...................................................................................................................... 83
4.4.7 Explanatory variables ................................................................................................... 84
4.4.8 Estimation strategy ...................................................................................................... 88
4.4.9 Estimation procedures .................................................................................................. 93
4.4.10 Comparison of cross-sectional and panel data model decomposition estimates ...... 103
4.4.11 Calculation of costs using decomposition estimates ................................................. 103
8

4.4.12 Specification tests ....................................................................................................... 104
4.4.13 Tests for attrition bias................................................................................................. 105
4.4.14 Construction of attrition weights ................................................................................ 107
4.4.15 Sensitivity analyses ..................................................................................................... 108
4.4.16 Diagnostic drivers of healthcare cost differences between preterm and full-term
groups .................................................................................................................................... 109
4.4.17 Total excess financial impact due to preterm births between 6 and 60 months of
age .................................................................................................................................... 111
4.5 RESULTS ................................................................................................................................ 111
4.5.1 Demographic and clinical characteristics at the time of birth .................................... 111
4.5.2 Sample characteristics during study period ................................................................ 112
4.5.3 Distribution of healthcare costs .................................................................................. 114
4.5.4 Interaction test results for the applicability of decomposition methods .................... 117
4.5.5 Statistical test results for the validity of fixed effects ................................................. 118
4.5.6 Pooled OLS and fixed effects OLS estimation results .................................................. 120
4.5.7 Results of two-part model regressions ....................................................................... 125
4.5.8 Blinder-Oaxaca decomposition results ....................................................................... 131
4.5.9 Specification test results ............................................................................................. 138
4.5.10 Attrition test results .................................................................................................... 140
4.5.11 Sensitivity analyses results ......................................................................................... 156
4.5.12 Results for drivers of healthcare cost difference between preterm and full-term
groups .................................................................................................................................... 163
4.5.13 Results for total excess financial impact due to preterm births between 6 and 60
months .................................................................................................................................... 166
9

4.6 DISCUSSION ........................................................................................................................... 167
4.7 LIMITATIONS .......................................................................................................................... 173
4.8 CONCLUSIONS ........................................................................................................................ 176
4.9 REFERENCES ........................................................................................................................... 177

List of Tables
Table 2-1 Summary of Parameters Used in the Cost-Effectiveness Model Comparing the Two
Feeding Strategies—Mother’s Milk Fortified with Bovine Milk-Based Versus Human Milk-
Based Human Milk Fortifier—Among Very Low Birth Weight Infants (≤1,250 Grams) ................... 22
Table 2-2 Demographic and Baseline Characteristics of Extremely Premature Infants in the
Final Sample Cohort Derived From OHSPD 2007 Discharge Data ..................................................... 27
Table 2-3 Neonatal Intensive Care Unit (NICU) Days and Initial Average Hospitalization
Costs Among Extremely Premature Infants Without Necrotizing Enterocolitis in the California
OSHPD 2007 ........................................................................................................................................ 29
Table 2-4 Multivariate Adjusted Incremental Costs and Length of Stay for Medical and
Surgical Necrotizing Enterocolitis (NEC) in the OSHPD 2007 Hospital Discharge Sample .............. 30
Table 2-5 Estimation of Expected Neonatal Intensive Care Unit (NICU) Costs Among
Extremely Premature Infants Fed with Bovine Milk-Based Human Milk Fortifier (HMF)
Versus 100% Human Milk Diets (Base-Case Scenario) ............................................................... ...... 32
Table 3-1 Procedures and ICD-9-CM Codes Used to Define Infants with Surgical NEC ........ ...... 48
10

Table 3-2 Comparison of Baseline Characteristics Between the NEC and Control Samples
Before and After Propensity Score Matching ................................................................................ ...... 52
Table 3-3 Prevalence of Chronic Developmental Health Conditions in the NEC and Matched
Control Groups Between 6 and 12 and 24 and 36 Months of Follow-up ...................................... ...... 55
Table 3-4 Comparison of Healthcare Utilization and Cost Estimates (Unadjusted) for
Medical, Surgical NEC and Matched Control Groups from 6 Months to 3 Years of Age ............ ...... 56
Table 3-5 Predicted Costs per 6-Months for Surgical NEC Children Versus Matched Controls
Over Time Across Birth-Weights .................................................................................................. ...... 61
Table 4-1 Mean differences in log healthcare (HC) costs between preterm and full-term
groups during study period ............................................................................................................ .... 117
Table 4-2 Mundlak (1978) test for fixed versus random effects using the pooled sample ........ .... 119
Table 4-3 Mean values, coefficients and difference in coefficients of observed characteristics
between very/extremely preterm (< 33 weeks of GA) vs. full-term born children, obtained from
pooled OLS and fixed effects regressions of log healthcare costs ................................................ .... 123
Table 4-4 Mean values, coefficients and difference in coefficients of observed characteristics
between moderate/late preterm (33-36 weeks of GA) vs. full-term born children, obtained from
pooled OLS and fixed effects regressions of log healthcare costs ................................................ .... 124
Table 4-5 Differences in proportion of children with healthcare costs greater than zero in a
given period ................................................................................................................................... .... 125
11

Table 4-6 Panel regressions showing marginal effects of observed characteristics on
probability of healthcare use and their differences between very/extreme preterm and full-term
born children .................................................................................................................................. .... 127
Table 4-7 Panel regressions showing marginal effects of observed characteristics on
probability of healthcare use and their differences between moderate/late preterm and full-term
born children .................................................................................................................................. .... 128
Table 4-8 Means and fixed effects coefficients of characteristics predicting log cost given Yit
> 0 in the very/extreme preterm and full-term groups ................................................................... .... 130
Table 4-9 Means and fixed effects coefficients of characteristics predicting log cost given Yit
> 0 in the moderate/late preterm and full-term groups .................................................................. .... 131
Table 4-10 Decomposition of total healthcare cost difference between very/extreme preterm
(<33 weeks of GA) and full-term children .................................................................................... .... 132
Table 4-11 Decomposition of total healthcare cost difference between moderate/late preterm
(33-36 weeks of GA) and full-term children ................................................................................. .... 137
Table 4-12 Linear panel data specification tests ...................................................................... .... 139
Table 4-13 Comparison of the overall log healthcare cost difference between preterm and
full-term born children in the ‘full’ versus ‘completing’ samples ................................................. .... 142
Table 4-14 Seemingly unrelated estimation results evaluating difference in coefficients of
ln-OLS model between full and completing samples [Dependent variable = log of overall
healthcare costs] ............................................................................................................................ .... 144
12

Table 4-15 Comparison of FE decomposition estimates using the full versus ‘completing’
samples ............................................................................................................................... .... 147
Table 4-16 Cross-sectional probit regressions to predict the probability of attrition based on
demographic characteristics and lagged outcomes ........................................................................ .... 149
Table 4-17 Becketti, Gould, Lilard and Welch test for non-random attrition in the control
and preterm groups ........................................................................................................................ .... 151
Table 4-18 Comparison of unweighted and weighted fixed effects OLS estimates [Note:
Attrition weights for children retained in the sample by the end 18 months (wave 2) applied to
waves 3 and 4 as well] ................................................................................................................... .... 154
Table 4-19 Decomposition of total healthcare cost difference between preterm subgroups
and full-term children [Estimates are based on two-part models adjusted for attrition] .................... 159
Table 4-20 Decomposition of total healthcare cost difference between preterm subgroups
and full-term children, after adjusting for lagged chronic conditions [Estimates are based on
two-part models adjusted for attrition] ............................................................................................... 162
List of Figures
Figure 2-1 Preparation of the Final Analytic Sample from 2007 California OSHPD
Discharge Data ..................................................................................................................................... 26
Figure 2-2 One-Way Sensitivity Analysis of Cost Savings to a 10% Change in Model
Parameters* ...................................................................................................................................... 33
13

Figure 3-1 NEC and Control Group Sample Sizes (after matching) From 6 Months to 3
Years of Age ...................................................................................................................................... 51
Figure 3-2 Adjusted Incremental Total Healthcare Costs per 6-Months Incurred by Medical
and Surgical NEC Survivors Over Matched Controls from 6 Months to 3 Years of Age .................... 59
Figure 3-3 Adjusted Incremental Inpatient and Ambulatory Care Costs per 6-Months (± SE)
Incurred by Surgical NEC Children Over Matched Controls, 6 Months to 3 Years of Age ................ 60
Figure 4-1 Mean inpatient and ambulatory healthcare costs during study period by group ........ 115
Figure 4-2 Distribution of log of total healthcare costs over time ............................................... 116
Figure 4-3 Comparison of percent healthcare cost difference due to returns to prematurity
from the perspective of full-term (ATU) and children born before 33 weeks of GA (ATET) ........... 134
Figure 4-4 Comparison of percent healthcare cost difference due to returns to prematurity
among untreated (full-term) and treated (born preterm 33-36 weeks of GA) children ...................... 136
Figure 4-5 Cumulative retention (%) in the full-term and preterm groups .................................. 141
Figure 4-6 Unweighted and weighted TPM incremental cost (ATET) estimates, <33 weeks
of GA vs. full-term ............................................................................................................................. 157
Figure 4-7 Unweighted and weighted TPM incremental cost (ATET) estimates, 33-36
weeks of GA vs. full-term .................................................................................................................. 157
Figure 4-8 Unweighted and weighted incremental cost (ATET) estimates, preterm sub-
groups vs. full-term ............................................................................................................................ 160
14

Figure 4-9 Comparison of incremental cost (ATET) estimates with and without adjusting
for chronic disease characteristics (lagged), between preterm sub-groups vs. full-term.
[Estimates are based on two-part models adjusted for attrition] ........................................................ 163
Figure 4-10 Healthcare cost difference due to differences in frequencies of chronic
conditions between children born < 33 weeks of GA and full-term (Sample: observations with
healthcare costs > 0) ........................................................................................................................... 165
Figure 4-11 Healthcare cost difference due to differences in frequencies of chronic
conditions between children born 33-36 weeks of GA and full-term (Sample: observations with
healthcare costs > 0) ........................................................................................................................... 166

15

1. INTRODUCTION
This dissertation is a compilation of 3 observational real world studies that deal with evaluation
of the economic consequences of preterm births (Section 4) and that of a devastating
complication of extreme prematurity, called necrotizing enterocolitis (Sections 2 and 3). Section 2
has two-fold objectives – i) to estimate the incremental impact of necrotizing enterocolitis on
hospital length of stay and costs during the initial hospitalization period and ii) to evaluate the
cost-effectiveness of a 100% human-milk based breast milk fortification strategy to initiate
enteral nutrition among very low birth-weight infants. In section 3, I compare the long-term
health care costs and risks of various chronic health conditions among infants who survived
necrotizing enterocolitis, to that of matched cohort of infants without necrotizing enterocolitis.
The purpose of this study is mainly to understand the financial impact of necrotizing enterocolitis
over phases of early growth and developmental period. In the last section (Section 4), I discuss
the evaluation of real world medical costs among preterm survivors, classified by gestational age
categories, over the period between 6 months to 5 years of age. The purpose of this study is to
estimate the long-term healthcare costs among preterm survivors in the Texas Medicaid
population, which is witnessing one of the highest rate of preterm births in the United States. The
study was also motivated by the fact that there is lack of US specific data on the long term
healthcare cost of preterm births, by gestational age, and that, the few studies that were available,
had methodological concerns, such as, not accounting for unobserved individual specific
heterogeneity while evaluating outcomes using observational databases.
The research works related to evaluating the economic impact of necrotizing enterocolitis,
described in sections 2 and 3 of this dissertation, were funded by an unrestricted educational grant
provided by Prolacta Bioscience, City of Industry, California, to the USC School of Pharmacy.
16

2. INITIAL HOSPITALIZATION COSTS OF NECROTIZING
ENTEROCOLITIS, AND COST-EFFECTIVENESS OF
EXCLUSIVELY HUMAN-MILK BASED PRODUCTS IN FEEDING
EXTREMELY PREMATURE INFANTS
1,2

2.1 Abstract
Objective: This study evaluated the cost-effectiveness of a 100% human milk-based diet
composed of mother’s milk fortified with a donor human milk-based human milk fortifier (HMF)
versus mother’s milk fortified with bovine milk-based HMF to initiate enteral nutrition among
extremely premature infants in the neonatal intensive care unit (NICU).
Methods: A net expected costs calculator was developed to compare the total NICU costs among
extremely premature infants who were fed either a bovine milk-based HMF-fortified diet or a
100% human milk-based diet, based on the previously observed risks of overall necrotizing
enterocolitis (NEC) and surgical NEC in a randomized controlled study that compared outcomes
of these two feeding strategies among 207 very low birth weight infants. The average NICU costs
for an extremely premature infant without NEC and the incremental costs due to medical and
surgical NEC were derived from a separate analysis of hospital discharges in the state of
California in 2007. The sensitivity of cost-effectiveness results to the risks and costs of NEC and
to prices of milk supplements was studied.
1
Ganapathy V, Hay JW, Kim JH. 2012. Breastfeeding Medicine 7(1):29-37.
2
Research was funded by an unrestricted educational grant from Prolacta Bioscience, City of Industry, CA.
17

Results: The adjusted incremental costs of medical NEC and surgical NEC over and above the
average costs incurred for extremely premature infants without NEC, in 2011 US$, were $74,004
(95% confidence interval [CI], $47,051–$100,957) and $198,040 (95% CI, $159,261–$236,819)
per infant, respectively. Extremely premature infants fed with 100% human milk-based products
had lower expected NICU length of stay and total expected costs of hospitalization, resulting in
net direct savings of 3.9 NICU days and $8,167.17 (95% confidence interval, $4,405–$11,930)
per extremely premature infant (p<0.0001). Costs savings from the donor HMF strategy were
sensitive to price and quantity of donor HMF, percentage reduction in risk of overall NEC and
surgical NEC achieved, and incremental costs of surgical NEC.
Conclusions: Compared with feeding extremely premature infants with mother’s milk fortified
with bovine milk-based supplements, a 100% human milk-based diet that includes mother’s milk
fortified with donor human milk-based HMF may result in potential net savings on medical care
resources by preventing NEC.

18

2.2 Background
Infants who are born extremely premature (EP) at ≤28 weeks of gestational age have an immature
gastrointestinal system that puts them at higher risk for NEC, an acute inflammatory condition of
the bowel that affects 6–7% of all premature infants with birth weights of <1,250 grams and is
also one of the leading causes of mortality among these infants (Horbar et al 2007; Fanaroff et al
2003; Petrosyan et al 2009; Thompson and Bizzarro 2008). The American Academy of Pediatrics
recommends initiating mother’s own milk early during feeding of EP infants owing to the
important health benefits that mother’s milk has to offer, including protection against NEC
(American Academy of Pediatrics 1997; Lucas and Cole 1990; Schanler et al 1999, Sisk et al
2007). Very often, though, mother’s milk is not available in sufficient quantities to premature
infants, thereby limiting the availability of this beneficial milk source (Henderson et al 2008).
Donor human milk is an available proxy for mother’s milk. It is believed that donor milk might
be offering putative benefits of preventing NEC, by supplying immunoprotective factors to the
immature gut mucosa (Schanler 2007; Updegrove 2004). However, it may be possible that the
protective effect is due to the absence of harmful antigens rather than the presence of beneficial
agents. A randomized controlled study by Schanler et al., (2005) comparing donor human milk to
preterm formula, did not find a significant protective effect of donor human milk on NEC, but a
significant difference in the study’s protocol was that it involved fortification of donor milk with
bovine milk-based HMF. Nevertheless, the study also noted a protective effect of mother’s own
milk on NEC. This study may have been confounded by the presence of negative effects of
exposure to bovine milk products.
In another randomized controlled study using an intent-to-treat analysis, Sullivan et al. (2010)
compared outcomes of fortifying mother’s milk using a human milk-based HMF,
19

Prolact+H2MF™ (Prolacta Bioscience, Monrovia, CA), versus fortification using bovine milk-
based HMF among 207 very low birth weight infants (<1,250 grams) in the NICU. Thus far, this
is the only study that compared the outcomes of fortifying mother’s milk using bovine milk-based
HMF versus donor human milk-based HMF. The duration of study was the earliest among the
following milestones achieved: discharge of the infant, 91 days of age, or attainment of 50% oral
feedings per day. Infants in both the groups received mother’s own milk when enteral nutrition
was initiated. However, they differed in the type of HMF used for fortifying mother’s milk to
which they were randomized—receiving either bovine milk-based HMF (BOV arm) or human
milk-based HMF (HUM arm)—and the type of nutritional supplementation given when mother’s
milk was no longer available (viz., preterm formula in the BOV arm and pasteurized donor
human milk in the HUM arm). It was observed in that study that infants in the HUM arm (n =
138) had significantly lower incidence of overall NEC and surgical NEC than infants in the BOV
arm (n = 69): eight of 138 versus 11 of 69 (p = 0.02) and two of 138 versus seven of 69 (p =
0.007) in the HUM versus BOV arms for overall NEC and surgical NEC, respectively. After
controlling for confounding factors, infants in the HUM arm were found to have a 77% reduction
in the odds of developing NEC (odds ratio = 0.23, 95% CI 0.08, 0.66) compared with infants in
the BOV arm. Major strengths of the study of Sullivan et al. (2010) were that the study was
randomized, the stratification scheme achieved a balance of patient characteristics across the
study groups, and there was good adherence to the study protocol. The study was, however,
limited in power to examine subgroups by sex or birth weight, and its sample may or may not be
representative of the universe of NICUs across the United States. The study of Sullivan et al.,
(2010) however, suggests that beyond the positive biologic benefits of human milk, there may be
factors in bovine milk products that negatively affect the premature intestine.
20

2.3 Objectives
The current article analyzes the cost-effectiveness of using a 100% human milk-based diet that
involves fortifying mother’s milk with human milk-based HMF versus using a bovine milk-based
HMF-fortified diet to initiate enteral nutrition among EP infants in the NICU. An important
secondary objective of this research was to estimate the average NICU costs of medical and
surgical NEC among EP infants in a separate analysis using hospital discharge database analysis.
2.4 Subjects and Methods
A net expected cost calculator for initial NICU hospitalization expenses among EP infants was
developed in Microsoft (Redmond, WA) Excel 2003, to compare total expected NICU costs
among infants who either received a donor human milk-based HMF-fortified diet (HUM arm) or
a bovine milk-based HMF-fortified diet (BOV arm). The parameters used to develop this
calculator include the following: (1) probabilities of overall NEC and surgical NEC in each group
(HUM arm vs. BOV arm), obtained from the study of Sullivan et al., (2005); (2) quantities and
prices of HMFs used in each group; (3) quantities and prices of donor human milk and preterm
formula used (when mother’s milk was no longer available to infants) in the HUM and BOV
arms, respectively; and (4) incremental costs of medical and surgical NEC over and above
average NICU costs incurred for an EP infant without NEC. Table 2-1 provides a summary of the
parameters 1, 2, and 3 as obtained from the randomized trial study (Sullivan et al, 2010).

21

Table 2-1 Summary of Parameters Used in the Cost-Effectiveness Model
Comparing the Two Feeding Strategies—Mother’s Milk Fortified
with Bovine Milk-Based Versus Human Milk-Based Human Milk
Fortifier—Among Very Low Birth Weight Infants (≤1,250 Grams)
Constituents of the diet fed to infants in each group, and their quantities
Products
(median quantity)
BOV arm
(Mother’s milk fortified with
bovine milk based HMF)
N = 69
HUM arm
(Mother’s milk fortified with
donor human milk based HMF)
N = 138
Mother’s own milk 5676 ml 4432 ml
Human milk fortifier
1

(25
th
, 75
th
percentiles)
128 packets
(25, 277)
1650.8 ml.
(915, 2528)
Preterm formula
2
961 ml -
Donor human milk
3
- 39.5 ml
Probabilities of NEC
Incidence of overall NEC 0.16 0.06
(% reduction in risk of overall
NEC vs. BOV arm
4
: 63%)
Incidence of surgical
NEC
0.10 0.01
(% reduction in risk of surgical
NEC vs. BOV arm
5
: 86%)
Average prices of milk products used
Average price per unit Avg. price of bovine HMF
used: $1.30 per packet
Avg. price of preterm formula
used:
$0.03 per ml.
Avg. price of Prolact+H2MF™:
$6.25 per ml.
Avg. price of donor human milk
used: $0.10 / ml.
1. Enfamil/Similac bovine HMF packets were used in the BOV arm and Prolact+H2MF™ was used in
the HUM arm for fortifying mother’s milk.
2. 53 out of 69 infants in the BOV arm received some preterm formula when their mothers were unable
to provide sufficient breast milk.
3. 82 out of 138 infants in the HUM arm received some donor human milk when their mothers were
unable to provide sufficient breast milk.
4. p-value for the difference in incidence of NEC between BOV and HUM arms is 0.02.
5. p-value for the difference in incidence of surgical NEC between BOV and HUM arms is 0.007.
To estimate the average hospitalization costs among EP infants without NEC and the average
incremental costs of medical and surgical NEC cases, we conducted multivariate analyses on
hospital discharge data collected during the year 2007 in the state of California, obtained from the
22

Office of Statewide Health Planning & Development (OSHPD). Preparation of the analytic
sample for cost analysis, consisting of discharges from among EP infants without NEC versus
those among infants with medical NEC and with surgical NEC, from the OSHPD 2007 database,
is explained as follows.
2.4.1 Analysis of NEC Hospitalization Costs
EP infant discharges among the 2007 OSHPD data were identified by International Classification
of Diseases, 9
th
Revision (ICD-9) diagnosis codes 765.0 (765.21–765.24). Diagnosis of NEC was
identified using ICD-9 codes 777.50–777.53 and NEC records with one or more of the following
surgical procedures: Exploratory laparotomy, bowel resection, stoma creation, and intestinal
anastomosis (all identified by ICD-9 procedure codes [Guner et al, 2009]) were classified as
surgical NEC discharges to differentiate from NEC cases that did not require surgical
management. Discharges from hospitals that did not report charges (approximately 12% of
discharges, mostly belonging to Kaiser Health Foundation hospitals) and discharges involving
transfers (approximately 26%) were excluded from analysis. Hospital specific cost-to-charge
ratios were used to adjust charges to obtain actual costs. Discharges with an average daily cost of
<$100 (approximately 0.3%) were considered improbable for premature infant stays and were
excluded from analysis. As a final step in the data preparation, we excluded infants who died
within the first 3 days of life. These infants were probably those born at gestational ages below
the limit of viability, often dying because of causes other than NEC, and therefore not likely
candidates to receive HMF therapy. All cost estimates were adjusted for inflation to 2011 US$
using the Medical Component of the Consumer Price Index (U.S. Bureau of Labor Statistics
2010). Average estimates of length of stay and costs for an EP infant in the sample, with and
without the last exclusion criterion (mortality at <3 days), were compared. We used the final
sample that excluded infants who died within 3 days of life to obtain multivariate adjusted
23

estimates for incremental costs of medical NEC and surgical NEC. Ordinary Least Squares
regression was used to obtain the incremental effects of medical NEC and surgical NEC on
hospitalization costs after adjusting for sex, race, payer status (Medicaid vs. private or other),
mortality, and comorbid perinatal conditions including respiratory distress, other infections,
anemia, jaundice, and endocrine, hemorrhagic, integumental, and other ill-defined conditions. We
also developed multivariate models using the Generalized Estimating Equations framework
(Liang and Zeger 1986) in which sample observations were weighted by predicted probability of
NEC occurrence using the inverse probability weighting technique (Kurth et al 2006). Because
the model coefficients did not vary significantly among the Ordinary Least Squares regression
and weighted Generalized Estimating Equations models, we used the Ordinary Least Squares
regression adjusted coefficients for medical and surgical NEC in the net cost calculator
(explained earlier in this section) because of their direct interpretability as incremental costs over
and above the costs of an infant without NEC.
2.4.2 Sensitivity Analyses
The sensitivity of cost-effectiveness results was analyzed using one-way and two-way percentage
changes in the parameters that were used to build the expected costs calculator. The impact of the
following scenarios on cost-effectiveness were considered: (1) price and quantity of donor human
milk-based HMF increased linearly; (2) both bovine HMF and preterm formula cost zero dollars
to account for the possibility that manufacturers distribute free samples of bovine HMFs and
preterm formulas in hospitals; and (3) percentage risk reductions of overall NEC and surgical
NEC achieved by donor milk HMF strategy were less than observed for the base-case (two-way
change).
24

2.5 Results
2.5.1 Costs of NEC
There were 2,560 EP infants in the final analytic sample derived from 2007 OSHPD data (Figure
2-1). Two hundred fifty-nine (10%) of them had NEC, and 82 of these infants suffered from
surgical NEC (3.2%). Table 2-2 provides a summary of demographic and other baseline
characteristics of the final analytic sample used for NEC cost analysis. Race was significantly
associated with NEC status when missing data were included. The NEC groups had significantly
higher rates of comorbid perinatal conditions such as respiratory distress and other respiratory
conditions, infections, hemorrhage, and endocrine and other ill-defined perinatal conditions than
the no NEC group. Also, there was a significantly higher rate of mortality before discharge
among medical NEC and surgical NEC groups compared with the no NEC group (rates of 20.9%,
23.2%, and 10.6% among medical NEC, surgical NEC, and no NEC, respectively; p<0.0001).
The mean length of stay for an EP infant without NEC obtained from the final sample was 64.5
days, and the mean cost was $207,378 in 2011 US$ (Table 2-3). The multivariate adjusted
estimates for the length of stay and cost of medical NEC over and above the cost for an EP infant
without NEC are 11.7 days (95% CI 6.9–16.5 days) and $74,004 (95% CI $47,051–$100,957),
respectively. The corresponding adjusted incremental length of stay and cost estimate for surgical
NEC are 43.1 days (95% CI 36.3–50 days) and $198,040 (95% CI $159,261–$236,819)
respectively. Table 2-3 and Table 2-4 also show that excluding infants who died within 3 days of
life had a significant impact on the baseline costs of an EP infant without NEC but did not have a
meaningful effect on the incremental costs of medical NEC and surgical NEC.
25

Figure 2-1 Preparation of the Final Analytic Sample from 2007 California
OSHPD Discharge Data

Step 1. Identifying All Extremely Premature Infant Discharges in OSHPD 2007 data
N = 5,502
Step 2. Excluding discharges where costs are not reported
‘n’ = 4,697
Step 3. Excluding discharges that are disposed to transfer to another hospital unit
n = 3,287
Step 5. Final sample after eliminating infants who died within 3 days of life
final ‘ n’ = 2,560
No NEC : 2,301;
Medical NEC : 177;
Surgical NEC : 82
Step 4. Excluding discharges with daily cost < $100
‘n’ = 3,274
Infants with no NEC: 3005;
Medical NEC: 185, Surgical NEC: 84;
Total Deaths: 1015 (31%) Deaths occurring at < 3 days: 714
26

Table 2-2 Demographic and Baseline Characteristics of Extremely Premature
Infants in the Final Sample Cohort Derived From OHSPD 2007
Discharge Data
Baseline
Characteristics
n (%)
Category Total
N = 2,560
No NEC

n= 2,301
Medical NEC
n (%)
n = 177
Surgical NEC
n (%)
n = 82
P-value
1

(2 sided)
Sex Male 1,110 1,003 (43.6) 74 (41.8) 33 (40.2) 0.0760*
Female 988 896 (38.9) 67 (37.9) 25 (30.5)
Unknown/Missing 462 402 (17.5) 36 (20.3) 24 (29.3)
Race/Ethnicity Hispanic White 565 499 (21.7) 47 (26.6) 19 (23.2) 0.0173*
Non-Hispanic
White
514 476 (20.7) 30 (16.9) 8 (9.8)
Black 151 140 (6.1) 9 (5.1) 2 (2.4)
Asian 110 100 (4.3) 6 (3.4) 4 (4.9)
Other 509 466 (20.3) 31 (17.5) 12 (14.6)
Unknown/Missing 711 620 (26.9) 54 (30.5) 37 (45.1)
Payer status Private 985 890 (38.7) 65 (36.7) 30 (36.6) 0.8915
Medicaid 1,167 1,042
(45.3)
84 (47.5) 41 (50)
Other /Unknown 408 369 (16) 28 (15.8) 11 (13.4)
Total deaths
before
discharge
Dead 301 245 (10.6) 37 (20.9) 19 (23.2) <0.0001*
Respiratory
distress
syndrome
Present 1,816 1,617 (70.3) 142 (80.2) 57 (69.5) 0.0185*
Other
respiratory
conditions of
newborn
2

Present 2,089 1,862 (80.9) 159 (89.8) 68 (82.9) 0.0123*
Infections
specific to
perinatal
period
3

Present 1,395 1,202 (52.2) 132 (74.6) 61 (74.4) <0.0001*
Neonatal
hemorrhage
Present 846 715 (31.1) 84 (47.5) 47 (57.3) <0.0001*
Neonatal
jaundice
Present 1,876 1,687 (73.3) 133 (75.1) 56 (68.3) 0.5075
Endocrine/
metabolic
disturbances
Present 1,184 1,026 (44.6) 106 (59.9) 52 (63.4) <0.0001*
27

Baseline
Characteristics
n (%)
Category Total
N = 2,560
No NEC

n= 2,301
Medical NEC
n (%)
n = 177
Surgical NEC
n (%)
n = 82
P-value
1

(2 sided)
Anemia Present 1,951 1,740 (75.6) 143 (80.8) 68 (82.9) 0.1037
Integument &
temperature
regulation
related
conditions
Present 193 173 (7.5) 12 (6.8) 8 (9.8) 0.6956
Other ill-
defined
perinatal
conditions
4

Present 1,631 1,448 (62.9) 126 (71.2) 57 (69.5) 0.0478*
1. p-value (2-sided) represents the significance of chi-square statistics obtained for the comparison of 3 groups – no
NEC, medical NEC and surgical NEC at 5% significance level. 2. Other respiratory conditions of newborn included
pneumonia, emphysema, apnea, chronic lung disease etc. (identified by ICD-9 codes 770.0 to 770.9). 3. Infections
specific to perinatal period included congenital rubella, cytomegalovirus infections, septicemia, urinary tract infections
etc. (771.0 – 771.8). 4. Other ill-defined perinatal conditions included central nervous system conditions, cardiac
dysrhythmias, periventricular leukomalacia etc. (779.0-779.9). *represents significance at p < .05.

28

Table 2-3 Neonatal Intensive Care Unit (NICU) Days and Initial Average
Hospitalization Costs Among Extremely Premature Infants Without
Necrotizing Enterocolitis in the California OSHPD 2007
Sample N
(No
NEC)
Label 25th
percentile
Mean Median 75
th

percentile
95th
percentile
99
th

percentile
Sample
without
excluding
early
deaths
3005
LOS
Costs (2011
US$)
2
10,402.10
49.4
159,473.40
51
109,207.80
79
233,583.80
122
517,971.70
159
836,777.229
Final
sample
excluding
deaths
within
the first 3
days of
life
2301
LOS
Costs (2011
US$)
39
77,780.75
64.5
207,378.00
65
165,610.80
88
280,569.90
129
578,052.70
164
872,513.65
29

Table 2-4 Multivariate Adjusted Incremental Costs and Length of Stay for Medical
and Surgical Necrotizing Enterocolitis (NEC) in the OSHPD 2007
Hospital Discharge Sample
Sample
NEC
Type
N
Adjusted
Mean
Length
of Stay
(in days)
1

95% CI
(low)
95%
CI
(high)
Adjusted
Mean Costs
(2011 US$)
1

95% CI
(low)
95% CI
(high)
p-value
Sample
without
excluding
early
deaths
Medical
NEC
Surgica
l NEC
185
84
11.0
41.9
6.7
35.7
15.3
48.2
71,645.00
195,795.00
47,953
161,370
95,337
230,220
<0.0001
<0.0001
Sample
excluding
deaths
within the
first 3 days
of life
Medical
NEC
Surgica
l NEC
177
82
11.7
43.1
6.9
36.3
16.5
50.0
74,004.00
198,040.00
47,051
159,261
100,957
236,819
<0.0001
<0.0001
1. The mean estimates for medical NEC and surgical NEC given in this table are the average incremental effects
of each condition when other parameters in the Ordinary Least Squares model such as sex, race, payer status
(Medicaid/private), death and other comorbid conditions during the perinatal period are held constant.
30

2.5.2 Cost-Effectiveness of HUM Arm Versus BOV Arm Feeding Strategies
The total expected NICU costs for an infant receiving exclusively human milk-based versus bovine
milk-based HMF-fortified diets are summarized in Table 2-5. Considering the risks and costs of
medical NEC and surgical NEC, the total expected NICU costs for an infant fed with bovine-fortified
diet versus human donor HMF-fortified diet were $231,954 and $223,787, respectively, resulting in
net direct savings of $8,167 (95% CI $4,405–$11,930; p < 0.0001) per EP infant treated in the NICU
by using the exclusively human milk-based feeding strategy. The corresponding expected NICU
lengths of stay for the two strategies were 69.5 and 65.6 days, respectively (data not shown), resulting
in the saving of 3.9 NICU days (95% CI 3.25–4.58 days; p<0.0001) per EP infant using the HUM
arm feeding strategy. Sensitivity of net savings to 10% changes in various model parameters is
depicted in Figure 2-2. Costs savings from the human milk-based HMF fortification strategy
decreased linearly with the increase in price or quantity of human milk-based HMF. A 50% increase
in the price (to $9.375/mL) of Prolact+H2MF resulted in net savings of $2,967 for the HUM arm
feeding strategy. Fortification using donor milk HMF was no longer a cost saving option when its
price exceeded $11.19/mL (a 78% increase from the current price). A zero dollar price on both bovine
HMF and preterm formula did not appreciably affect the cost savings ($7,972, which is a 2% decrease
from the original estimate), showing that net savings are not affected by changes in the prices of
bovine milk products. Cost savings were sensitive to changes in the degree of risk reduction in overall
NEC and surgical NEC achieved by the HUM arm feeding strategy and incremental costs of surgical
NEC, with elasticity ratios of >1. With other parameters held fixed, a minimum 54% percentage risk
reduction in both overall NEC and surgical NEC would have to be achieved by the donor HMF
strategy in order to remain cost-effective (the net savings being $3,039.57 per EP infant treated in the
NICU).
31

Table 2-5 Estimation of Expected Neonatal Intensive Care Unit (NICU) Costs
Among Extremely Premature Infants Fed with Bovine Milk-Based
Human Milk Fortifier (HMF) Versus 100% Human Milk Diets (Base-
Case Scenario)
Costs Formula BOV HUM
Total costs for supplements
used per infant
C
1
= [(Median quantity of HMF used
x Unit cost of HMF) + (Median
quantity of supplement used when
mother’s milk was not available x
Unit cost of supplement)]
1

$195.23 $10,321.45
Average Baseline
Hospitalization costs per
EP infant
C
2
=[1 x Mean Baseline cost estimates
of EP infants without NEC
2
]
207,378.00 207,378.00
Expected Incremental costs
of Medical NEC per EP
infant
C
3
= [p
1
1
x mean adjusted incremental
costs of NEC
3
]
$11,797.74 $4,290.09
Expected Incremental costs
of Surgical NEC per EP
infant
C
4
= [p
2
1
x (mean adjusted
incremental costs of surgical NEC –
mean adjusted incremental costs of
medical NEC)
3
]
$12,583.36 $1,797.62
Total expected NICU costs
per EP Infant
TC = C
1
+ C
2
+ C
3
+ C
4

$231,954.33
(95%CI
4
:
226,457.74-
237,450.93)
$223,787.16
(95%CI
4
:
222,053.28-
225,521.04)
Net savings in hospital
costs per EP infant owing
to HUM over BOV
TC
BOV
- TC
HUM

$8,167.17
(95% CI
4
: 4,405– 11,930)
1. Data obtained from table 1. 2. Data obtained from table 3 (second row). 3. Data obtained from table 4
(second row).4. Represents the estimates calculated using 95% confidence interval limits for adjusted
incremental costs of medical and surgical NEC as given in table 4 (second row). HMF = Human milk fortifier;
EP = Extremely Preterm (GA ≤ 28 weeks); HUM = donor human milk HMF fortified diet; BOV = Bovine milk
HMF fortified diet; NEC = necrotizing enterocolitis.
32

Figure 2-2 One-Way Sensitivity Analysis of Cost Savings to a 10% Change in Model
Parameters*

*IC = incremental costs, *HUM = Prolacta HMF fortified diet, *NEC = necrotizing enterocolitis, *p
nec
=
probability of NEC, *p
surg nec
= probability of surgical NEC.
*There was a net savings of $8,167 per infant. OSHPD, California Office of Statewide Health Planning and
Development.
2.6 Discussion
The cost-effectiveness analysis presented in this article serves to demonstrate to the neonatal
community the value of preventing NEC, which is a serious and potentially life-threatening neonatal
gastrointestinal complication among EP infants. Management of NEC could be highly resource
intensive, requiring greater length of NICU stay. Often respiratory and circulatory support is provided
in severe cases, and about a third of infants affected with NEC require surgical intervention (Blakely
2008; Horbar et al 2007). In our analysis, using California statewide discharge data from over 257
hospitals, we found that after adjusting for potential confounders, EP infants with medical NEC
stayed in the hospital for an additional 11.7 days and incurred additional costs of US $74,004
33

compared with EP infants without NEC; infants with surgical NEC stayed for an additional 43.1 days
and incurred additional costs of US $198,040 compared with the average EP infant without NEC.
Initiating enteral nutrition among EP infants by feeding mother’s milk has been shown to be highly
beneficial for these infants and to offer protection against NEC (Boyd et al 2007; Lucas and Cole
1990; McGuire and Anthony 2003; Schanler et al 1999, Schanler 2007; Sisk et al 2007). Researchers
have observed a threshold effect of human milk in the reduction of incidence of NEC (American
Academy of Pediatrics 1997; Furman et al 2003; Meinzen-Derr et al 2009; Schanler 1999). Key
findings of the randomized controlled study conducted by Sullivan et al. (2010) are that a 100%
human milk-based diet is associated with a 63% lesser rate of NEC than that observed for the
alternative bovine milk-based fortification of mother’s milk and that the corresponding reduction in
incidence of surgeries for NEC is more profound (86%). A net cost calculator developed based on
these results shows a direct net savings of $8,167 per infant receiving an exclusively human milk-
based feeding strategy. The main drivers of cost savings in this model are the reduction in the rate of
NEC and surgical NEC and their associated NICU costs. These savings could only be conservative
for many reasons. For instance, we have not accounted for physician fees and post-treatment care
costs of NEC in our model. Treating complications of surgery (for NEC) such as short bowel
syndrome could be highly resource intensive because of the prolonged dependence of these patients
on total parenteral nutrition (Patel et al 1998). Survival of infants with intestinal failure following
surgical NEC is associated with significantly prolonged hospitalization and higher utilization of
hospital resources especially if small bowel transplantation, which might improve the chances of
survival in such patients, is required (Vennarecci et al 2000). The expected costs of intestinal
transplantation have been estimated to be in the range of $130,000–$250,000 depending on the type
of allograft (Sudan 2006) besides other significant costs for pre- and post-transplant evaluation and
testing, physician fees, lifelong immunosuppressant drugs, rehabilitation costs, etc. These costs are
34

likely to rise further as a greater number of NEC/intestinal failure survivors are generated by the
current survival trends among EP infants. Given the huge reduction in the incidence of NEC and
surgery for NEC (and hence short bowel syndrome) among infants who were fed a 100% human
milk-based fortified diet, significant costs of post-NEC complications can be averted.
2.6.1 Economic Value of NEC Risk Reduction
Approximately 1.5% of live births in the United States, or 64,500 births, were very low birth weight
(weighing <1,500 grams) in 2007, and this represents a 20% increase in incidence of prematurity
from 1990 (Hamilton et al 2009). The mean incidence of NEC is re-ported to be 7–10% among EP
infants (Thompson and Bizzarro 2008). This implies that if we can achieve at least a 55% reduction
each in the rate of NEC and surgeries for NEC, we can avert 2,483–3,548 NEC cases among EP
births annually. The corresponding savings that could be achieved for the U.S. population would be
US $212 million in hospitalization costs alone. Apart from these direct savings, prevention of NEC
also entails indirect savings through prevention of NEC-associated mortality among newborns. The
mortality rate due to NEC in the United States is estimated at 13.4 deaths per 100,000 live births
(Horbar et al 2007). At the given mortality rate, about 587–839 newborns every year are expected to
die owing to NEC. If we could achieve a 55% reduction in the rate of NEC (and conservatively
assuming that feeding practices do not have additional direct benefit on improving mortality), about
323–461 lives could be saved annually. The economic value of prevention of mortality at birth is
estimated to be very high considering that the survival rates are higher for infants. Studies on value of
statistical life have placed a value of $5,000,000 for the life of an infant for even small reductions in
mortality risk (Murphy and Topel 1999). Thus the economic value of reduction in the risk of
mortality due to NEC, given by the formula (number of lives saved x the economic value of life at
birth), would be US $1.6–2.3 billion. Also, the total economic value incurred by the reduction in the
risk of NEC, considering the direct costs and mortality costs averted, would be US $1.8–2.5 billion.
35

2.7 Limitations
Although this article carries important information to the neonatal community, the cost-effectiveness
results should be interpreted with some caution. The results of our study are based on the only
randomized trial conducted so far comparing bovine and donor HMFs. It could be possible that the
incidence of NEC and surgical NEC observed among the 12 NICU units (11 in the United States and
one in Austria) in the study of Sullivan et al. (2010) may or may not be representative of the universe
of NICU units catering to the healthcare needs of EP infants. Given the lower frequencies of EP births
and NEC in the general population, the costs of conducting such huge trials are prohibitive.
Nevertheless, the primary driver of costs savings is, in fact, the percentage reduction in rates of
overall NEC and surgical NEC in the HUM arm and not the absolute NEC rates. The highly
significant results obtained for incidence of surgical NEC (p=0.007) would mean that the results have
less than 1 in 140 chance of being random. Second, the use of statewide hospital discharge data to
evaluate the costs of NEC may be perceived as a potential limitation of the study in that the feeding
patterns among infants in the discharge sample cannot be ascertained. At the time the clinical trial
started, however, Prolacta HMF was not being used routinely in the United States; there were perhaps
fewer than 10 NICUs that were using Prolacta HMF on a sporadic basis. Although real world data
through chart review may have provided better data on proportions of milk use, these data would have
been helpful only to magnify the cost difference we observed with a comparison to an exclusive
human milk diet because the control group from the study of Sullivan et al. (2010) likely had a higher
human milk intake compared with previous clinical experience or against national data. This
observation was likely due to a positive effect that human milk research has on human milk usage.
Overall we felt that the power in having large statewide data outweighed the benefit of greater
resolution on feeding patterns that a significantly smaller and costly multicenter chart-review study
would provide.
36

2.8 Conclusions
The NICU cost burden of NEC among EP infants is huge. Provision of an exclusively human milk
diet composed of mother’s own milk, or donor human milk when mother’s milk is not adequately
available, and fortified by donor HMF can result in saving net NICU resources and produce societal
value by preventing infant mortality. The analyses presented in this article may assist healthcare
providers and institutions to justify an increased use of human milk and human milk products to
promote better health outcomes in EP infants.
2.9 References
1. American Academy of Pediatrics. 1997. Breastfeeding and the use of human milk. Work Group
on Breastfeeding. Pediatrics 100:1035–1039.
2. Blakely ML. 2008. Surgical management of necrotizing enterocolitis and isolated intestinal
perforation in premature neonates. Semin Perinatol 32:122–126.
3. Boyd CA, Quigley MA, Brocklehurst P. 2007. Donor breast milk versus infant formula for
preterm infants: Systematic review and meta-analysis. Arch Dis Child Fetal Neonatal Ed
92:F169–F175.
4. Fanaroff A, Hack M, Walsh M. 2003. The NICHD Neonatal Re-search Network: Changes in
practice and outcomes during the first 15 years. Semin Perinatol 27:281–287.
5. Furman L, Taylor G, Minich N, et al. 2003. The effect of maternal milk on neonatal morbidity of
very low-birth weight infants. Arch Pediatr Adolesc Med 157:66–71.
37

6. Guner YS, Friedlich P, Wee CP, et al. 2009. State-based analysis of necrotizing enterocolitis
outcomes. J Surg Res 157:21–29.
7. Hamilton BE, Martum JA, Ventura SJ. Births: Preliminary Data for 2007. National Center for
Health Statistics, Hyattsville, MD, 2009.
8. Henderson JJ, Hartmann PE, Newnham JP, et al. 2008. Effect of preterm birth and antenatal
corticosteroid treatment on lactogenesis in women. Pediatrics 121:e92–e100.
9. Horbar JD, Carpenter J, Kenny M, eds. 2007. Very Low Birthweight Database Summary.
Vermont Oxford Network, Burlington, VT.
10. Kurth T, Walker AM, Glynn RJ, et al. 2006. Results of multivariate logistic regression,
propensity matching, propensity adjustment, and propensity-based weighting under conditions of
nonuniform effect. Am J Epidemiol 163:262–270.
11. Liang K, Zeger SL. 1986. Longitudinal data analysis using generalized linear models. Biometrika
73:13–22.
12. Lucas A, Cole TJ. 1990. Breast milk and neonatal necrotizing enterocolitis. Lancet 336:1519–
1523.
13. McGuire W, Anthony M. 2003. Donor human milk versus formula for preventing necrotizing
enterocolitis in preterm infants: Systematic review. Arch Dis Child Fetal Neonatal Ed 88:F11–
F14.
14. Meinzen-Derr J, Poindexter B, Wrage L, et al. 2009. Role of human milk in extremely low birth
weight infants’ risk of necrotizing enterocolitis or death. J Perinatol 29:57–62.
38

15. Murphy KM, Topel RH. 1999. The Economic Value of Medical Research. University of Chicago
Press, faculty .chicagobooth.edu/kevin.murphy/research/murphy & topel .pdf (accessed
December 2010).
16. Patel JC, Tepas JJ, Huffman SD, et al. 1998. Neonatal necrotizing enterocolitis: The long-term
prospective. Am Surg 64: 575–579.
17. Petrosyan M, Guner YS, Williams M, et al. 2009. Current concepts regarding the pathogenesis of
necrotizing enterocolitis. Pediatr Surg Int 25:309–318.
18. Schanler RJ, Lau C, Hurst NM, et al. 2005. Randomized trial of donor human milk versus
preterm formula as substitutes for mother’s own milk in feeding of extremely premature infants.
Pediatrics 116:400–406.
19. Schanler RJ, Shulman RJ, Lau C. 1999. Feeding strategies for premature infants: Beneficial
outcomes of feeding fortified human milk vs. preterm formula. Pediatrics 103:1150–1157.
20. Schanler RJ. 2007. Evaluation of the evidence to support current recommendations to meet the
needs of premature infants: The role of human milk. Am J Clin Nutr 85(Suppl): 625S–628S.
21. Sisk PM, Lovelady CA, Dillard RG, et al. 2007. Early human milk feeding is associated with a
lower risk of necrotizing enterocolitis in very low birth weight infants. J Perinatol 27:428–433.
22. Sudan D. 2006. Cost and quality of life after intestinal transplantation. Gastroenterology 130(2
Suppl 1):S158–S162.
39

23. Sullivan S, Schanler RJ, Kim JH, et al. 2010. An exclusively human milk-based diet is associated
with a lower rate of necrotizing enterocolitis than a diet of human milk and bovine milk-based
products. J Pediatr 156:562–567.
24. Thompson AM, Bizzarro MJ. 2008. Necrotizing enterocolitis in newborns: Pathogenesis,
prevention and management. Drugs 68:1227–1238.
25. U.S. Bureau of Labor Statistics. Consumer Price Index—All Urban Consumers—Medical Care.
data.bls.gov:8080/PDQ/outside.jsp?survey=cu (accessed December 2010).
26. Updegrove K. 2004. Necrotizing enterocolitis: The evidence for use of human milk in prevention
and treatment. J Hum Lact 20:335–339.
27. Vennarecci G, Kato T, Misiakos EP, et al. 2000. Intestinal transplantation for short gut syndrome
attributable to necrotizing enterocolitis. Pediatrics 105:E25.
40

3. LONG TERM HEALTHCARE COSTS OF INFANTS WHO
SURVIVED NEONATAL NECROTIZING ENTEROCOLITIS: A
RETROSPECTIVE LONGITUDINAL STUDY AMONG INFANTS
ENROLLED IN TEXAS MEDICAID
3,4

3.1 Abstract
Background: Infants who survive advanced necrotizing enterocolitis (NEC) at the time of birth are at
increased risk of having poor long term physiological and neurodevelopmental growth. The economic
implications of the long term morbidity in these children have not been studied to date. This paper
compares the long term healthcare costs beyond the initial hospitalization period incurred by medical
and surgical NEC survivors with that of matched controls without a diagnosis of NEC during birth
hospitalization.
Methods: The longitudinal healthcare utilization claim files of infants born between January 2002
and December 2003 and enrolled in the Texas Medicaid fee-for-service program were used for this
research. Propensity scoring was used to match infants diagnosed with NEC during birth
hospitalization to infants without a diagnosis of NEC on the basis of gender, race, prematurity,
extremely low birth weight status and presence of any major birth defects. The Medicaid paid all-
inclusive healthcare costs for the period from 6 months to 3 years of age among children in the
medical NEC, surgical NEC and matched control groups were evaluated descriptively, and in a
3
Ganapathy V, Hay JW, Kim JH, Lee ML, Rechtman DJ. 2013. BMC Pediatrics 13:127-138.
4
Research was funded by an unrestricted educational grant from Prolacta Bioscience, City of Industry, CA.
41

generalized linear regression framework in order to model the impact of NEC over time and by birth
weight.
Results: Two hundred fifty NEC survivors (73 with surgical NEC) and 2,909 matched controls were
available for follow-up. Medical NEC infants incurred significantly higher healthcare costs than
matched controls between 6–12 months of age (mean incremental cost = US $5,112 per infant). No
significant difference in healthcare costs between medical NEC infants and matched controls was
seen after 12 months. Surgical NEC survivors incurred healthcare costs that were consistently higher
than that of matched controls through 36 months of age. The mean incremental healthcare costs of
surgical NEC infants compared to matched controls between 6–12, 12–24 and 24–36 months of age
were US $18,274, $14,067 (p<0.01) and $8,501 (p=0.06) per infant per six month period,
respectively. These incremental costs were found to vary between subgroups of infants born with
birth weight <1,000 grams versus ≥ 1,000 grams (p<0.05).
Conclusions: The all-inclusive healthcare costs of surgical NEC survivors continued to be
substantially higher than that of matched controls through the early childhood development period.
These results can have important treatment and policy implications. Further research in this topic is
needed.

42

3.2 Background
Neonatal necrotizing enterocolitis (NEC) is a gastro-intestinal inflammatory condition in neonates
that has a detrimental effect on the survival and long term development of affected infants. The
disease is most commonly seen in premature infants, although up to 25% of NEC cases have been
observed among full-term babies (Neu and Walker 2011; Ng 2001). The overall incidence of
necrotizing enterocolitis in the United States among infants born with birth weights <1500 grams is
estimated to be 7 – 12% (Grave et al 2007; Horbar and Carpenter 2008; Pierro and Hall 2003;
Petrosyan et al 2009). The etiology of NEC is understood to be multi-factorial; ischemic injury and
aberrant microbial colonization, among other factors, have been found to play a very important role in
the disease process (Neu and Walker 2011).
Initial management of NEC is often highly complicated and resource intensive. About 44 to 70% of
neonates diagnosed with NEC show signs of advanced disease (commonly categorized as Bell’s stage
III) that requires surgical management (Bell et al 1978; Cikrit et al 1986; Pierro and Hall 2013). The
economic burden of NEC has been estimated to be approximately 500 million to 1 billion US dollars
annually (Bisquera et al 2002; Neu and Walker 2011). These estimates are probably conservative
given the strong evidence to support the many adverse consequences of surgical NEC beyond the
initial hospitalization period. In particular, the outcomes of bowel resection among survivors of the
procedure has been well studied and has been found to be associated with complications such as short
gut syndrome and prolonged administration of parenteral nutrition (Blakely et al 2008; Cikrit et al
1986; Ladd et al 1998). There also is growing evidence for a possible link between surgical NEC and
growth and neurodevelopmental impairment, leading to poor health outcomes in the long term (Cikrit
et al 1986; Hintz et al 2005; Rees et al 2007; Schulzke et al 2007; Simon 1994; Tejani et al 1978).
The long term economic outcomes, such as healthcare or special education costs, of NEC survivors
43

have not been well studied and there are very few published studies on this topic. There is a clear
need for research on this topic to understand the full spectrum of disease burden, and also to evaluate
the cost-effectiveness of novel therapeutic strategies that can mitigate the risk of developing NEC or
reduce its impact on infant growth and development in the long term.
3.3 Objectives
The primary goal of this study is to compare the healthcare costs between 6 and 36 months of
chronological age, among survivors of medical and surgical NEC to that of matched infants without a
diagnosis of NEC during birth hospitalization. This study was driven by the hypothesis that NEC
survivors on average will continue to have greater demand for healthcare services throughout the
early developmental period due to the increased risk of long term complications in these children as
compared to similar infants without a diagnosis of NEC. Furthermore, the incremental demand for
healthcare services will vary depending on the type of NEC (medical or surgical).
3.4 Study Design and Methods
This is a retrospective cohort study using claims database analyses. Infants enrolled in the Texas
Medicaid fee-for-service program, born between January 2002 and December 2003, were selected for
this study. Greater than 50% of live births in Texas were covered by Medicaid (Medicaid Coverage
2012). Also, unlike in other large states such as California, Illinois or New York, a significant
proportion of Texas Medicaid clients were managed under traditional fee-for-service arrangements
during the study period (2002 to 2006), making this data source ideal for our research (Janek and
Ghahremani 2013).
44

3.4.1 Sample Preparation and Propensity Score Matching
Infants with NEC diagnoses were identified using a primary or secondary ICD-9-CM code of 777.50-
53 in the hospital or physician claims at the time of initial hospitalization. NEC infants who had
undergone procedures such as exploratory laparotomy involving bowel resection, peritoneal drainage,
enterostomy with creation of stoma, etc. were defined as having “surgical NEC” (Table 3-1). NEC
infants without claims for any of these surgical procedures were defined as having “medical NEC”.
Based on the theoretical work of Rosenbaum & Rubin (1985) and Imbens & Rubin (2009) on the role
of propensity score matching in the design of observational studies, each NEC infant was matched to
infants without NEC diagnoses at a 1:10 ratio without replacement. Matching using propensity scores
help to achieve a balance among the baseline covariate distributions of the treated and counterfactual
groups even at the design stage.
5
Propensity score matching also helps to improve the precision of
estimates of the treatment effect being studied [Imbens & Rubin, 2009]. The application of the
propensity score matching design to match each infant with NEC to infants from a control pool,
consisting of all infants without NEC, was conducted in a two-step process as described below.
Step 1: Propensity score estimation
A stepwise logistic regression was fitted to model the probability of occurrence of necrotizing
enterocolitis (0 = no NEC, 1 = NEC) using a set of predictor variables, that were chosen based on
information available in the literature regarding their association with necrotizing enterocolitis. The
effects of the following variables and interaction terms were tested for statistical significance at the
5% level and entered into the logistic model by using a step-wise selection procedure in logistic
regression, using the SAS software version 9.2., (SAS; Cary, NC) : 1) prematurity (0 = born fully
5
For a detailed discussion about the role of design versus analysis in observations studies four causal effects
refer Rubin 2006, and for a discussion on the theoretical properties of propensity scores and its application in
matching methods refer Imbens & Rubin (2009).
45

mature, 1 = born premature); 2) extreme low birth-weight status (ELBW; defined as birth weight
<1,000 grams using the 5th digit of ICD-9 codes: 765.0x – 765.1x); 3) African-American race; 4)
male gender; 6) presence of major birth defects (0/1 variable), which was defined as having one or
more of the following conditions: congenital heart disease (CHD) including patent ductus arteriosus
(PDA), neural tube defects (NTD), hereditary CNS conditions and cleft lip or palate
6
; 7) an
interaction term between prematurity and African-American Race; and 8) an interaction term between
prematurity and birth defects. The propensity score (pscore) is the model predicted probability of
NEC occurrence. The logit of the predicted propensity scores was then calculated using the formula








− ) 1 (
log
pscore
pscore
. The univariate distributions of the logit of the propensity scores, for the NEC
and the non-NEC groups were analyzed before the appropriate propensity score matching technique
was chosen.
Step 2: Propensity score matching:
A hybrid matching technique that combines the best features of nearest neighbor and matching within
caliper methods was chosen for constructing the control group
7
. The metric chosen for the distance
between the two groups was the Mahalanobis distance, described by Cochran and Rubin (1973) and
studied in detail by Carpenter (1977) and Rubin (1979,1980). The matching technique was
implemented using the macro program, PSMatching (Coca-Perraillon M, 2007) in SAS 9.2. The
program implements the matching technique in the following number of steps:
6
Identified using ICD-9-CM codes: congenital heart conditions (ICD-9 CM codes 745-747), neural tube defects
(741,742), hereditary CNS conditions (330-337) and cleft lip or palate (749.X).
7
A detailed discussion on the construction of matched sample within calipers can be found in Althauser RP and
Rubin DB (1971).
46

1) Infants in the NEC group and the control pool were randomly ordered.
2) The program finds all available infants without NEC with estimated propensity scores that
differ from the propensity score value for the NEC infant by less than a specified constant c
(caliper size). Using recommendation of Rosenbaum and Rubin (1985) to use a caliper size that is
one-fourth of the standard deviation of the logit of propensity score of the treatment group, a c
value of 0.5 was chosen to implement matching within caliper.
8

3) From the subset of infants identified by step 2, the program selects M matches with the
closest Mahalanobis distance within caliper. A 1:10 match, without replacement chosen to
construct the matched sample of controls.
4) Once the suitable matches within caliper are found for the first infant, the NEC infant and the
matched controls are removed from the lists of treatment and control pools. The process is
repeated for the next NEC infant beginning from step 2 until each NEC infant in the treatment
pool has ≤ M suitable matches from the control pool.
After propensity matching NEC infants to controls, the covariate distributions of the NEC sample and
the matched controls are compared for statistically significant differences to check their balance
properties.
8
The standard deviation of the logit of the propensity score in the NEC group was 2.05.
47

Table 3-1 Procedures and ICD-9-CM Codes Used to Define Infants with Surgical
NEC
Surgical Procedures ICD-9-CM Codes
Intestinal resection procedures
45.02-3, 45.1, 45.29, 45.3-4, 45.41, 45.49,
45.50-2, 45.60-3, 45.70-9, 45.8, 46.99
Procedures related to stoma creation 46.0 – 46.64
Intestinal anastomosis 45.9 – 94, 46.73-79, 46.93-4
Exploratory laparotomy 45.0, 45.00, 54.11
Percutaneous abdominal drainage 54.91
3.5 Outcomes & Statistical Analyses
The longitudinal inpatient, outpatient, physician and prescription claim files of all infants were
followed from 6 months up to 3 years after birth. Follow-up beyond this period was restricted by the
extremely small sample size (<30 children) in the surgical NEC cohort. Descriptive tests for attrition
were performed by comparing the sample characteristics of children retained in each group at the end
of 36 months to the characteristics at baseline. The prevalence of chronic developmental health
conditions in the NEC and control groups were studied over the follow-up period. The combined
Mantel-Haenszel odds ratios (Mantel and Haenszel 1959) adjusted for ELBW status were reported for
these conditions for the NEC and matched control groups. The hospital utilization, inpatient,
ambulatory care including home healthcare, and the all-inclusive (grand total) healthcare costs in the
NEC and matched control groups were evaluated descriptively. The total all-inclusive healthcare
costs represent the inflation-adjusted amounts paid by Medicaid for inpatient, ambulatory care
including home healthcare, prescription and professional services (all in 2009 US$).
48

Generalized linear regressions with a log link function and gamma distribution assumption were used
to model the incremental costs for medical and surgical NEC over matched controls between 6–36
months of age (Manning and Mullahy 2001). Cluster robust standard errors were used to account for
correlation within subject. Three different models were explored to estimate the total healthcare costs
per six months: 1) allowing the impact of NEC to vary over time by fitting an interaction term
between NEC type (no NEC, medical NEC and surgical NEC) and age-period (6–12, 12–24 and 24–
36 months of age); 2) allowing both NEC and ELBW effects to vary over time; and 3) adding an
interaction term between NEC type and ELBW status in addition to specifying the time-varying slope
terms for NEC and ELBW status. Race, gender and hospitalization status in the previous period
(Yes/No) were included as covariates in all 3 specifications. The three model specifications were
tested for over-fitting using standard goodness-of-fit criteria such as the deviance and the AIC
(Akaike information criteria). The average incremental costs of medical and surgical NEC over
matched controls at each age-period were estimated using the margins command in STATA 11.0
(College Station, TX) and predicted costs were estimated (Williams 2012). Statistical significance of
all descriptive between-group comparisons and model-based coefficients was deter-mined using an
alpha level of 0.05.
This study was approved by the Health Sciences Review Board at University of Southern California.
3.6 Results
3.6.1 Baseline Characteristics
Three hundred and sixteen infants in the 2002–2003 fee-for-service Medicaid sample had NEC
diagnoses, 111 of which had surgical NEC (Figure 3-1). The effects of prematurity, birth defects,
VLBW status, and interaction between prematurity and birth defects met the model entry criteria (p <
0.05) in the step-wise logistic regression, and only these effects entered the final logistic model that
49

was used to predict the probability of occurrence of NEC. Propensity matching at baseline resulted in
2,909 controls well matched on all baseline characteristics except ELBW status, PDA and presence of
neural tube defects (Table 3-2). There were 101 (32%) and 759 (26%) ELBW infants in the NEC and
matched control group, respectively (p<0.05). PDA was the most common birth defect observed
among NEC infants (30%) followed by neural tube defects (10%). The proportions of infants with
PDA and neural tube defects among matched controls were 17% and 7%, respectively (p<0.05).
However, the two groups were comparable in the proportion of infants with any major birth defects as
defined in the methods section (48% and 44% among NEC and matched controls, respectively,
p=0.13).
50

Figure 3-1 NEC and Control Group Sample Sizes (after matching) From 6 Months
to 3 Years of Age

At Baseline :
All NEC: 316
- Surgical NEC: 111
Matched controls: 2,909
No. died, 0-6 months:
All NEC: -66 (21%)
- Surgical NEC: -38 (34%)
Matched controls: -260 (9%)
SURVIVORS AT 6 MOS
All NEC : 250 (79%)
- Surgical NEC: 73 (66%)
Matched controls: 2,649 (91%)
SAMPLE AT 12 MOS*
All NEC: 203 (81%)
- Surgical NEC: 62 (85%)
Matched controls: 2,109 (80%)
SAMPLE AT 24 MOS*
All NEC: 144 (58%)
- Surgical NEC: 42 (58%)
Matched controls: 1,326 (50%)
SAMPLE AT 36 MOS*
All NEC: 105 (42%)
- Surgical NEC: 35 (50%)
Matched controls: 951 (36%)
51

Table 3-2 Comparison of Baseline Characteristics Between the NEC and Control
Samples Before and After Propensity Score Matching
NEC No NEC
Characteristics (n = 316)
Before Matching
(n = 122,929)
After Matching
(n = 2,909)
Prematurity 212 (67%) 15,451 (13%)** 1869 (64%)
ELBW status 101 (32%) 921 (0.8%)** 759 (26%)*
Race

African-American 36 (11%) 7,777 (6%)** 350 (12%)
White 35 (11%) 13,964 (11%) 257 (9%)
Hispanic 111 (35%) 68,361 (56%) 1,407 (48%)
Other / Unknown 134 (43%) 32,827 (27%) 895 (31%)
Male 173 (55%) 62,252 (51%) 1,480 (51%)
Birth defects
Any major birth
defects
‡

153 (48%) 8874 (7%)** 1,279 (44%)
Patent ductus
arteriosus
95 (30%) 2284 (2%)** 501 (17%)**
Neural tube defects 33 (10%) 1101 (1%)** 200 (7%)*
ELBW = extremely low birth-weight (BW <1000g); NEC = necrotizing enterocolitis; **p<0.01 and *p<0.05;

‡
Includes congenital heart disease (CHD), neural tube defects, hereditary degenerative CNS conditions and cleft lip or palate.
Thirty-eight infants in the surgical NEC (34%), 28 in the medical NEC (14%) and 260 infants in the
control groups (9%) died before 6 months of age (Figure 3-1). This left 250 NEC survivors (73
among them with surgical NEC) and 2,649 survivors among matched controls for follow-up beyond 6
months of age. Attrition was high among survivors in both NEC and control groups, mainly due to
drop-out from the Medicaid program.
52

Comparison of the characteristics of infants who were retained in the NEC and control groups at 36
months showed that there was no significant change within and between groups for the characteristics
that were matched at baseline.
3.6.2 Prevalence of chronic conditions in the medical and surgical NEC groups vs.
controls
Table 3-3 lists the proportion of survivors in the medical NEC, surgical NEC and matched control
groups with diagnoses of various chronic conditions observed during 6–12 and 24–36 months of age.
Adjusting for ELBW status, the risk of developing bronchopulmonary dysplasia (BPD) was
significantly higher in the medical and surgical NEC groups through 36 months of age (p<0.01); the
risk of malabsorption syndrome, metabolic disorders, failure to thrive (FTT) and neurodevelopmental
delay (NDD) were significantly higher in the surgical NEC group than matched controls through 36
months of age (p<0.05). Also, a significant difference was observed in the proportion of children
receiving care for feeding difficulties and gastrointestinal ostomies between the surgical NEC and
matched control groups through 36 months (p<0.05). Medical NEC infants faced a significantly
higher risk of FTT, feeding difficulties, NDD and open gastrointestinal ostomies between 6–12
months of age, but not in the subsequent periods of evaluation.
3.6.3 Univariate distribution of healthcare utilization and costs
The univariate distributions of healthcare utilization and costs in the NEC and matched control groups
are presented in Table 3-4. Healthcare utilization and cost estimates were highly skewed and the bulk
of utilization and healthcare costs were concentrated in the upper right tails in all 3 groups. Medical
NEC infants on average had 3 additional hospital days than matched controls between 6–12 months
of age (p<0.01) but the corresponding inpatient costs did not reach significance between the two
groups (p=0.056). The total ambulatory care cost, including home healthcare costs, was significantly
53

higher in the medical NEC group compared to matched controls for the 6–12 months of age period
(p<0.01). The difference in healthcare utilization and costs between medical NEC and control groups
after 12 months of age was not statistically significant.
54

Table 3-3 Prevalence of Chronic Developmental Health Conditions in the NEC and
Matched Control Groups Between 6 and 12 and 24 and 36 Months of
Follow-up
Age: 6 - 12 Months

Chronic Conditions
Controls
(N = 2,109) n
(%)
Medical NEC
(N = 141) n
(%)
Surgical NEC
(N = 62) n (%)
Adjusted odds
ratio‡
(95% CI)
Medical NEC
vs. Controls
Adjusted odds
ratio‡
(95% CI)
Surgical NEC vs.
Controls
BPD 98 (5%) 19 (14%) 16 (26%) 3 (1.8 – 5.6)** 4 (2 - 7)**
Malabsorption syndrome† 13 (1%) 2 (1%) 13 (21%) 2.3 (0.5 – 10) 47 (19 - 116)**
Failure to Thrive 132 (6%) 22 (16%) 18 (29%) 3 (1.7 – 5)** 4 (2 - 6)**
NDD 168 (8%) 21 (15%) 16 (26%)
1.9 (1.2 –
3.2)**
2.4 (1.4 – 4.5)**
Cerebral palsy 17 (1%) 3 (2%) 2 (3%) 2.6 (0.7 – 9) 2.3 (0.5 – 10.8)
GI artificial openings present 42 (2%) 9 (6%) 15 (24%)
3.2 (1.5 –
6.9)**
9 (5 - 21)**
Metabolic Disturbances
§
58 (3%) 5 (4%) 8 (13%) 1.2 (0.5 – 3) 4.8 (2.1 – 10.7)**
Feeding Difficulties 33 (2%) 8 (6%) 14 (23%) 4 (1.7 – 8.4)** 11 (5.2 – 21.8)**
Age: 24 - 36 Months

Chronic Conditions
Controls
(N = 951)
n (%)
Medical NEC
(N = 70)
n (%)
Surgical
NEC
(N = 35)
n (%)
Adjusted odds
ratio‡ (95%
CI)
Medical NEC
vs. Controls
Adjusted odds
ratio‡ (95% CI)
Surgical NEC vs.
Controls
BPD 14 (1%) 4 (6%) 5 (14%)
4.6 (1.4 –
15)**
5.5 (2 - 16)**
Malabsorption syndrome† 4 (0.4 %) 0 (0%) 7 (20%) - 62 (15 - 249)**
Failure to thrive 97 (10%) 6 (9%) 11 (31%) 0.8 (0.3 – 2) 3 (1.3 - 6)*
NDD 133 (14%) 13 (19%) 13 (37%) 1.5 (0.8 – 2.9) 2.6 (1.2 – 5.6)*
Cerebral palsy 35 (4%) 1 (1%) 4 (11%) 0.4 (0.05 – 3) 2.1 (0.7 – 6.4)
GI artificial openings present 35 (4%) 4 (6%) 10 (29%) 1.8 (0.6 – 5.3) 6 (3 - 14)**
Metabolic Disturbances
§
29 (3%) 3 (4%) 4 (11%) 1.4 (0.4 – 5) 3.4 (1.1 -10)*
Feeding Difficulties 19 (2%) 3 (4%) 4 (11%) 2.4 (0.7 – 9) 4 (1.2 - 13)*
BPD = Bronchopulmonary dysplasia; ELBW = extremely low birth weight; GI = gastro-intestinal; NDD = Neurodevelopmental delay; **p<0.01 and *p<0.05.
‡
Odds ratios were obtained after adjusting for extremely low birth weight status using Mantel-Haenszel chi-squared tests.
†
Malabsorption syndrome includes post-surgical non-absorption (ICD-9 579.3 commonly used to code for short bowel syndrome) and other unspecified
intestinal malabsorption (ICD-9 579.9).
§
Metabolic disturbances include disorders of amino-acid, carbohydrate, lipid and mineral metabolism etc. (ICD-9 codes 270-279).
55

Table 3-4 Comparison of Healthcare Utilization and Cost Estimates (Unadjusted) for Medical, Surgical NEC
and Matched Control Groups from 6 Months to 3 Years of Age
Age -> 6-12 months 12-24 months 24-36 months
Controls Medical NEC Surgical NEC Controls Medical NEC Surgical NEC Controls Medical NEC Surgical NEC
N -> 2109 141 62 1326 102 42 951 70 35
Mean Hospital
admissions
a
(SD)

0.1 (0.4) 0.2 (1)** 1 (1)** 0 (1) 0 (0.5) 1 (2)** 0 (1) 0 (0.3) 1 (1)**
IQR 0 0 1 0 0 1 0 0 1
90th percentile 0 1 2 1 1 3 0 0 1
Mean Hospital days

b
(SD)

1 (9) 4 (17)** 16 (33)** 1 (4) 1 (4) 10 (33)** 0.5 (3) 0.5 (3) 2 (6)**
IQR 0 0 10 0 0 4 0 0 2
90th percentile 2 6 59 2 3 13 0 0 3
Inpatient costs (SD)
b
2,922 (28,056) 8,068 (33,452)
35,867
(79,511)**
1,942 (12,319) 2,557 (9,823) 23,102 (71,101)** 1,046 (8,737) 1,039 (6,395) 7,842 (29,930)**
IQR 0 0 24,150 0 0 7,589 0 0 2,615
90th percentile 2,167 8,769 126,901 3,435 6,296 26,085 0 0 11,786
Total home
healthcare
b
costs
(SD)
741 (5,575) 2,882 (14,251)** 3,564 (9,895)** 2,539 (17,372) 3,102 (16,412) 11,364 (34,901)** 2,237 (14,806) 2,380 (13,167) 9,485 (23,184)**
IQR 0 0 1,741 0 0 3,533 0 0 4,947
90th percentile 160 1,536 7,761 184 386 36,309 199 128 31,976
Total ambulatory
care costs
b,c
(SD)
2,332 (7,388) 5,129 (16,391)** 8,764 (14,920)** 5,961 (21,091) 6,392 (18,719) 21,715 (38,769)** 4,346 (19,294) 3,946 (14,009) 16,461 (34,291)**
IQR 1,247 3,414 10,572 2,154 4,144 21,573 1,372 1,874 11,320
90th percentile 5,958 10,846 20,126 14,984 17,382 61,962 5,691 4,357 61,559
Overall healthcare
costs
b,c
(SD)
5,598(30,654
)
13,610
(38,264)**
45,213 (87,497)** 8,726 (28,039) 9,856 (23,111) 46,378 (91,535)** 6,279 (24,018) 5,809 (16,966) 26,055 (52,637)**
IQR 2,238 7,157 41,957 3,987 8,921 32,077 2,466 1,706 17,753
90th percentile 8,661 25,432 143,132 18,075 22,827 106,997 8,779 10,290 84,338
SD = standard deviation; IQR = inter-quartile range (difference between the 75
th
and 25
th
percentile estimates); *p<0.05 and **p<0.01.
a) Outcome was assumed to be negative binomial distributed while evaluating statistical significance of difference estimates between NEC versus control groups. b) Outcome was assumed to be t-distributed while
evaluating statistical significance of difference estimates between NEC versus control groups. c) Total ambulatory care costs included costs of laboratory, medical consultation, preventive care, ambulatory surgery,
home health services and professional payments. Overall healthcare costs include inpatient, ambulatory as well as prescription medication costs.
56

Surgical NEC survivors had significantly higher in-patient utilization and inpatient costs than
matched controls during all time periods (Table 3-4). The difference in the mean unadjusted
inpatient costs between surgical NEC and control groups was US $32,945, $21,160 and $6,796
per child corresponding to the difference in mean hospital days of 15, 9 and 1.5 days for the
periods 6–12, 12–24 and 24–36 months of age, respectively (p<0.01). The mean unadjusted
ambulatory care cost among surgical NEC survivors was US $6,432, $15,754 and $12,115 more
than that for matched controls, between 6–12, 12–24 and 24–36 months of age, respectively
(p<0.01). Home healthcare costs accounted for 44 to 60% of the difference in ambulatory care
costs observed over the follow-up period between surgical NEC and control groups. The
difference in the mean unadjusted all-inclusive healthcare costs between surgical NEC and
control groups was US $39,615, $37,652 and $19,776 for the period 6–12, 12–24 and 24–36
months of age, respectively (p<0.01).
3.6.4 Multivariate generalized linear modeling results
Multivariate modeling demonstrated that the incremental effect of the two NEC types and ELBW
status on the all-inclusive healthcare cost decreased over time (p<0.01 for the slope terms NEC
type and ELBW status by age-period). The healthcare cost models also predicted a decrease in the
incremental effect of NEC on healthcare costs in ELBW infants (p=0.09 and 0.02 for the
interaction effect of medical and surgical NEC, respectively, with ELBW status). The
specification that included all three interaction terms in the model (as shown in Equation 1 below)
had a superior fit over the other models tested, and, therefore, was used to obtain the adjusted
incremental healthcare cost estimates of each NEC type over matched controls.
57

Equation 1:
Healthcare Costs / 6 Months α f (NEC type, ELBW status, race gender, age-
period, NEC type Χ age-period, ELBW status Χ age-period, NEC type Χ ELBW
status, prior hospitalization)
The model-adjusted difference in all-inclusive health-care cost between each NEC type and
matched controls is shown in Figure 3-2. The adjusted mean incremental cost of medical NEC
survivors was $5,112 between 6–12 months of age (95% confidence interval (CI): $274 -$9,950;
p<0.05). The adjusted healthcare cost difference between medical NEC and matched control
groups was not statistically significant after 12 months of age. The adjusted mean incremental
healthcare cost per six months in the surgical NEC group over matched controls was US $18,274
(95% CI: 7,315 - 29,234; p<0.01), $14,067 (95% CI: 3,906 - 24,228; p<0.01) and $8,501 (95%
CI: -475 - 17,448; p=0.06) for the periods 6–12, 12–24 and 24–36 months of age, respectively.
Ambulatory care cost was the main driver of the healthcare cost differences between surgical
NEC and control groups beyond 12 months of age as shown in Figure 3-3. The predicted mean
healthcare costs of surgical NEC survivors and matched controls among sub-groups of children
born with birth-weights <1,000 grams and ≥1,000 grams are reported in Table 3-5.
58

Figure 3-2 Adjusted Incremental Total Healthcare Costs per 6-Months Incurred
by Medical and Surgical NEC Survivors Over Matched Controls from
6 Months to 3 Y ears of Age

59

Figure 3-3 Adjusted Incremental Inpatient and Ambulatory Care Costs per 6-
Months (± SE) Incurred by Surgical NEC Children Over Matched
Controls, 6 Months to 3 Y ears of Age

$14,894
$6,546
$1,570
$3,334
$8,018
$3,833
$0
$5,000
$10,000
$15,000
$20,000
$25,000
6-12 12-24 24-36
Δ Costs/6 mos (US$)
Age period (months)
Inpatient costs Ambulatory costs
60

Table 3-5 Predicted Costs per 6-Months for Surgical NEC Children Versus
Matched Controls Over Time Across Birth-Weights
Birth-weight <1000g Birth-weight ≥ 1000g
Age 6-12 12-24 24-36 6-12 12-24 24-36
Control group

Predicted costs 8,540 15,559 10,347 1,526 3,795 3,356
95% CI (low) 6,210 11,830 7,287 1,086 3,026 2,156
95% CI (high) $10,869 19,289 13,407 1,966 4,565 4,554
Surgical NEC group

Predicted costs 34,020 27,013 14,476 17,311 18,764 13,367
95% CI (low) 12,597 9,329 2,945 5,603 7,158 3,143
95% CI (high) 55,443 44,696 26,006 29,020 30,371 23,590
3.7 Discussion
Neonatal necrotizing enterocolitis has a high fatality rate among infants affected by the condition.
The morbidity and long term health outcomes among NEC survivors are highly influenced by the
pathological stage of NEC and the extent of damage to the intestines (Georgeson and Breaux
1992; Goulet et al 1991; Ladd et al 1998; Wilmore 1972). Surgical intervention is an important
surrogate for severity of NEC and the associated high risk of mortality and poor developmental
outcomes, regardless of the surgical procedure used (Adams-Chapman and Stoll 2006; Bisquera
et al 2002; Blakely et al 2008; Cikrit et al 1986; De-Souza et al 2001; Dzakovic et al 2001; Hintz
2005; Ladd et al 1998; Ricketts and Jerles 1990; Schulzke et al 2007; Simon 1994; Tejani et al
1978). Previous studies of healthcare costs associated with NEC have shown that both medically
and surgically treated NEC infants incur significantly higher inpatient hospital expenditures than
similar infants without NEC due to longer length of stay in neonatal intensive care units
(Bisquera et al 2002; Ganapathy et al 2012; Russell et al 2007; Underwood et al 2007). However,
61

the healthcare costs of NEC survivors over the long term have not been studied to date despite
increasing evidence of poor health out-comes among these children.
In this study we compared the healthcare utilization of 250 NEC survivors in the Medicaid
population to that of controls matched on prematurity and ELBW status, black race and presence
of birth defects, from 6 months to 3 years of age. When matching NEC infants to controls, we
found that patent ductus arteriosus was more frequently observed in the NEC group. The
association of PDA and NEC is well known and is thought to be due to excessive left to right
shunting leading to systemic hypoperfusion, a known risk factor for NEC (Hamrick and
Hansmann 2010). We found that 28 infants in the NEC group (9%) and 119 infants in the
matched control group (4%) had undergone surgical PDA ligation or division procedures. While
additional costs due to PDA surgery can be incurred during the initial hospitalization period, the
cost of PDA over the long-term should be no different in infants with or without NEC. In other
words, a PDA-NEC association that was observed in the data would not have an impact on the
main findings of this paper.
We found that medical NEC survivors incurred $5000 more in healthcare costs on average than
matched controls between 6 to 12 months of age. These incremental costs were mainly driven by
ambulatory care expenses, possibly attributable to management of artificial GI openings and
follow-up care received for other develop-mental problems observed during this period. These
included failure to thrive, feeding difficulties, BPD and NDD. However, the healthcare costs of
medical NEC survivors did not differ from matched controls after 12 months of age. These results
indicate that the likelihood of experiencing developmental complications leading to increased
utilization of healthcare resources over the long term (>1 year of age) is not significant in
comparing medical NEC survivors to matched controls.
62

On the other hand, the all-inclusive healthcare costs among surgical NEC survivors continued to
be higher than matched controls beyond 6 months with the adjusted incremental costs being
statistically significant up to 2 years of age. The incremental costs of surgical NEC between 2 to 3
years were still substantial and the lack of statistical significance could be due to the very small
number of surgical NEC infants remaining in this time period (n=35). Our findings show that
surgical NEC survivors incurred an average $60,000 more in healthcare costs than matched
controls over the period from 6 months to 3 years of age.
The high costs among surgical NEC survivors were initially driven by inpatient expenditures.
However, the frequency of hospital admissions and level of inpatient expenditures decreased over
time. Home healthcare and other ambulatory care expenditures were the main drivers of costs
among surgical NEC children from 1 to 3 years of age. The net difference in costs between
surgical NEC children and matched controls was smaller in the extremely low birth weight group
(BW <1,000 grams) compared to the cost difference found in children born with BW ≥1,000
grams. A similar trend was noted in the risk of chronic health conditions wherein the odds ratios
for the association of surgical NEC with developmental health conditions were smaller for infants
with BW <1,000 grams compared to BW ≥1,000 grams (though only the combined odds ratios
were reported due to the very small number of NEC infants with BW <1,000 grams). These
results are to be expected given the already higher rate of complications in extremely low birth
weight status and the consequent decrease in the marginal effect of surgical NEC on healthcare
costs in these children. Nevertheless, the results clearly show that children born with extremely
low birth weight and who survive severe NEC incurred higher healthcare costs than children with
only one of these risk factors.
63

The intensity of healthcare use and costs among surgical NEC survivors could be driven by one
or more factors such as: treatment for post-surgical complications (e.g., short bowel syndrome
(SBS) / intestinal malabsorption which was seen more often in the surgical NEC group), costs
associated with nutrition (e.g., length of total parenteral nutrition (TPN) support required by
survivors and the complications associated with TPN), treatment of infections associated with
ostomies (a significantly higher proportion of surgical NEC survivors lived with open ostomies
for a significant period of time), and costs of care for very frequently reported conditions such as
failure to thrive, NDD and nutritional and metabolic disturbances.
In a study among infants with short bowel syndrome, Spencer, et al. reported the average costs of
SBS between year 1 to year 5 to be US $250,000-300,000 per year and that parenteral nutrition
alone contributed to roughly $200,000 each year (Spencer et al 2008). These costs appear to be
much higher than the 90th percentile of costs among surgical NEC infants that we observed in the
Medicaid cohort. The higher costs in the Spencer, et al. (2008) study are partly due to the use of
billable Medicaid charges for home healthcare services, whereas we used the actual Medicaid
paid amounts which typically represent very low reimbursement rates for these services. Also,
children with SBS represent those who have the most severe health status among surgical NEC
survivors. Our surgical NEC sample had 13 out of 62 patients (21%) between 6–12 months and 7
out of 35 patients (20%) between 24–36 months receiving care for intestinal malabsorption and
the average costs in this sub-group would have been much higher. Also, it could not be
ascertained from the claims data how many patients had severe SBS as defined in the Spencer, et
al. (2008) study (i.e. loss of ≥70% of small intestinal length, ≥2 months of parenteral nutrition
dependence, etc.).
64

Regardless of the data source, the healthcare costs reported in this paper are to be treated as
highly conservative, since very expensive treatments such as small bowel transplantation were
not accounted for in the analyses (because these procedures were not covered under Texas
Medicaid during the study period). Transplant procedures may be required for long term survival
in a small proportion of surgical NEC survivors with failure of intestinal function (Vennarecci et
al 2000). Also, a significant proportion of surgical NEC infants in the Medicaid sample were
found to have NDD and the odds ratios were comparable to those reported in other studies (Hintz
et al 2005; Rees et al 2007; Schulzke et al 2007). The extent to which NDD influences direct
healthcare costs is not very clear, although NDD can have a significant impact on diagnostic
screening tests, physical and occupational therapy, and special education costs. More research
using multiple data sources is needed to specifically understand the economic impact of
neurodevelopmental delay among survivors of surgical NEC.
While the healthcare costs discussed above are highly relevant from a payer perspective, it should
be remembered that the long term costs from a societal perspective would also account for the
value of lives lost due to mortality attributed to NEC during the first 6 months of life. The value
of a statistical life (VSL) is estimated at $7.4 million according to the 2002 Environmental
Protection Agency (EPA) estimates (Johansson 2006). Medicaid covers 40% of live births in the
U.S. If the NEC mortality rate observed in the current study were to be applied to the overall
Medicaid population in the U.S., approximately 700–900 Medicaid infants would be expected to
die be-cause of NEC annually and the total economic value of lives lost to NEC would be 5.2 –
6.6 billion US$.
This paper investigates the real world utilization of a longitudinal cohort of NEC survivors.
Unfortunately, attrition in the sample resulted in smaller cohorts over time. However, the extent
65

of attrition in the NEC and control groups was comparable and the balance in the baseline
characteristics between the NEC and control cohorts was maintained over time. Additionally, a
set of exploratory analyses (not shown in this paper) that was conducted to evaluate the
probability of attrition over time showed that demographics and health outcomes in the previous
time periods (e.g., healthcare costs, disability status and hospitalizations) together explained less
than 10% of total variation in attrition in the Medicaid cohort. This suggests that attrition in the
Medicaid cohort is predominantly caused by extrinsic factors that do not impact health outcomes
(e.g., loss of Medicaid eligibility due to income changes, availability of employer insurance,
migration to a different state, etc.).
3.8 Limitations
A significant limitation of this study could be that our findings may not be generalizable to the
universe of NEC survivors. This is because the study’s findings were derived from a sample of
children that belonged to low-income families with a higher proportion of Hispanic children and,
possibly, with a higher baseline risk for poor health status than NEC survivors in the
commercially insured population. Additional research using healthcare utilization data obtained
from a representative sample of commercially insured NEC survivors is needed in order to
improve the generalizability of our findings. Nevertheless, given the fact that Medicaid is one of
the largest payers of healthcare for children in the US, the estimates from this population are very
useful in understanding the overall economic burden of surgical NEC from a US public payer
perspective. Be-sides generalizability, the study also suffers from some of the classical limitations
of using claims data that are not collected for research purposes. Most importantly, these data do
not contain specific information that may be of potential research interest, such as birth order,
maternal characteristics such as education, income and breast feeding practices. Considering that
66

data on the NEC population may be significantly hard to find in practice or simply too expensive
to collect prospectively, we consider that the benefit of finding easy to collect longitudinal
economic data outweighs the significant challenges of using retrospective claims data. Further
research is needed on long term costs that are not captured by medical claims, such as special
education costs and caregiver productivity costs.
3.9 Conclusions
The healthcare costs of children who survived surgical necrotizing enterocolitis during birth
hospitalization are substantial over the early childhood development period. Understanding the
economic burden of NEC in the long term would aid healthcare providers, policy makers and
payers to make informed decisions in pro-viding care for infants at high risk for NEC. Further re-
search on this topic is needed.
3.10 References
1. Adams-Chapman I, Stoll BJ. 2006. Neonatal infection and long-term neurodevelopmental
outcome in the preterm infant. Curr Opin Infect Dis 19:290–297.
2. Althauser RP., Rubin D. The computerized construction of a matched sample. Am J Sociol
(1970): 325-346.
3. Bell MJ, Ternberg JL, Feigin RD, Keating JP, Marshall R, Barton L, Brotherton T. 1978.
Neonatal necrotizing enterocolitis. Therapeutic decisions based upon clinical staging. Ann
Surg 187:1–7.
4. Bisquera JA, Cooper TR, Berseth CL. 2002. Impact of necrotizing enterocolitis on length of
stay and hospital charges in very low birth weight infants. Pediatrics 109:423–428.
67

5. Blakely ML, Gupta H, Lally KP. 2008. Surgical management of necrotizing enterocolitis and
isolated intestinal perforation in premature neonates. Semin Perinatol 32:122–126.
6. Carpenter RG. Matching when covariables are normally distributed. Biometrika 64, no. 2
(1977): 299-307.
7. Cikrit D, West KW, Schreiner R, Grosfeld JL. 1986. Long-term follow-up after surgical
management of necrotizing enterocolitis: sixty-three cases. J Pediatr Surg 21:533–535.
8. Coca-Perraillon M. 2007. Local and global optimal propensity score matching. In Statistics
and Data Analysis. SAS Global Forum 1–9. Downloaded from:
http://www2.sas.com/proceedings/forum2007/185-2007.pdf.
9. Cochran WG., Rubin DB. Controlling bias in observational studies: A review. Sankhyā:
Indian J Statistics, Series A (1973): 417-446.
10. De-Souza JC, Da-Motta UI, Ketzer CR. 2001. Prognostic factors of mortality in newborns
with necrotizing enterocolitis submitted to exploratory laparotomy. J Pediatr Surg 36:482–
486.
11. Dzakovic A, Notrica DM, Smith EO, Wesson DE, Jaksic T. 2001. Primary peritoneal
drainage for increasing ventilatory requirements in critically ill neonates with necrotizing
enterocolitis. J Pediatr Surg 36:730–732.
12. Fanaroff AA, Hack M, Walsh MC. 2003.The NICHD neonatal research network: changes in
practice and outcomes during the first 15 years. Semin Perinatol 27:281–287.
68

13. Ganapathy V, Hay JW, Kim JH. 2012. Costs of necrotizing enterocolitis and cost-
effectiveness of exclusively human milk-based products in feeding extremely premature
infants. Breastfeed Med 7(1):29–37.
14. Georgeson KE, Breaux CW. 1992. Outcome and intestinal adaptation in neonatal short-bowel
syndrome. J Pediatr Surg 27:344–350.
15. Goulet OJ, Revillon Y, Jan D, De-Potter S, Maurage C, Lortat-Jacob S, Martelli H, Nihoul-
Fekete C, Ricour C. 1991. Neonatal short bowel syndrome. J Pediatr 119:18–23.
16. Grave GD, Nelson SA, Walker WA, Moss RL, Dvorak B, Hamilton FA, Higgins R, Raju
TNK. 2007. New therapies and preventive approaches for necrotizing enterocolitis: report of
a research planning workshop. Pediatr Res 62:510–514.
17. Hamrick SEG, Hansmann G. 2010. Patent ductus arteriosus in the preterm infant. Pediatrics
125:19–26.
18. Hintz SR, Kendrick DE, Vohr BR, Poole WK, Higgins RD. 2005. Changes in
neurodevelopmental outcomes at 18 to 22 months’ corrected age among infants of less than
25 weeks’ gestational age born in 1993–1999. Pediatrics 115:1645–1651.
19. Horbar JD, Carpenter J. 2008. 2007 Very low birth weight database summary. Burlington,
VT: Vermont Oxford Network.
20. Imbens GW, Rubin DB. 2009. Design in observational studies: matching to ensure balance in
covariate distributions. In Causal Inference Part II, Chapter 15. Cambridge University Press.
69

21. Janek KL, Ghahremani K. 2013. Texas Medicaid Managed Care. In Texas Medicaid and
CHIP in Perspective, Chapter 7. Austin, TX: Texas Health and Human Services Commission.
22. Johansson P. 2006. On the definition and estimation of the value of a statistical life.
Departmental Working Papers 2006-23, Department of Economics, Management and
Quantitative Methods at Universita degli Studi di Milano. Downloaded from:
http://wp.demm.unimi.it/tl_files/wp/2006/DEMM- 2006_023wp.pdf.
23. Ladd AP, Rescorla FJ, West KW, Scherer L, Engum SA, Grosfeld JL. 1998. Long-term
follow-up after bowel resection for necrotizing enterocolitis: factors affecting outcome. J
Pediatr Surg 33:967–972.
24. Manning WG, Mullahy J. 2001. Estimating log models: to transform or not to transform? J
Health Econ 20:461–494.
25. Mantel N, Haenszel W. 1959. Statistical aspects of the analysis of data from retrospective
studies of disease. J National Canc Inst 22:719–748.
26. Medicaid coverage of births in Texas (March Of Dimes) 2012. Downloaded from:
http://www.marchofdimes.com/peristats on 12/6/2012.
27. Neu J, Walker WA. 2011. Necrotizing enterocolitis. New Eng J Med 364:255–264.
28. Ng S. 2001. Necrotizing enterocolitis in the full-term neonate. J Paediatr Child Health 37:1–
4.
29. Petrosyan M, Guner YS, Williams M, Grishin A, Ford HR. 2009. Current concepts regarding
the pathogenesis of necrotizing enterocolitis. Pediatr Surg Int 25:309–318.
70

30. Pierro A, Hall N 2003. Surgical treatments of infants with necrotizing enterocolitis. Seminars
in Neonatology. SN 8:223–232.
31. Rees CM, Pierro A, Eaton S. 2007. Neurodevelopmental outcomes of neonates with
medically and surgically treated necrotizing enterocolitis. Arch Dis Child Fetal Neonatal Ed
92:F193–F198.
32. Ricketts RR, Jerles ML. 1990. Neonatal necrotizing enterocolitis: experience with 100
consecutive surgical patients. World J Surg 14:600–605.
33. Rosenbaum PR, Rubin DB. 1985. Constructing a control group using multivariate matched
sampling methods that incorporate the propensity score. Am Stat 39:33–38.
34. Rubin DB. Using multivariate matched sampling and regression adjustment to control bias in
observational studies. Journal of the American Statistical Association 74, no. 366a (1979):
318-328.
35. Rubin DB. Bias reduction using Mahalanobis-metric matching. Biometrics (1980): 293-298.
36. Russell RB, Green NS, Steiner CA, Meikle S, Howse JL, Poschman K, Dias T, Potetz L,
Davidoff MJ, Damus K, Petrini JR. 2007. Cost of hospitalization for preterm and low birth
weight infants in the United States. Pediatrics 120:e1–e9.
37. Schulzke SM, Deshpande GC, Patole SK. 2007. Neurodevelopmental outcomes of very low-
birth-weight infants with necrotizing enterocolitis: a systematic review of observational
studies. Arch Pediatr Adolescent Med 161:583–590.
71

38. Simon NP. 1994. Follow-up for infants with necrotizing enterocolitis. Clinics In Perinatol
21:411–424.
39. Spencer AU, Kovacevich D, McKinney-Barnett M, Hair D, Canham J, Maksym C,
Teitelbaum DH. 2008. Pediatric short-bowel syndrome: the cost of comprehensive care. Am J
Clin Nutr 88:1552–1559.
40. Tejani A, Dobias B, Nangia BS, Mahadevan R. 1978. Growth, health, and development after
neonatal gut surgery: a long-term follow-up. Pediatrics 61:685–693.
41. Underwood MA, Danielsen B, Gilbert WM. 2007. Cost, causes and rates of rehospitalization
of preterm infants. J Perinatol 27:614–619.
42. Vennarecci G, Kato T, Misiakos EP, Neto AB, Verzaro R, Pinna A, Nery J, Khan F,
Thompson JF, Tzakis AG. 2000. Intestinal transplantation for short gut syndrome attributable
to necrotizing enterocolitis. Pediatrics 105:E25.
43. Williams R. 2012. Using the margins command to estimate and interpret adjusted predictions
and marginal effects. Stata J 12:308–331.
44. Wilmore DW. 1972. Factors correlating with a successful outcome following extensive
intestinal resection in newborn infants. J Pediatr 80:88–95.
72

4. Application of Counterfactual Decomposition Methods with Fixed
Effects to Evaluate the Impact of Prematurity on Long-Term
Healthcare Costs of Children Belonging to Low-Income Families
4.1 Abstract
Background: The long-term healthcare cost among preterm survivors, classified by gestational
age, has not been studied well. Previous studies have largely ignored the role of unobserved
heterogeneity, leading to biased inferences regarding the treatment effect of preterm births.
Objectives: To evaluate the incremental healthcare costs and drivers of costs, among preterm
survivors, within a US public payer population from 6 months to 5 years of age.
Study Design & Methods: Children born moderate to late preterm (33-36 weeks of gestational
age), very to extremely preterm (< 33 weeks of GA) or full-term and enrolled in the Texas
Medicaid fee-for-service program were identified from their birth hospitalization claims. An
unbalanced panel structure was used to summarize data on time-varying covariates and outcomes
within each period. Two-part model, with fixed effects in each step, was used to estimate
logarithm of total healthcare costs, separately, among preterm and control groups. The
incremental effect of preterm births was obtained using Blinder-Oaxaca type decompositions and
standard errors were estimated by bootstrapping. Attrition weighting was used to account for non-
random attrition. Cumulative costs per child were estimated after applying a discount rate of 3%
to costs incurred beyond 12 months of age.
Results: The incremental impact of moderate to late preterm birth, per se, was not significant
beyond 6 months, after controlling for unobserved child specific factors. When combined with an
73

additional risk factor, birth weight < 2,500 grams, the average incremental cost of M/L preterm
birth was US$ 560 per child for the total period between 6 and 60 months of age (p < 0.05). Over
the same period, the average incremental cost of preterm birth at < 33 weeks of GA was $4,800
per child (p < 0.001). Neurodevelopmental delay, asthma/bronchial disorders, respiratory
symptoms and refractory vision disorders were top drivers of the healthcare cost difference
between preterm and full-term survivors. Weighting for attrition had a significant impact on the
treatment effect estimates.
Conclusions: Controlling for unobserved heterogeneity while evaluating the long-term
consequences of preterm birth should be considered. M/L preterm with birth weight < 2,500
grams and V/E preterm born children bear significant financial impact to Texas Medicaid
program through 5 years of age. These findings have important policy implications.

74

4.2 Introduction
The impact of prematurity (infant births occurring at < 37 completed weeks of gestational age)
and its complications on long-term direct medical care costs, at best, have not been well studied
(Zupancic JA. 2006; Stavros Petrou, Eddama, and Mangham 2011). Longitudinal health
insurance claims databases provide a cost-effective approach to evaluate the medical costs of a
large number of preterm infants in comparison with that of a cohort of infants born full-term.
Claims data, however, do not contain information on maternal, socio-economic or genetic factors
and these omitted variables may lead to biased and inconsistent estimates of the impact of
prematurity on healthcare cost. More specifically, the impact of subject-level heterogeneity (due
to a combination of hereditary, familial and socio-economic factors) on prematurity and pediatric
developmental outcomes has been well studied in the outcomes literature (Rosenzweig and
Wolpin 1988; Brown et al. 2013). Therefore, ignoring such heterogeneity that is mostly
unobserved causes an “identification” problem, leading to biased inferences about the treatment
effect.
When suitable identifying instruments to eliminate omitted variable bias are not easily found,
panel data fixed effects modeling techniques help to adequately control for time-invariant
unobserved heterogeneity. However, if the treatment of interest in the model changes over the
individual dimension but not over the time dimension (as in the case of prematurity), inference
regarding the treatment effect cannot be made using the fixed-effects technique unless we impose
restrictions on the fixed effects term, allowing it to be only partially correlated with the observed
characteristics, or not correlated with the particular time invariant treatment variable of interest
(Hsiao 2003; pp 51-53). Rather we address this econometric problem by estimating separate fixed
effects models of healthcare costs for preterm and full-term groups and use counterfactual
75

decomposition analyses of the Blinder-Oaxaca type to estimate the difference in healthcare costs
between preterm and full-term born children. Blinder-Oaxaca type decomposition methods allow
us to segregate the mean difference in healthcare cost difference between the preterm and full-
term groups into a part that is totally explained by differences in the group characteristics
(endowments effect), and a part that is due to coefficients (coefficients effect), which captures
any residual group-level differences in healthcare costs (Blinder 1973; R. Oaxaca 1973; Bauer
and Sinning 2008). Upon controlling for bias due to unobserved individual heterogeneity in the
healthcare cost estimation, the coefficients effect obtained from the B-O decomposition may
provide a consistent estimate of the treatment effect of prematurity, provided the conditional
distributions of all other time invariant factors that may not be controlled for in the estimation
(possibly those that are not causally related to prematurity), remains similar between the preterm
and full-term groups (Fortin, Lemieux, and Firpo 2011).
4.2.1 Background & Brief Review of Literature
4.2.1.1 Prematurity and the initial care cost dilemma
“Prematurity” is a term that is used to classify infant births occurring at less than 37 completed
weeks of gestational age (WHO 2012). Every year, nearly 500,000 babies, i.e., 1 of every 8
infants born in the United States are born premature (JA Martin, Hamilton, and Sutton 2010). The
survival of preterm infants, especially those born extremely preterm, has drastically improved
over the past few decades due to significant advances in neonatology. For instance, early use of
antenatal corticosteroids, assisted ventilation techniques and provision of exogenous surfactants
that prevent respiratory failure among premature babies have drastically reduced mortality rates
and have extended the lower limit of viability to children born as early as < 23 weeks of
gestational age (Saigal and Doyle 2008; Ho and Saigal 2005). Currently greater emphasis is being
placed on therapeutic strategies and early childhood interventions that target tertiary prevention,
76

i.e. interventions aimed at reducing chances of long-ranged morbidity and improving health
outcomes among children surviving beyond the neonatal period (Iams et al. 2008).
While these developments in premature infant care hold huge promise, they also present with
challenges in healthcare decision making that concerns with the financial impact of these new
interventions on a constrained healthcare budget (Zupancic JA. 2006). For instance, Texas
Medicaid paid nearly $446 million, only for the neonatal intensive care expenses of premature
infants in 2009 (Seuhs 2012). In a healthcare market place that is increasingly moving towards a
value-based reimbursement model, payer and policy makers demand evidence for the cost-
effectiveness of NICU or early childhood interventions that are used to prevent long-term
morbidity. However, these decisions are not easily made due to lack of data on the economic
outcomes and healthcare cost drivers among preterm infants, over the long term.
As part of its policy imperatives the Institute of Medicine (IOM) task force for Understanding
Prematurity and Assuring Healthy Outcomes, has identified the evaluation of long term clinical
and economic outcomes of preterm birth as high unmet needs where further research is required
(Behrman and Butler 2007, Section IV).
4.2.1.2 Gaps in the literature on long-term economic costs of preterm births
Preterm birth may expose an infant to multiple short-term neonatal complications including
respiratory distress syndrome, neonatal infections, feeding intolerance, patent ductus arteriosus
and central nervous system injuries (Ho and Saigal 2005; Saigal and Doyle 2008; Clark et al.
2001; Adams-Chapman and Stoll 2006; Smith et al. 2005). A commonly used proxy for severity
of complications is the gestational age at birth, with very or extremely preterm infants (typically
born at <33 weeks of gestational age) often facing a higher risk of respiratory, gastrointestinal,
infectious disease and neurodevelopmental complications as compared to full-term or other
77

preterm infants. Despite the vast literature on clinical outcomes, there are only handful of studies
which have evaluated the costs of prematurity beyond the early survival period (Stavros Petrou,
Eddama, and Mangham 2011; Zupancic JA. 2006).
Available economic studies have mostly used birth weights as a proxy for prematurity rather than
gestational age, focusing mostly on children born with birth-weights < 1,500g. While low birth-
weight is an important risk factor for neonatal mortality and morbidity, gestational age at birth is
a more relevant proxy for preterm births and consequent long-term morbidity (Behrman and
Butler 2007). The few studies available in the literature varied widely with respect to geography,
sample size, methodology and time horizon over which costs were reported. The available studies
mostly report potential, rather than the actual costs, based on costs of disabling conditions that
may be attributable to extreme immaturity (e.g., cerebral palsy or mental retardation) (Yeargin-
Allsopp et al. 1992; Stavros Petrou, Eddama, and Mangham 2011; Mangham et al. 2009). Recent
studies have shown that children born moderate to late preterm (33-36 weeks of gestational age),
who constitute the vast majority of preterm born infants, have a higher risk of long-term
respiratory and/or neurodevelopmental problems than compared to full-term infants (Engle,
Tomashek, and Wallman 2007; Moss 2006; Clements et al. 2007; Morse et al. 2009; Boyle and
Boyle 2013). However the healthcare costs of this sub-group beyond infancy have not been
studied. Further, the drivers of long-term healthcare costs among very/extremely preterm born
and moderate or late preterm born children, at best, has not been well studied.
Finally, previous studies have mostly ignored the possibility of selection bias due to potential
confounders that are not observed among observational databases. For instance, studies have
shown that the combined influence of socioeconomic and mother’s health status may influence
neurodevelopmental outcomes in young children (McGauhey et al. 1991; Downey and Coyne
78

1990) though, such information is hard to find among most retrospective databases. Statistical
approaches that ignore such heterogeneity at the individual level while evaluating long-term
economic outcomes, may lead to biased and inconsistent estimates of the impact of preterm births
(Brown et al. 2013). Other methodological considerations, when dealing with real world
healthcare cost data, such as, handling highly skewed cost distributions, zero mass, and non-
random attrition in the case of longitudinal studies, are mostly not addressed in previous works on
this topic with the exception of a few studies (S Petrou, Sach, and Davidson 2001; Stavros Petrou
2005; Stavros Petrou, Eddama, and Mangham 2011; Stavros Petrou and Khan 2012). These
methodological weaknesses, further, could adversely affect the precision and/or consistency of
incremental cost estimates of preterm births in the short and long term.
4.3 Study Objectives
We studied the real world medical costs of preterm born children among two gestational age
categories - children born before 33 weeks and between 33-36 weeks of gestational age, in
comparison to children born full-term within a US public payer population for the period between
6 and 60 months of age. Using regression decomposition methods we evaluated the treatment
effects of prematurity on the healthcare cost difference between preterm and full-term born
children after controlling for individual fixed effects. We hypothesize that preterm birth may have
a heterogeneous impact on healthcare costs among children belonging to similar socio-economic
background, even after accounting for differences in the prevalence of disability and unobserved
child specific effects. We also studied the diagnostic drivers of healthcare cost differences
between preterm and full-term born children.
79

4.4 Study Design & Methods
4.4.1 Description of data
We used the longitudinal de-identified Medicaid health insurance claims data of children, born in
Texas between January 2002 and December 2003, and enrolled in the Texas Medicaid fee-for-
service (FFS) program up to December 2008. Non-disabled children under the age of 21,
belonging to low-income families constituted the majority of the Medicaid case-load (>50%)
during any given financial year (March of Dimes 2012). The benefits covered by Texas Medicaid,
for children are comprehensive with almost no limits to these benefits provided under age 21.
Greater than 50% of live births in Texas were covered by Medicaid (March of Dimes 2012).
Also, unlike in other large states such as California, Illinois or New York, significant proportion
of Texas Medicaid clients were managed under traditional fee-for-service arrangements during
the study period (2002 to 2008), making this data source ideal for our research (“Managed Care”
2013).
The enrollment, inpatient, outpatient, prescription, physician and other healthcare provider (HCP)
service claims, up to 5 years of age of children in the 2002-2003 cohort were obtained from the
claim header files. Enrollment data contained information on periods of eligibility, demographics
(Race, gender, year of birth, residence zip-code at birth) and category under which a child was
eligible for coverage (e.g. aid for families with dependent children (AFDC), children receiving
social security income (SSI), covered under an expansion program, etc.). The claim header files
contained claim level information on –ICD-9-CM (International classification of Diseases,
Clinical Modification) diagnoses, procedures, DRG (Diagnostic related group) codes (among
inpatient claims), type of outpatient/provider service (e.g. consultation, early periodic screening,
vision, lab & radiology, etc.), prescriptions filled (identified through National Drug Codes),
80

prescription refill dates, and the paid dollar amounts for all claims. Mortality data was captured
only when deaths took place within hospital (using discharge status of hospitalization).
4.4.2 Identification and classification of preterm infants
All singleton live births that occurred between January 2002 and December 2003 were identified
using ICD-9-CM code V30.0 from a newborn’s birth hospitalization claim. Infants who were
born preterm were identified using ICD-9-CM codes 765.0 (“disorders related to extreme
immaturity of infant”) or 765.1 (“disorders related to other preterm infants”) from inpatient
claims at the time of birth (Appendix 1). A 50% random sample of infants with neither 765.xx
nor 764.xx (“slow fetal growth and malnutrition”) diagnosis was used, and additional exclusions
were applied to obtain a cohort of healthy full-term newborn infants (see Appendix 2).
The gestational age at birth is an indicator of the degree of prematurity that predicts the long-term
morbidity among preterm infants, with most studies showing an inverse relationship between
gestational age and outcomes during the first year of birth (Smith et al. 2005; Kneyber, Plötz, and
Kimpen 2005; Krilov et al. 2009; Fledelius and Greisen 1993; Yeargin-Allsopp et al. 1992). An
additional diagnosis code 765.2x, that indicated the gestational age of preterm infant, is generally
used while recording prematurity on insurance claims data (Staff, HSS Inc. 2003). The fifth digit
of this code was used to classify preterm infants into two gestational age categories as per the
WHO classification of preterm births (WHO 2012): a) preterm infants born at ≤ 32 completed
weeks of gestational age (“very to extremely preterm infants”); and b) preterm infants born
between 33 and 36 completed weeks of gestation (“moderate to late preterm infants”).
4.4.3 Exclusions
Preterm infants whose gestational age was missing or greater than 36 weeks in the inpatient files
at the time of birth were excluded from the dataset. Infants with diagnoses of fetal growth
81

retardation and/or malnutrition (ICD-9-CM 764.XX) and heavy-for-gestational age (766.XX) at
the time of birth were also excluded, as these conditions may be linked to a different set of causal
effects on health outcomes of a child. Similarly, infants with the following rare newborn
conditions that are not expected to be different between preterm and full-term babies were
excluded: birth trauma (767), intra-uterine hypoxia (768), neonatal hemorrhage (772), newborn
hematological disorders (776) and other ill-defined newborn conditions (779). Miscellaneous
exclusions that had to do with missing eligibility records and date of birth mismatch were also
applied.
4.4.4 Criteria for determining attrition
Children in the original sample were continuously enrolled for different lengths of time over the
60-month period. Since there was insufficient data on the enrollment file to determine when an
individual dropped out of the fee-for-service cohort, attrition in the cohorts was determined using
healthcare utilization information, at the end of 0-6, 7-18, 19-36 and 37-60 month periods since
the time of birth. The criteria for retention (non-attrition) during each of the four time intervals
was that - a child must have had at least one outpatient visit, or a healthcare provider claim, at any
point during each time interval. The criteria was based on the fact that, newborn children who are
alive, would have at least one office or healthcare provider visit during each of these time
intervals for a developmental screening test, (or) an immunization shot that is considered
mandatory per standard recommendations (CDC 2014). An exception to this criteria was made if
death of a child occurred during a particular time-period, in which case, an inpatient hospital
claim in that time period was sufficient to determine non-attrition.
4.4.5 Database set-up
The raw dataset was ultimately converted into an unbalanced panel framework comprising of a
patient-per-six-month structure, starting from birth up to 60 months of age. A six-month
82

observation period was considered to be optimal for following the incidence of various chronic
conditions in the cohort, as the child grew older. The choice of a six-month unit of observation
was also motivated by the fact that, the average time to renewal of insurance eligibility in this
population was approximately 180 days. Therefore, a 6-month unit of time was a natural choice
to follow any change in the status, and category of insurance coverage due to disability or change
in family income of the individual. A rich set of dummy variables for diagnoses during each 6-
month period, including early childhood respiratory, central nervous system (CNS),
neurodevelopmental, endocrine, nutritional, metabolic, gastrointestinal and infectious disease
conditions was developed using the ICD-9-CM classification of diagnoses. As a final step in data
preparation, the cumulative inpatient, outpatient, provider and prescription paid dollar amounts
were estimated per each six-month period and were adjusted for inflation to constant 2011 US
dollars using the medical consumer price index.
4.4.6 Outcomes
The main outcome of interest in this study was the incremental all-inclusive total healthcare cost
per child, for the period from 6 to 60 months of age. This is a grand summary of total inpatient
and ambulatory care costs within each 6-month period and the cumulative healthcare cost over
the study duration (as explained in section 4.4.16). The total inpatient hospital cost included
Medicaid paid amounts for inpatient hospitalization and inpatient prescription medication claims
in each 6-month period. Total ambulatory care cost was the sum of Medicaid paid amounts for
outpatient, physician, other healthcare provider (including home health care) services and
outpatient prescription medications during each 6-month period.
83

4.4.7 Explanatory variables
Prematurity by gestational age: The treatment of interest in this analysis was prematurity,
classified into two groups - “very preterm/extremely preterm infants” and “moderate/late preterm
infants”, based on gestational ages as explained in Section 4.2.
9

The following time-varying variables were developed for the analyses of healthcare cost
differences between preterm and full-term children:
4.4.7.1 SSI vs. other eligibility categories
When children became disabled, they qualified for Medicaid coverage as recipients of social
security income (SSI), whereas, a vast majority of children qualified for Medicaid coverage
primarily based on their family income and/or structure (under the Aid for Families with
Dependent Children (AFDC) program) or through various other (income based) expansion
programs.
10
A dummy variable that indicated whether a child was eligible for Medicaid coverage
through SSI or other programs (0 = other, 1 = SSI) during each period was used as a proxy for
disability status. A more detailed categorical variable indicating the eligibility category at time ‘t’
(1 = SSI, 2 = AFDC/cash grant and 3 = missing eligibility status in that period vs. the reference
group “children eligible through expansion programs”) was used in regression analyses to study
the impact of change in poverty levels on healthcare utilization during the course of the panel
(e.g. cash grant recipient to 133 percent poverty level, i.e. coverage through expansion).
9
Relevant sub-groups were also analyzed as part of sensitivity analyses (see section 4.5.11)
10
The Ribicoff program implemented in 1984 and the passage of the Omnibus Budget Reconciliation Acts (OBRA) of 1986, 1987
and 1990 have expanded Medicaid eligibility to infants, and young children up to 18 years of age over and above the income
eligibility limits for AFDC (e.g. “the 133 percent” and “100 percent level” expansions). OBRA 1987 also required states to cover
children who lived in families with incomes below the AFDC income threshold, regardless of family structure (for e.g. children in
foster care, adopted children, children in institutions or inpatient psychiatric facilities, etc.)
84

4.4.7.2 Hospitalization in the previous period
Since prior hospitalization may have a spillover effect on healthcare utilization in the current
period, hospitalization status in the previous period (t-1) was used as additional control variable
in regression analyses. Since all children had hospitalization records at the time of birth, only
rehospitalizations were considered for the period where ‘t-1’ represented the first 6 months of
birth.
4.4.7.3 Age
Age of the child corresponds with the panel period. Age was specified as a categorical variable
with 6-12 months as the reference category.
4.4.7.4 Congenital heart defects
Since the Medicaid population had a high incidence of congenital heart defects, where most
except patent ductus arteriosus, may not be directly associated with preterm births, a dummy
variable indicating claims for congenital heart defects other than patent ductus arteriosus at time
‘t’ was included as a control variable in the regressions analyses.
4.4.7.5 Diagnostic drivers
Number of medical conditions, chiefly, the ones related to respiratory, central nervous system
(CNS) or neurodevelopmental problems drive high healthcare costs among young children
(Newacheck and Taylor 1992; Newacheck 1994). In the analyses to evaluate the diagnostic
drivers of healthcare cost difference between preterm and full-term born children (described in
Section 4.4.14) the impacts of following list of chronic conditions, observed during the early
childhood developmental period on healthcare costs were studied (a complete list of the
conditions and their ICD-9-CM codes is provided in Appendix 1).
85

4.4.7.5.1 Respiratory conditions
We studied the impacts of four sets of conditions that were of interest in this category - asthma &
bronchial disorders (bronchiolitis and bronchitis), wheezing & other common respiratory
symptoms frequently occurring during childhood, lower respiratory tract infections (includes
respiratory syncytial virus and pneumonia infections) and upper respiratory infections (otitis
media, laryngitis, sinusitis & rhinitis).
4.4.7.5.2 Central nervous system (CNS) disorders
Three sets of disorders were studied in this class - a set of rare diseases including cerebral
degenerations manifest during early childhood, cerebrovascular disorders, hemiplegia and
hemiparesis; seizures (epilepsies) and motor disorders (cerebral palsy, paralytic syndromes).
4.4.7.5.3 Neurodevelopmental delay
Motor delay, coordination disorders, speech, learning delay and mental retardation were studied
in this category.
4.4.7.5.4 Vision disorders
Visual field disturbances, nystagmus and refractory disorders were studied in this category.
4.4.7.5.5 Delay in physiological growth
Delayed milestones and failure to thrive were studied in this category.
4.4.7.5.6 Gastrointestinal disorders
GI functional disorders such as gastroesophageal reflux, malabsorption, colic, constipation and
diarrhea were studied in this category.
86

4.4.7.5.7 Miscellaneous conditions
The impact of conditions such as, nutritional deficiencies
11
, endocrine and metabolic disorders,
miscellaneous infections
12
, hernias and gastrointestinal mechanical defects13 were also studied.
Indicator variables were used to represent whether or not each condition (or set of conditions)
was observed during a given time period (0 = “not observed” and 1 = “observed” at time ‘t’ that
allows to study the impact of the change in status of this variable from 0 to 1). This approach was
preferred over static coding of the disease variables (i.e., coding disease as “observed” from the
time it was first diagnosed to the end of the panel), in order to account for differences in
utilization pattern of the various chronic conditions that is likely to vary over the time period of
the panel.14 We conducted additional sensitivity analyses to test the impact of the disease status
in the previous period on health care costs in the current period to examine whether there were
any spillover effects for some of these diseases or group of conditions.
4.4.7.6 Time invariant characteristics
The following time-invariant characteristics were employed in additional analyses involving
population-averaged and random effect models – i) gender (1 = male versus “female” as
reference), ii) Race-ethnicity (1 = White, 2 = African-American, 3 = Asian/Other and 4 =
Race/ethnicity information undisclosed or missing versus Hispanic as the reference group) and
11
Nutritional deficiencies mostly included anemia. Other conditions in this category were vitamin and other unspecified nutritional
deficiencies.
12
Miscellaneous conditions included mostly urinary tract infections; Other conditions in this category were sepsis, CNS, skin and
unspecified viral infections.
13
Conditions include those requiring surgery such as, bowel obstruction / congenital abdominal wall or GI tract defects.
14
Another disadvantage of the static coding approach is that, due to the fixed effects nature of the current analyses the indicator
variables are more likely to represent the disease effect at the time when the switch (from 0 to 1) occurred and therefore does not
provide an effect of the disease over the total panel duration.

87

iii) birth cohort indicator (1 = children born in year 2003 vs. born in year 2002 as the reference
group). The Texas region in which a child’s place of residence was located was used as an
auxiliary variable in analyses using attrition weights (13 categories, viz. Metroplex, Gulf coast,
Upper east, Alamo, Central, Capital, West, Upper Rio Grande, North-west, High plains, South
east and miscellaneous regions versus South Texas as reference group).
4.4.8 Estimation strategy
4.4.8.1 Conceptual framework
Preterm birth exposes an infant to multiple short-term neonatal complications, the severity of
which depends upon the gestational age at birth. Among survivors, these short-term effects may
lead to a higher rate and/or severity of chronic diseases resulting in differences in the healthcare
cost distribution between the preterm and full-term groups. We represent the healthcare cost
functions of preterm and full-term groups as:
𝐘𝐘 𝐆𝐆 ( 𝐢𝐢 , 𝐭𝐭 )
∝ 𝛉𝛉 𝐆𝐆 ( 𝛍𝛍 𝐢𝐢 , 𝐙𝐙 𝐢𝐢 𝐭𝐭 − 𝟏𝟏 , 𝐂𝐂 𝐢𝐢 𝐭𝐭 − 𝟏𝟏 , 𝐀𝐀𝐀𝐀 𝐀𝐀 𝐭𝐭 , 𝛆𝛆 𝐢𝐢 𝐭𝐭 ) , 𝐆𝐆 ∈ ( 𝐏𝐏 , 𝐂𝐂 ) ……(1)
where, G is a group indicator for a set of preterm (P) and full-term (C) children in the sample, (P,
F), Y
it
= healthcare cost for individual ‘i’ at time ‘t’, θ
G
is the functional form, Z
it-1
is the
eligibility category in the previous period (0 = Expansions, 1 = SSI (proxy for disabled), 2 =
AFDC/cash grant and 3 = missing eligibility status), C
it-1
is a set of other time-varying variables
that include hospitalization in the previous period and diagnosis of congenital heart defects
(except patent ductus arteriosus) and or other chronic conditions occurring in period ‘t-1’, Age
t
is
the age at time ‘t’, μ
i
is the individual fixed-effects term (explained in Section 4.4.11), and 𝜀𝜀 𝑖𝑖 𝑖𝑖 is
the contemporaneous error term that is assumed to be conditionally independent of X
it
(the set of
covariates Z
it-1
, C
it
, Age
t
), i.e, 𝐸𝐸 � 𝜀𝜀 ( 𝐺𝐺 )𝑖𝑖 𝑖𝑖 � 𝑋𝑋 𝑖𝑖 𝑖𝑖 �= 0 (see next two sections for a full discussion on the
identification condition).
88

The mean difference in healthcare cost between the two groups after running the group
regressions, separately as represented in equation (1), is given by,
𝐃𝐃 �
= 𝐘𝐘 �
𝐢𝐢 𝐭𝐭 ( 𝐏𝐏 )
− 𝐘𝐘 �
𝐢𝐢 𝐭𝐭 ( 𝐂𝐂 )
, which can be decomposed as,
𝐃𝐃 �
= � 𝛃𝛃 �
𝐏𝐏 𝐏𝐏 − 𝛃𝛃 �
𝐂𝐂 𝐏𝐏 �+ ∑ 𝐗𝐗 �
𝐂𝐂𝐂𝐂
𝐊𝐊 𝐂𝐂 = 𝟏𝟏 � 𝛃𝛃 �
𝐏𝐏 𝐂𝐂 − 𝛃𝛃 �
𝐂𝐂𝐂𝐂
�
� � � � � � � � � � � � � � � � � � � � � � � � �
( 𝐂𝐂𝐂𝐂 𝐀𝐀 𝐂𝐂𝐂𝐂 𝐢𝐢 𝐂𝐂 𝐢𝐢𝐀𝐀 𝐂𝐂 𝐭𝐭 𝐂𝐂 𝐀𝐀𝐂𝐂𝐂𝐂𝐀𝐀𝐂𝐂 𝐭𝐭 )
+ ∑ ( 𝐗𝐗 �
𝐏𝐏 𝐂𝐂 𝐊𝐊 𝐂𝐂 = 𝟏𝟏 − 𝐗𝐗 �
𝐂𝐂𝐂𝐂
) 𝛃𝛃 �
𝐏𝐏 𝐂𝐂 � � � � � � � � � � � � � � �
(𝐄𝐄 𝐂𝐂 𝐄𝐄 𝐂𝐂 𝐄𝐄𝐄𝐄 𝐀𝐀 𝐂𝐂𝐭𝐭 𝐂𝐂 𝐀𝐀 𝐂𝐂𝐂𝐂𝐀𝐀 𝐂𝐂𝐭𝐭 )
……(2)
where, 𝛽𝛽 ̂
𝐺𝐺 0
and 𝛽𝛽 ̂
𝐺𝐺𝐺𝐺
(k = 1,….,K) are the estimated intercept and slope coefficients, respectively,
of regression models for groups G = P, C. The formula in equation (2) represents the standard
decomposition technique originally proposed by Blinder and Oaxaca (Blinder 1973; R. Oaxaca
1973), to study wage disparities between gender or among Races. The Blinder-Oaxaca (BO)
decomposition technique has ever since, found application to study group differences across
multiple disciplines including, education and health outcomes (Barrera-Osorio and Garcia-
Moreno 2011; Donnell et al. 2007; Wagstaff, Van Doorslaer, and Watanabe 2003; García-Altés,
Pinilla, and Ortún 2011).
Using the two-fold BO decomposition shown in equation (2), we are able to study: 1) the
aggregate difference in healthcare cost due to differences in the mean level of group
characteristics between the two groups (commonly referred to as “endowment effects” or
“composition effects” in the literature) and 2) aggregate difference in the effects of covariates
between the two groups that provides an estimate of the total returns to group membership
(“coefficient effects”). The latter effect (also commonly referred to as “structural effect” in the
wage decomposition literature) can be used to provide an estimate of the sum of the effect of
prematurity on preterm or full-term born children and thus, can be thought of as the average
treatment effect on the treated (ATET) or the untreated (ATU), respectively (Fortin, Lemieux,
89

and Firpo 2011, pages 4-5). The interpretation of the coefficients effect as treatment effect of
prematurity may need additional assumptions (see Section 4.4.8.3).
In the context of B-O decomposition procedures, the interpretation of the coefficients effect as
ATET or ATU is made when the counterfactual weights used to estimate the decomposition are
the estimated coefficients of the control or the preterm group (the case in equation 2),
respectively. Some economists have argued against using a set of coefficients obtained
exclusively from one group as the counterfactual weights because, such an approach may lead to
overvaluation of one group and undervaluation of the other (see Cotton 1988). Therefore, several
researchers have subsequently modified the standard BO decomposition to make use of a set of
alternate counterfactual weights, generally written as β
∗
= W β
P
+ (I − W) β
C
, where W = I
corresponds to β* = β
P
, W = 0 corresponds to β* = β
C
and W = ωI could reflect a weighting
corresponding to the share of the two groups in the population, i.e, the arithmetic or sample size
weighted average of the coefficients obtained from each group (Reimers 1983; Cotton 1988).
15

Also, by assuming a linear relationship between the outcomes and covariates, the standard and
weighted BO decomposition methods allow for easy estimation of the contribution of each
explanatory characteristic to the aggregate difference in endowment effects (detailed
decomposition).
4.4.8.2 Identification condition in decompositions
The two-fold decomposition technique explained above relies on estimating the unconditional
counterfactual distributions of the outcome variable (Edoka 2012; Fortin, Lemieux, and Firpo
2011). The unconditional counterfactual distribution is constructed to simulate what the log
15
Reimers (1983) has proposed using 𝛽𝛽 ∗ �
= 0.5 𝛽𝛽 𝑃𝑃 �
+ 0.5 𝛽𝛽 𝐶𝐶 �
and Cotton (1988) has proposed using 𝛽𝛽 ∗ �
=
𝑛𝑛 𝑃𝑃 𝑛𝑛 𝑃𝑃 + 𝑛𝑛 𝐶𝐶 𝛽𝛽 𝑃𝑃 �
+
𝑛𝑛 𝐶𝐶 𝑛𝑛 𝑃𝑃 + 𝑛𝑛 𝐶𝐶 𝛽𝛽 𝐶𝐶 �
as
weights.
90

healthcare cost distribution of children in the preterm group would be if they had the same level
of characteristics as the full-term group (or vice-versa). The unconditional counterfactual
distribution of log healthcare costs is generated by integrating the conditional distribution of log
healthcare costs given a set of covariates in one group over the marginal distribution of covariates
in the other group. For example, using the full-term group as reference group, we can estimate the
unconditional counterfactual distribution of log healthcare cost, 𝑭𝑭 𝑷𝑷 𝒀𝒀 𝑪𝑪 , by replacing 𝐹𝐹 𝑌𝑌 𝑃𝑃 | 𝑋𝑋 ( 𝑌𝑌 | 𝑋𝑋 𝑖𝑖 𝑖𝑖 =
𝑥𝑥 𝑖𝑖 𝑖𝑖 ) with 𝐹𝐹 𝑌𝑌 𝐶𝐶 | 𝑋𝑋 ( 𝑌𝑌 | 𝑋𝑋 𝑖𝑖 𝑖𝑖 = 𝑥𝑥 𝑖𝑖 𝑖𝑖 ), such that
𝐅𝐅 𝐘𝐘 𝐏𝐏 𝐂𝐂 ( 𝐘𝐘 𝐢𝐢 𝐭𝐭 ) = ∫ 𝐅𝐅 𝐘𝐘 𝐂𝐂 | 𝐗𝐗 ( 𝐘𝐘 𝐢𝐢 𝐭𝐭 | 𝐗𝐗 𝐢𝐢 𝐭𝐭 = 𝐱𝐱 𝐢𝐢 𝐭𝐭 ). 𝐄𝐄𝐅𝐅
𝐗𝐗 𝐏𝐏 ( 𝐱𝐱 𝐢𝐢 𝐭𝐭 ) ……(3)
where, 𝐹𝐹 𝑌𝑌 𝐶𝐶 | 𝑋𝑋 ( 𝑌𝑌 𝑖𝑖 𝑖𝑖 | 𝑋𝑋 𝑖𝑖 𝑖𝑖 = 𝑥𝑥 𝑖𝑖 𝑖𝑖 ) is the conditional distribution of log healthcare cost in the full-term
group and 𝐹𝐹 𝑋𝑋 𝑃𝑃 ( 𝑥𝑥 𝑖𝑖 𝑖𝑖 ) is the marginal distribution of X
it
in the preterm group. The unconditional
counterfactual distribution in the above equation represents the distribution of log healthcare costs
that would be observed in the preterm group if the distribution of characteristics were similar to
the full-term control group. Using this technique, consistent estimate of the effect of prematurity
can be identified by imposing the following (weak) ignorability assumption:
for G = (P, C),
if (D
G
, X
it
, ε
it
) has a joint distribution,
where, D
G
is a group indicator and X
it
is a set of covariates; then, for all x
it
in X
it
,
ℰ
𝑖𝑖 𝑖𝑖 is independent of D
G
given X
it
= x
it
or equivalently,
𝐃𝐃 𝐆𝐆 ⊥ 𝓔𝓔 𝐢𝐢 𝐭𝐭 | 𝐗𝐗 𝐢𝐢 𝐭𝐭 ……(4)
91

In other words, the conditional independence/ignorability assumption in (4) implies that, the
conditional distribution of ℰ
𝑖𝑖 𝑖𝑖 given X
it
is the same for both groups and is independent of group
membership (Fortin, Lemieux, and Firpo 2011).
4.4.8.3 Unobserved heterogeneity and fixed effects
The assumption of ignorability stated above for the decomposition technique holds true only
when there are no omitted variables in the structural model, i.e. where there are no unobserved
variables that are correlated with preterm births as well as health outcomes in children, thereby
rendering some or all of the observed covariates in the model potentially endogenous (Fortin,
Lemieux, and Firpo 2011). Unobserved individual level factors such as, method of delivery
(spontaneous, instrumental or surgical C-section), potential malnourishment, small or large for
gestational age, birth order, undiagnosed congenital problems and genetic propensities to attain
normal developmental milestones or to catch-up growth may cause the structural parameters in
equation (1) to be different between the preterm (especially the very/extremely low birth weight)
and full-term groups (Ip et al. 2007; Carey 2001; Rosenzweig and Wolpin 1988). Further, studies
have shown that poverty level and maternal education may be highly associated with poor
pregnancy outcomes (preterm birth for example) as well as developmental outcomes in children
(Behrman and Butler 2007; Kramer 2003; Violato et al. 2011).
Variables that can adequately instrument for preterm birth were difficult to obtain for this
research. However, the availability of panel data offers an advantage of using the fixed-effects
modeling strategy that can be used to eliminate all subject-specific omitted variables from
equation (1), so long as they are time-invariant (Mundlak 1978; Hsiao 2003b). The relevance of
controlling for individual fixed effects in pediatric outcomes research can also be found
elsewhere in the literature (Brown et al. 2013). By introducing a fixed-effects term in equation (1)
92

(represented by μ
i
), the ignorability assumption in (4) holds when, E � ϵ
it( G)
| µ
i
, X
it
�= 0. Or
alternately,
𝐃𝐃 𝐆𝐆 ⊥ 𝓔𝓔 𝐢𝐢 𝐭𝐭 | 𝝁𝝁 𝒊𝒊 , 𝐗𝐗 𝐢𝐢 𝐭𝐭 ……(5)
Also, the following condition has to be satisfied for the coefficients obtained from a fixed-effects
model to be unbiased and consistent -
E � 𝜇𝜇 𝑖𝑖 , 𝜖𝜖 𝑖𝑖 𝑖𝑖 ( 𝐺𝐺 )
� ≠ 0
Thus, accounting for fixed effects followed by applying the B-O decomposition technique may
provide consistent estimates of treatment effect of preterm births. It should be noted that, the
fixed effects term μ
i
here represents the net impact of factors that may be related to both
prematurity and future healthcare costs. However, it has to be acknowledged that the use of fixed
effects technique may not fully control for all time-invariant factors, whether directly related or
unrelated, to preterm births and future healthcare costs (see Limitations, section 4.7). Further, the
ignorability assumption in (5) can hold only when the distribution of other time invariant factors
that may not be controlled by using the fixed effects estimation (within each group), are similar
between the two groups. By invoking this assumption, the portion of the cost differential
explained by coefficients following Blinder-Oaxaca type decompositions may provide consistent
estimate for returns to prematurity, after correcting for unobserved heterogeneity among preterm
and full-term groups.
4.4.9 Estimation procedures
We began first by testing the validity of using Oaxaca type decompositions and child fixed effects
for this research. We then specified a two-part model with fixed effects to estimate healthcare
costs, separately among preterm and full-term groups. Finally we estimated the treatment and
93

endowment effects, using the two-fold Blinder and Oaxaca decomposition and an extension of
this technique that applies to non-linear models (Fairlie 2005). We compared the decomposition
estimates (endowment and treatment effects) obtained from the main model to those obtained
from the cross-sectional ln-OLS and random effects regression models. We conducted statistical
tests to assess model specification.
4.4.9.1 Interaction test for applicability of decomposition methods
The ability to use fixed effects to study the impact of time invariant variables is the primary
motivation for estimating separate regression models among full-term and preterm groups.
However, we formally assessed the applicability of Oaxaca type decompositions to study
difference in healthcare costs between the groups. Accordingly, we ran pooled ordinary least
squares regression of log healthcare costs (ln-OLS) involving dummy variables for gestational
age (P
1
: 33-36 weeks of GA, P
2
: <33 weeks of GA, vs. full-term) and interaction of these groups
with observed characteristics - eligibility category in the previous period (SSI, AFDC or missing
vs. expansion programs at ‘t-1’) and all chronic conditions (observed at ‘t’). An F-test for the
joint significance of coefficients on interaction terms was conducted to test the null hypothesis
that the coefficients are jointly equal to zero, versus the coefficients were not all equal to zero.
4.4.9.2 Statistical tests for the validity of child fixed effects
The role of fixed effects in this research was statistically tested using two procedures that
constitute running augmented random effects regressions to test the random effects assumption,
that the observed characteristics in equation (1) are not correlated with unit effects (in our case
the child fixed effects), i.e. 𝐸𝐸 ( 𝑋𝑋 𝑖𝑖 𝑖𝑖 ,
𝑢𝑢 𝑖𝑖 ) = 0. The testing implied in these procedures is
asymptotically equivalent to the traditional Hausman test. However, unlike the Hausman type test
for fixed versus random effects, both the procedures discussed below allow for robust testing that
can be extended to heteroskedastic and cluster robust versions (Arellano, 1993).
94

4.4.9.2.1 Mundlak (1978) procedure and test
The Mundlak procedure was run by augmenting the pooled OLS regression with group means of
variables that vary within individual over time (Mundlak 1978; J. M. Wooldridge 2002; Brown et
al. 2013). Suppose Y
it
is the log of healthcare costs given by the equation
𝑌𝑌 𝑖𝑖 𝑖𝑖 = 𝛽𝛽 0
+ 𝛽𝛽 1
𝑋𝑋 𝑖𝑖 𝑖𝑖 − 1
+ 𝛽𝛽 2
𝐶𝐶 𝑖𝑖 + 𝛾𝛾 𝑖𝑖 + 𝜇𝜇 𝑖𝑖 + 𝜀𝜀 𝑖𝑖 𝑖𝑖
……. (6)
X
it-1
in the above equation collectively represents the time-varying factors (eligibility category
t-1
,
hospitalization
t-1
, diagnosis of congenital heart disease), C
i
represents time-invariant factors
(gestational age, male gender, Race and birth cohort indicator) and γ
t
captures time effect. The
fixed effects term is given by 𝜇𝜇 𝑖𝑖 = 𝜑𝜑 ( 𝑥𝑥 𝚤𝚤 )
� � � �
, where 𝑥𝑥 𝚤𝚤 � is the sample mean of x
it
across all periods
during the study period for child i. Controlling for child fixed effects, we maintain the assumption
that E(ε
it
|μ
i
, x
it
) = 0. However, in order for the random effects or pooled OLS estimator to be
consistent, a much stronger assumption that E(μ
i
|x
it
) = 0 is also needed. We tested the random
effects assumption using the null hypothesis (H
0
) that, φ = 0. Rejection of null hypothesis leads to
the conclusion that fixed effects matter and random effects or OLS estimators are inconsistent.
An advantage of the Mundlak procedure is that it provides the ability to test the validity of fixed
effects after accounting for observed time-invariant factors (and vice-versa), which is important
in the context of this research since the treatment variable of interest, i.e., prematurity, does not
vary over time.
16

16
This advantage entails, in a sense that the Mundlak procedure allows simultaneous estimation of time-invariant random effects and
fixed effects. However the procedure may not be used for estimation of treatment effect because the addition of within-group means
alone may not guarantee exclusion of all unit specific effects (especially the ones that may be correlated with time-invariant factors
such as prematurity but not with time varying characteristics).
95

4.4.9.2.2 Sargan-Hansen (S-H) test
The S-H test treats the fixed versus random effects evaluation as a test of over identification using
a generalized method of moments framework (Schaffer and Stillman 2006; Arellano, 1993). A
key difference between the S-H and the Mundlak procedures is that, in the S-H test the random
effects regression is re-estimated after augmenting with additional variables consisting of the
original regressors transformed into deviations-from-mean form (instead of group means alone).
This ensures removal of any additional unit specific effects that are correlated with observed time
invariant characteristics in the model. Similar to the Mundlak test, the S-H procedure uses a Wald
chi-squared test of significance for the estimated coefficients on the additional transformed
(deviation-from-mean) variables and rejection of null hypothesis leads to the conclusion that
random effects or pooled OLS estimators are inconsistent (Arellano, 1993).
4.4.9.3 Model specification: Two-part model with fixed effects
Since cost data were highly skewed with a greater proportion of zero utilization observed with
increasing age, healthcare costs in the preterm and full-term groups were separately estimated
using a two-part model specification (Mihaylova et al. 2011; Manning and Mullahy 2001). A
two-part specification helps to model healthcare demand in a separate step, thereby explicitly
accounting for the probability of healthcare use in the estimation of quantity of healthcare
expenditures within each gestational age group. A log-linear model without the two-part
specification may not be well suited to handle the high density of zeros, unless by resorting to a
potentially erroneous solution like adding small non-zero values to the zero cost (see footnote).
17

More importantly, a single equation ln-cost model does not explicitly account for healthcare
17
The solution of adding small non-zero values is done so that observations with zero utilization can be included in the estimation as
if they had very little healthcare use. This practice may mask any systematic differences in the probability of healthcare use. Also the
non-zero values, however small they may be numerically (e.g., 0.01, 0.0001 or 0.000001), may not be close to each other in logs and
hence can be influential in log-linear regressions.
96

demand when estimating treatment effects, which might be also subject to influence of
unobserved factors. A two-part specification is also considered a more flexible and attractive
approach as compared to traditional censored regression modeling approaches since it relaxes the
assumption of independence of error terms from the two equations, and also does not require an
exclusion restriction to model the unconditional expected cost.
18
Hence a two-part system
specification was used to combine the predicted probability of incurring any healthcare cost
obtained from a first step equation, with the healthcare cost density conditional on Pr (Y
it
>0),
modeled in the second step equation. i.e,
𝐸𝐸 𝐺𝐺 ( 𝑌𝑌 𝑖𝑖 𝑖𝑖 | 𝑋𝑋 𝑖𝑖 𝑖𝑖 ) = 𝐸𝐸 𝐺𝐺 ( 𝑌𝑌 𝑖𝑖 𝑖𝑖 | 𝑌𝑌 𝑖𝑖 𝑖𝑖 > 0)
� � � � � � � � � � �
𝑆𝑆 𝑖𝑖 𝑆𝑆𝑆𝑆 2
∗ 𝐹𝐹 𝐺𝐺 ( 𝑍𝑍 𝑖𝑖 𝑖𝑖 ′
𝛾𝛾 )
� � � � �
𝑆𝑆 𝑖𝑖 𝑆𝑆𝑆𝑆 1
, …………… (7)
where, G = (P, C) is the group indicator; P represents one of the two preterm groups and C refers
to children born full-term (controls); 𝐹𝐹 𝐺𝐺 ( 𝑍𝑍 ′
𝛾𝛾 ) is the predicted probability of Y
it
>0 estimated for
each group.
Step I: Panel data Conditional Fixed-effects Regression
The first step equation was estimated using a panel data conditional fixed effects logit regression
procedure in STATA that accounts for unobserved heterogeneity in the healthcare use probability
estimation (version 13.1, College Station, TX, USA). This procedure fits a conditional likelihood
model with a dichotomous dependent variable that takes the value of ‘0’ k
1i
times and a value of
‘1’ k
2i
times where ‘i’ is the panel indicator, so that the total number of observations, T
i
= k
1i
+ k
2i
18
A significant advantage of this feature being that it permits the use of same set of explanatory variables in both the equations
(Mullahy 2009).
97

(Greene 2012)
.
19

Let i = 1, 2,..,,n denote the panels and t = 1,2,…,T
i
denote the observations for
the i
th
panel. Let y
it
be the dependent variable taking on values 0 or 1. Let 𝑘𝑘 1 𝑖𝑖 = ∑ 𝑦𝑦 𝑖𝑖 𝑖𝑖 𝑇𝑇 𝑖𝑖 𝑖𝑖 = 1
be the
observed number of ones for the dependent variable in the i
th
group. Then, the probability of a
possible value of y
i
conditional on ∑ 𝑦𝑦 𝑖𝑖 𝑖𝑖 𝑇𝑇 𝑖𝑖 𝑖𝑖 = 1
= 𝑘𝑘 1 𝑖𝑖 is given by (Hosmer, Lemeshow, and
Sturdivant 2013),
Pr � 𝑦𝑦 𝑖𝑖 � ∑ 𝑦𝑦 𝑖𝑖 𝑖𝑖 𝑇𝑇 𝑖𝑖 𝑖𝑖 = 1
= 𝑘𝑘 1 𝑖𝑖 �=
ex p ( ∑ 𝑦𝑦 𝑖𝑖𝑖𝑖
𝑇𝑇 𝑖𝑖 𝑖𝑖 = 1
𝑥𝑥 𝑖𝑖𝑖𝑖
𝛽𝛽 )
∑ ex p (
𝑑𝑑 𝑖𝑖 ∈ 𝑆𝑆 𝑖𝑖 ∑ 𝑑𝑑 𝑖𝑖𝑖𝑖
𝑥𝑥 𝑖𝑖𝑖𝑖
𝛽𝛽 )
𝑇𝑇 𝑖𝑖 𝑖𝑖 = 1
…………….(8)
where, d
it
is equal to 0 or 1 with ∑ 𝑦𝑦 𝑖𝑖 𝑖𝑖 𝑇𝑇 𝑖𝑖 𝑖𝑖 = 1
= 𝑘𝑘 1 𝑖𝑖 , and S
i
is the set of all possible combinations of
k
1i
ones and k
2i
zeros. The conditional likelihood is:
𝑙𝑙𝑙𝑙 𝑙𝑙 = ∑ { ∑ 𝑦𝑦 𝑖𝑖 𝑖𝑖 𝑇𝑇 𝑖𝑖 𝑖𝑖 = 1
𝑥𝑥 𝑖𝑖 𝑖𝑖 𝛽𝛽 − 𝑙𝑙𝑙𝑙𝑙𝑙 𝑓𝑓 𝑖𝑖 𝑛𝑛 𝑖𝑖 = 1
( 𝑇𝑇 𝑖𝑖 , 𝑘𝑘 1 𝑖𝑖 ) ……………….(9)
where x
it
represents a set of time varying covariates and 𝑓𝑓 𝑖𝑖 ( 𝑇𝑇 𝑖𝑖 , 𝑘𝑘 1 𝑖𝑖 ) in (9) is the function in the
denominator in (8). Eligibility category through which a child qualified for Medicaid coverage
(SSI, AFDC or missing program of eligibility information vs. expansions), hospital admission, a
claim for congenital heart defects (except patent ductus arteriosus) in the previous time period
and dummy variables for age-period were the time varying characteristics used in the model. The
marginal effects of characteristics obtained using the conditional fixed effects procedure was
compared with the marginal effect estimates obtained using population-averaged probit model.
Time invariant characteristics such as, gender, Race, and birth cohort (year 2002 vs. 2003) were
19
A potential drawback is that the predicted probability obtained using this procedure is grossly under-estimated due to the
conditioning on one positive outcome within the panel (J. M. Wooldridge 1995). Hence while calculating decomposition of Step I of
the two-part model we calculated the total difference in probability of a non-zero outcome using the raw difference in probabilities of
the estimation sample in both groups. The estimation sample consists of all observations that changes outcome from ‘0’ to ‘1’ at least
once within the panel.

98

used as additional controls in the population-average model. Due to the repeated measures
structure of the data, we assumed an “unstructured” within-panel correlation and estimated the
cluster robust standard errors for efficient estimation of the coefficients (Cameron and Trivedi
2009).
Step II: Fixed effects Model for healthcare costs conditional on Y
it
> 0
In equation (7) 𝐸𝐸 ( 𝑌𝑌 𝑖𝑖 𝑖𝑖 | 𝑌𝑌 𝑖𝑖 𝑖𝑖 > 0) is the expected healthcare cost given Y
it
>0. Healthcare costs
(conditional on Y
it
> 0) were log-transformed due to the positive skew and heavy right tails
inherent in the distribution that renders statistical testing and inference very difficult. After log
transformation healthcare cost in the second step of the equation was estimated using a fixed-
effects regression model. The fixed effects model is given by the equation –
( 𝑙𝑙𝑙𝑙 𝑌𝑌 𝑖𝑖 𝑖𝑖 | 𝑌𝑌 𝑖𝑖 𝑖𝑖 > 0) = 𝛽𝛽 0
+ 𝛽𝛽 1
𝑋𝑋 𝑖𝑖 𝑖𝑖 − 1
+ 𝛾𝛾 𝑖𝑖 + 𝜇𝜇 𝑖𝑖 + 𝜀𝜀 𝑖𝑖 𝑖𝑖
…………….. (10)
where, X
it-1
collectively represents the time-varying factors viz., eligibility categories: SSI (which
is proxy for disability status), AFDC or missing program of eligibility information vs.
expansions, hospitalization and diagnosis of congenital heart defects (except patent ductus
arteriosus), γ
t
represents dummy variables for age-period and 𝜇𝜇 𝑖𝑖 represents the fixed effects term
for child i. The fixed effects procedure in STATA uses within-transformation to provide
consistent estimates of the coefficients free from the influence of any unobserved individual level
factors (affecting group membership) that are time-invariant (Cameron and Trivedi 2009). The
procedure in STATA estimates the following equation under the constraint 𝜇𝜇 ̅ = 0.
( 𝑦𝑦 𝑖𝑖 𝑖𝑖 − 𝑦𝑦 𝚤𝚤 � + 𝑦𝑦 � ) = 𝛽𝛽 0
+ 𝛽𝛽 1
( 𝑥𝑥 𝑖𝑖 𝑖𝑖 − 1
− 𝑥𝑥 𝚤𝚤 � + 𝑥𝑥 ̿) + 𝛾𝛾 𝑖𝑖 + ( 𝜀𝜀 𝑖𝑖 𝑖𝑖
− 𝜖𝜖 𝚤𝚤 � + 𝜇𝜇 ̅) + 𝜖𝜖 ̿ …………..(11)
99

where, 𝑦𝑦 𝚤𝚤 �, 𝑥𝑥 𝚤𝚤 � � � , 𝜖𝜖 𝚤𝚤 � are the within panel (‘i’) averages of 𝑦𝑦 𝑖𝑖 𝑖𝑖 , 𝑥𝑥 𝑖𝑖 𝑖𝑖 𝑎𝑎 𝑙𝑙 𝑎𝑎 𝜀𝜀 𝑖𝑖 𝑖𝑖 , respectively; 𝑦𝑦 � , 𝑥𝑥 ̿, 𝜖𝜖 ̿, are the
grand averages 𝑦𝑦 𝑖𝑖 𝑖𝑖 , 𝑥𝑥 𝑖𝑖 𝑖𝑖 𝑎𝑎 𝑙𝑙 𝑎𝑎 𝜀𝜀 𝑖𝑖 𝑖𝑖 , respectively over the time and unit dimensions. Note that the
intercept term β
0
here represents the baseline group effect after eliminating the fixed effects term,
but also contains the average of other time-invariant factors that cannot be fully controlled by
using the individual dummies alone. Thus, in order for the coefficients effect, following B-O
decomposition, to be interpreted as the true returns to preterm birth, the conditional distributions
of these other time-invariant unobserved factors among the preterm and full-term groups should
be similar (See Limitations, section 4.7).
The chronic conditions explained in Section 4.6.1 (V) were not controlled for within the
decomposition framework since some of these conditions could be causally associated with
prematurity (especially extreme immaturity) and hence controlling for these conditions could
underestimate the true impact of prematurity. However, we conducted sensitivity analyses to
estimate the treatment effect of preterm births after controlling for chronic medical conditions
observed during period t-1 (see Section 4.4.14.3). Also refer to cost driver analyses in Section
4.4.15.
4.4.9.4 Blinder-Oaxaca decomposition and estimation of treatment effects
Following model estimation using the two-part model, decomposition of the overall difference in
probability of healthcare use (Step I) between the preterm and full-term groups was carried out in
STATA using a modification of the Blinder-Oaxaca procedure proposed by Robert Fairlie (Fairlie
100

2005).
20
Following Fairlie (2005), if P = preterm and C = full-term (control) groups, the
decomposition for the conditional fixed-effects model can be obtained from:
𝑌𝑌 �
𝑃𝑃 − 𝑌𝑌 �
𝐹𝐹 = � ∑
𝐹𝐹 � 𝑋𝑋 𝑖𝑖 𝑃𝑃 𝛽𝛽 �
𝑃𝑃 �
𝑁𝑁 𝑃𝑃 𝑁𝑁 𝑃𝑃 𝑖𝑖 = 1
− ∑
𝐹𝐹 � 𝑋𝑋 𝑖𝑖 𝐶𝐶 𝛽𝛽 �
𝑃𝑃 �
𝑁𝑁 𝐶𝐶 𝑁𝑁 𝐶𝐶 𝑖𝑖 = 1
� + � ∑
𝐹𝐹 � 𝑋𝑋 𝑖𝑖 𝐶𝐶 𝛽𝛽 �
𝑃𝑃 �
𝑁𝑁 𝐶𝐶 𝑁𝑁 𝐶𝐶 𝑖𝑖 = 1
− ∑
𝐹𝐹 � 𝑋𝑋 𝑖𝑖 𝐶𝐶 𝛽𝛽 �
𝐶𝐶 �
𝑁𝑁 𝐶𝐶 𝑁𝑁 𝐶𝐶 𝑖𝑖 = 1
� ……(12)
with N
j
being the sample size for group G, and 𝐹𝐹 ( 𝑋𝑋 𝑖𝑖 𝐺𝐺 𝛽𝛽 ̂
𝐺𝐺 ) the cumulative normal density function.
𝑌𝑌 �
𝑃𝑃 − 𝑌𝑌 �
𝐹𝐹 is the overall difference in outcome, calculated by taking the difference in mean
probabilities between the preterm and full-term groups in the clogit (conditional fixed-effects
logit) estimation sample. The first term on the right-hand side of equation (12) is the difference in
probability of healthcare use that is due to observed group characteristics (endowment effects).
To calculate this term we multiplied the difference in means of characteristics between the two
groups by the average marginal effects of covariates in the preterm group. The second term in
equation (12) represents the contribution that is due to difference in the coefficients (coefficients
effect) and was estimated by subtracting the endowment effects from the overall difference in
probabilities between the two groups (term on LHS). This term provides an estimate of the
average treatment effect on untreated (ATU), whereas using coefficients in the control group (β
C
)
as counterfactual weights provides the average treatment effect on treated (ATET).
The overall difference in 𝐸𝐸 ( 𝑌𝑌 𝑖𝑖 𝑖𝑖 | 𝑌𝑌 𝑖𝑖 𝑖𝑖 > 0) between the preterm and full-term groups obtained from
the second step of the two-part model, was decomposed using the standard Blinder-Oaxaca
decomposition for log-linear models in STATA as given in equation (2) (Jann 2008a). The
difference in the unconditional expected healthcare cost between the preterm and control groups
is given by:
20
A decomposition of the outcome variable similar to equation (2) is not applicable in the non-linear case because the conditional
expectations E(Y itg |X itg) may differ from 𝑋𝑋 �
𝑔𝑔 𝛽𝛽 ̂
𝐺𝐺 𝑔𝑔 .
101

𝐸𝐸 ( 𝑌𝑌 𝑖𝑖𝑖𝑖
| 𝑋𝑋 𝑖𝑖𝑖𝑖
, 𝑃𝑃 ) − 𝐸𝐸 ( 𝑌𝑌 𝑖𝑖𝑖𝑖
| 𝑋𝑋 𝑖𝑖𝑖𝑖
, 𝐶𝐶 ) = 𝛥𝛥 𝐸𝐸 ( 𝑌𝑌 𝑖𝑖𝑖𝑖
| 𝑋𝑋 𝑖𝑖𝑖𝑖
) = 𝛥𝛥 𝐸𝐸 ( 𝑋𝑋 𝑖𝑖𝑖𝑖
′
𝛽𝛽 ) ∗ 𝐹𝐹 𝐶𝐶 ( 𝑍𝑍 𝑖𝑖𝑖𝑖
′
𝛾𝛾 ) + 𝐸𝐸 𝑃𝑃 ( 𝑋𝑋 𝑖𝑖𝑖𝑖
′
𝛽𝛽 ) ∗ 𝛥𝛥 𝐹𝐹 ( 𝑍𝑍 𝑖𝑖𝑖𝑖
′
𝛾𝛾 ) ……(13)
where, ‘Δ’ represents difference in each quantity observed between the preterm and control
groups. Using equations for steps I and II of TPM in (2) and (12), and using the formula for
incremental healthcare cost in (13), the aggregate decomposition of the two-part system was
calculated using the following formula –
𝚫𝚫 𝐄𝐄 ( 𝐘𝐘 𝐢𝐢 𝐭𝐭 | 𝐗𝐗 𝐢𝐢 𝐭𝐭 ) = � 𝚫𝚫 𝐄𝐄 ( ∆ 𝐗𝐗 𝐢𝐢 𝐭𝐭 ′
𝛃𝛃 ) + 𝚫𝚫 𝐄𝐄 ( 𝐗𝐗 𝐢𝐢 𝐭𝐭 ′
∆ 𝛃𝛃 ) �
� � � � � � � � � � � � � � � � �
𝐃𝐃 𝐀𝐀 𝐂𝐂𝐂𝐂 𝐄𝐄 𝐃𝐃 𝐂𝐂 𝐂𝐂 𝐢𝐢𝐭𝐭𝐢𝐢𝐂𝐂 𝐂𝐂 𝐂𝐂𝐂𝐂 𝐒𝐒 𝐒𝐒 𝐄𝐄𝐏𝐏 𝟐𝟐 ∗ 𝐅𝐅 𝐂𝐂 ( 𝐙𝐙 ′
𝛄𝛄 ) + 𝐄𝐄 𝐏𝐏 ( 𝐗𝐗 𝐢𝐢 𝐭𝐭 ′
𝛃𝛃 ) ∗ � 𝐅𝐅 ( ∆ 𝐙𝐙 𝐢𝐢 𝐭𝐭 ′
𝛄𝛄 ) + 𝐅𝐅 ( 𝐙𝐙 𝐢𝐢 𝐭𝐭 ′
∆ 𝛄𝛄 ) �
� � � � � � � � � � � � � � �
𝐃𝐃 𝐀𝐀 𝐂𝐂𝐂𝐂 𝐄𝐄 𝐃𝐃 𝐂𝐂 𝐂𝐂 𝐢𝐢 𝐭𝐭𝐢𝐢𝐂𝐂 𝐂𝐂 𝐂𝐂𝐂𝐂 𝐒𝐒 𝐒𝐒 𝐄𝐄𝐏𝐏 𝟏𝟏 ……(14)
Thus the aggregate decomposition (endowments and coefficients) for step 1, step 2 and the
combined two-part model, of the differences in healthcare cost between preterm infants born at
33-36 weeks of GA and controls, and preterm infants born at < 33 weeks of GA and controls
were derived using coefficients of the preterm (to estimate ATU) and control group (to estimate
ATET) as counterfactual weights. The standard errors and 95% confidence intervals for the
various decomposition estimates were obtained using non-parametric cluster bootstrap procedure
in STATA (Poi 2004; MacKinnon 2006). Bootstrapped data for the entire two-part system were
constructed by repeated random sampling with replacement over 100 replications of the
estimation data (Gould and Pitblado 2010). The cluster bootstrap samples were then used to
approximate the sampling distribution of the various decomposition estimates.
The endowment and treatment effects (ATET and ATU) were finally expressed as percentages
drawn over the overall difference in log healthcare cost between the preterm and full-term groups.
The percentage contribution of coefficients can be interpreted as the covariate and fixed effects
adjusted treatment effect of prematurity. However, as explained previously, this interpretation
further requires the assumption that, the conditional distribution of other time-invariant factors
102

that cannot be fully controlled by using individual fixed effects within each group, is the same
between the preterm and full-term groups.
4.4.10 Comparison of cross-sectional and panel data model decomposition
estimates
The decomposition estimates obtained from the two-part fixed effects model were compared with
the estimates obtained from cross-sectional pooled OLS, Hausman-Taylor random effects and
single equation fixed effect models. The Hausman-Taylor (H-T) approach uses an error
component specification that allows some of the variables to be correlated with the individual
error terms while specifying other variables exogenously. Accordingly, eligibility category (SSI,
AFDC versus expansion program categories) was specified as endogenous while hospitalization
and diagnosis of CHD in the previous period were specified as exogenous (given that the first lag
of these variables may not be correlated with ε
it
, i.e. the error term in the future period). The
variables gender, Race and year of birth were treated as exogenous as well. The set of time-
varying and time-invariant exogenous variables were used as instruments to obtain unbiased and
efficient estimates of the endogenous variables in the model using a flexible two-stage least
squares estimation procedure (Hausman and Taylor 1981; Baltagi and Khanti-Akom 1990).
4.4.11 Calculation of costs using decomposition estimates
The overall difference in healthcare cost per 6-months (in US$) between preterm and full-term
groups was calculated by multiplying the raw mean healthcare costs of control group by the cost
ratio (exponent of the difference in log healthcare costs between groups), using observations in
the estimation sample. The coefficient (ATU or ATET) and endowment effects in terms of
healthcare dollars were estimated by multiplying the corresponding percentage contributions
103

(decomposition estimates) by the mean overall healthcare cost difference between the two
groups.
21

4.4.12 Specification tests
The tests for the validity of fixed effects are already discussed in Section 4.4.9.1. In addition to
using pooled observations, these tests were also conducted on the preterm and full-term groups
separately. The Wooldridge-Drukker test for autocorrelation in panel data was performed for each
group separately as well as for the combined sample. This is a serial correlation test based on the
OLS residuals of the first-differenced model (Drukker 2003). Under the null hypothesis of no
serial correlation of idiosyncratic errors, the residuals from the regression of the first-differenced
variables should have an autocorrelation of -0.5 and this is a Wald (F-statistic) test. However
simulation studies have suggested that this test is not robust against temporal heteroskedasticity
and may have less power in the fixed effects case than in the random effects case in samples with
low levels of serial correlation (Drukker 2003). Hence we also performed additional tests that
have increased power (i.e. performs better for any sample size and time dimension) and specific
for testing autocorrelation in fixed-effect models. The Baltagi-Wu (1999) locally best invariant
statistic (LBI) and a modified version of the Bhargava, Franzini and Narendranathan (1982)
Durbin-Watson statistic were calculated (Baltagi and Wu 1999; Bhargava, Franzini, and
Narendranathan 1982; Baltagi et al. 2007). A value of the modified BFN Durbin-Watson statistic
or Baltagi-Wu LBI-statistic close to 2 indicates no autocorrelation (the value ranges from 0 to 4).
Finally, linear trend of the time fixed effect (i.e. age effect) was tested against the specification
including the time fixed effects as dummy variables (with 6-12 months as the reference group).
The models were compared using the Bayesian Information Criteria (BIC).
21
Note: Given small differences in estimation sample sizes used for the different models (pooled OLS, H-T random effects, single
equation FE and two-part model FE, the overall healthcare cost difference, to which the decomposition percentages were applied
varied slightly between the different models.
104

4.4.13 Tests for attrition bias
Children who are enrolled in the insurance program leave the cohort over a period of time owing
to changes in parental income, migration from a state, availability of alternate source of
healthcare or other unknown reasons. Systematic or non-random attrition, if present, could lead to
biased and inconsistent estimates of the treatment effect, though often in large databases such
biases have been shown to be small and insignificant (Wooldridge 2010 pp. 827-837; Fitzgerald,
Gottschalk, and Moffitt 1998; Outes-leon and Dercon). The time invariant characteristics
(demographic and clinical characteristics before the study period) of the samples retained at 6-18
and 37-60 months of age were descriptively compared. However, a change in the composition of
the samples over time does not necessarily imply that attrition causes bias unless formally tested.
Therefore, we conducted the following tests in order to understand whether attrition impacted the
estimates of our main model coefficients:
4.4.13.1 Statistical test for the difference in model coefficients: Full versus
Completing samples
We first ran separate ln-OLS regressions of total healthcare cost for full-term and preterm
children using all-available sample (preterm and full-term children available in the sample
ignoring attrition, hereafter called “full” sample) and the sample of children who remained in the
fee-for-service cohort up to 60 months of age (hereafter called “completing sample”). We then
tested whether the difference in coefficients of each covariate (including constant term) between
full versus completing samples for statistical significance using seemingly unrelated estimation (-
suest) procedure in STATA (Weesie 1999). The -suest procedure calculates the combined robust
(or cluster robust) covariance matrix allowing statistical comparison of coefficients among
overlapping samples by performing a Wald test (Clogg, Petkova, and Haritou 1995; J. M.
Fitzgerald 2011; Cameron and Trivedi 2009). Finally we compared the BO decomposition
105

estimates of the healthcare cost differential between preterm and full-term born children obtained
using full versus completing samples.
4.4.13.2 Attrition test based on selection on observables
To further probe whether attrition in the full sample was random or not, we used the probit test
developed by Fitzgerald, Gottschalk and Moffitt (1998) for testing attrition bias in panel data.
Both these methodologies are based on selection on observable, but endogenous, lagged
dependent variables that can be correlated with both outcomes as well as attrition. Fitzergald et al.
defines selection on observables by the conditional independence condition that
Pr(A=0 |Y, z, x) = Pr(A=0 |x, z), ……..(15)
where ‘A’ stands for attrition (0 = non-attrition/1 = attrition) and z is a vector of lagged dependent
variables and other variables predicting attrition that can be excluded from the main model
(referred to as auxiliary variables or instruments). The validity of this approach depends on
whether this conditional assumption holds and attrition can be treated as ignorable, once z is
controlled for. Under the weak ignorability assumption, consistent treatment effect estimates can
be obtained by weighting the observed data by the inverse of the probability of non-attrition
conditional on observed covariates.
We used all available lagged values of total healthcare costs (for example, Y
t-1
during 6-18
months, Y
t-1
and Y
t-2
during 19-36 months, Y
t-1
, Y
t-2
and Y
t-3
during 37-60 months of age) and
region of residence variables to predict attrition during each wave, using separate cross-sectional
probit regressions for full-term and preterm groups. The statistical significance of the coefficients
on the auxiliary variables and the pseudo-R-squared values obtained from the probit models were
used to determine whether attrition was random or not.
106

The Becketti, Gould, Lillard and Welch (BGLW) test was implemented by regressing the first lag
of total healthcare costs on region of residence, other demographic variables (Race, gender and
year of birth), a dummy variable for attrition and interaction of attrition dummy with other
explanatory variables. An F-test of the joint significance of the attrition dummy and the
interaction variables was then conducted to determine whether the coefficients from the
explanatory variables differ between infants who were retained in the cohort and infants who
attrit from the panel during each wave (Becketti et al. 1988; Outes-leon and Dercon).
4.4.14 Construction of attrition weights
We first estimated cross-sectional probit models to predict non-attrition during each wave (6-18,
19-36 months and 37-60 months), for the full-term (control) and preterm groups separately, by
using a set of lagged total healthcare cost variables [y
t-1
, y
t-2
, …,y
0
], dummy variables for region of
residence and other demographics characteristics (Race, gender and birth cohort).
A
∗
= δ
0
+ δ
1
∗ race + δ
2
∗ birth cohort + δ
3
∗ region + ∑ ϕ
T
y
i, t − k
+ v
t
......(16)
such that, A = 1 if A* ≥ 0 (means attrition) and A = 0 if A* < 0 (means no attrition). Restricted
probit models excluding the auxiliary instruments (Y
i,t-k
and region) were subsequently estimated.
The inverse probability weights were calculated using the formula
w(z, x) = �
Pr ( A = 0| z, x)
Pr ( A = 0| x)
�
− 1
……….(17)
We applied these weights to the fixed effects models for log healthcare costs. We finally
performed a Hausman type test to compare the weighted and unweighted estimates. Fixed effects
coefficients were compared after applying constant attrition weights, i.e., using panel attrition
weights derived for wave 2 that were assumed to be constant within panel (‘i’) through wave 4, or
107

non-constant attrition weights where, attrition weights were derived for each wave and applied to
the observations (i.e. weights within panel varying over time).
4.4.15 Sensitivity analyses
4.4.15.1 Estimation of incremental effects of preterm births using weighted
estimates
We repeated the two-part model estimation and decomposition following estimation as discussed
in Sections 4.4.9.2 and 4.4.9.3 after accounting for attrition in the panel data, using weights as
explained in the previous section. Weighted estimates were obtained by using attrition weights
calculated for children retained in wave 2 and applying these weights through end of wave 4
(constant weights).
4.4.15.2 Analyses of incremental effect of moderate/late preterm born with
birth weights < 2,500 g
For the purpose of this research preterm children were identified based on gestational ages in the
insurance claim records at the time of birth. Due to the possibility of errors in accurately
determining gestational age, there is potential for misclassification of preterm births especially
towards the upper end of the gestational age spectrum (35-36 weeks of gestational age). We
repeated the analyses among the moderate/late preterm group by restricting this group to only
children with birth weights <2,500 grams. The two criteria together (i.e. being born as early as
33-36 weeks and having a birth weight <2,500 grams) may provide greater validity for the
classification of infants into the moderate/late preterm group than any one condition alone
(Eworuke et al. 2012). Further, birth weight also being a predictor of morbidity, the combined
criteria may help to better identify late preterm children with significant risk for long term
morbidity and higher healthcare costs.
108

4.4.15.3 Analyses of incremental effects of very preterm and extremely
preterm born children separately
We also analyzed the healthcare costs in the < 33 weeks of GA group by splitting them into two
separate groups: children born between 29-32 weeks of GA (very preterm) and children born at ≤
28 weeks of GA (extremely preterm), since clinically these sub-groups represent different risk
profiles. This may also throw more light into the diagnostic drivers of healthcare costs, if they are
different between these two groups (see Section 4.4.15).
4.4.15.4 Analyses of residual effects of preterm births after adjusting for
chronic disease characteristics
The first lag of chronic conditions (listed in Section 4.6.1 (V)) experienced over time in the
cohort were added to the list of regressors in the two-part model estimation, regardless of whether
these conditions were caused by preterm birth or not. Lagged values of the dummy variables for
chronic conditions were used, since they occur earlier in time (t-1) and hence may not be directly
involved in the causal pathway and/or correlated with the idiosyncratic error at time ‘t’. The
percentage of healthcare cost difference due to coefficients was estimated. The coefficient effect
was interpreted as the residual heterogeneity due to group membership effects, i.e. any residual
effects due to preterm birth after accounting for chronic disease characteristics (that have
occurred prior to exposure period ‘t’) and other explained characteristics.
4.4.16 Diagnostic drivers of healthcare cost differences between preterm
and full-term groups
Many chronic conditions were more prevalent among preterms than children born full-term. The
difference in frequencies of these diagnoses, within a given period, may drive significant portion
of the healthcare cost differential between the two groups in that period. We used the detailed
decomposition framework to study the percent contribution of each chronic condition to the
observed component (endowments) of the healthcare cost difference between the two groups.
109

Identifying the contribution of characteristics to explained differential is straightforward because
the total component is a simple sum over the individual contributions. i.e, (X
P
� � � �
− X
F
� � �
) β
P
�
=
(X
1 P
� � � � �
− X
1 F
� � � � �
) β
1 P
�
+ (X
2 P
� � � � �
− X
2 F
� � � � �
) β
2 P
�
+ ⋯ .
22
In order to evaluate the percent contributions, we ran
fixed effects log-linear models, separately for the full-term and preterm groups, using
observations with non-zero costs (i.e. second step of the two-part model). Since the goal of the
analysis was to identify the diagnoses driving costs, models were specified using chronic
conditions (explained in Section 4.4.7(V)) observed at time ‘t’ and their first lags, along with
other covariates (eligibility category, hospitalization status in the previous period and age-period).
The overall healthcare cost difference was decomposed into a part explained by linear
combination of differences in the mean frequencies of each chronic condition and a part owing to
residual group membership effects (coefficient effects). Due to differences in the prevalence of
chronic conditions that may be causal, the explained portion of the decomposition is expected to
be larger than the endowments effect obtained using the main model specification in equation
(10), which excludes the dummy variables for chronic diagnoses at time ‘t’.
In order to estimate the net contribution of chronic diagnoses towards returns to preterm birth, we
estimated two-part models by including chronic diagnoses at times ‘t’ as well as ‘t-1’. The
difference in coefficients effect obtained from these models and the coefficients effect obtained
using the model without including chronic diagnoses (equation 10) provided the net contribution
of chronic diagnoses to coefficients effect.
22
However, the detailed decomposition of the coefficient effects, when there are categorical variables, is complicated since the results
depend on the choice of the omitted base category. Changing the base category changes the contribution of the categorical variable as
a whole (Jann 2008b)
110

4.4.17 Total excess financial impact due to preterm births between 6 and 60
months of age
The cumulative incremental cost per preterm child in each gestational age category, for the study
period between 6 and 60 months of age (54 months) was estimated after applying 3% discount
rate to costs incurred beyond a 1 year time horizon. The present value of the costs incurred
beyond one year time horizon was calculated using the formula 𝑃𝑃𝑃𝑃 = 𝐹𝐹 𝐶𝐶 � 1 −
1
( 1 + 𝑟𝑟 )
𝑖𝑖 � / 𝑟𝑟 where,
PV = present value, FC = future costs, r = discount rate (3%) and t = time horizon. The total
excess cumulative costs due to each gestational age group were calculated for the cohort of
children born in year 2003, after accounting for attrition over time.
4.5 Results
4.5.1 Demographic and clinical characteristics at the time of birth
Preparation of the analytic sample is shown in Appendix 2. After applying exclusions, the number
of survivors at 6 months of age was, 68,501, 2,229 and 1,008 in the full-term, moderate/late
preterm and very/extremely preterm infant groups. The number of children dead, attrited or
retained during each wave is shown in Appendix 3. By the end of 60 months of age there were
38,243, 1,161 and 623 children in the full-term, moderate/late and very/extreme preterm groups,
respectively. Though mortality during the first 6 months period was substantially high among
preterm born, the number of deaths during the study period (6-60 months of age) was not
significant in any of the groups.
A comparison of the sample baseline characteristics (demographic and clinical characteristics)
over time is shown in Appendix 4. A majority of children in the analytic sample (> 45% across all
3 groups) were of Hispanic origin. The proportion of African-American children was higher in
111

the preterm groups as compared to full-term group (p < 0.001). There weren’t notable differences
between the 3 groups in terms of gender or geographical distribution across the 13 Texas regions.
Birth-weights in the moderate/late preterm group were mainly distributed in the upper quartile of
the low birth-weight spectrum (2,000 to ≥ 2,500 grams), and towards the lower end (<1,250
grams) among very/extremely preterm infants. Eleven percent of infants in the moderate/late
preterm and 23% of infants in the very/extremely preterm born groups had birth-weights in the
range of 1,500 to 2,000 grams. As newborns, higher rates of congenital heart defects, perinatal
comorbid conditions such as, respiratory distress and failure, neonatal infections and feeding
difficulties (including failure to thrive) were observed more frequently among preterm infants
than among controls. Higher incidence of perinatal gastrointestinal disorders (necrotizing
enterocolitis/ peritonitis, 8%), retinopathy (21%) and patent ductus arteriosus (29%) was
observed in the very/extremely preterm group, which may have long term implications.
4.5.2 Sample characteristics during study period
The means of time varying characteristics at each time-point, and the proportion of children
having each long-term chronic condition (prevalence) are explained in Appendix 5. Focusing on
the mid-point of the panel (i.e. 19-36 months of age period), it was observed that most children
were eligible for Medicaid under expansion programs (i.e. coverage for children that do not meet
the traditional AFDC criteria but otherwise belonging to low income families) and less than 15%
of children were eligible under the various AFDC programs in all 3 groups. The proportion of
children receiving social security income (which is a proxy for disability) was different between
the 3 groups, with the proportion being higher among very/extremely preterm born (30%) as
compared to moderate/late preterm (2%) or full-term born children (1%). The number of chronic
conditions per an average child in a 6-month period in the full-term, moderate/late preterm and
very/extremely preterm groups were 0.4 (± 0.7), 0.4 (± 0.8) and 1.1 (± 1.5), respectively. Also,
112

the proportion of children hospitalized in the past 6-month period among full-term, moderate/late
preterm and very/extremely preterm born children were 3%, 4% and 11%, respectively.
4.5.2.1 Prevalence of long-term pediatric chronic conditions
As shown in Appendix 5, at any time-point during the panel, respiratory diseases (bronchitis,
asthma, wheezing/other respiratory symptoms and respiratory infections) were the most prevalent
chronic conditions among all 3 groups, followed by neurodevelopmental problems and
gastrointestinal functional disorders. Focusing again on the mid-point of the panel (i.e. 19-36
months of age period) the prevalence of asthma, bronchitis and lower respiratory tract infections
were similar between full-term (11%, 16% and 8.5%, for asthma, bronchitis and LRTI
respectively) and moderate/late preterm born (12%, 16% and 8%), but were higher among
very/extremely preterm born children (26%, 21% and 18% for asthma, bronchitis and LRTI
respectively).
23
Long-term CNS conditions (epilepsies and motor disorders including cerebral
palsy) were more common among preterm children (1-7% depending on gestational age) as
compared to children born full-term (<1%). A higher proportion of children in the very/extremely
preterm group experienced neurodevelopmental delay (27%) and delayed physiological
milestones (15%), as compared to children in the full-term (4% and 2%, NDD and delayed
milestones, respectively) or moderate/late preterm (7% and 3%, NDD and delayed milestones,
respectively) groups. Children born very/extremely preterm also had a higher prevalence of
refractory vision disorders (18.5% as compared to 1.5% and 2% in children born full-term and
moderate/late preterm, respectively), and attention/conduct disorders (3% versus 1% among
23
Prevalence of respiratory conditions were the highest during 6-18 months of age and decreased gradually with age. The trend was
opposite for long term CNS and neurodevelopmental conditions that increased during the course of time in the panel in all 3 groups.
113

children born full-term and moderate/late preterm)
24
. Most other conditions including
gastrointestinal disorders, upper respiratory tract infections
25
, miscellaneous (non-respiratory)
infections, endocrine and metabolic disorders and nutritional deficiencies were comparable across
the 3 groups.
4.5.3 Distribution of healthcare costs
Mean unadjusted healthcare costs per 6-months incurred by children in the full-term,
moderate/late preterm and very/extreme preterm groups, between 6 and 60 months of age, were
US$ 735 (standard deviation = 5,004; interquartile range = 557), 1,174 (SD: 7,345; IQR = 629)
and 5,974 (SD: 25,117; IQR = 2,931), respectively.
The mean cumulative unadjusted dollar amounts paid for inpatient and ambulatory care services
in each group are shown in Figure 4-1. The mean unadjusted cumulative healthcare costs incurred
by a full-term born child were US$ 2,193 (SD = ±10,846), 2,001 (11,419) and 1,940 (13,478),
during 6-18, 19-36 and 36-60 months of age, respectively. On average the cumulative healthcare
costs among moderate/late preterm and very/extreme preterm born children were 60% and 680%
higher than healthcare costs among full-term born children, respectively. The mean cumulative
costs per child corresponding to the three time periods among moderate/late preterm born
children were, (US$) 3,288 (±13,416), 3,021 (17,198) and 3,637 (24,837), respectively. Similarly,
the cumulative costs (per child) corresponding to the three time periods among very/extremely
preterm born children were, (US$) 22,523 (±70,884), 13,014 (37,260) and 13,218 (44,543),
respectively.
24
Prevalence of attention and conduct disorders increased with age. For example the prevalence of these conditions during 37-60
months of age in the full-term, moderate/late preterm and very/extremely preterm sub-groups were 2%, 3% and 6%, respectively.
25
Includes otitis media, allergic rhinitis, pharyngitis and laryngitis
114

Greater than 70% of medical costs among full-term and moderate/late preterm born children were
incurred for ambulatory care services between 6 and 60 months age. Among very/extreme
preterm born children, inpatient costs contributed to 50% of total medical cost during 6-18
months, and only 20% or lesser of total costs beyond 18 months of age, while the remaining were
incurred for ambulatory care services.
Figure 4-1 Mean inpatient and ambulatory healthcare costs during study period
by group

The logarithm of total healthcare cost distributions in the full-term and preterm group between 6
and 60 months of age are shown in Figure 4-2.
0
2,500
5,000
7,500
10,000
12,500
15,000
17,500
20,000
22,500
25,000
6-18 18-36 36-60 6-18 18-36 36-60 6-18 18-36 36-60
Controls 33-36 weeks of GA <=32 weeks of GA
115

Figure 4-2 Distribution of log of total healthcare costs over time

As shown in Figure 4-2, the log-transformed distributions were bimodal with a density of zeros,
followed by positive values and long right tails (especially in the preterm groups). The proportion
of children with zero utilization in a certain 6-month period was ≥10% beyond 12 months of age.
The proportion of zeros rose up to 30% in the full-term and moderate/late preterm groups beyond
42 months of age. Taking the full-term group as reference group, the mean difference in log
healthcare costs per child per 6-months during the study period was 0.12 and 1.26 among
moderate/late preterm and very/extremely preterm groups, respectively (Table 4-1).
116

Table 4-1 Mean differences in log healthcare (HC) costs between preterm and
full-term groups during study period
A. <33 weeks of GA vs. full-term B. 33-36 weeks of GA vs. full-term
Mean log HC cost
among <33 weeks of
GA group
5.91 (0.06)
Mean log HC cost in the
33-36 weeks of GA
group
4.77 (0.03)
Mean log HC cost
among full-term
4.65 (0.005)
Mean log HC cost
among full-term
4.65 (0.005)
Overall difference in
log HC costs
1.26*** (0.06)
Overall difference in log
HC costs
0.12*** (0.03)
***p<0.001; HC = healthcare
4.5.4 Interaction test results for the applicability of decomposition methods
Appendix 6 shows coefficients from a pooled ln-OLS regression of healthcare costs involving
interaction of each preterm category (versus full-term group) with eligibility category in the
previous period (SSI, AFDC or missing vs. expansion programs at ‘t-1’) and all chronic
conditions (observed at ‘t’). The intercept estimates for the moderate/late preterm (coefficient =
0.139, p<0.05) and very/extremely preterm groups (0.310, p<0.0001) and one or more interaction
terms involving each group were statistically significant. A test of the null hypothesis, that the
interaction terms are jointly zero was rejected [F (48, 476903) = 15.59 (p < 0.05)]. These results
lend further support to the use of Oaxaca type decompositions for studying the difference in
healthcare cost between preterm and full-term born children, indicating that the impact of
observed characteristics may differ between the gestational age categories.
117

4.5.5 Statistical test results for the validity of fixed effects
Table 4-2 reports the results of the Mundlak (1978) test and shows a comparison of the
coefficient estimates obtained from the random effects and Mundlak regressions for healthcare
costs conditional on Y
it
>0. The coefficients on each of the additional estimated mean transformed
variables were statistically significant. Further, the null hypothesis that the coefficients on the
augmented group mean variables are jointly zero was rejected (χ2 (13) = 1449471.9; p<0.001),
indicating that the assumption that observed characteristics are not correlated with the child fixed
effects is invalid.
Similarly, the results of the Sargan-Hansen over-identification test for fixed versus random
effects also resulted in rejection of null hypothesis that the random effects estimator is consistent
(| χ2 (21) | = 5335.5; p < 0.0001). The conclusion that the random effects assumption is invalid
remained true even when these tests were conducted in the full-term and preterm sub-groups
separately (See Table 4-12).

118

Table 4-2 Mundlak (1978) test for fixed versus random effects using the pooled
sample
Characteristics
Dependent var. = (Y
it
) = Log(healthcare
costs)|Yit > 0 Mundlak test and test
parameters (1978)
Random Effects
estimator
Mundlak-Random
Effects estimator
Constant
6.203*** 5.430***
Individual fixed effects:
μi = Φ(Xibar)
H0: Φ = 0
chi-squared(13, 385964) =
14497.91
p-value >| chi-sq|= <0.0001
[0.00822] [0.0174]
<33 weeks of GA(vs.
full-term)
0.674*** 0.318***
[0.0414] [0.0252]
33-36 weeks of GA
0.147*** 0.102***
[0.0197] [0.0164]
Male
0.0976*** 0.0771***
[0.00601] [0.00561]
Year of birth (2003 vs.
2002)
-0.143*** -0.106***
[0.00607] [0.00571]
White (vs. Hispanics)
-0.0703*** -0.0429***
[0.00850] [0.00832]
African-American
-0.254*** -0.253***
[0.0116] [0.0110]
Other Race
-0.270*** -0.201***
[0.0394] [0.0360]
Unknown Race
-0.190*** -0.0183*
[0.00872] [0.00867]
Congenital heart disease
1.165*** 0.859***
mean__chdx2
0.770***
[0.0262] [0.0250] [0.0591]
Hospitalization (t-1)
0.200*** -0.114*** mean__anyhos
p2_l1
2.105***
[0.0101] [0.00949] [0.0266]
SSI (vs. expansions)
1.346*** 0.573***
mean__tpcat2
1.468***
[0.0370] [0.0242] [0.0393]
AFDC
0.237*** 0.109***
mean__tpcat3
0.323***
[0.00575] [0.00613] [0.0113]
Missing program
category
0.120*** 0.240***
mean__tpcat4
-
0.634***
[0.0133] [0.0142] [0.0305]
12-18 (vs. 6-12 mos)
-0.241*** -0.286***
mean__p2
0.377***
[0.00531] [0.00595] [0.0307]
18-24 mos
-0.362*** -0.527***
mean__p3
0.818***
[0.00664] [0.00727] [0.0303]
24-30 mos
-0.612*** -0.797***
mean__p4
0.931***
[0.00726] [0.00757] [0.0322]
30-36 mos
-0.940*** -1.147***
mean__p5
1.467***
[0.00795] [0.00787] [0.0321]
36-42 mos
-0.936*** -1.144***
mean__p6
0.794***
[0.00867] [0.00864] [0.0368]
42-48 mos
-1.158*** -1.384***
mean__p7
1.357***
[0.00933] [0.00894] [0.0394]
48-54 mos
-1.090*** -1.301***
mean__p8
0.959***
[0.00929] [0.00885] [0.0376]
54-60 mos
-1.122*** -1.336***
mean__p9
1.320***
[0.00957] [0.00906] [0.0375]
SSI = Social security income; AFDC = Aid for families with dependent children; *p<0.05; +p<0.10;
119

4.5.6 Pooled OLS and fixed effects OLS estimation results
4.5.6.1 <33 weeks of GA versus full-term
Table 4-3 is showing the mean values and coefficients of observed characteristics of children
born before 33 weeks of GA versus full-term using pooled OLS and fixed effects (FE) panel
regressions of log healthcare costs. Columns E & H are showing the difference in the pooled OLS
and FE regression coefficients, respectively, between the two groups weighted by the mean
values of characteristics of full-term group (i.e., these columns show the incremental impact of
preterm birth on healthcare cost, any, with respect to these characteristics).
A comparison of the mean values showed a higher proportion of African-American children
(16% vs. 8%), recipients of SSI (32% vs. 1%) and children with hospital admission in the
previous period (10% vs. 4%) among very/extreme preterm versus full-term groups. The
remaining observed characteristics were similar between the groups (Table 4-3, columns A & C).
Within each group, disability status (or children receiving SSI) had the highest impact on
healthcare costs followed by, congenital heart defects, hospitalization, coverage under AFDC
category in the previous period, male gender and race-ethnicity.
The proportion of variance in log healthcare costs due to individual fixed effects terms (ρ) were
0.47 and 0.34 in the very/extreme preterm and full-term born groups, respectively, thus showing
substantial variation due to unobserved child specific effects in these groups.
As shown in Table 4-3, column E, the difference in pooled OLS coefficients between
very/extreme preterm and full-term groups, weighted by mean characteristics of the control
group, were statistically significant for gender (∆ = 0.12, p = 0.025), eligibility under AFDC
(versus income expansion programs; ∆ = 0.28, p < 0.001) and hospital admission in the previous
period (∆ = 0.023, p < 0.001). The direction of these differences was in favor of the full-term
120

group, showing a relatively greater impact of these characteristics among children born
very/extreme preterm. Also the difference in pooled OLS intercepts between very/extreme
preterm and full-term groups was 0.40 (p = 0.008), showing a significant baseline difference
between the two groups (see column E).
After controlling for unobserved fixed effects, there was a reduction in the size of the coefficient
estimates of observed time-varying characteristics in both the groups (columns F and G). The
differences in FE coefficients of observed time-varying characteristics between very/extremely
preterm and full-term groups were negligible, except for the difference in coefficients on AFDC
category (∆ = 0.103, p < 0.001), that is in favor of full-term, and age-period (∆ = -0.07 at mid-
point of age; p < 0.001) that is in favor of very/extreme preterm group. Using the fixed effects
model, the difference in intercepts between the two groups was 1.31 (p < 0.001) thus attributing a
significant part of the observed healthcare difference to the baseline group differences between
very/extreme preterm and full-term birth (see column H).
4.5.6.2 33-36 weeks of GA versus full-term
Table 4-4 is showing the mean values, coefficients and weighted difference in coefficients of
observed characteristics between moderate/late preterm (33-36 weeks of GA) and full-term
groups, using pooled OLS and fixed effects regressions of log healthcare costs. Observed
characteristics were comparable between the two groups, with the exception of a higher
proportion of children born in year 2003 (vs. 2002) in the moderate/late preterm group. Disability
status (or children receiving SSI) had the highest impact on healthcare costs followed by,
congenital heart defects, hospitalization, and coverage under AFDC in the previous period, male
gender and race-ethnicity.
121

Using pooled OLS model assumptions, the differences in coefficients (weighted by mean
characteristics of full-term group) between moderate/late preterm and full-term groups were small
for most observed characteristics, except birth cohort (∆ = 0.15, p = 0.004) and age-period (∆ = -
0.03 at mid-point of age, p < 0.001). The difference in intercept terms between the two groups
was not statistically significant (∆ = - 0.18, p = 0.159) using pooled OLS model.
Controlling for unobserved fixed effects had a significant impact on the coefficients of observed
characteristics causing the size of estimated coefficients to be smaller than the pooled OLS
coefficients. The differences in FE coefficients between the two groups were not statistically
significant except for age-period (∆ = -0.05 at mid-point of age, p < 0.001) that is in favor of the
moderate/late preterm group. After controlling for fixed effects, there was a positive and
statistically significant difference in the intercept terms between the two groups (∆ = 0.16, p <
0.001), thus attributing a significant part of the observed healthcare difference to the baseline
group differences between moderate/late preterm and full-term birth (see column H, Table 4-4).
122

Table 4-3 Mean values, coefficients and difference in coefficients of observed characteristics between
very/extremely preterm (< 33 weeks of GA) vs. full-term born children, obtained from pooled OLS
and fixed effects regressions of log healthcare costs
Observed Characteristics
(A) Means of
predictors (<33
weeks of GA)
(B) Coefficients
(Pooled OLS)
(<33 weeks of GA)
(C) Means of
predictors
(Full-term)
(D) Coefficients
(Pooled OLS)
(Full-term)
(E)
Difference
in
coefficients
* 𝐗𝐗 �
𝐂𝐂 (Pooled
OLS)
(F) Coefficients
(Fixed effects)
(<33 weeks of GA)
(G)
Coefficients
(Fixed effects)
(Full-term)
(H)
Difference
in
coefficients
* 𝐗𝐗 �
𝐂𝐂 (Fixed
effects)
N = 6987 R-squared: 0.25 N = 455,665 R-squared:0.17
R-squared: 0.19
rho = 0.47
R-squared:
0.13 rho = 0.34
Constant

6.49***

6.09*** 0.40** 7.58*** 6.28*** 1.31***
Male 0.52 0.34** 0.51 0.10*** 0.12* - - -
Year of birth (2003 vs. 2002) 0.59 -0.12 0.48 -0.21*** 0.05 - - -
White (vs. Hispanic) 0.13 -0.21 0.15 -0.26*** 0.008 - - -
African-American 0.16 -0.66*** 0.08 -0.52*** -0.01 - - -
Other Race 0.01 -0.575 0.01 -0.43*** -0.001 - - -
Race unknown 0.18 -0.32* 0.13 -0.76*** 0.06** - - -
SSI
t-1
(vs. expansions) 0.32 1.89*** 0.01 2.94*** -0.01*** 0.97*** 1.30*** -0.004***
AFDC
t-1
0.25 1.21*** 0.30 0.31*** 0.28*** 0.38*** 0.04*** 0.103***
Elig. cat. missing
t-1
0.03 -2.33*** 0.04 -1.61*** -0.03*** -1.16*** -0.41*** -0.032***
CHD
t
0.02 1.81*** 0.01 1.83*** -0.0001 0.92*** 1.21*** -0.002
Hospitalization
t-1
0.10 1.30*** 0.04 0.71*** 0.023*** -0.012 -0.097*** 0.003
12-18 (vs. 6-12 mos) 0.14 -1.19*** 0.15 -0.74*** -0.07*** -1.20*** -0.80*** -0.06***
18-24 months 0.12 -1.06*** 0.12 -0.84*** -0.03 -1.59*** -1.23*** -0.04***
24-30 months 0.12 -1.65*** 0.12 -1.21*** -0.05*** -2.33*** -1.68*** -0.08***
30-36 months 0.12 -2.27*** 0.12 -1.9*** -0.04** -3.01*** -2.44*** -0.07***
36-42 months 0.09 -2.08*** 0.08 -1.85*** -0.02 -2.92*** -2.29*** -0.05***
42-48 months 0.09 -2.23*** 0.08 -2.22*** -0.001 -3.16*** -2.73*** -0.04***
48-54 months 0.09 -1.96*** 0.08 -2.0*** 0.003 -2.90*** -2.51*** -0.03***
54-60 months 0.09 -2.07*** 0.08 -2.31*** 0.02 -3.03*** -2.82*** -0.02
***p<0.001 **p<0.01 *p<0.05 FE: Fixed effects rho = variance due to FE; SSI: Social security income AFDC: Aid for families with dependent children CHD: Congenital heart disease OLS: Ordinary least squares
123

Table 4-4 Mean values, coefficients and difference in coefficients of observed characteristics between
moderate/late preterm (33-36 weeks of GA) vs. full-term born children, obtained from pooled OLS
and fixed effects regressions of log healthcare costs
Observed
Characteristics
(A) Means of
predictors
(33-36 weeks
of GA)
(B) Coefficients
(Pooled OLS)
(33-36 weeks of
GA)
(C) Means
of predictors
(Full-term)
(D)
Coefficients
(Pooled OLS)
(Full-term)
(E) Difference
in coefficients
* 𝐗𝐗 �
𝐂𝐂 (Pooled
OLS)
(F) Coefficients
(Fixed effects)
(33-36 weeks of
GA)
(G) Coefficients
(Fixed effects)
(Full-term)
(H)
Difference
in
coefficients
* 𝐗𝐗 �
𝐂𝐂 (Fixed
effects) N = 14,320 R-squared: 0.19 N = 455,665
R-
squared:0.17
R-squared:
0.15 rho = 0.36
R-squared: 0.13
rho = 0.34
Constant

5.90***

6.09*** -0.18 6.44*** 6.28*** 0.16***
Male 0.50 0.14* 0.51 0.10*** 0.02 - -

Year of birth (2003 vs.
2002)
0.91 0.098 0.48 -0.21*** 0.15** - -

White (vs. Hispanic) 0.16 -0.14 0.15 -0.26*** 0.02 - -

African-American 0.11 -0.46*** 0.08 -0.52*** 0.004 - -

Other Race 0.01 0.035 0.01 -0.43*** 0.003 - -

Race unknown 0.12 -0.58*** 0.13 -0.76*** 0.02 - -

SSI t-1 (vs. expansions) 0.02 3.05*** 0.01 2.94*** 0.001 1.14*** 1.30*** -0.002
AFDC t-1 0.33 0.40*** 0.30 0.31*** 0.028 0.10 0.04*** 0.02
Elig. cat. missing t-1 0.04 -1.59*** 0.04 -1.61*** 0.0006 -0.34* -0.41*** 0.003
CHD t 0.01 1.78*** 0.01 1.83*** -0.0003 1.32*** 1.21*** 0.001
Hospitalization t-1 0.05 0.82*** 0.04 0.71*** 0.004 -0.16 -0.097*** -0.002
12-18 (vs. 6-12 mos) 0.16 -0.88*** 0.15 -0.74*** -0.02* -0.97*** -0.80*** -0.025**
18-24 months 0.12 -0.63*** 0.12 -0.84*** 0.025** -1.12*** -1.23*** 0.013
24-30 months 0.12 -1.33*** 0.12 -1.21*** -0.01 -1.89*** -1.68*** -0.03***
30-36 months 0.12 -2.22*** 0.12 -1.9*** -0.03*** -2.8*** -2.44*** -0.05***
36-42 months 0.08 -2.24*** 0.08 -1.85*** -0.033*** -2.69*** -2.29*** -0.03***
42-48 months 0.08 -2.08*** 0.08 -2.22*** 0.012 -2.65*** -2.73*** 0.007
48-54 months 0.08 -1.90*** 0.08 -2.0*** 0.008 -2.44*** -2.51*** 0.006
54-60 months 0.08 -2.21*** 0.08 -2.31*** 0.008 -2.72*** -2.82*** 0.008
***p<0.001 **p<0.01 *p<0.05 FE: Fixed effects rho = variance due to FE; SSI: Social security income AFDC: Aid for families with dependent children CHD: Congenital heart disease
OLS: Ordinary least squares
124

4.5.7 Results of two-part model regressions
4.5.7.1 Results from Step I of two-part models
The unadjusted mean differences in proportion of children with non-zero healthcare use (Y
it
> 0)
in a given period, between very/extreme preterm, moderate/late preterm and full-term born
groups were 6% (p < 0.001), and 0.6% (p = 0.147), respectively (Table 4-5).
Table 4-5 Differences in proportion of children with healthcare costs greater
than zero in a given period
A. <33 weeks of GA vs. full-term B. 33-36 weeks of GA vs. full-term
Mean proportion (<33
weeks of GA group) (N =
6,987)
0.863 (0.344)
Mean proportion (33-36
weeks of GA group) (N =
14,320)
0.811 (0.39)
Mean proportion (Full-term
cohort) (N = 455,665)
0.808 (0.394)
Mean proportion (Full-term
cohort) (N = 455,665)
0.808 (0.394)
Raw difference in
proportions (<33 weeks vs.
full-term)
0.06*** (0.006)
Raw difference in
proportions (33-36 weeks
vs. full-term)
0.006 (0.004)
***p<0.001
Table 4-6 below is showing the marginal effects of covariates on the probability of healthcare
utilization, using population averaged probit and conditional fixed effects logit models for
very/extreme preterm and full-term groups. Within each group, disabled children (children
receiving SSI) had a highly significant impact on the probability of healthcare use in both groups,
followed by, congenital heart defects, hospitalization and coverage under AFDC category in the
previous time period, male gender and race-ethnicity.
Using a pooled probit model, the differences in marginal effects between very/extreme preterm
and full-term groups were significant for coverage under AFDC category (∆ = 0.01, p < 0.001)
and age-period (∆ = 0.01 at mid-point of age, p < 0.001). After controlling for unobserved fixed
125

effects, the marginal effects of characteristics were significantly reduced in both groups such that,
only the marginal impact of age-period was significant in the very/extreme preterm group (∆ = -
0.31 at mid-point of age, p < 0.001). Using conditional fixed effects model, the differences in
marginal effects between the two groups were not statistically significant for any of the
characteristics. These results suggest that any difference in probability of healthcare use between
the groups were mostly due to differences in frequencies of these characteristics (endowments
effect) and/or due to the underlying group effect (i.e. very/extreme prematurity).
Table 4-7 shows the results for step I of two-part model for moderate/late preterm group in
comparison with full-term. These results were much similar to those observed for very/extreme
preterm children discussed above. After controlling for fixed effects, only marginal effects of
CHD (∆ = 0.09, p < 0.05) and age-period (∆ = -0.32 at mid-point of age, p < 0.001) remained
significant among children in the moderate/late preterm group. Using conditional fixed effects
model, the differences in marginal effects between the two groups were not significant for any
observed characteristics.

126

Table 4-6 Panel regressions showing marginal effects of observed characteristics
on probability of healthcare use and their differences between
very/extreme preterm and full-term born children
Group
characteristics
(A) Marginal
effects (Pooled
Probit) Pr[Y
it
> 0|
X
it
]/ ∆ X
it
(<33
weeks of GA)
(B) Marginal
effects
(Pooled
Probit) Pr[Y
it

> 0| X
it
]/ ∆ X
it

(Full-term)
(C)
Difference
in ME * X
c

(Pooled
Probit)
(D) Marginal Effects
(Conditional FE logit)
Pr[Y
it
> 0| µ
i
,X
it
] / ∆ X
it

(<33 weeks of GA)
(E) Marginal Effects
(Conditional FE
logit) Pr[Y
it
> 0|
µ
i
,X
it
] / ∆ X
it
(Full-
term)
(F)
Difference in
ME * X
c

(Conditional
FE logit)
N 6,987 455,665

3,265 281,516

Model Statistics - -

Pseudo R2: 0.1647
Log pseudo-likelihood:
-1053.44
Pseudo R2:0.1416
Log pseudo-likelihood:
-94121.999
-
Male 0.017 0.006*** 0.006 - - -
Year of birth (2003
vs. 2002)
-0.006 -0.015*** 0.004 - - -
White (vs.
Hispanic)
-0.035* -0.03*** 0.000 - - -
African-American -0.054*** -0.05*** 0.000 - - -
Other Race -0.032 -0.042*** 0.000 - - -
Race unknown -0.075*** -0.12*** 0.005** - - -
SSI
t-1
(vs.
expansions)
0.101*** 0.22*** -0.001*** 0.0428 0.086*** 0.000
AFDC
t-1
0.044*** 0.008*** 0.011** -0.002 -0.016*** 0.004
Elig. cat. Missing
t-1
-0.152*** -0.14*** 0.000 -0.035* -0.008*** -0.002
†

CHD
t-1
0.059* 0.06*** 0.000 0.04 0.02** 0.000
Hospitalization
t-1
0.052*** 0.04*** 0.000 0.02 0.015*** 0.000
12-18 (vs. 6-12
mos)
-0.082*** -0.09*** 0.002 -0.15*** -0.13*** -0.003
18-24 months -0.075*** -0.11*** 0.004* -0.19*** -0.19*** -0.001
24-30 months -0.107*** -0.15*** 0.005* -0.25*** -0.23*** -0.002
30-36 months -0.168*** -0.24*** 0.01*** -0.31*** -0.29*** -0.002
36-42 months -0.154*** -0.22*** 0.007** -0.30*** -0.27*** -0.002
42-48 months -0.174*** -0.27*** 0.008* -0.33*** -0.31*** -0.002
48-54 months -0.147*** -0.23*** 0.007*** -0.31*** -0.29*** -0.001
54-60 months -0.162*** -0.29*** 0.010*** -0.32*** -0.32*** 0.000
SSI = Eligible due to receipt of social security income; AFDC = Aid for Families with Dependent Children; CHD = congenital heart disease; ‘period’ refers to each 6-month
age-period and t-1 refers to the previous 6-month time period. ***p<0.001; **p<0.01; *p<0.05;
†
p<0.10;
127

Table 4-7 Panel regressions showing marginal effects of observed characteristics
on probability of healthcare use and their differences between
moderate/late preterm and full-term born children
Group
characteristics
(A) Marginal
Effects (Pooled
Probit) Pr[Y
it
>
0| X
it
] / ∆ X
it
(33-
36 weeks of GA)
(B) Marginal
Effects
(Pooled
Probit) Pr[Y
it

> 0| X
it
] / ∆ X
it

(Full-term)
(C)
Difference
in ME * X
c

(Pooled
Probit)
(D) Marginal Effects
(Conditional FE logit)
Pr[Y
it
> 0| µ
i
,X
it
] / ∆ X
it

(33-36 weeks of GA)
(E) Marginal
Effects
(Conditional FE
logit) Pr[Y
it
> 0|
µ
i
,X
it
] / ∆ X
it
(Full-
term)
(F)
Difference
in ME * X
c

(Condition
al FE logit)
N 14,320 455,665

8,595 281,516

Model Statistics - -

Pseudo R2: 0.1525
Log pseudo-likelihood:
-2842.75
Pseudo R2:0.1416
Log pseudo-
likelihood: -
94121.999
-
Male 0.01 0.006*** 0.003 - - -
Year of birth
(2003 vs. 2002)
0.006 -0.015*** 0.010 - - -
White (vs.
Hispanic)
-0.009 -0.03*** 0.004* - - -
African-American -0.04*** -0.05*** 0.000 - - -
Other Race 0.01 -0.042*** 0.000 - - -
Race unknown -0.10*** -0.12*** 0.002
†
- - -
SSI
t-1
(vs.
expansions)
0.16*** 0.220*** -0.001 0.011 0.09*** 0.000
AFDC
t-1
0.012 0.008*** 0.001 -0.007 -0.016*** 0.003
Elig. cat.
Missing
t-1

-0.14*** -0.14*** 0.000 -0.006 -0.008*** 0.000
CHD
t-1
0.09** 0.06*** 0.000 0.09* 0.02** 0.002
Hospitalization
t-1
0.042* 0.04*** 0.000 0.004 0.015*** 0.000
12-18 (vs. 6-12
mos)
-0.10*** -0.09*** 0.000 -0.14*** -0.13*** -0.001
18-24 months -0.080*** -0.11*** 0.005* -0.17*** -0.19*** 0.003*
24-30 months -0.15*** -0.15*** 0.000 -0.24*** -0.23*** -0.001
30-36 months -0.26*** -0.24*** -0.002 -0.32*** -0.29*** -0.003
†

36-42 months -0.28*** -0.22*** -0.003
†
-0.32*** -0.28*** -0.004*
42-48 months -0.25*** -0.27*** 0.001 -0.31*** -0.31*** 0.000
48-54 months -0.20*** -0.23*** 0.002 -0.28*** -0.29*** 0.001
54-60 months -0.26*** -0.29*** 0.002 -0.31*** -0.32*** 0.001
SSI = Eligible due to receipt of social security income; AFDC = Aid for Families with Dependent Children; CHD = congenital heart disease; ‘period’ refers to each 6-
month age-period and t-1 refers to the previous 6-month time period.
128

4.5.7.2 Results from Step II of two-part models
4.5.7.2.1 <33 weeks of GA versus full-term
Table 4-8 below is showing coefficients from fixed effects models of log healthcare costs
conditional on Y
it
> 0 for very/extreme preterm versus full-term groups. The difference in
coefficients on AFDC category and hospitalization in the previous period, between the two
groups were 0.08 (p < 0.001) and 0.006 (p < 0.05), respectively, the difference being in favor of
the full-term group. Conversely, the differences in coefficients on disability status in the previous
period ∆ = -0.002, p < 0.001) and age-period were negative (i.e. in favor of preterm group) and
were statistically significant (average difference in the impact of age-period on healthcare costs
between the groups being -4% to -9%). The fixed effects adjusted difference in the intercept
coefficients between very/extreme preterm and full-term groups was 1.39 (p < 0.001; see column
E, Table 4-8) that indicates a substantial baseline difference between the two groups.
4.5.7.2.2 33-36 weeks of GA versus full-term
Table 4-9 is showing coefficients from fixed effects models of log healthcare costs conditional on
Y
it
> 0 for moderate/late preterm versus full-term groups. The differences in coefficients on SSI
and AFDC categories were 0.002 (p < 0.05) and 0.02 (p < 0.01), in favor of the full-term group.
Conversely, a difference of -1% to -4% in the coefficients of age-period on healthcare costs
(favoring preterm) was observed between the two groups until 42 months of age. The fixed
effects adjusted difference in the intercept coefficients between moderate/late preterm and full-
term groups was 0.20 (p < 0.001).

129

Table 4-8 Means and fixed effects coefficients of characteristics predicting log
cost given Yit > 0 in the very/extreme preterm and full-term groups
Observed
Characteristics
(A) Means of
predictors (<33
weeks of GA)
(B) Coefficients
(Fixed Effects)
E[Y
it
| Y
it
> 0, µ
i
,
X
it
] (<33 weeks of
GA)
(C) Means of
predictors
(Full-term)
(D) Coefficients
(Fixed Effects)
E[Y
it
| Y
it
> 0, µ
i
,
X
it
] (Full-term)
(E) Difference in
coefficients* X
C

(Fixed effects)
Model Statistics N = 6028
R-squared: 0.155
rho = 0.579
N =368,317
R-squared:0.113
rho = 0.447
-
Constant
†
- 7.807***

6.416*** 1.39***
SSI
t-1
(vs.
expansions)
0.34 0.538*** 0.01 0.667***
-0.002***
AFDC
t-1
0.26 0.352*** 0.33 0.107*** 0.08***
Elig. cat. Missing
t-1
0.01 0.205 0.02 0.236*** -0.001
CHD
t
0.03 0.711*** 0.007 0.849*** -0.001
Hospitalization
t-1
0.11 0.009 0.04 -0.119*** 0.006*
12-18 (vs. 6-12 mos) 0.15 -0.670*** 0.16 -0.275*** -0.06***
18-24 months 0.13 -0.992*** 0.13 -0.518*** -0.06***
24-30 months 0.12 -1.527*** 0.12 -0.779*** -0.09***
30-36 months 0.11 -1.902*** 0.11 -1.124*** -0.08***
36-42 months 0.09 -1.955*** 0.08 -1.125*** -0.065***
42-48 months 0.08 -2.006*** 0.07 -1.375*** -0.046***
48-54 months 0.08 -1.881*** 0.08 -1.291*** -0.045***
54-60 months 0.08 -1.876*** 0.07 -1.327*** -0.039***
CHD = congenital heart disease; ‘period’ refers to each 6-month age-period and t-1 refers to the previous 6-month time period.

130

Table 4-9 Means and fixed effects coefficients of characteristics predicting log
cost given Yit > 0 in the moderate/late preterm and full-term groups
Observed
Characteristics
(A) Means of
predictors (33-36
weeks of GA)
(B) Coefficients
(Fixed Effects)
E[Y
it
| Y
it
> 0, µ
i
, X
it
]
(33-36 weeks of
GA)
(C) Means of
predictors
(Full-term)
(D) Coefficients
(Fixed Effects)
E[Y
it
| Y
it
> 0, µ
i
,
X
it
] (Full-term)
(E) Difference in
coefficients* X
C

(Fixed effects)
Model Statistics N = 11,619
R-squared: 0.126
rho = 0.493
N =368,317
R-squared:0.113
rho = 0.447
-
Constant
†
- 6.617*** - 6.416*** 0.20***
SSI
t-1
(vs. expansions) 0.02 0.834*** 0.01 0.667*** 0.002*
AFDC
t-1
0.36 0.156*** 0.33 0.107*** 0.02**
Elig. cat. Missing
t-1
0.02 0.369*** 0.02 0.236*** 0.003*
CHD
t
0.01 0.880*** 0.007 0.849*** 0.0002
Hospitalization
t-1
0.06 -0.159** 0.04 -0.119*** -0.002
12-18 (vs. 6-12 mos) 0.17 -0.443*** 0.16 -0.275*** -0.03***
18-24 months 0.13 -0.570*** 0.13 -0.518*** -0.007
24-30 months 0.12 -0.987*** 0.12 -0.779*** -0.03***
30-36 months 0.11 -1.504*** 0.11 -1.124*** -0.04***
36-42 months 0.07 -1.324*** 0.08 -1.125*** -0.02***
42-48 months 0.07 -1.341*** 0.07 -1.375*** 0.003
48-54 months 0.08 -1.348*** 0.08 -1.291*** -0.004
54-60 months 0.07 -1.372*** 0.07 -1.327*** -0.003
CHD = congenital heart disease; ‘period’ refers to each 6-month age-period and t-1 refers to the previous 6-month time period.
4.5.8 Blinder-Oaxaca decomposition results
4.5.8.1 <33 weeks of GA versus full-term
Table 4-10 shows a comparison of the Blinder-Oaxaca decomposition estimates for the healthcare
cost difference between very/extreme preterm and full-term children, obtained using cross-
sectional and panel data approaches. The treatment effects from the perspective of a full-term
child (ATU) and preterm child (ATET) are shown in panels A and B, respectively.

131

Table 4-10 Decomposition of total healthcare cost difference between
very/extreme preterm (<33 weeks of GA) and full-term children
A. Using characteristics of full-term child as reference (treatment effect = ATU)
Decompositi
on term
Estimates of differentials
(<33 weeks of GA vs. full-
term)
(A) Ln-OLS
(B) Hausman-
Taylor RE
estimator
(C) FE-OLS
(within
estimator)
(D) TPM with FE
(FE logit & within
estimators)
Endowments
effect (X
p
-
X
c
)*β
p

Mean diff. (±SE) in log HC
costs due to difference in
group characteristics
0.57*** (0.05) 0.31*** (0.05) 0.28*** (0.02) 0.17** (0.05)
% of total difference due to
diff. in group characteristics
45%***
[37% - 53%]
27%***
[19% - 36%]
22%***
[18% - 25%]
20%**
[8% - 31%]
Mean diff. in healthcare costs
per 6-months (US$)
a
[95% CI]
$835***
[$687 - $984]
$416***
[$293 - $555]
$408***
[$334 - $464]
$204**
[$82 - $316]
Coefficients
effect (ATU)
(β
p
- β
c
)* 𝑿𝑿 𝑪𝑪 � � � �

Mean diff. (±SE) in log HC
costs due to coefficients
0.69*** (0.06) 0.82*** (0.07) 0.98*** (0.01) 0.70*** (0.08)
% of total difference due to
coefficients
55%*** [45%
- 64%]
73%***
[60% - 85%]
78%***
[76% - 79%]
80%***
[63% - 98%]
Mean diff. in healthcare costs
per 6-months (US$)
a
[95% CI]
$1,021*** [$835
- $1,188]
$1,124***
[$924 - $1,309]
$1,448***
[$1,411 -
$1,466]
$816***
[$642 - $999]
a
Calculated using formula: overall healthcare cost difference (X)% contribution due to characteristics (or coefficients); V/E = very/extremely; GA = gestational age; RE
= random effects; FE = fixed effects; TPM = two-part model; CI = confidence intervals; SE = standard errors. ***p<0.001; **p<0.01; *p<0.05; †p<0.10.
B. Using characteristics of preterm child as reference (treatment effect = ATET)
Decomposition
term
Estimates of differentials
(<33 weeks of GA vs. full-
term)
(A) Ln-OLS (B) Hausman-
Taylor RE
estimator
(C) FE OLS
(within
estimator)
(D) TPM with FE
(FE logit & within
estimators)
Endowments
effect (X
p
-
X
c
)*β
C

Mean diff. (±SE) in log HC
costs due to diff. in group
characteristics
0.87*** (0.04) 0.53*** (0.04) 0.39*** (0.02) 0.28*** (0.03)
% of total difference due to
diff. in group characteristics
69%***
[62% - 76%]
47%***
[41% - 53%]
31%***
[28% - 34%]
32%***
[25% - 40%]
Mean diff. in healthcare costs
per 6-months (US$)
a
[95% CI]
$1,281***
[$1,151-$1,411]
$724***
[$632- $816]
$575***
[$520 - $631]
$326**
[$255 - $408]
Coefficients
effect (ATET)
(β
p
- β
c
)* 𝑿𝑿 𝑷𝑷 � � � �

Mean diff. (±SE) in log HC
costs due to coefficients
0.39*** (0.06) 0.60*** (0.06) 0.87*** (0.01) 0.59*** (0.06)
% of total difference due to
coefficients
31%***
[21% - 40%]
53%***
[43% - 63%]
69%*** [67%
- 71%]
68%***
[53% - 82%]
Mean diff. in healthcare costs
per 6-months (US$)
a
[95% CI]
$575***
[$390 - $742]
$816***
[$662 - $970]
$1,281***
[$1,244 - $1,318]
$693***
[$540 - $836]
a
Calculated using formula: overall healthcare cost difference (X) % contribution due to characteristics (or coefficients); V/E = very/extremely; GA = gestational age; RE =
random effects; FE = fixed effects; TPM = two-part model; CI = confidence intervals; SE = standard errors. ***p<0.001; **p<0.01; *p<0.05;
†
p<0.10
132

As shown in Table 4-10 panels A and B, the cross-sectional ln-OLS model that ignores
unobserved heterogeneity, underestimates the true impact of preterm birth occurring at < 33
weeks of gestational age by attributing more explanatory power to the group characteristics.
Using characteristics of the full-term group as reference, 45% of the 1.257 ln-healthcare cost
difference is due to differences among group characteristics (95% CI: 37% - 53%) and 55% of the
mean cost differential is due to coefficients (95% CI: 45% - 64%). When individual heterogeneity
was accounted for in the estimation, the average treatment effect on untreated (ATU) was higher
than the corresponding pooled OLS based estimate of ATU by 18-25%, depending on model
type. Similarly, depending on model type, accounting for child fixed effects led to an increase in
average treatment effect on treated (ATET) by 22-37%, as compared to the corresponding pooled
OLS based estimate of ATET (columns B to D vs. column A).
The average treatment effect on treated using the within-estimator (panel B) was 16% more than
the corresponding estimate of ATET obtained using the random effects Hausman-Taylor model.
The ATU and ATET estimates derived from two-part model of ln-healthcare costs, adjusted for
fixed effects in both steps were, 0.70 (80% of total healthcare cost differential; 95% CI: 63% -
98%) and 0.59 (68%; 95% CI: 53% - 82%), respectively.
Another point of distinction between the cross-sectional and panel data models is the gap between
the ATU and ATET estimates. While there was a significant gap (~24%) between the two
estimates using the cross-sectional approach, the gap between these estimates became smaller
(~10%) after accounting for fixed effects, indicating that the panel estimate of coefficients effects
is closer to the true effect of prematurity (see Figure 4-3).
133

Figure 4-3 Comparison of percent healthcare cost difference due to returns to
prematurity from the perspective of full-term (ATU) and children
born before 33 weeks of GA (ATET)

[Error bars represent bootstrapped 95% confidence intervals; ***p<0.001; H-T: Hausman-Taylor; FE =
Fixed effects; TPM = Two-part model; OLS = Ordinary Least Squares; ATT = Average treatment effect on
treated; ATU = Average treatment effect on untreated]
Table 4-10 is also showing the predicted incremental healthcare costs per 6-months,
corresponding with the endowments (i.e. portion explained by differences in group
characteristics) and coefficients effects (portion explained by differences in coefficients), for each
model type. Using two-part model with fixed effects, the average treatment effect of very/extreme
prematurity on treated is $693 per 6-months per child (95% CI: $ 540 - 836) and the part of
healthcare cost differential due to differences in group characteristics is $326 (95% CI: $ 255 –
405), for the period between 6 and 60 months of age. The average treatment effect of
very/extreme prematurity on untreated (controls) is $816 (95% CI: $ 642 – 999) and the
***
***
***
***
***
***
***
***
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
Ln-OLS H-T RE FE (within) TPM with FE
% Difference due to Coefficients
Model Type
ATU
ATET
134

corresponding part of healthcare cost differential due to differences in group characteristics is
$204 (95% CI: $ 82 – 316) per patient per 6-months, for the period between 6 and 60 months.
4.5.8.2 33-36 weeks of GA versus full-term:
Using pooled OLS and random effects H-T model, the ATU and ATET estimates of preterm birth
occurring between 33-36 weeks of GA were separated by a large gap (see Figure 4-4). Further,
the treatment effect estimates had wider confidence intervals suggesting that, there may be no
incremental financial impact of preterm birth occurring between 33-36 weeks compared to
children born full-term. However, the estimates of ATU and ATET converged after elimination
of bias due to fixed effects (Figure 4-4). Using two-part model with fixed effects, the contribution
of returns to preterm birth (coefficients effect) towards healthcare cost difference between
moderate/late preterm and full-term children, were not statistically significant (ATU = 55%, p-
value = 0.231 and ATET = 63% p-value = 0.164 of the 0.07 TPM adjusted cost differential; Table
4-11). The endowments effect due to differences among group characteristics for the period
between 6 and 60 months of age was $18 per child per 6-months using 𝛽𝛽 𝑃𝑃 as counterfactual
weight (95% CI: $ 7 - 31) or $22 per child per 6-months using 𝛽𝛽 𝐶𝐶 as counterfactual weight (95%
CI: $ 7 - 38).
135

Figure 4-4 Comparison of percent healthcare cost difference due to returns to
prematurity among untreated (full-term) and treated (born preterm
33-36 weeks of GA) children

[Error bars represent bootstrapped 95% confidence intervals; ***p<0.001; H-T: Hausman-Taylor; FE =
Fixed effects; TPM = Two-part model; OLS = Ordinary Least Squares]
***
***
***
***
-130%
-80%
-30%
20%
70%
120%
170%
220%
Ln-OLS H-T RE FE (within) TPM with FE
% Difference due to
Coefficients
Model Type
ATU
ATET
136

Table 4-11 Decomposition of total healthcare cost difference between
moderate/late preterm (33-36 weeks of GA) and full-term children
A. Using characteristics of full-term child as reference (treatment effect = ATUT)
Decomposition
estimates
Estimates of differentials
(33-36 weeks of GA vs. full-term)
(A) FE OLS
(within estimator)
(B) TPM with FE
(within estimator)
Endowments
effect
(X
p
- X
c
)*β
p

Mean diff. (±SE) in log HC costs due
to group characteristics
0.04*** (.01) 0.03*** (0.008)
% of total difference due to group
characteristics
33%***
[17% - 50%]
45%***
[15% - 77%]
Mean diff. in predicted healthcare
costs per 6-months (US$)
a
[95% CI]
$31***
[$16 - $47]
$22***
[$7 - $38]
Coefficients
effect
(ATU)
(β
p
- β
c
)*X
c

Mean diff. (±SE) in log HC costs due
to coefficients
0.08*** (0.001) 0.04 (0.03)
% of total difference due to
coefficients
67%***
[66% - 75%]
55%
[-35% - 145%]
Mean diff. in predicted healthcare
costs per 6-months (US$)
a
[95% CI]
$63***
[$62 - $70]
$27
[-$17 - $72]
a
Calculated by applying cost ratio between groups in the estimation sample to the unadjusted reference (full-term) group costs; GA =
gestational age; RE = random effects; FE = fixed effects; TPM = two-part model; CI = confidence intervals; SE = standard errors.
***p<0.001; **p<0.01; *p<0.05;
†
p<0.10
B. Using characteristics of preterm child as reference (treatment effect = ATET)
Decomposition
estimates
Estimates of differentials (33-36
weeks of GA vs. full-term)
(A) FE OLS
(within estimator)
(B) TPM with FE
(within estimator)
Endowments
effect (X
p
-
X
c
)*β
C

Mean diff. (±SE) in log HC costs due
to group characteristics
0.04*** (0.01) 0.024** (0.008)
% of total difference due to group
characteristics
33%***
[17% - 50%]
37%**
[15% - 62%]
Mean diff. in predicted healthcare
costs per 6-months (US$)
a
[95% CI]
$31***
[$16 - $47]
$18** [$7 - $31]
Coefficients
effect (ATET)
(β
p
- β
c
)*X
P

Mean diff. (±SE) in log HC costs due
to coefficients
0.08*** (0.001) 0.04 (0.03)
% of total difference due to
coefficients
67%*** [66% -
75%]
63% [-31% - 154%]
Mean diff. in predicted healthcare
costs per 6-months (US$)
a
[95% CI]
$63*** [$62 - $70] $31 [-$15 - $76]
a
Calculated by applying cost ratio between groups in the estimation sample to the unadjusted reference (full-term) group costs; GA =
gestational age; RE = random effects; FE = fixed effects; TPM = two-part model; CI = confidence intervals; SE = standard errors;
***p<0.001; **p<0.01; *p<0.05; †p<0.10
137

4.5.9 Specification test results
Table 4-12 provides a summary of the specification tests conducted on the linear panel data
models. The results for the various tests of fixed versus random effects are explained in detail in
Section 4.5.5.
In summary, all tests of fixed versus random effects showed that the random effects assumption is
invalid. The Wooldridge F-test for serial correlation among pooled as well as individual samples
led to rejecting the null hypothesis that there is no serial correlation in all samples. However, the
more powerful tests such as the modified BFN Durbin-Watson and the Baltagi-Wu LBI tests
showed a value of the statistic that is very close to 2, which suggests that there may be no
autocorrelation in the panel. The specification with the time dummies had the lowest BIC and
further, the difference in BIC statistics between the two models (i.e. the model with time variable
as a continuous variable versus the model with time dummies) was > 10 (271.65), showing strong
evidence in favor of specifying the time fixed effects as dummy variables. The null hypothesis for
the pooled interaction test, that the coefficients on the two preterm categories and their interaction
with explanatory variables are jointly zero was rejected [F (48, 476903) = 15.59 (p < 0.05)]. This
suggests that the Oaxaca type decomposition approach might be suitable for studying the
difference in healthcare cost between preterm and full-term born children given the possible
interaction effects between gestational age, disability status and chronic conditions.
138

Table 4-12 Linear panel data specification tests
Test Model Test hypotheses Sample Test statitistic p-value
Mundlak (1978) test
Random effects
regression augmented
with group means of
X
it

H
0
: Random effects is
consistent
H
1
: random effects is
inconsistent
Pooled χ
2
(28, 476972): 10304.44 <0.001
Controls only χ2 (28, 455665): 15119.15 <0.001
33-36 weeks of GA χ2 (28, 14320): 459.21 <0.001
GA: <33 weeks χ2 (28, 6987): 315.09 <0.001
Overidentification test for
fixed vs. random effects
(Schaeffer & Stillman,
2006)
Random effects
regressions augmented
with deviation-from-
mean transformed X
it

H0: Individual effects are
orthogonal to regressors
H1: random effects is
inconsistent
Pooled χ2 (35, 476972): 110000 <0.001
Controls only χ2 (35, 455665):8669.828 <0.001
33-36 weeks of GA χ2 (34, 14320):852.284 <0.001
GA: <33 weeks χ2 (34, 6987 ): 472.421 <0.001
Wooldridge-Drukker panel
autocorrelation (Drukker
2003)
Panel data pooled OLS
model with robust
standard errors
H0: no first order
correlation of
idiosyncratic errors
H1: serial correlation of
error terms
Pooled sample F(1, 67607): 1411.35 <0.001
controls only F(1, 64599): 1313.45 <0.001
GA: 33-36 weeks F(1, 2070): 52.687 <0.001
GA: <33 weeks F(1, 936): 497.457 <0.001
Bhargava et al.,Modified
Durbin-Watson (Bhargava,
et al. 1982; Baltagi, et al.
2007)
Fixed effects model
with AR(1)
disturbances
H0: no serial correlation
H1: serial correlation
Pooled sample modified DBW:1.775

Controls only modified DBW:1.779 -
GA: 33-36 weeks modified DBW:1.767 -
GA: <33 weeks modified DBW:1.663 -
Baltagi-Wu (1999)
(Locally Best Invariant)
Fixed effects model
with AR(1)
disturbances
H0: no serial correlation
H1: serial correlation
Pooled sample LBI: 2.135 -
Controls only LBI: 2.135 -
GA: 33-36 weeks LBI: 2.14 -
GA: <33 weeks LBI: 2.029 -
Time variable
specification (Bayesian
Information Criteria)
Fixed effects model
with cluster roubst
errors
Linear time trend is
specified correctly
Pooled sample
BIC (linear time spec.) =
1897501
BIC (time dummies spec.) =
1891622; BIC difference = 5879
-
Interaction test to check
applicability of Blinder-
Oaxaca technique
Ln-OLS regression of
healthcare costs on
preterm categories and
interaction with
explanatory variables
H
0
: Interaction terms are
jointly zero
H
1
: Interaction terms
jointly not equal to zero
Pooled F (48, 476903): 15.59 <0.05
139

4.5.10 Attrition test results
4.5.10.1 Descriptive results
Figure 4-5 is showing the percent children who remained in the fee-for-service sample at the end
of each wave of follow-up (cumulative retention), and Appendix 7 is showing the cross-sectional
attrition rates by wave in each group. As shown in these figures, the pattern of attrition over time
was mostly similar between children in the control and 33-36 weeks of GA cohorts. The
cumulative retention rate was initially lower among children born at <33 weeks of GA by end of
wave 2 (i.e. by end of 18 months of age), but this is mainly attributed to a high rate of mortality
occurring in the previous wave (see Appendix 3 for cumulative mortality rates in all 3 groups).
However, effective attrition during each wave (for reasons other than death) was much less in the
cohort born <33 weeks of GA as compared to the other two groups (see Appendices 3 and 7). By
the end of 60 months of age, the proportions of children who remained in the cohort relative to
the sample size at time 0 were, 36%, 34% and 35% in the control, moderate/late and
very/extremely preterm groups, respectively.
Comparison of demographic characteristics among the non-attrited between wave 2 (6-18
months) and the last wave (37-60 months) (Appendix 4) showed small changes in each group’s
constitution over time with respect to these characteristics. The proportion of children of Hispanic
origin in the sample increased with time, accompanied by a proportional decrease in children
whose race/ethnicity information was missing. The percentage of children belonging to other race
or ethnic backgrounds remained stable over time. Compared to the rest of the Texas regions, a
higher percentage of children born in the Metroplex and Gulf coast areas remained in the sample
up to end of 60 months. Similarly children born in the year 2002 were more likely to remain in
the FFS sample than children born in the year 2003. In the < 33 weeks of GA cohort, the
proportion of children with extremely low birth weight (≤ 1,249g) increased over time. Also there
140

were small differences in distribution of preterm children who suffered various complications at
the time of birth (for example children who experienced respiratory failure during the neonatal
period were more likely to stay).
26

Figure 4-5 Cumulative retention (%) in the full-term and preterm groups

Table 4-13 is showing the differences in overall log healthcare cost between the preterm and full-
term groups derived from the ‘full’ (all-available) versus ‘completing’ (i.e. those who remained in
the cohort up to 60 months) samples. The difference in overall log cost differential (difference-in-
difference) between the two estimation samples was 0.20 log-points in the case of very/extremely
preterm group and 0.06 log-points for moderate/late preterm group. In other words, using only the
data for infants who remained in the cohort up to completion of 60 months of age, very/extremely
preterm infants incurred an average 4.3 times and moderate/late preterm infants incurred an
average of 1.20 times the healthcare costs of a child born full-term, whereas, using the full
26
An exception being proportion of extremely preterm children with retinopathy of prematurity that decreased drastically from wave
2 to 3.
0 6 18 36 60
Full-term 100% 93% 64% 52% 36%
33-36 weeks of GA 100% 94% 66% 51% 34%
<33 weeks of GA 100% 90% 56% 46% 35%
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
Cumulative retention in sample
Full-term
33-36 weeks of GA
<33 weeks of GA
141

sample, extremely preterm infants incurred an average of 3.5 times and moderate/late preterm
infants incurred an average of 1.13 times more healthcare costs as compared to full-term. These
results suggest that the completing sample might be less healthy than children in the full sample
that attrit over time.
Table 4-13 Comparison of the overall log healthcare cost difference between
preterm and full-term born children in the ‘full’ versus ‘completing’
samples
Groups
compared
Full (all-available) sample Completing sample*
Log HC cost
(SE)
Cost ratio
[95% CI]
Log HC cost
(SE)
Cost ratio [95% CI]
<33 weeks of GA
vs. full-term
1.26 3.5 1.46 4.3
(0.06) [3.1 - 3.95] (0.02) [4.1 - 4.5]
33-36 weeks of
GA vs. full-term
0.12 1.13 0.18 1.20
(0.03) [1.06 - 1.21] (0.01) [1.17 - 1.22]
*Completing sample refers to infants who remained in the cohort up to completion of 60 months of age
4.5.10.2 Model based attrition test results
4.5.10.2.1 Results for comparison of coefficients between full and
completing samples
Comparison of the coefficients between full and completing samples showed statistically
significant differences for one or more covariates, although the impacts were minimal for
very/extremely preterm born children (see Table 4-14). The intercept estimates (constants) of all
3 groups derived from the completing sample were higher than the corresponding estimates
142

obtained from the full sample (α
full
– α
completing
= -0.22, -0.34 (p<0.05) and -0.12 (p = 0.3312) log-
points among full-term, moderate/late preterm and extremely preterm groups, respectively).
Statistically significant differences in the size of the slope estimates were mostly observed only in
the full-term group, except for differences in coefficients on CHD and age-period (see Table
4-14). Across all 3 groups, a statistically significant and higher impact of congenital heart disease
was seen in the full sample as compared to the completing sample (β
CHD(full)
– β
CHD(completing)
=
0.36, 0.48 and 0.60 log-points among full-term, moderate/late and very/extreme preterm groups,
respectively (p<0.05)). Conversely, the magnitude of reduction in healthcare cost over time, was
smaller among the completing sample as compared to the full sample, consistently across all 3
groups and these differences were statistically significant. In all 3 groups, the chi-squared test for
the equality of coefficients between the full and completing samples failed to reject the null
hypothesis, thus showing that using only the observations completing 60 months of age, in each
group, may be biased.
Table 4-15 is showing the FE decomposition estimates of the total healthcare cost differential
between preterm and full-term born children using full and completing samples. The
decomposition estimates showed very little variation between the full and completing samples in
the case of very/extremely preterm born group. The decomposition estimates for the cost
differential between moderate/late preterm born children and controls varied substantially
between the two samples. The treatment effect of preterm birth occurring between 33-36 weeks
was seen to be larger when using the completing sample as compared to the full sample. These
results taken together (Table 4-13, 4-14, and 4-15) show that attrition might have significant
influence on the results.

143

Table 4-14 Seemingly unrelated estimation results evaluating difference in coefficients of ln-OLS model
between full and completing samples [Dependent variable = log of overall healthcare costs]
Dependent
variable =
ln (overall
healthcare
costs)
Full-term Preterm: <33 weeks of GA Preterm: 33-36 weeks of GA
Joint test
for equality
of
coefficients
between
full and
completing
groups
Chi-squared statistic 3345.30
p-value <0.001
Chi-squared statistic
96.77
p-value <0.001

Chi-squared
statistic 129.79
p-value <0.001

Characteristi
cs
Full (b) Completing (B) (b-B) p >|chi2|
Full
(b)
Completing
(B)
(b-B)
p
>|chi2|
Full
(b)
Completing
(B)
(b-
B)
p
>|chi2|
Constant
6.085 6.305
-0.22 <0.001
6.48
5
6.607 -0.12 0.3312
5.90
4
6.247 -0.34 <0.05
Male 0.104 0.0932 0.01 0.2723 0.33
8
0.443 -0.11 0.1921 0.13
6
0.135 0.00 0.9837
Year of
birth (2003
vs. 2002)
0.214 -0.230 0.02 0.1138
-
0.12
0
-0.124 0.00 0.9538
0.09
78
-
0.0113
0.11 0.3531
White (vs.
Hispanic)
0.262 -0.353 0.09 <0.001
-
0.20
8
-0.273 0.07 0.5465
-
0.14
2
-0.287 0.15 0.0599
African-
American
0.515 -0.631 0.12 <0.001
-
0.66
1
-0.674 0.01 0.9088
-
0.46
3
-0.584 0.12 0.2306
Other Race 0.430 -0.449 0.02 0.8342
-
0.57
5
-0.604 0.03 0.9705
0.03
47
0.159 -0.12 0.7591
Race
unknown
0.756 -0.559 -0.20 <0.001
-
0.31
7
-0.135 -0.18 0.1962
-
0.57
9
-
0.0372
-0.54 <0.01
SSI
t-1
(vs.
expansions)
2.943 2.791 0.15 <0.001
1.89
2
1.950 -0.06 0.5079
3.05
5
3.374 -0.32 0.1784
144

Dependent
variable =
ln (overall
healthcare
costs)
Full-term Preterm: <33 weeks of GA Preterm: 33-36 weeks of GA
Joint test
for equality
of
coefficients
between
full and
completing
groups
Chi-squared statistic 3345.30
p-value <0.001
Chi-squared statistic
96.77
p-value <0.001

Chi-squared
statistic 129.79
p-value <0.001

Characteristi
cs
Full (b) Completing (B) (b-B) p >|chi2|
Full
(b)
Completing
(B)
(b-B)
p
>|chi2|
Full
(b)
Completing
(B)
(b-
B)
p
>|chi2|
AFDC
t-1
0.307 0.299 0.01 0.4719
1.21
3
1.523 -0.31 <0.01
0.39
8
0.424 -0.03 0.718
Elig. cat.
Missing
t-1

1.606 -1.983 0.38 <0.001
-
2.33
4
-2.469 0.14 0.4388
-
1.59
2
-1.966 0.37 <0.05
CHD
t-1
1.829 1.469 0.36 <0.001
1.80
7
1.206 0.60 <0.001
1.78
0
1.300 0.48 <0.01
Hospitalizati
on
t-1

0.709 0.744 -0.04 <0.05
1.29
5
1.165 0.13 0.1225
0.82
3
0.832 -0.01 0.9174
12-18 (vs. 6-
12 mos)
0.738 -0.353 -0.39 <0.001
-
1.19
1
-0.779 -0.41 <0.001
-
0.87
7
-0.330 -0.55 <0.001
18-24
months
0.841 -0.718 -0.12 <0.001
-
1.06
4
-0.977 -0.09 0.4269
-
0.63
4
-0.541 -0.09 0.2786
24-30
months
1.214 -1.035 -0.18 <0.001
-
1.65
5
-1.348 -0.31 <0.01
-
1.33
3
-1.204 -0.13 0.1505
30-36
months
1.942 -1.576 -0.37 <0.001
-
2.26
8
-1.914 -0.35 <0.01
-
2.21
7
-1.990 -0.23 0.0194
36-42
months
1.847 -1.526 -0.32 <0.001
-
2.08
2
-1.822 -0.26 <0.05
-
2.23
8
-1.862 -0.38 <0.001
42-48
months
2.218 -2.197 -0.02 0.1308
-
2.23
4
-2.274 0.04 0.696
-
2.08
2
-1.872 -0.21 0.0283
48-54 1.996 -2.055 0.06 <0.001 - -1.995 0.04 0.7364 - -2.005 0.11 0.2443
145

Dependent
variable =
ln (overall
healthcare
costs)
Full-term Preterm: <33 weeks of GA Preterm: 33-36 weeks of GA
Joint test
for equality
of
coefficients
between
full and
completing
groups
Chi-squared statistic 3345.30
p-value <0.001
Chi-squared statistic
96.77
p-value <0.001

Chi-squared
statistic 129.79
p-value <0.001

Characteristi
cs
Full (b) Completing (B) (b-B) p >|chi2|
Full
(b)
Completing
(B)
(b-B)
p
>|chi2|
Full
(b)
Completing
(B)
(b-
B)
p
>|chi2|
months 1.96
0
1.89
6
54-60
months
2.307 -2.337 0.03 <0.05
-
2.06
8
-2.125 0.06 0.6026
-
2.21
1
-2.220 0.01 0.9264
SSI = Social security income; AFDC = Aid for families with dependent children; CHD = congenital heart disease; GA = gestational age

146

Table 4-15 Comparison of FE decomposition estimates using the full versus
‘completing’ samples
<33 weeks of GA vs. Full-term
Decomposition of
overall HC costs
FE-OLS Two-part model with FE
Full (all
available) sample
Completing
sample
Full (all available)
sample
Completing
sample
Endowments (%) 31%*** 34%*** 32%*** 37%***
ATET (%) 69%*** 66%*** 68%*** 63%***
33-36 weeks of GA
FE-OLS Two-part model with FE
Decomposition of
overall HC costs
Full (all
available) sample
Completing
sample
Full (all available)
sample
Completing
sample
Endowments (%) 33%*** 14%** 37%** 9%
ATET (%) 67%*** 86%*** 63%
†
91%
***p<0.001; **p<0.01;
†
not significant at α = 5% level; ATET = Average treatment effect on treated
4.5.10.2.2 Results for attrition test based on selection on observables
framework
Table 4-16 is showing cross-sectional probit regressions that predict the probability of attrition
based on demographic variables and a set of lagged outcomes, used as auxiliary instruments.
These models, using a combination of demographic variables with lagged outcomes significantly
explained probability of attrition although, the explanatory power decreased for the last two
waves (pseudo R
2
values ranging between 20-26% for attrition during 6-18 months, versus 7-13%
during the last wave). The first lag of overall healthcare costs was a significant predictor of
147

attrition over time in all three groups, with the exception of attrition during 6-18 months of age in
the <33 weeks of GA group. Observations whose race/ethnicity was unknown had a higher
probability of attrition over time among all 3 groups. Also, the place of residence was found to be
an important predictor of attrition over time as some of the regions had a higher impact on
attrition, probably due to managed care penetration.
Table 4-17 is showing the results of the BGLW test for attrition (a test that is considered as the
reverse of probit test). The coefficients on the dummy variable for attrition were statistically
significant in the full-term and < 33 weeks GA groups, and during wave 2 among children in the
33-36 weeks of GA group showing that attrition may not be a random process. The F-test for joint
significance of coefficients on the attrition dummy and its interactions with demographic
characteristics failed to reject the null hypothesis that attrition in the sample was random.

148

Table 4-16 Cross-sectional probit regressions to predict the probability of
attrition based on demographic characteristics and lagged outcomes
Dependent
variable
P(Attrit=1)
Controls (full-term
infants)
Preterm infants, 33-36
weeks of GA
Preterm infants, <33 weeks
of GA
6-18
mos
19-36
mos
37-60
mos
6-18
mos
19-36
mos
37-60
mos
6-18
mos
19-36
mos
37-60
mos
N
99,13
0
68,477 55,234 1245 1001 825 3,180 2,229 1,739
Pseudo R-sq 0.299 0.205 0.085 0.203 0.179 0.129 0.264 0.162 0.068
Log-likelihood
(initial)
-
62451
.4
-
41828.
4
-
38227.
1
-663.6 -571.3 -546.1 -1972.9 -1385.5 -1203.6
Log-likelihood
(final)
-
43776
.5
-
33251.
1
-
34979.
5
-528.9 -469.2 -475.9 -1452.9 -1160.9 -1121.3
Covariates
Coeff
[SE]
Coeff.[
SE]
Coeff.[
SE]
Coeff.[
SE]
Coeff.[
SE]
Coeff.[
SE]
Coeff.[
SE]
Coeff.[
SE]
Coeff.[
SE]
Auxillary Variables
Ln (overall HC
costs)
t-1

-
0.2**
*
-0.4*** -0.1*** -0.1*** -0.2*** -0.2*** -0.06 -0.3*** -0.1***
[0.007
]
[0.006] [0.005] [0.03] [0.03] [0.03] [0.03] [0.03] [0.03]
Ln (overall HC
costs)
t-2

-
-0.006
-
0.03**
*
-
-0.01 -0.007
-
-0.08* -0.03
[0.008] [0.004] [0.03] [0.03] [0.04] [0.02]
Ln (overall HC
costs)
t-3

- -
-
0.05**
*
- -
-0.1**
- -
-0.05
[0.007] [0.04] [0.04]
Demographic Variables
Male (vs
female)
-0.005
0.05**
*
-0.02 -0.02 0.1 0.1 0.02 0.008 -0.03

[0.010
]
[0.01] [0.01] [0.09] [0.09] [0.10] [0.05] [0.06] [0.06]
Birth cohort
(2003 vs. 2002)
0.02* 0.2*** 0.4*** -0.07 0.02 0.09 -0.03 0.1 0.1
[0.010
]
[0.01] [0.01] [0.09] [0.10] [0.10] [0.10] [0.1] [0.1]
White (vs.
Hispanic)
-
0.6**
*
0.02 -0.04* -0.3* -0.03 0.007 -0.4*** -0.02 0.03
[0.02] [0.02] [0.02] [0.2] [0.2] [0.2] [0.1] [0.09] [0.10]
African-
American
-
0.4**
*
-0.2*** -0.1*** -0.2 -0.3 -0.2 -0.2* -0.03 -0.004
[0.02] [0.02] [0.02] [0.1] [0.2] [0.1] [0.10] [0.1] [0.1]
Other
Race/ethnicity
0.1* 0.3*** 0.2* -0.1 0.8 -0.4 0.06 0.04 -0.3
[0.05] [0.07] [0.08] [0.4] [0.5] [0.5] [0.3] [0.4] [0.4]
Race/ethnicity
unknown
0.4**
*
0.9*** 0.8*** 0.2 0.6*** 0.5*** 0.4*** 0.7*** 0.9***
149

Dependent
variable
P(Attrit=1)
Controls (full-term
infants)
Preterm infants, 33-36
weeks of GA
Preterm infants, <33 weeks
of GA
6-18
mos
19-36
mos
37-60
mos
6-18
mos
19-36
mos
37-60
mos
6-18
mos
19-36
mos
37-60
mos
[0.01] [0.01] [0.02] [0.1] [0.1] [0.1] [0.07] [0.08] [0.1]
Region of residence (vs. South Texas)
Metroplex
1.7**
*
0.7*** 0.2*** 1.5*** 0.5** 0.1 1.7*** 0.9*** 0.06
[0.02] [0.02] [0.02] [0.2] [0.2] [0.2] [0.09] [0.1] [0.1]
Gulf coast
1.8**
*
0.8*** 0.3*** 1.6*** 0.9*** 0.2 1.7*** 0.9*** -0.01
[0.02] [0.02] [0.02] [0.2] [0.2] [0.2] [0.09] [0.1] [0.1]
Upper east
0.04 0.1*** 0.3*** 0.2 -0.07 0.3 0.1 0.2 0.2
[0.03] [0.02] [0.02] [0.3] [0.2] [0.2] [0.1] [0.1] [0.1]
Alamo
1.0**
*
0.3*** -0.1*** 1.2*** 0.4 0.3 0.9*** 0.2 -0.4**
[0.02] [0.03] [0.03] [0.2] [0.2] [0.2] [0.1] [0.1] [0.1]
Central
0.2**
*
0.2*** 0.2*** 0.4 0.002 0.2 0.2 0.1 -0.04
[0.03] [0.02] [0.02] [0.2] [0.2] [0.2] [0.2] [0.1] [0.1]
Capital
1.7**
*
0.7*** 0.07* 1.4*** 0.8*** -0.2 1.6*** 0.8*** -0.3
[0.02] [0.03] [0.04] [0.2] [0.2] [0.3] [0.1] [0.2] [0.2]
West
-
0.2**
*
0.09**
*
0.2*** 0.07 0.03 0.2 -0.3 0.1 0.2
[0.03] [0.02] [0.02] [0.4] [0.3] [0.2] [0.2] [0.1] [0.1]
Upper Rio
Grande
1.5**
*
0.7*** 0.4*** 1.3*** 0.3 0.7 1.6*** 0.4 0.4
[0.03] [0.04] [0.04] [0.3] [0.3] [0.4] [0.1] [0.2] [0.2]
North-west
-0.08* 0.02 0.1*** 0.3 0.3 -0.08 0.2 0.2 0.01
[0.04] [0.03] [0.03] [0.3] [0.2] [0.3] [0.2] [0.2] [0.2]
High Plains
0.3**
*
0.08** 0.2*** 0.3 0.08 0.2 0.4* 0.2 0.07
[0.03] [0.03] [0.03] [0.3] [0.3] [0.3] [0.2] [0.2] [0.2]
Southeast
0.5**
*
0.1*** 0.2*** 0.1 -0.1 0.3 0.4 0.4 0.2
[0.03] [0.03] [0.03] [0.3] [0.3] [0.3] [0.2] [0.2] [0.2]
Miscellaneous
regions
0.7**
*
0.1 -0.3 0 0 0 0 0.4 0.3
[0.2] [0.2] [0.2] [.] [.] [.] [.] [0.7] [0.9]
Constant
0.07 1.6*** 1.1*** -0.4 1.0* 2.6*** -0.9** 1.5*** 1.5***
[0.05] [0.06] [0.06] [0.4] [0.4] [0.5] [0.3] [0.4] [0.4]
Standard errors in brackets; * p<0.05; ** p<0.01; *** p<0.001
150

Table 4-17 Becketti, Gould, Lilard and Welch test for non-random attrition in the
control and preterm groups
Dependent
variable = 1 lag
of total
healthcare
costs
Controls (full-term born)
Preterm born, 33-36
weeks of GA
Preterm born, <33 weeks
of GA
6-18
mos
19-36
mos
37-60
mos
6-18
mos
19-36
mos
37-60
mos
6-18
mos
19-36
mos
37-60
mos
N
9913
0
6847
7
5523
4
3183 2229 1739 1246 1002 826
F-statistic for
joint
significance of
Att and IattX*
69.84
304.6
9
76.44 1.46 9.09 5.66 2.61 5.89 6.86
|p-value| > F
<0.00
1
<0.00
1
<0.00
1
0.093
3
<0.00
1
<0.00
1
<0.00
1
<0.00
1
<0.00
1

Coeff.
[SE]
Coeff.
[SE]
Coeff.
[SE]
Coeff.
[SE]
Coeff.
[SE]
Coeff.
[SE]
Coeff.
[SE]
Coeff.
[SE]
Coeff.
[SE]
Att (1 =
attrition)
-
0.4**
*
-
0.8**
*
-
0.4**
*
-0.1
-
0.7**
-0.3 0.2
-
1.3**
*
-0.6*
[0.02] [0.02]
[0.02
]
[0.2] [0.2] [0.2] [0.5] [0.4] [0.3]
Att*Male (vs.
female)
-
0.03*
-
0.04*
0.01 -0.03 0.08 -0.04 -0.3 -0.1 -0.2
[0.01] [0.02]
[0.02
]
[0.06
]
[0.1] [0.1] [0.2] [0.3] [0.2]
Att*Year of
birth (2003 vs.
2002)
-
0.004
0.01
0.2**
*
0.09 -0.02 0.05 0.2 0.01 -0.07
[0.01] [0.02]
[0.02
]
[0.1] [0.2] [0.2] [0.3] [0.3] [0.3]
Att*White (vs.
Hispanic)
0.01
0.08*
*
-
0.09*
*
0.1 0.02 -0.02 0.6 0.4 0.3
[0.03] [0.03] [0.03] [0.1] [0.2] [0.2] [0.4] [0.4] [0.4]
Att*African-
American
0.02 0.04 -0.01 0.4** 0.3 -0.1 0.2 -0.05 -0.08
[0.03] [0.04] [0.04] [0.1] [0.2] [0.2] [0.4] [0.5] [0.4]
Att*Other race
0.2** 0.2 0.1 -0.3 1.1 0.06 1.5*** -0.6 -0.5
[0.06] [0.1] [0.1] [0.4] [0.8] [0.8] [0.3] [1.0] [1.2]
Att*Uknown
race
0.02
-
0.000
5
-
0.4**
*
0.1 -0.06
-
1.2**
*
-0.3 -0.4 -0.2
[0.01] [0.02] [0.04]
[0.08
]
[0.2] [0.3] [0.3] [0.3] [0.4]
Att*Metroplex
(vs. South
Texas)
0.2**
*
0.1**
*
-0.02 -0.1 -0.1 -0.5* -0.7 0.10
-
1.5**
*
[0.02] [0.03] [0.04] [0.2] [0.2] [0.2] [0.6] [0.4] [0.4]
151

Dependent
variable = 1 lag
of total
healthcare
costs
Controls (full-term born)
Preterm born, 33-36
weeks of GA
Preterm born, <33 weeks
of GA
6-18
mos
19-36
mos
37-60
mos
6-18
mos
19-36
mos
37-60
mos
6-18
mos
19-36
mos
37-60
mos
Att*Gulf coast
0.1**
*
0.2**
*
-
0.003
-0.2 -0.2 0.2 -0.7 0.4 0.2
[0.02] [0.03] [0.04] [0.2] [0.2] [0.2] [0.5] [0.5] [0.5]
Att*Upper East
0.06 0.07* 0.008 -0.3
-
0.008
-0.5 -0.8 0.4 -0.7
[0.05] [0.04] [0.04] [0.2] [0.2] [0.3] [0.8] [0.9] [0.5]
Att*Alamo
-0.03
-
0.1**
-
0.3**
*
-0.02 -0.5* -0.7* -0.6 0.8 -0.6
[0.03] [0.04] [0.05] [0.2] [0.2] [0.3] [0.6] [0.5] [0.5]
Att*Central
0.08* 0.1** -0.07 -0.1 0.10 0.05 -1.3 -0.3 -0.4
[0.04] [0.04] [0.04] [0.3] [0.3] [0.3] [1.1] [0.5] [0.5]
Att*Capital
0.3**
*
0.09
-
0.2**
-0.03 0.04 0.3 -0.6 -0.4 -0.1
[0.03] [0.05] [0.06] [0.3] [0.4] [0.4] [0.7] [0.6] [0.6]
Att*West
0.07
0.1**
*
-0.02 0.2 -0.1 -0.07 0.7 1.9* 0.09
[0.07] [0.04] [0.04] [0.3] [0.3] [0.3] [0.6] [0.8] [0.6]
Att*Upper Rio
Grande
0.3**
*
0.02 0.06 0.09 -0.1
1.3**
*
-0.3 2.0 -0.09
[0.04] [0.06] [0.08] [0.2] [0.3] [0.4] [0.9] [1.1] [1.0]
Att*North west
0.05 0.1* -0.05 0.1 0.1 -0.09 -0.3 0.5 -0.4
[0.07] [0.05] [0.05] [0.3] [0.3] [0.3] [1.2] [0.6] [0.7]
Att*High Plains
-0.06 0.05 -0.10 0.03 -0.2 0.4 0.4 0.3 -0.9
[0.05] [0.05] [0.06] [0.3] [0.3] [0.4] [0.6] [0.9] [0.6]
Att*South East
0.2**
*
0.03 0.009 -0.2 0.2 0.3 -2.9 0.1 -0.4
[0.05] [0.05] [0.06] [0.3] [0.3] [0.4] [3.3] [0.7] [0.7]
Att*Miscellane
ous regions
-0.4* -0.3 -0.8* 0 -0.8* 1.8* 0 0 0
[0.2] [0.3] [0.4] [.] [0.3] [0.7] [.] [.] [.]
Constant
7.7**
*
7.4**
*
7.3**
*
9.0**
*
7.4**
*
7.3**
*
11.2**
*
8.9**
*
8.4**
*
[0.00
7]
[0.00
9]
[0.01]
[0.06
]
[0.1] [0.2] [0.1] [0.1] [0.2]
N
9913
0
68477
5523
4
3183 2229 1739 1246 1002 826
R-sq 0.053 0.188 0.136 0.024 0.159 0.150 0.089 0.171 0.209
Att = attrition dummy (0 = no attrition / 1 = attrition during time each wave); IattX* = interaction of attrition
dummy with X variables; Standard errors in brackets; * p<0.05; **p<0.01; ***p<0.001;

152

4.5.10.2.3 Results for comparison of unweighted and weighted
model coefficients
Table 4-18 is showing a qualitative comparison between the unweighted and weighted coefficient
estimates obtained from fixed effects group regressions of ln-total healthcare costs. Weighted
estimates were obtained by using attrition weights calculated for children retained in wave 2 and
applying these weights through end of wave 4 (constant weights within panel). The magnitude of
bias in coefficients, expressed as percentages, is shown in Columns C, F and I in Table 4-18.
Accordingly, the unweighted estimates were found to overestimate the baseline group level
effects (constant terms) by 26% among controls, 16% among children born between 33-36 weeks
of GA and 15% among children born before 33 weeks of gestational age. The differences in the
unweighted and weighted slope estimates, especially among the preterm groups, were mostly
smaller in magnitude (<10%). The Hausman test for the equality of weighted and unweighted
estimates (or consistency of unweighted estimates under null) led to rejecting the null hypothesis
in all 3 groups (chi2(14) = controls: 12,947.87 (p<0.001), moderate to late preterm: 172.85
(p<0.001) and very/extreme preterm cohort: 51.10 (p<0.001)), thereby showing that the
unweighted estimates are inconsistent.
When weighted estimates were calculated based on non-constant attrition weights derived for
each wave, the results did not vary much from the results shown in Table 4-18 (see Appendix 8).
Hence only constant weights were used among two-part models to derive decomposition
estimates that account for non-random attrition.

153

Table 4-18 Comparison of unweighted and weighted fixed effects OLS estimates
[Note: Attrition weights for children retained in the sample by the end
18 months (wave 2) applied to waves 3 and 4 as well]
Dependent
variable =
ln(overall
healthcare
costs)
(A) (B) (C) (D) (E) (F) (G) (H)
(I)
Controls 33-36 weeks of GA <33 weeks of GA
Unweig
hted
Coeff.
[SE]
Weig
hted
Coeff.
[SE]
%
Differe
nce in
coeffici
ents
(UW -
W)
Unweig
hted
Coeff.
[SE]
Weig
hted
Coeff.
[SE]
%
Differe
nce in
coeffici
ents
(UW -
W)
Unweig
hted
Coeff.
[SE]
Weig
hted
Coeff.
[SE]
%
Differe
nce in
coeffici
ents
(UW -
W)
N 455665 455665 14320 14320 6987 6987
R-sq 0.445 0.437 0.483 0.472 0.584 0.582
adj. R-sq 0.325 0.315 0.367 0.354 0.502 0.500
Constant
6.28***
6.05**
*
26%
6.44***
6.29**
*
16%
7.58***
7.44**
*
15%
[0.012]
[0.012
]
[0.070]
[0.070
]
[0.10] [0.10]
SSI
t-1
(vs.
expansions)
1.30***
1.36**
*
-6%
1.14***
1.06**
*
8%
0.97***
0.94**
*
3%
[0.054]
[0.056
]
[0.25] [0.26] [0.11] [0.11]
AFDC
t-1

0.042**
*
0.099*
**
-6%
0.10 0.12
-2%
0.38***
0.38**
*
0%
[0.011]
[0.012
]
[0.064]
[0.066
]
[0.10] [0.10]
Elig. cat.
Missing
t-1

-0.41***
-
0.23**
*
-17%
-0.34** -0.15
-17%
-1.16***
-
1.12**
*
-4%
[0.019]
[0.018
]
[0.11] [0.11] [0.18] [0.18]
CHD
t

1.21***
1.30**
*
-9%
1.32***
1.38**
*
-6%
0.92***
0.96**
*
-4%
[0.052]
[0.057
]
[0.23] [0.25] [0.20] [0.22]
Hospitaliza
tion
t-1

-
0.097**
*
-
0.10**
*
0%
-0.16 -0.13
-3%
-0.012 0.0093
-2%
[0.019]
[0.020
]
[0.093]
[0.098
]
[0.098] [0.10]
12-18
months
(vs. 6-12
mos)
-0.80***
-
0.74**
*
-6%
-0.97***
-
0.96**
*
-1%
-1.20***
-
1.22**
*
2%
[0.011]
[0.012
]
[0.064]
[0.065
]
[0.096]
[0.097
]
18-24
months
-1.23***
-
1.21**
*
-2%
-1.12***
-
1.16**
*
1%
-1.59***
-
1.59**
*
0%
[0.014]
[0.014
]
[0.079]
[0.081
]
[0.11] [0.11]
154

Dependent
variable =
ln(overall
healthcare
costs)
(A) (B) (C) (D) (E) (F) (G) (H)
(I)
Controls 33-36 weeks of GA <33 weeks of GA
Unweig
hted
Coeff.
[SE]
Weig
hted
Coeff.
[SE]
%
Differe
nce in
coeffici
ents
(UW -
W)
Unweig
hted
Coeff.
[SE]
Weig
hted
Coeff.
[SE]
%
Differe
nce in
coeffici
ents
(UW -
W)
Unweig
hted
Coeff.
[SE]
Weig
hted
Coeff.
[SE]
%
Differe
nce in
coeffici
ents
(UW -
W)
N 455665 455665 14320 14320 6987 6987
R-sq 0.445 0.437 0.483 0.472 0.584 0.582
adj. R-sq 0.325 0.315 0.367 0.354 0.502 0.500
24-30
months
-1.68***
-
1.61**
*
-7%
-1.89***
-
1.87**
*
-2%
-2.32***
-
2.32**
*
0%
[0.014]
[0.015
]
[0.081]
[0.083
]
[0.11] [0.11]
30-36
months
-2.44***
-
2.32**
*
-11%
-2.82***
-
2.72**
*
-10%
-3.01***
-
2.95**
*
-6%
[0.014]
[0.015
]
[0.082]
[0.084
]
[0.11] [0.12]
36-42
months
-2.29***
-
2.24**
*
-5%
-2.69***
-
2.65**
*
-4%
-2.92***
-
2.92**
*
0%
[0.016]
[0.016
]
[0.092]
[0.092
]
[0.12] [0.12]
42-48
months
-2.73***
-
2.63**
*
-10%
-2.65***
-
2.61**
*
-4%
-3.16***
-
3.09**
*
-7%
[0.016]
[0.016
]
[0.094]
[0.094
]
[0.12] [0.13]
48-54
months
-2.51***
-
2.35**
*
-15%
-2.44***
-
2.34**
*
-10%
-2.90***
-
2.83**
*
-7%
[0.016]
[0.016
]
[0.094]
[0.095
]
[0.12] [0.13]
54-60
months
-
2.81***
-
2.66**
*
-14%
-
2.71***
-
2.63**
*
-8%
-
3.02***
-
2.95**
*
-7%
[0.016]
[0.016
]
[0.094]
[0.094
]
[0.13] [0.13]
***p<0.001; **p<0.01; *p<0.05; t = current 6-month period t-1 = previous 6-month period; SSI: Social security income
AFDC: Aid for families with dependent children CHD: Congenital heart disease SE = standard error

155

4.5.11 Sensitivity analyses results
4.5.11.1 Results for treatment effects using attrition weighted estimates
Figure 4-6 and Figure 4-7 are showing a comparison of the weighted and unweighted estimates of
average treatment effects on treated for preterm births occurring at < 33 weeks of GA and 33-36
weeks of GA, respectively. The net effects of attrition weighting among the two groups were
opposite in direction, i.e. weighting decreased the treatment effect of birth occurring at < 33
weeks of GA and increased the treatment effect of birth occurring between 33-36 weeks of GA as
compared to the unweighted treatment effect estimates. After weighting, the incremental
healthcare cost per 6 months, between 6 and 60 months of age, incurred for children born before
33 weeks of gestational age was $557 (95% confidence intervals: $411 - $694, p < 0.001), which
is 20% lesser than the incremental costs obtained for the same sample without weighting. The
unweighted and weighted treatment effect estimates of preterm births occurring between 33-36
weeks of GA were not significant at the 5% level, but the weighted estimates were significant at
the 10% level ((∆ = $40 per child per 6-months, p = 0.079). Also, the weighted incremental cost
estimate was 29% higher than the unweighted estimate.
156

Figure 4-6 Unweighted and weighted TPM incremental cost (ATET) estimates,
<33 weeks of GA vs. full-term

[Error bars represent 95% confidence intervals TPM = Two-part model ATET = average treatment effect
on treated].
Figure 4-7 Unweighted and weighted TPM incremental cost (ATET) estimates,
33-36 weeks of GA vs. full-term

$693
$557
$0
$100
$200
$300
$400
$500
$600
$700
$800
$900
Incremental HC cost per 6-months
(US$)
Unweighted vs. Weighted Treatment Effects, < 33 weeks of GA
Unweighted
Weighted
$31
$40
-$40
-$20
$0
$20
$40
$60
$80
$100
Incremental HC cost per 6-months (US$)
Unweighted vs. Weighted Treatment Effects, 33-36 weeks of GA
Unweighted
Weighted
157

[Error bars represent 95% confidence intervals TPM = Two-part model ATET = average treatment effect
on treated]
4.5.11.2 Results for incremental effects of moderate/late preterm, birth
weight < 2,500 grams, versus full-term
The number of moderate/late preterm born survivors, born with weight < 2,500g, at 6 months of
age was 1,236. By the end of 60 months of age there were 647 children in this group. Table 4-19
is showing the endowments and ATET estimates for the various sub-groups of children born
preterm having a high risk of long-term morbidity, as compared to children born at term,
corrected for non-random attrition. Moderate/late preterm born infants who also weighed <2,500
grams at the time of birth incurred on average $87 more in healthcare expenditures per 6-months
than controls, for the time period between 6 and 60 months of age. Seventy-four percent ($64, p <
0.05) of the cost difference is due to returns to prematurity (95% CI: $14 - $113) and the
remaining 26% ($23, p < 0.01) is due to differences among group characteristics.
4.5.11.3 Results for incremental effects of very and extremely preterm born
children separately, versus full-term
There were 493 and 515 children at the beginning of the cohort that were born very preterm (29 -
32 weeks of GA) and extremely preterm (≤ 28 weeks of GA), respectively. At the end of 60
months, 275 in the very preterm and 348 in the extremely preterm group remained in the cohort.
The additional healthcare cost incurred by the sub-group of children born very preterm, when
compared with children born at term is $593 (Table 4-19). Seventy-two percent of the cost
difference is attributed to difference in coefficients between the two groups (ATET = $427, 95%
CI: $291 - $563; p < 0.001) and the portion of cost differential explained by group characteristics
was $166 (p < 0.001).
Children who were born extremely preterm (≤ 28 weeks of GA) incurred an average of $1,270
more than the average healthcare cost among full-term born children. Fifty-three percent of this
158

difference is attributed to difference in coefficients between the two groups (ATET = $673 per
child per 6-months, 95% CI: $ 419 - $940, p < 0.001) and the remaining 47% of the cost
difference ($597; p < 0.001) is due to differences among group characteristics.
Figure 4-8 shows a comparison of the unweighted and weighted treatment effects, in terms of
incremental costs (US$) for the three sub-groups.
Table 4-19 Decomposition of total healthcare cost difference between preterm
subgroups and full-term children [Estimates are based on two-part
models adjusted for attrition]
Decomposition
estimates
Estimates
(A) 33-36 weeks of
GA, <2500g
(B) 29-32 weeks
of GA
(C) <=28 weeks of
GA
Endowments
effect (X
p
-
X
c
)*β
p

Mean diff. (±SE) in log
HC costs due to group
characteristics
0.03** (0.01) 0.18*** (0.03) 0.51*** (0.07)
% of total difference due
to group characteristics
26%** [9% - 48%]
28%***
[19% - 37%]
47%***
[35% - 59%]
Mean diff. in predicted
healthcare costs per 6-
months (US$)
a
[95% CI]
$23** [$8 - $42]
$166***
[$113 - $219]
$597***
[$445 - $750]
Coefficients effect
(ATET) (β
p
-
β
c
)*X
c

Mean diff. (±SE) in log
HC costs due to
coefficients
0.094* (0.04) 0.48*** (0.08) 0.59*** (0.12)
% of total difference due
to coefficients
74%* [16% - 131%]
72%***
[49% - 95%]
53%***
[33% - 74%]
Mean diff. in predicted
healthcare costs per 6-
months (US$)
a
[95% CI]
$64* [$14 - $113]
$427***
[$291 - $563]
$673***
[$419 - $940]
a
Calculated by applying cost ratio between groups in the estimation sample to the unadjusted reference (full-term) group
costs; GA = gestational age; RE = random effects; FE = fixed effects; TPM = two-part model; CI = confidence intervals;
SE = standard errors; ***p<0.001; **p<0.01; *p<0.05; †p<0.10
159

Figure 4-8 Unweighted and weighted incremental cost (ATET) estimates,
preterm sub-groups vs. full-term
[Error bars represent 95% confidence intervals TPM = Two-part model ATET = average treatment effect
on treated]
4.5.11.4 Results for treatment effects adjusted for chronic disease
characteristics
After adjusting for lagged chronic conditions, returns to prematurity contributed 44% of the 0.89
difference in ln-total healthcare cost between very/extreme preterm and full-term (p < 0.001)
groups, and 57% of the 0.13 log-point difference in total healthcare costs between moderate/late
preterm, with birth weight < 2,500g and full-term (p = 0.052) groups. The endowment and ATET
estimates of total healthcare cost difference between various preterm sub-groups and full-term
group are provided in Table 4-20.
160

Figure 4-9 is showing a comparison between the ATET estimates, obtained with and without
adjusting for lagged chronic disease characteristics, in dollar scale. As shown in this figure, when
adjusted for disease characteristics, the incremental costs, attributed to prematurity, were lesser
by 23% among moderate/late preterm born, with birth weight < 2,500g (ATET = $49 per child
per 6 months (p = 0.052)) and 28%, among very/extreme preterm (ATET = $402 (95% CI: $256
- $539)), as compared to ATET estimates not adjusted for diseases characteristics ($64 and 557,
respectively). After adjusting for disease characteristics, incremental costs of preterm birth
occurring at ≤ 28 weeks of GA, was lesser than the corresponding treatment effect estimate
obtained without adjusting for disease characteristics by 36% (ATET = $429 (95% CI: $ 174 –
684)). This estimate represents only 34% of the 1.10 log-point difference in total healthcare costs
between extremely preterm and full-term groups (p < 0.001, 95% CI: 14% - 54%). The remaining
66% of the ln-total healthcare cost difference between the two groups is due to differences in
disease and other characteristics between the two groups. (p < 0.001; 95% CI: 53% - 80%).
Similarly, the incremental cost of preterm birth occurring between 29 – 32 weeks of GA adjusted
for disease characteristics was 20% lesser than the corresponding estimate without adjustment for
disease characteristics (ATET = $341 (95% CI: $202 – 480)). This estimate represents 58% of the
0.66 ln-healthcare cost differences between the two groups (p < 0.001; 95% CI: 34 – 81%), and
the remaining 42% were explained by differences in frequency of disease and other
characteristics (p < 0.001; 95% CI: 31% - 53%).
161

Table 4-20 Decomposition of total healthcare cost difference between preterm
subgroups and full-term children, after adjusting for lagged chronic
conditions [Estimates are based on two-part models adjusted for
attrition]
Decomposition
estimates
Estimates of
differentials
(33-36 weeks
of GA vs.
full-term)
(A) 33-36
weeks of
GA
(B) < 33
weeks of
GA
(C) 33-36
weeks of
GA, BW
≤ 2,500g
(D) 29-32
weeks of
GA
(E) <=28
weeks of
GA
Endowments
effect (X
p
-
X
c
)*β
p

Mean diff.
(±SE) in log
HC costs due
to group
characteristics
0.04**
(0.01)
0.50***
(0.05)
0.055***
(0.01)
0.28***
(0.04)
0.66***
(0.08)
% of total
difference due
to group
characteristics
49%**
[20% -
75%]
56%***
[45% -
67%]
43%***
[20% -
65%]
42%***
[31% -
53%]
66%***
[53% -
80%]
Mean diff. in
predicted
healthcare
costs per 6-
months (US$)
a
[95% CI]
$29** [$12
- $44]
$511***
[$411 -
$612]
$37***
[$17 - $56]
$249***
[$184 -
$314]
$837***
[$672 -
$1,015]
Coefficients
effect (ATET)
(β
p
- β
c
)*X
c

Mean diff.
(±SE) in log
HC costs due
to coefficients
0.05 (0.03)
0.39***
(0.07)
0.07' (0.04)
0.38***
(0.08)
0.37**
(0.11)
% of total
difference due
to returns to
prematurity
51% [-25%
- 129%]
44%***
[28% -
59%]
57%
†
[-
0.4% -
115%]
58%***
[34% -
81%]
34%**
[14% -
54%]
Mean diff. in
predicted
healthcare
costs per 6-
months (US$)
a
[95% CI]
$30 [- $15 -
$75]
$402***
[$256 -
$539]
$49
†
[-$35
- $100]
$341***
[$202 -
$480]
$429***
[$174 -
$684]
a
Calculated by applying cost ratio between groups in the estimation sample to the unadjusted reference (full-term)
group costs; GA = gestational age; RE = random effects; FE = fixed effects; TPM = two-part model; CI =
confidence intervals; SE = standard errors; ***p<0.001; **p<0.01; *p<0.05; †p<0.10

162

Figure 4-9 Comparison of incremental cost (ATET) estimates with and without
adjusting for chronic disease characteristics (lagged), between
preterm sub-groups vs. full-term. [Estimates are based on two-part
models adjusted for attrition]
[Error bars represent 95% confidence intervals TPM = Two-part model ATET = average treatment effect
on treated]
4.5.12 Results for drivers of healthcare cost difference between preterm
and full-term groups
4.5.12.1 <33 weeks of GA versus full-term
Among observations with positive healthcare costs in the very to extreme preterm group,
endowments, including disease and other characteristics, accounted for 62% of total healthcare
cost difference between very to extreme preterm and full-term groups ($1,019 per child per 6-
months). Figure 4-10 is showing the contribution of chronic conditions to endowments effect of
preterm births occurring at < 33 weeks of GA. A significant portion of the endowments effect
$64
$49
$557
$402
$427
$341
$673
$429
-$400
-$200
$0
$200
$400
$600
$800
$1,000
$1,200
Not adjusted
for chronic
dis.
Adjusted for
chronic dis.
Not adjusted
for chronic
dis.
Adjusted for
chronic dis.
Not adjusted
for chronic
dis.
Adjusted for
chronic dis.
Not adjusted
for chronic
dis.
Adjusted for
chronic dis.
33-36 weeks of GA,
<2,500 g
<33 weeks of GA 29-32 weeks of GA ≤28 weeks of GA
Incremental HC cost per 6-months (US$)
Average treatment effect on treated, adjusted for lagged chronic
conditions
p=0.052 p<0.05
p<0.001
p<0.001
p<0.001
p<0.001
p<0.001
p<0.001
163

was due to disabilities that qualify these children to receive Medicaid insurance (% endowments
effect = 30% [95% CI: 27% - 33%]). After accounting for disability, the following are the percent
contributions, in decreasing order, of each medical condition group to the endowments effect:
neurodevelopmental delay (14%; 95% CI: 12% - 15%), vision disorders (11%; 10% - 13%),
asthma & bronchial disorders (10%; 9 -12%), other respiratory symptoms (8%; 7 - 9%), failure to
thrive (6%; 5 - 7%), motor disorders (5%; 3 - 6%), lower respiratory tract infections (4%; 3 -
5%), GI functional disorders (3%; 3-4%) and endocrine/metabolic disorders (2%; 1-3%).
The coefficients portion of the total healthcare cost difference (mean coefficients effect) after
accounting for disability and medical conditions is 38% in the < 33 weeks of GA group ($679 per
child per 6-months). Difference in constant terms (i.e. group effect) is the major contributor to
coefficient effects (mean diff. = +211%; 95% CI: 191% – 232%). AFDC category accounted for
+13% and hospitalization in the previous period contributed +3% of coefficients effect (95% CI:
10% - 17%). Among medical conditions, statistically significant contributions to coefficients
effect, were observed for respiratory symptoms (mean diff. = 4%; 95% CI: 0 – 8%) and motor
disorders (3%).
4.5.12.2 33-36 weeks of GA, birth weight < 2,500 grams versus full-term:
Among observations with positive healthcare expenditures, disease and other group
characteristics accounted for 34% of total healthcare cost difference ($65 per child per 6-months).
Disabled status accounted for 9% of the endowments effect. Difference in rates of
neurodevelopmental delay contributed 19% (95% CI: 12 – 26%), asthma & bronchial disorders
14% (4 – 25%), other respiratory symptoms 11% (5 – 17%), failure to thrive 9% (5 – 13%), lower
respiratory tract infections 7% (1 – 13%), motor disorders 6% (0 – 11%) and vision disorders 6%
(2 – 10%), towards endowments effect (Figure 4-11). The net unexplained portion of the total
healthcare cost difference (mean coefficients effect) after accounting for disability and medical
164

conditions is 65% ($122 per 6-month period) groups. Difference in constant terms is the
predominant contributor to coefficients effect (mean diff. = 189%; 95% CI: 142% – 237%).
AFDC category accounted for +27% of coefficients effect (95% CI: 10% - 17%).
Figure 4-10 Healthcare cost difference due to differences in frequencies of chronic
conditions between children born < 33 weeks of GA and full-term
(Sample: observations with healthcare costs > 0)

30%
2%
10%
8%
4%
2%
14%
5%
1%
11%
6%
2%
3% 1%
1% 1%
HC cost difference due to endowments (Total pie = $1,019 / 6-month
period) Disability
CHD
Asthma
Other respiratory
LRTI
CNS
NDD
Motor dis.
Seizure dis.
Vision dis.
FTT
Endocrine/Metab.
GI function
GI defects
Hernias
Misc. conditions
165

Figure 4-11 Healthcare cost difference due to differences in frequencies of chronic
conditions between children born 33-36 weeks of GA and full-term
(Sample: observations with healthcare costs > 0)

4.5.13 Results for total excess financial impact due to preterm births
between 6 and 60 months
For the cohort of children born in 2003, after accounting for attrition over time and applying a
discount rate of 3% for costs incurred beyond 1 year, the excess financial costs to Texas Medicaid
due to preterm birth at 33-36 weeks of GA and birth weight < 2,500g, for the cumulative period
between 6 to 60 months of age (54 months), is $456,800 (roughly $560 per child). For the same
birth cohort, the excess healthcare costs due to preterm births in the 29-32 weeks of GA and
extremely preterm (≤ 28 weeks of GA) categories, for the cumulative period between 6 to 60
months of age are, US$ 977,650 ($3700 per child) and 1,133,458 ($5,830 per child), respectively.
9%
4%
14%
11%
7%
19%
6%
6%
9%
4%
4%
HC cost difference due to endowments (Total pie = $65 / 6-month period)
Disability
CHD
Asthma
Other respiratory
LRTI
NDD
Motor dis.
Vision dis.
FTT
GI function
166

4.6 Discussion
We studied the healthcare cost differences between preterm and full-term born children beyond 6
months up to 5 years of age, by gestational age categories, using statistical methods that control
for unobserved heterogeneity. Many restricted panel estimators have been used in applied
research to evaluate the impact of a time-invariant variable on time-varying outcomes, including
random effects, Mundlak (1978), Hausman-Taylor (1981), two stage GLS (R. L. Oaxaca and
Geisler 2003) and fixed effects vector decomposition (FEVD) estimators (Pl
̈ mper and Troeger
2007). These techniques have a common drawback that, the potential correlation of the time-
invariant variable with the individual fixed effects is excluded from the computation of variance,
leading to standard errors that are biased (Chatelain and Ralf 2015). Also, these techniques either
lack a formal identification strategy, or have identification conditions that are weak in the absence
of suitable instruments, making inference about the treatment effect of time-invariant
characteristics difficult.
We evaluated the treatment effect of preterm births, by gestational age category, on healthcare
costs using traditional unit fixed-effects estimation followed by, Blinder-Oaxaca type
counterfactual decomposition of the total healthcare difference between preterm and full-term
groups. A similar treatment of unobserved heterogeneity, with counterfactual decomposition
methods in health economics, can be found in a paper by Andrew M. Jones and Angel Lopez
Nicolas, where they studied inequality in health between genders using the British Household
Panel Survey (Jones and López Nicolás 2006). Other references, where the use of fixed effects
with counterfactual decompositions has been discussed include papers in labor economics by,
Busch and Holst (Busch and Holst 2011) and, Kilbourne B, England P and Beron K (Kilbourne,
England, and Beron 1994). This methodology has the following advantages: i) The method
167

segregates the portion of healthcare cost difference that is due to differences in group
characteristics, from the portion of differential that is due to coefficients (elasticities). The latter
part of the decomposition is equivalent to the treatment effect of preterm birth, under certain
assumptions discussed in sections 4.4.8.3 and 4.7. (Kline 2014; Fortin, Lemieux, and Firpo 2011).
The impact of a time invariant variable, such as gestational age on time varying outcomes can
thus be determined, while simultaneously eliminating unobserved individual specific factors,
related to prematurity, from the estimation. iii) The endowments component of the Blinder-
Oaxaca decomposition provides a useful way of studying the contributions of various disease
conditions to the healthcare cost difference between preterm and full-term born children. This is
an important distinction between the treatment effects and B-O decomposition methods. While
the latter allows for the estimation of ‘composition’ or ‘endowments’ effect, treatment effect
methods only use the group characteristics as covariates to control for in the estimation (Fortin,
Lemieux, and Firpo 2011).
We showed that unobserved heterogeneity plays an important role in the comparison of
healthcare cost outcomes between preterm and full-term born children over the long term,
regardless of gestational age category. This has been demonstrated clearly using the gap between
ATU and ATET estimates of preterm birth, especially of those occurring at 33-36 weeks of GA.
Whereas, the cross-sectional approaches provide treatment effect estimates that vary widely
between ATU and ATET, estimates derived using fixed effects models are comparable to one
another. This shows that, cross-sectional approaches may not be appropriate in studying
healthcare cost outcomes over the long term. The impact of preterm birth, perhaps, could be
under-estimated (as in the case of < 33 weeks of GA), or over-estimated (as in the case of 33-36
weeks of GA), by ignoring subject-specific effects in the estimation. The findings of our study,
therefore, have an important methodological implication, for applying panel data methods to
168

observational research evaluating health outcomes among preterm born children. A similar
observation, emphasizing the role of child fixed effects while dealing with observational research,
has been made in a study of differences in cognitive development among children belonging to
indigenous versus non-indigenous populations in Australia (Brown et al. 2013).
Attrition in the fee-for-service (FFS) cohort was mainly due to children dropping out of Medicaid
entirely. Among those who remained enrolled, a potential source of attrition was the switching of
children from FFS to managed care plan, called Primary Care Case Management (PCCM),
although, the penetration of PCCM in Texas Medicaid was not substantial until 2007 (“Managed
Care” 2013), whereas our study duration is from 2002 to 2008. A detailed analysis on all possible
factors that could be influencing enrollment in the Texas Medicaid cohort is outside the scope of
this dissertation. Nevertheless, an in depth review of the literature was conducted to understand
whether any policy changes during 2002-2008 could affect Medicaid enrollment as well as
healthcare utilization among preterm infants in particular. No such evidence of any policy
amendment or legislation was found, except for the dramatic increase in the penetration of
managed care plans, called Primary Care Case Management (PCCM), in Texas Medicaid
program during 2007. We accounted for non-random attrition in the sample by using established
weighting procedures based on a selection on observarbles framework. We did not test attrition
from a selection on unobservables framework for this research. However, it is also worth
mentioning that linear fixed effects estimators are consistent in the presence of selection on
unobservables, so long as the non-ignorable attrition is due to time invariant unobservables
(Nijman and Verbeek 1992).
Depending on the degree of prematurity, the initial hospitalization cost of premature infants could
be in the range of US$ 60-220,000 per infant - at least 10 times more than the costs incurred for
169

infants born at term (Rogowski 1999; Gilbert, Nesbitt, and Danielsen 2003; Schmitt, Sneed, and
Phibbs 2006; St John et al. 2000; Phibbs and Schmitt 2006; Adams et al. 2003; Chollet, Newman,
and Sumner 1996). The first year costs among moderate/late preterm births were evaluated in two
earlier studies. The definition of moderate/late preterm births used in these studies matched with
the definition used for our study. Both these studies showed higher odds of re-hospitalization
following initial hospital discharge among moderate to late preterm infants, during the first year
of birth (Bird et al. 2010; McLaurin et al. 2009). Using commercial claims data, McLaurin et al.,
showed that the first year costs beyond initial discharge among moderate to late preterms were 3
times greater as compared to term infants ($12,247 vs. $4,069), but the study did not adjust for
any confounding variables. Using Arkansas state Medicaid claims, MacBird et al, had found a
modest increase in adjusted costs among moderate to late preterm infants, through the first year of
life (nearly a $500 increase in first year costs beyond initial discharge). When cross-sectional
methods are used, our estimates are closer to those observed in the MacBird et al. study.
However, we found that the overall treatment effect of preterm birth in the 33-36 weeks of GA
category, was not statistically significant beyond 6 months of age, after accounting for
unobserved heterogeneity in both healthcare demand and intensity of healthcare use estimation
steps. This shows that, the impact of moderate to late preterm births on future healthcare costs, as
observed in the previous studies, could be confounded by unobserved factors.
However, when combined with an additional risk factor, viz., birth weight < 2,500g, moderate to
late preterm birth had a small but statistically significant incremental impact, equal to $65 per
child per 6-months, during the study period. While these incremental costs (per child) may appear
very low, the overall financial impact is quite substantial, since at least 50% of preterm children
(or 5% of all live births) are born moderate to late preterm and low birth-weight (Davidoff et al.
2006; Joyce a Martin et al. 2005). The $456,800 estimate for the additional financial cost beyond
170

6 months of age, to Texas Medicaid due to preterm births in this category, is probably
conservative since a number of preterm children, who potentially can be classified under this
category could not be included in the analysis, due to lack of gestational age data for these
children.
Preterm children in the < 33 weeks of GA category continue to bear significant financial burden
beyond infancy. The excess cumulative financial costs to Texas Medicaid due to preterm births in
the < 33 weeks of GA category, between 6 and 60 months of age, were close to $ 2 million for a
given birth cohort in Texas Medicaid ($4,800 per child). The net cost burden figures could be
underestimated due to the inability to classify many preterm children, potentially born at 29-32
weeks of GA, due to lack of data on gestational age category.
A study by Mangham, et al., modeled the childhood costs of preterm born infants in the < 33
weeks of GA category (Mangham et al. 2009) using a Markov model (Mangham et al. 2009). The
Markov health states were based on level of disability (no, mild, moderate and severe) that
provided an overall assessment of motor, sensory and cognitive abilities. Costs were estimated
from resource utilization data collected prospectively from multiple cohorts in the UK and
Ireland. The average incremental cost per preterm survivor between discharge from initial
hospitalization up to 5 years of age, in 2006 prices, were £1,694 in the very preterm category
(equivalent to $2,033 in US, obtained using price level ratio of purchasing power parity
conversion factor to market exchange rate in 2006) and £4,042 ($4,850) in the extremely preterm
category. The healthcare costs among very to extremely preterm born children in our study were
much higher, despite the fact that, we studied medical costs alone, whereas, the costs reported in
Mangham et al. study included expenses for social service contacts and special education for
disabled children as well. While there may be differences in the rate of medical resources utilized
171

between the two countries (UK versus USA), however, these results unequivocally show that, the
real world healthcare cost burden of preterm births occurring at < 33 weeks of GA could be under
estimated in previous studies that ignored unobserved heterogeneity. Another drawback of
Mangham et al., study is that, the definition of disability has been restricted to few, severe,
neurosensory, motor and cognitive problems. This might have also led to underestimation of the
true healthcare costs among preterm born children without these disabling conditions, but who
may have other significant problems. However, under the social security program through which
disabled infants qualify for Medicaid, disability status is determined by more diverse set of
medical conditions including CNS, respiratory, cardiovascular and developmental disorders,
some of which, may occur more frequently among preterm than full-term born children (Perrin,
Bloom, and Gortmaker 2007). We showed that, neurodevelopmental delay (including
coordination disorders, motor delay, apraxia, speech/language delay and mental retardation) is a
major contributor of differences in mean healthcare costs between very to extremely preterm and
full-term children over the long term, followed by group of conditions, such as, vision problems
(retinopathy, visual disturbances, nystagmus, myopia, pediatric cataract and other vision
disorders), asthma, other bronchial conditions and respiratory symptoms (mainly wheeze). The
observation that there were no significant differences between preterm and full-term groups in the
contributions of these chronic diagnoses to coefficient effects might lead us to think that, the
severity of these conditions probably are not different between preterm and full-term children.
However we may not be able to conclude so about the severity using only healthcare costs as an
outcome. It may be possible that the heterogeneity in terms of severity does not reflect among
healthcare costs but might have an impact on quality of life outcomes.
Finally, our analyses showed that, preterm birth, per se, continued to have a significant partial
effect on healthcare utilization, even after accounting for differences in the rates of disability and
172

a comprehensive set of chronic diseases. However, there are no notable differences between
preterm and full-term groups in the returns to each characteristic or chronic condition (except
AFDC status). It may be possible that some of the developmental problems caused by preterm
birth, may not be classifiable under existing ICD-9 codes, and are instead, simply coded as
disorders associated with preterm birth. Or, it is also possible that many preterm born children
have more than one chronic disease resulting in poorer health status, which, is not adequately
captured in the model specification. In any case, these results show that preterm survivors
continue to have poorer health status that may lead to increased demand for healthcare services,
as compared to full-term born children, during the early childhood development period. Formal
health status assessments among pre-school children have been mostly restricted to studying
neurodevelopmental outcomes. A few studies have shown differences in health status between
pre-school children born very to extremely preterm and full-term, where pre-school children born
preterm or with very low birth weight have been shown to perform more poorly than their peers
in physical, emotional, and/or social functioning (Zwicker and Harris 2008). Data collection
efforts to integrate data on quality of life with healthcare utilization can therefore be helpful to
understand the impact of underlying health status in the difference in healthcare utilization and
costs between preterm and full-term born children.
4.7 Limitations
This is the first study to have addressed the issue of unobserved heterogeneity in evaluating
economic outcomes among preterm children using observational claims data. However, it has to
be recognized that the methodology that is used to evaluate the impact of preterm births on
healthcare costs may have some limitations. Controlling for fixed effects within preterm and full-
term groups respectively, controls for time-invariant factors that may confound the estimation of
173

treatment effect of preterm births on healthcare costs. Nevertheless, to the extent that there could
be factors that are correlated with preterm births, that could affect the treatment effect, the B-O
decomposition methodology could be biased. For example, if preterm children are more likely to
be born in families with a history of premature births, the child specific fixed effects technique
cannot fully eliminate the bias due to this particular factor and the treatment effect estimates
obtained from the B-O decomposition procedure could be biased. Additional sources of bias can
possibly include factors such as maternal health seeking behaviors, education and awareness
about prenatal care, life-style characteristics, comorbidities and family circumstances. Since
prematurity has a multi-factorial etiology it might be impossible to control for all predisposing
factors that can potentially bias the treatment effects. In such a case, considering the unit effects
as representing the net impact of at least some of these factors could serve as a fair approximation
of the extent of unobserved omitted variable bias (Allison 2005). Approaches such as the use of
sibling fixed effects, if data on family structure were available, could provide the ability to
compare preterm and full-term children born within the same family, thereby offering the
potential for more robust control of additional unobserved factors related to individual family
circumstances and maternal health seeking behaviors. Further, given the observational nature of
claims data, even time-invariant factors that may not be causally associated with preterm births
might induce bias in the treatment effect estimates, if the conditional distribution of such time
invariant unobserved factors happened to vary between the two groups. However, such factors
may be rare given the sample sizes dealt with in this study. Nevertheless, the reader of this paper
is cautioned against making causal interpretation of the treatment effect estimates obtained using
the application of Blinder-Oaxaca decomposition technique to the panel data models, as used in
this study. Also the dummy variable specification of chronic conditions averages out the impact
of these conditions across various severity levels. Though decomposition methods based on
174

quantile regressions would be useful in understanding the differential impact of preterm births
across severity levels, such methods are currently difficult to implement with fixed effects.
The study suffers from a few limitations that are specific to the dataset, either due to lack of
sufficient data or those that are typical of observational research conducted using insurance
claims data. Gestational age and birth weight information used in this study were based on ICD-
9-CM codes associated with hospital claims at the time of birth. The lack of additional resources,
such as being able to verify data on claims with gestational age data on the birth certificates may
be perceived as a limitation. But studies have shown that the use of multiple criteria, such as, a
combination of birth-weight and gestational age information increases the specificity of
classification obtained solely based on claims records (Eworuke et al. 2012). Alternative ways of
classification of preterm births, such as, according to whether preterm children were born small or
large for gestational age could have provided further useful findings of clinical relevance, as
recent studies have shown that such classifications could be better indicators of long term
morbidity (Saigal and Doyle 2008). Secondly, data on health status severity (for e.g., as defined
by quality of life scores) and/or characteristics that reflect developmental milestones over time,
such as head circumference, body weight, height and functional level, could potentially serve to
explain some of the unexplained differences in coefficients effect of preterm births if such
information were available. Texas Medicaid is one of the largest state Medicaid programs and it
provides coverage for > 50% of births in Texas. However, the sample demographics may not be
representative of the overall Medicaid population due to a large proportion of children of
Hispanic origin. The findings of our study thus may not be generalizable to the overall low-
income population in the US. Despite these limitations, it has to be recognized that the
availability of longitudinal data on paid medical and pharmacy claims provides for a cost-
175

effective way to evaluate the real world costs among preterm infants, since prospective studies
are too expensive to conduct and have a potential to underestimate the true financial impact.
4.8 Conclusions
The Centers for Disease Control and Prevention (CDC) has estimated that the incidence of
preterm births has increased over 20% between 1990 and 2007 and there is a drastic decline in the
mortality rates. Hence understanding the financial impact of preterm births have become more
important than before to aid in making important clinical and policy decisions. Our study
provides estimates of real world medical costs among preterm children beyond 1 year of age,
reporting cost estimates by gestational age categories. We showed that unobserved individual
specific factors play a significant role in determining long-term outcomes among preterm infants,
and not accounting for fixed effects leads to biased inference regarding the treatment effect of
preterm births. We offered a potential solution to estimate the treatment effect of a time-invariant
variable, such as preterm birth, without resorting to restricted panel estimators and demonstrated
the use of Blinder-Oaxaca type decompositions to accomplish this.
After accounting for child specific fixed effects, there is no difference in the long-term healthcare
costs among children born moderate/late preterm, but with normal birth weights, when compared
to the healthcare costs among children born at term. The sub-group of moderate/late preterm
infants with birth weight < 2,500 g have a modest increase in healthcare costs between 6 and 60
months of age. Very/extremely preterm infants continue to bear substantial financial impact to
Texas Medicaid through 5 years of age. We showed that neurodevelopmental delay,
asthma/bronchial disorders, other respiratory symptoms (including wheezing) and vision related
disorders, are the top drivers of incremental expenditures among preterm children, for the period
176

between 6 and 60 months of age. These findings have important policy implications to fund
perinatal interventions that are aimed at reducing long-term morbidity among preterm survivors.
4.9 References
1. Adams, E Kathleen, Brenda Nishimura, Robert K Merritt, and Cathy Melvin. 2003. “Costs of
Poor Birth Outcomes among Privately Insured.” Journal of Health Care Finance 29 (3): 11–27.
2. Adams-Chapman, Ira, and Barbara J Stoll. 2006. “Neonatal Infection and Long-Term
Neurodevelopmental Outcome in the Preterm Infant.” Current Opinion in Infectious Diseases 19
(3): 290–97. http://www.ncbi.nlm.nih.gov/pubmed/16645492.
3. Allison, Paul D. 2005. “Introduction to Fixed Effects Methods.” In Fixed Effects Regression
Models for Longitudinal Data Using SAS, 1–8.
4. Arellano, Manuel. 1993. “On the Testing of Correlated Effects with Panel Data.” Journal of
Econometrics. doi:10.1016/0304-4076(93)90040-C.
5. Baltagi, Badi H, and Sophon Khanti-Akom. 1990. “On Efficient Estimation with Panel Data: An
Empirical Comparison of Instrumental Variables Estimators.” Journal of Applied Econometrics 5
(4). Wiley Subscription Services, Inc., A Wiley Company: 401–6.
http://resolver.scholarsportal.info/resolve/08837252/v05i0004/401_oeewpdecoive.
6. Baltagi, Badi H., Seuck Heun Song, Byoung Cheol Jung, and Won Koh. 2007. “Testing for Serial
Correlation, Spatial Autocorrelation and Random Effects Using Panel Data.” Journal of
Econometrics 140 (1): 5–51.
7. Baltagi, Badi H., and Ping X. Wu. 1999. “Unequally Spaced Panel Data Regressions with ar(1)
Disturbances.” Econometric Theory. doi:10.1017/S0266466699156020.
8. Barrera-Osorio, Felipe, and Vicente Garcia-Moreno. 2011. Using the Oaxaca-Blinder
Decomposition Technique to Analyze Learning Outcomes Changes over Time. Policy Research
177

Working Paper.
http://search.proquest.com/docview/864938759?accountid=14533\nhttp://elibrary.worldbank.org/
deliver/5584.pdf?itemId=/content/workingpaper/10.1596/1813-9450-5584&mimeType=pdf.
9. Bauer, Thomas K., and Mathias Sinning. 2008. “An Extension of the Blinder–Oaxaca
Decomposition to Nonlinear Models.” AStA Advances in Statistical Analysis.
doi:10.1007/s10182-008-0056-3.
10. Becketti, Sean, William Gould, Lee Lillard, and Finis Welch. 1988. “The Panel Study of Income
Dynamics after Fourteen Years: An Evaluation.” Journal of Labor Economics.
11. Behrman, Richard E, and Adrienne Stith Butler. 2007. Preterm Birth: Causes, Consequences, and
Prevention. Edited by R E Behrman and A S Butler. Prevention. Vol. 10. National Academies
Press (US). http://www.ncbi.nlm.nih.gov/pubmed/20669423.
12. Bhargava, A., L. Franzini, and W. Narendranathan. 1982. “Serial Correlation and the Fixed
Effects Model.” The Review of Economic Studies 49 (4): 533–49. doi:10.2307/2297285.
13. Bird, T Mac, Janet M Bronstein, Richard W Hall, Curtis L Lowery, Richard Nugent, and Glen P
Mays. 2010. “Late Preterm Infants: Birth Outcomes and Health Care Utilization in the First
Year.” Pediatrics 126 (2): e311–19.
14. Blinder, Alan S. 1973. “Wage Discrimination: Reduced Form and Structural Estimates.” The
Journal of Human Resources 8 (4). University of Wisconsin Press: 436–55. doi:10.2307/144855.
15. Boyle, J D, and E M Boyle. 2013. “Born Just a Few Weeks Early: Does It Matter?” Archives of
Disease in childhood.Fetal and Neonatal Edition 98 (1). Hull York Medical School, University of
Hull, Hull, UK.: F85–88. doi:10.1136/archdischild-2011-300535 [doi].
16. Brown, Richard, Bethany Davies, Prabha Prayaga, and Simon Quinn. 2013. Fixed Effects
Matter : The Importance of Panel Data for Understanding Cognitive Development in Australian
178

Indigenous Children ^. http://www.younglives.org.uk/files/others/inequalities-conference-
papers/quinn-importancepaneldataforunderstandingcogdevelopmentaustralianindigenouschildren.
17. Busch, Anne, and Elke Holst. 2011. “Gender-Specific Occupational Segregation, Glass Ceiling
Effects, and Earnings in Managerial Positions: Results of a Fixed Effects Model.” SSRN
Electronic Journal, January. doi:10.2139/ssrn.1750260.
18. Cameron, A C, and P K Trivedi. 2009. Microeconometrics Using Stata. Edited by StataCorp Lp.
Stata Press Books. Vol. 5. Stata Press. http://stata.biz/news/statanews.23.4.pdf.
19. Carey, William B. 2001. “From Neurons to Neighborhoods: The Science of Early Childhood
Development.” Journal of Developmental & Behavioral Pediatrics. doi:10.1097/00004703-
200104000-00011.
20. CDC. 2014. “Recommended Immunizations for Children from Birth Through 6 Years Old -
Parent-Ver-Sch-0-6yrs.pdf.” http://www.cdc.gov/vaccines/parents/downloads/parent-ver-sch-0-
6yrs.pdf.
21. Chatelain, Jean-Bernard, and Kirsten Ralf. 2015. “Inference on Time-Invariant Variables Using
Panel Data: A Pre-Test Estimator with an Application to the Returns to Schooling.” Accessed
January 29. https://halshs.archives-ouvertes.fr/hal-00492039/.
22. Chollet, D J, J F Newman, and A T Sumner. 1996. “The Cost of Poor Birth Outcomes in
Employer-Sponsored Health Plans.” Medical Care 34 (12): 1219–34.
http://www.ncbi.nlm.nih.gov/pubmed/8962588.
23. Clark, R H, D R Gerstmann, A H Jobe, S T Moffitt, A S Slutsky, and B A Yoder. 2001. “Lung
Injury in Neonates: Causes, Strategies for Prevention, and Long-Term Consequences.” The
Journal of Pediatrics 139 (4): 478–86. doi:10.1067/mpd.2001.118201.
179

24. Clements, Karen M, Wanda D Barfield, M Femi Ayadi, and Nancy Wilber. 2007. “Preterm Birth-
Associated Cost of Early Intervention Services: An Analysis by Gestational Age.” Pediatrics 119
(4): e866–74. http://www.ncbi.nlm.nih.gov/pubmed/17339387.
25. Clogg, Clifford C., Eva Petkova, and Adamantios Haritou. 1995. “Statistical Methods for
Comparing Regression Coefficients Between Models.” American Journal of Sociology.
doi:10.1086/230638.
26. Cotton, Jeremiah. 1988. “On the Decomposition of Wage Differentials.” The Review of
Economics and Statistics 70 (2): 236. doi:10.2307/1928307.
27. Davidoff, Michael J., Todd Dias, Karla Damus, Rebecca Russell, Vani R. Bettegowda, Siobhan
Dolan, Richard H. Schwarz, Nancy S. Green, and Joann Petrini. 2006. “Changes in the
Gestational Age Distribution among U.S. Singleton Births: Impact on Rates of Late Preterm
Birth, 1992 to 2002.” In Seminars in Perinatology, 30:8–15.
28. Donnell, Owen O, Eddy Van Doorslaer, Adam Wagstaff, and Magnus Lindelow. 2007. Analyzing
Health Equity Using Household Survey Data Analyzing : A Guide toTechniques and Their
Implementation. Measurement.
29. Downey, G, and J C Coyne. 1990. “Children of Depressed Parents: An Integrative Review.”
Psychological Bulletin 108 (1): 50–76.
30. Drukker, David M. 2003. “Testing for Serial Correlation in Linear Panel-Data Models.” Stata
Journal 3 (2): 168–77.
31. Edoka, Ijeoma P. 2012. Decomposing Differences in Cotinine Distribution between Children and
Adolescents from Different Socioeconomic Backgrounds. WP 12/29. WP 12/29, HEDG,
University of York. York. york.ac.uk/res/herc/hedgwp.
32. Engle, William A, Kay M Tomashek, and Carol Wallman. 2007. “‘Late-Preterm’ Infants: A
Population at Risk.” Pediatrics 120 (6): 1390–1401. doi:10.1542/peds.2007-2952.
180

33. Eworuke, Efe, Christian Hampp, Arwa Saidi, and Almut G. Winterstein. 2012. “An Algorithm to
Identify Preterm Infants in Administrative Claims Data.” Pharmacoepidemiology and Drug
Safety 21 (6): 640–50.
34. Fairlie, Robert W. 2005. “An Extension of the Blinder-Oaxaca Decomposition Technique to
Logit and Probit Models.” Journal of Economic & Social Measurement 30 (4). IOS Press: 305–
16.
http://search.ebscohost.com/login.aspx?direct=true&AuthType=ip,shib&db=buh&AN=20237092
&site=ehost-live.
35. Fitzgerald, John, Peter Gottschalk, and Robert Moffitt. 1998. “An Analysis of Sample Attrition in
Panel Data: The Michigan Panel Study of Income Dynamics.” The Journal of Human Resources,
Journal of Human Resources, 33 (2). National Bureau of Economic Research Cambridge, Mass.,
USA: 251–99. doi:10.2307/146433.
36. Fitzgerald, John M. 2011. “Attrition in Models of Intergenerational Links Using the PSID with
Extensions to Health and to Sibling Models.” The B.E. Journal of Economic Analysis & Policy.
doi:10.2202/1935-1682.2868.
37. Fledelius, H C, and G Greisen. 1993. “Very Pre-Term Birth and Visual Impairment. A
Retrospective Investigation of 411 Infants of Gestational Age 30 Weeks or Less, 1983-89
Rigshospitalet, Copenhagen.” Acta Ophthalmologica Supplement, no. 210: 63–65.
38. Fortin, Nicole, Thomas Lemieux, and Sergio Firpo. 2011. Decomposition Methods in Economics.
Handbook of Labor Economics. Vol. 4.
39. García-Altés, Anna, Jaime Pinilla, and Vicente Ortún. 2011. “The Evolution of Health Status and
Chronic Conditions in Catalonia, 1994-2006: The Paradox of Health Revisited Using the Blinder
- Oaxaca Decomposition.” BMC Health Services Research 11: 116.
181

40. Gilbert, William M, Thomas S Nesbitt, and Beate Danielsen. 2003. “The Cost of Prematurity:
Quantification by Gestational Age and Birth Weight.” Obstetrics and Gynecology 102 (3).
Department of Obstetrics and Gynecology, University of California at Davis, California, USA.
wmgilbert@ucdavis.edu: 488–92.
http://linkinghub.elsevier.com/retrieve/pii/S0029784403006173.
41. Gould, William (StataCorp), and Jeffrey (StataCorp) Pitblado. 2010. “Stata | FAQ: Guidelines for
Bootstrap Samples.” http://www.stata.com/support/faqs/statistics/bootstrapped-samples-
guidelines/.
42. Greene, William H. 2012. Econometric Analysis, Chapter 17, Seventh Edition. Upper Saddle
River, New Jersey: Prentice Hall.
43. Hausman, Jerry A, and William E Taylor. 1981. “Panel Data and Unobservable Individual
Effects.” Econometrica 49 (6). The Econometric Society: 1377–98. doi:10.2307/1911406.
44. Ho, S., and S. Saigal. 2005. “Current Survival and Early Outcomes of Infants of Borderline
Viability.” NeoReviews 6 (3): e123–32. doi:10.1542/neo.6-3-e123.
45. Hosmer, DW, SA Lemeshow, and RX Sturdivant. 2013. Applied Logistic Regression. 3rd Ed.
Hoboken, New Jersey: Wiley.
46. Hsiao, Cheng. 2003a. Analysis of Panel Data. Cambridge University Press.
http://books.google.com/books?hl=en&lr=&id=i9iPG7C3EP4C&pgis=1.
47. ———. 2003b. “Analysis of Panel Data.” Edited by Cambridge University Press. International
Journal of Forecasting, Econometric Society monographs, 20 (1). Cambridge University Press:
142–43. doi:10.1016/j.ijforecast.2003.11.004.
48. Iams, Jay D, Roberto Romero, Jennifer F Culhane, and Robert L Goldenberg. 2008. “Primary,
Secondary, and Tertiary Interventions to Reduce the Morbidity and Mortality of Preterm Birth.”
Lance 371 (9607). Elsevier Ltd: 164–75. doi:10.1016/S0140-6736(08)60108-7.
182

49. Ip, Stanley, Mei Chung, Gowri Raman, Priscilla Chew, Nombulelo Magula, Deirdre DeVine,
Thomas Trikalinos, and Joseph Lau. 2007. “Breastfeeding and Maternal and Infant Health
Outcomes in Developed Countries.” Evidence Report/technology Assessment, no. 153: 1–186.
50. Jann, Ben. 2008a. “A Stata Implementation of the Blinder-Oaxaca Decomposition A Stata
Implementation of the Blinder-Oaxaca Decomposition.” Stata Journal, ETH Zurich Sociology
Working Paper, 8 (5): 453–79. http://www.ssc.wisc.edu/~jmuniz/jann_oaxaca.pdf.
51. ———. 2008b. “The Blinder-Oaxaca Decomposition for Linear Regression Models.” Stata
Journal 8: 453–79.
52. Jones, Andrew M., and Ángel López Nicolás. 2006. “Allowing for Heterogeneity in the
Decomposition of Measures of Inequality in Health*.” The Journal of Economic Inequality 4 (3):
347–65. doi:10.1007/s10888-005-9019-z.
53. Kilbourne, B., P. England, and K. Beron. 1994. “Effects of Individual, Occupational, and
Industrial Characteristics on Earnings: Intersections of Race and Gender.” Social Forces 72 (4):
1149–76. doi:10.1093/sf/72.4.1149.
54. Kline, Patrick. 2014. “A Note on Variance Estimation for the Oaxaca Estimator of Average
Treatment Effects.” Economics Letters 122 (3): 428–31.
55. Kneyber, Martin CJ, Frans B Plötz, and Jan LL Kimpen. 2005. “Bench-to-Bedside Review:
Paediatric Viral Lower Respiratory Tract Disease Necessitating Mechanical Ventilation – Should
We Use Exogenous Surfactant?” Critical Care 9 (6). BioMed Central: 550–55.
http://ccforum.com/content/pdf/cc3823.pdf.
56. Kramer, MS. 2003. “The Epidemiology of Adverse Pregnancy Outcomes: An Overview.” The
Journal of Nutrition. http://jn.nutrition.org/content/133/5/1592S.short.
57. Krilov, Leonard R, Leonard B Weiner, Ram Yogev, Jaime Fergie, Ben Z Katz, Kelly J
Henrickson, and Robert C Welliver. 2009. “The 2009 COID Recommendations for RSV
183

Prophylaxis: Issues of Efficacy, Cost, and Evidence-Based Medicine.” Pediatrics 124 (6).
American Academy of Pediatrics: 1682–84. doi:10.1542/peds.2009-2681.
58. MacKinnon, James G. 2006. “Bootstrap Methods in Econometrics.” In Economic Record. Vol.
82.
59. “Managed Care.” 2013. In Texas Health and Human Services Commission, Pink Book. Austin,
TX. http://www.hhsc.state.tx.us/medicaid/reports/PB9/8_PB_9th_ed_Chapter7.pdf.
60. Mangham, Lindsay J, Stavros Petrou, Lex W Doyle, Elizabeth S Draper, and Neil Marlow. 2009.
“The Cost of Preterm Birth throughout Childhood in England and Wales.” Pediatrics 123 (2).
AMER ACAD PEDIATRICS: e312–27. http://discovery.ucl.ac.uk/792082/.
61. Manning, Willard G, and John Mullahy. 2001. “Estimating Log Models: To Transform or Not to
Transform?” Journal of Health Economics 20 (4): 461–94. doi:10.1016/S0167-6296(01)00086-8.
62. March of Dimes. 2012. “Medicaid Coverage of Births Texas, Peristats; March Of Dimes.”
http://www.marchofdimes.com/peristats/ViewSubtopic.aspx?reg=48&top=11&stop=154&lev=1
&slev=4&obj=1&dv=cr.
63. Martin, JA, BE Hamilton, and PD Sutton. 2010. “Births: Final Data for 2007.” National Vital
Statistics ….
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.172.2665&rep=rep1&type=pdf.
64. Martin, Joyce a, Brady E Hamilton, Paul D Sutton, Stephanie J Ventura, Fay Menacker, and
Martha L Munson. 2005. “Births: Final Data for 2003.” National Vital Statistics Reports : From
the Centers for Disease Control and Prevention, National Center for Health Statistics, National
Vital Statistics System 54 (2): 1–116. http://www.ncbi.nlm.nih.gov/pubmed/16176060.
65. McGauhey, P J, B Starfield, C Alexander, and M E Ensminger. 1991. “Social Environment and
Vulnerability of Low Birth Weight Children: A Social-Epidemiological Perspective.” Pediatrics
88 (5): 943–53.
184

66. McLaurin, Kimmie K, Caroline B Hall, E Anne Jackson, Oksana V Owens, and Parthiv J
Mahadevia. 2009. “Persistence of Morbidity and Cost Differences between Late-Preterm and
Term Infants during the First Year of Life.” Pediatrics 123 (2): 653–59. doi:10.1542/peds.2008-
1439.
67. Mihaylova, Borislava, Andrew Briggs, Anthony O’Hagan, and Simon G Thompson. 2011.
“Review of Statistical Methods for Analysing Healthcare Resources and Costs.” Health
Economics 20 (8): 897–916.
68. Morse, Steven Benjamin, Hao Zheng, Yiwei Tang, and Jeffrey Roth. 2009. “Early School-Age
Outcomes of Late Preterm Infants.” Pediatrics 123 (4): e622–29. doi:10.1542/peds.2008-1405.
69. Moss, Timothy J M. 2006. “Respiratory Consequences of Preterm Birth.” Clinical and
Experimental Pharmacology Physiology 33 (3): 280–84.
70. Mullahy, John. 2009. “Econometric Modeling of Health Care Costs and Expenditures: A Survey
of Analytical Issues and Related Policy Considerations.” Medical Care 47 (7 Suppl 1). From the
Department of Population Health Sciences, University of Wisconsin-Madison, Madison,
Wisconsin.: S104–8. http://www.ncbi.nlm.nih.gov/pubmed/19536020.
71. Mundlak, Yair. 1978. “On the Pooling of Time Series and Cross Section Data.” Econometrica 46
(1). The Econometric Society: 69–85. http://www.jstor.org/stable/1913646.
72. Newacheck, P W. 1994. “Poverty and Childhood Chronic Illness.” Archives of Pediatrics &
Adolescent Medicine 148 (11): 1143–49. doi:10.1001/archpedi.1994.02170110029005.
73. Newacheck, P W, and W R Taylor. 1992. “Childhood Chronic Illness: Prevalence, Severity, and
Impact.” American Journal of Public Health 82 (3): 364–71.
74. Nijman, Theo, and Marno Verbeek. 1992. “Testing for Selectivity Bias in Panel Data Models.”
International Economic Review 33 (3): 681–703. doi:10.2307/2527133.
185

75. Oaxaca, Ronald. 1973. “Male-Female Wage Differentials in Urban Labor Markets.” International
Economic Review, International Economic Review, 14 (3). JSTOR: 693–709.
doi:10.2307/2525981.
76. Oaxaca, Ronald L., and Iris Geisler. 2003. “Fixed Effects Models with Time Invariant Variables:
A Theoretical Note.” Economics Letters 80 (3): 373–77. doi:10.1016/S0165-1765(03)00121-6.
77. Outes-leon, Ingo, and Stefan Dercon. Survey Attrition and Attrition Bias in Young Lives. 5.
78. Perrin, James M, Sheila R Bloom, and Steven L Gortmaker. 2007. “The Increase of Childhood
Chronic Conditions in the United States.” JAMA 297 (24). American Medical Association: 2755–
59. doi:10.1001/jama.297.24.2755.
79. Petrou, S, T Sach, and L Davidson. 2001. “The Long-Term Costs of Preterm Birth and Low Birth
Weight: Results of a Systematic Review.” Child Care Health and Development 27 (2): 97–115.
80. Petrou, Stavros. 2005. “The Economic Consequences of Preterm Birth during the First 10 Years
of Life.” BJOG an International Journal of Obstetrics and Gynaecology 112 Suppl (March): 10–
15. doi:10.1111/j.1471-0528.2005.00577.x.
81. Petrou, Stavros, Oya Eddama, and Lindsay Mangham. 2011. “A Structured Review of the Recent
Literature on the Economic Consequences of Preterm Birth.” Archives of Disease in Childhood.
Fetal and Neonatal Edition 96 (3): F225–32. doi:10.1136/adc.2009.161117.
82. Petrou, Stavros, and Kamran Khan. 2012. “Economic Costs Associated with Moderate and Late
Preterm Birth: Primary and Secondary Evidence.” Seminars in Fetal & Neonatal Medicine 17 (3).
Elsevier Ltd: 170–78. doi:10.1016/j.siny.2012.02.001.
83. Phibbs, Ciaran S, and Susan K Schmitt. 2006. “Estimates of the Cost and Length of Stay Changes
That Can Be Attributed to One-Week Increases in Gestational Age for Premature Infants.” Early
Human Development 82 (2): 85–95.
186

http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Citation&list
_uids=16459031.
84. Pl
̈ mper, Thomas, and Vera E. Troeger. 2007. “Efficient Estimation of Time-Invariant and Rarely
Changing Variables in Finite Sample Panel Analyses with Unit Fixed Effects.” Political Analysis
15 (2): 124–39.
85. Poi, Brian P. 2004. “From the Help Desk : Some Bootstrapping Techniques.” The Stata Journal 4
(3): 312–28.
86. Reimers, Cordelia W. 1983. “Labor Market Discrimination Against Hispanic and Black Men.”
The Review of Economics and Statistics 65 (4): 570–79. doi:10.2307/1935925.
87. Rogowski, J. 1999. “Measuring the Cost of Neonatal and Perinatal Care.” Pediatrics 103 (1 Suppl
E): 329–35.
88. Rosenzweig, Mark R, and Kenneth I Wolpin. 1988. “Heterogeneity, Intrafamily Distribution, and
Child Health.” Journal of Human Resources 23: 437–61.
89. Saigal, Saroj, and Lex W Doyle. 2008. “An Overview of Mortality and Sequelae of Preterm Birth
from Infancy to Adulthood.” Lance 371 (9608). England: 261–69.
http://www.ncbi.nlm.nih.gov/pubmed/18207020.
90. Schaffer, M. E, and S. Stillman. 2006. “XTOVERID: Stata Module to Calculate Tests of
Overidentifying Restrictions after Xtreg, Xtivreg, xtivreg2, Xthtaylor.” Statistical Software
Components. http://ideas.repec.org/c/boc/bocode/s456779.html.
91. Schmitt, Susan K, LaShika Sneed, and Ciaran S Phibbs. 2006. “Costs of Newborn Care in
California: A Population-Based Study.” Pediatrics 117 (1): 154–60.
http://www.ncbi.nlm.nih.gov/pubmed/16396873.
92. Seuhs, Thomas. 2012. “Medicaid Initiative Seeks to Reduce Risk of Premature Births,” July 3.
http://www.hhs.state.tx.us/news/release/070312-Premature-Births.shtml.
187

93. Smith, Vincent C, John A F Zupancic, Marie C McCormick, Lisa A Croen, John Greene, Gabriel
J Escobar, and Douglas K Richardson. 2005. “Trends in Severe Bronchopulmonary Dysplasia
Rates between 1994 and 2002.” The Journal of Pediatrics 146 (4): 469–73.
doi:10.1016/j.jpeds.2004.12.023.
94. St John, E B, K G Nelson, S P Cliver, R R Bishnoi, and R L Goldenberg. 2000. “Cost of Neonatal
Care according to Gestational Age at Birth and Survival Status.” American Journal of Obstetrics
and Gynecology 182 (1 Pt 1): 170–75. http://www.ncbi.nlm.nih.gov/pubmed/10649175.
95. Staff, HSS Inc. 2003. “Obstetric and Newborn Coding Guidelines Reviewed for ICD-9-CM
Coding Issues on ADVANCE for Health Information Professionals.” http://health-
information.advanceweb.com/Article/Obstetric-and-Newborn-Coding-Guidelines-Reviewed-for-
ICD-9-CM-Coding-Issues.aspx.
96. Violato, Mara, Stavros Petrou, Ron Gray, and Maggie Redshaw. 2011. “Family Income and Child
Cognitive and Behavioural Development in the United Kingdom: Does Money Matter?” Health
Economics 20 (10): 1201–25.
97. Wagstaff, Adam, Eddy Van Doorslaer, and Naoko Watanabe. 2003. “On Decomposing the
Causes of Health Sector Inequalities with an Application to Malnutrition Inequalities in
Vietnam.” Journal of Econometrics 112 (1): 207–23.
98. Weesie, J. 1999. “Seemingly Unrelated Estimation and the Cluster-Adjusted Sandwich
Estimator.” Stata Technical Bulitin 52: 34–47.
99. WHO. 2012. “Preterm Birth Fact Sheet, Novemeber 2012.”
http://www.who.int/mediacentre/factsheets/fs363/en/.
100. Wooldridge, Jeffrey. 2010. “Corner Solution Responses.” In Econometric Analysis of Cross
Section and Panel Data , 2nd ed., 667–730. Cambridge: MIT Press.
188

101. Wooldridge, Jeffrey M. 2002. Econometric Analysis of Cross Section and Panel Data. Edited
by MA Cambridge. Booksgooglecom. Vol. 58. MIT Press Books. MIT Press.
doi:10.1515/humr.2003.021.
102. Wooldridge, Jeffrey M. 1995. “Selection Corrections for Panel Data Models under Conditional
Mean Independence Assumptions.” Journal of Econometrics. doi:10.1016/0304-4076(94)01645-
G.
103. Yeargin-Allsopp, Marshalyn, Godfrey P. Oakley, Catherine C. Murphy, R. Keith Sikes, and The
Metropolitan Atlanta Developmental Disabilities Study Staff. 1992. “A Multiple-Source Method
for Studying the Prevalence of Developmental Disabilities in Children: The Metropolitan Atlanta
Developmental Disabilities Study.” Pediatrics 89 (4): 624–30.
http://pediatrics.aappublications.org/content/89/4/624.short.
104. Zupancic JA. 2006. A Systematic Review of Costs Associated with Prematurity. .
105. Zwicker, Jill Glennis, and Susan Richardson Harris. 2008. “Quality of Life of Formerly Preterm
and Very Low Birth Weight Infants from Preschool Age to Adulthood: A Systematic Review.”
Pediatrics 121 (2): e366–76.

189

Appendix 1. ICD-9-CM codes used in the classification of premature infants into gestational age
and birth-weight categories
ICD-9-CM code Description Comments
765.0x Disorders related to
extreme immaturity of
infant
Fifth digit is used for coding birth-
weight category
0: unspecified [weight]
1: less than 500g
2: 500-749g
3: 750-999g
4: 1,000-1,249g
5: 1,250-1,499g
6: 1,500-1,749g
7: 1,750-1,999g
8: 2,000-2,499g
9: 2,500g or over
765.1x Disorders related to
other preterm infants
765.2x Weeks of gestation Fifth digit is used for coding
gestational age category
0: unspecified weeks of gestation
1: < 24 completed weeks of GA
2: 24 completed weeks of GA
3: 25-26 completed weeks of GA
4: 27-28 completed weeks of GA
5: 29-30 completed weeks of GA
6: 31-32 completed weeks of GA
7: 33-34 completed weeks of GA
8: 35-36 completed weeks of GA
9: 37 or more completed weeks of GA

190

Appendix 2. Sample Preparation Diagram.
191

Appendix 3. Deaths, attrition and cumulative retention among full-term and preterm cohorts.
Control group
(A)
Age
in
mo
nths
(B)
No.
rejoin
ed
cohort
(C) No.
retaine
d
(cumul
ative N)
(D) %
Remainin
g in
cohort
(cumulati
ve)
(E) No.
of
Deaths
(F) %
Deaths
(G)
Attritio
n
(H)
Effectiv
e
attrition
( G
t
-B
t-
1
)
(I)
Effectiv
e
attrition
rate
(%)
(G)
Cumulative
attrition

0
0
107,045 100% 0 0% 0 0 0% 0%

6
1,509
99,336 93% 206 0.2% 7,709 7,709 7% 7%

18
7,306
68,501 64% 24 0.2% 32,138 30,629 30% 36%

36
9,358
55,242 52% 8 0.2% 20,541 13,235 17% 48%

60
-
38,243 36% 3 0.2% 26,349 16,991 26% 64%

GA: 33-36 weeks

(A)
Age
in
mo
nths
(B No.
rejoined
cohort
(C) No.
retained
(cumulat
ive N)
(D) %
Remainin
g in
cohort
(cumulati
ve)
(E) No.
of
Deaths
(F) %
Deaths
(G)
Attritio
n
(H)
Effectiv
e
attritio
n ( G
t
-
B
t-1
)
(I)
Effective
attrition
rate (%)
(G)
Cumul
ative
attritio
n

0 0 3,393 100% 0 0 0 0 0% 0%

6 37 3,190 94% 7 0.2% 203 203 6% 6%

18 209 2229 66% 0 0.2% 991 954 30% 34%

36 331 1740 51% 1 0.2% 698 489 20% 49%

60 - 1161 34% 1 0.3% 909 578 28% 66%

GA: < 33 weeks

(A)
Age
in
mont
hs
(B)
No.
rejoin
ed
cohort
(C) No.
retained
(cumulativ
e N)
(D) %
Remaining
in cohort
(cumulative)
(E) No.
of
Deaths
(F) %
Deaths
(G)
Attrition
(H)
Effective
attrition
( G
t
-B
t-
1
)
(I)
Effecti
ve
attritio
n rate
(%)
(G)
Cumulati
ve
attrition

0 0 1804 100% 0 0 0 0 0% 0%

6 42 1628 90% 382 21% 176 176 10% 10%

18 86 1008 56% 6 22% 280 238 14% 23%

36 107 830 46% 4 22% 258 172 16% 32%

60 - 623 35% 0 22% 310 203 22% 44%

392

192

Appendix 4. Comparison of baseline characteristics between full-term and preterm cohorts over
time.

Time period (Age)
→
Wave 2 (6-18 months) Wave 3 (19-36 months) Wave 4 (37-60 months)

Classification by
GA →
Controls
33-36
weeks
< 33
weeks
Controls
33-36
weeks
< 33
weeks
Controls
33-36
weeks
< 33
weeks
N → 68,501 2,229 1008 55,242 1,740 830 38,243 1,161 623
Gender
Male
0.51 0.51 0.52 0.51 0.51 0.52 0.51 0.50 0.52
Female
0.49 0.49 0.48 0.49 0.49 0.48 0.49 0.50 0.48
Race-
ethnicity
White, non-
hispanic
0.15 0.16 0.13 0.15 0.16 0.13 0.14 0.14 0.13
Black, non-
hispanic
0.07 0.11 0.15 0.08 0.11 0.16 0.08 0.12 0.17
Hispanic
0.56 0.53 0.48 0.63 0.59 0.51 0.71 0.68 0.56
Other race-
ethnicity
0.005 0.006 0.008 0.005 0.006 0.008 0.006 0.01 0.008
Missing race-
ethnicity
0.21 0.19 0.24 0.13 0.13 0.19 0.06 0.05 0.13
Texas
region of
residence
South Texas
0.35 0.30 0.23 0.37 0.33 0.23 0.34 0.29 0.24
Metroplex
0.10 0.12 0.15 0.10 0.11 0.15 0.14 0.17 0.18
Gulf coast
0.09 0.09 0.12 0.09 0.09 0.11 0.13 0.14 0.14
Upper east Texas
0.10 0.10 0.11 0.09 0.10 0.11 0.07 0.07 0.10
Alamo
0.05 0.06 0.06 0.05 0.06 0.07 0.06 0.08 0.06
Central Texas
0.07 0.08 0.11 0.07 0.08 0.11 0.05 0.06 0.09
Capital
0.03 0.02 0.04 0.03 0.02 0.04 0.04 0.04 0.05
West Texas
0.06 0.07 0.04 0.06 0.06 0.04 0.04 0.04 0.04
Upper Rio Grande
0.02 0.02 0.02 0.02 0.03 0.02 0.02 0.03 0.01
Northwest Texas
0.05 0.05 0.04 0.04 0.04 0.03 0.04 0.04 0.03
High plains
0.04 0.05 0.03 0.04 0.05 0.03 0.03 0.04 0.03
Southeast Texas
0.04 0.03 0.03 0.04 0.03 0.04 0.03 0.02 0.03
Miscellaneous
0.001 0.001 0.001 0.001 0.001 0.001 0.002 0.001 0.002
Year of
birth
2002
0.49 0.08 0.41 0.51 0.09 0.40 0.55 0.09 0.42
2003
0.51 0.92 0.59 0.49 0.91 0.60 0.45 0.91 0.58
Birth-
weight
class
BW ≥ 2,500g
0 0.45 0 0 0.44 0 0 0.44 0
BW: 2,000-2,499g
0 0.38 0.05 0 0.38 0.04 0 0.38 0.04
BW: 1,250-2,000g
0 0.11 0.23 0 0.11 0.23 0 0.11 0.20
BW ≤ 1,249g
0 0 0.69 0 0 0.7 0 0 0.74
BW missing
1.0 0.07 0 1.0 0.07 0 1.0 0.07 0
Birth
defects CHD
0.05 0.09 0.28 0.05 0.10 0.29 0.06 0.09 0.3
Neonatal
RDS
0.03 0.16 0.68 0.03 0.17 0.67 0.03 0.14 0.70
193

Time period (Age)
→
Wave 2 (6-18 months) Wave 3 (19-36 months) Wave 4 (37-60 months)

Classification by
GA →
Controls
33-36
weeks
< 33
weeks
Controls
33-36
weeks
< 33
weeks
Controls
33-36
weeks
< 33
weeks
N → 68,501 2,229 1008 55,242 1,740 830 38,243 1,161 623
conditions
observed
within few
months of
birth*
Respiratory failure
0.02 0.04 0.25 0.03 0.05 0.26 0.03 0.06 0.29
PDA
0.02 0.02 0.29 0.02 0.02 0.30 0.02 0.03 0.33
ROP
0 0.002 0.21 0 0.001 0.10 0 0 0.08
Neonatal infections
0.07 0.11 0.20 0.07 0.13 0.20 0.07 0.10 0.22
NEC/ Peritonitis/
Other GI
0.002 0.006 0.08 0.002 0.009 0.08 0.002 0.009 0.09
Newborn feeding
difficulties & FTT
0.04 0.08 0.18 0.04 0.08 0.18 0.04 0.09 0.20
GA = gestational age; BW = birth-weight; CHD = congenital heart disease; CV = cerebrovascular; RDS = respiratory distress syndrome; PDA = patent ductus arteriosus; ROP
= retinopathy of prematurity; NEC = necrotizing enterocolitis; GI = gastrointestinal; FTT = failure to thrive. *Note: The numbers reported for these major perinatal conditions
are not prevalence numbers. Since these conditions emerged during the perinatal period, they were treated as time invariant characteristics.
194

Appendix 5. Summary of time-varying characteristics during each wave.

Time period
(Age) → 6-18 months 19-36 months 37-60 months

Classificatio
n by
gestational
age →
Controls
33-36
weeks
< 33
weeks
Controls
33-36
weeks
< 33
weeks
Controls
33-36
weeks
< 33
weeks

N → 68,501 2,229 1008 55,242 1,740 830 38,243 1,161 623

Medicaid program of eligibility at time 't' (6 categories)

Expansion
s
0.68 0.67 0.41 0.77 0.77 0.51 0.77 0.74 0.47

SSI 0.006 0.01 0.36 0.01 0.02 0.30 0.03 0.05 0.36

AFDC/TA
NF
0.28 0.29 0.2 0.12 0.13 0.11 0.10 0.12 0.10

Category
missing
0.029 0.027 0.039 0.097 0.08 0.078 0.10 0.09 0.07

Number of medical conditions treated per 6-month period
1
(± S.D.)

t 0.8 (0.9) 0.8 (1.0) 1.7 (1.95) 0.4 (0.7) 0.4 (0.8) 1.1 (1.5) 0.1 (0.4) 0.2 (0.6) 0.6 (1.1)
t-1 0.9 (1.0) 1.1 (1.1) 2.6 (2.3) 0.6 (0.85) 0.6 (0.95) 1.4 (1.7) 0.2 (0.5) 0.2 (0.6) 0.8 (1.3)

Comorbidities at time 't' (Yes/no)
Chronic
respiratory
conditions
BPD 0.001 0.003 0.15 0.0005 0.001 0.05 0.0001 0 0.02
Pediatric
asthma
0.12 0.15 0.25 0.11 0.12 0.26 0.07 0.08 0.23
Bronchitis 0.21 0.20 0.24 0.16 0.16 0.21 0.06 0.07 0.13
Other
respiratory
symptoms
(Wheeze,
apnea, cough)
0.18 0.20 0.40 0.12 0.13 0.24 0.06 0.06 0.19
LRT
infections
0.12 0.14 0.20 0.085 0.08 0.18 0.03 0.04 0.12
URT
infections
0.74 0.72 0.68 0.59 0.56 0.6 0.34 0.36 0.47
CNS
disorders
Perinatal
CNS
2
/CV
disorders
0.002 0.002 0.04 0.001 0.002 0.02 0.001 0.001 0.010
Epileptic
disorders
0.003 0.004 0.01 0.003 0.008 0.02 0.003 0.007 0.03
Motor
disorders
3

0.005 0.006 0.04 0.006 0.01 0.07 0.005 0.01 0.07
Growth
delay &
Developmen
tal disorders
NDD 0.02 0.04 0.28 0.04 0.07 0.27 0.016 0.02 0.09
Delayed
milestones/F
TT
0.016 0.03 0.16 0.02 0.03 0.15 0.007 0.015 0.08
195

Time period
(Age) → 6-18 months 19-36 months 37-60 months

Classificatio
n by
gestational
age →
Controls
33-36
weeks
< 33
weeks
Controls
33-36
weeks
< 33
weeks
Controls
33-36
weeks
< 33
weeks

N → 68,501 2,229 1008 55,242 1,740 830 38,243 1,161 623
Attention/con
duct disorders
0.002 0.003 0.02 0.008 0.01 0.03 0.02 0.03 0.06
Vision
problems
0.01 0.02 0.26 0.015 0.02 0.185 0.02 0.03 0.17

GI functional
problems
4

0.07 0.09 0.21 0.05 0.05 0.10 0.03 0.03 0.12

GI defects
5
0.005 0.009 0.03 0.004 0.006 0.01 0.002 0.003 0.006

Nutritional
deficiency,
anemias
0.04 0.035 0.06 0.04 0.04 0.05 0.02 0.015 0.035

Endocrine &
metabolic
disorders
6

0.037 0.04 0.06 0.02 0.03 0.05 0.009 0.01 0.04

Hernias 0.007 0.008 0.05 0.004 0.005 0.02 0.002 0.002 0.006

Misc.
infections
7

0.15 0.13 0.17 0.12 0.11 0.13 0.06 0.05 0.14

Receipt of early childhood intervention /comprehensive care services in the past (Yes/no)

t-1 0.09 0.09 0.25 0.07 0.09 0.30 0.04 0.04 0.1
t-2 0.09 0.09 0.2 0.08 0.09 0.32 0.04 0.06 0.16

Hospitalization in the previous 6-month period (Yes / no)
8

t-1 0.08 0.11 0.17 0.03 0.04 0.11 0.01 0.02 0.04
TANF = Temporary assistance for needy families; SSI = Social security income; AFDC = Aid for families with dependent children; PDA = patent ductus
arteriosus; CHD = congenital heart diseases; BPD = bronchopulmonary dysplasia; NEC = necrotizing enterocolitis; CV = cerebrovascular; LRT = lower
respiratory tract; CNS = central nervous system; NDD = neurodevelopmental delay; FTT = failure to thrive; GI = gastro-intestinal; t-1 = past 6-month period (1
lag); t-2 = one but previous 6-month period (2 lags);
1. Include all conditions listed above except miscellaneous infections. 2. Conditions including cerebral degenerations manifest in childhood, spastic hemiplegia,
hemiparesis and cerebrovascular diseases. 3. Cerebral palsy, paralytic syndromes, abnormal gait, incoordination, tics or repetitive movements. 4. Malabsorption,
gastroesophageal reflux, colic, diarrhea or constipation. 5. Bowel obstruction, abdominal wall or GI tract defects. 6. Hypothyroidism, diabetes, pituitary
disorders, disorders of carbohydrate, fat or protein metabolism. 7. Urinary tract, skin, CNS, sepsis or unspecified disorders. 8. Hospital episode during the first 6
months period refers to an episode other than the birth hospitalization.

196

Appendix 6. Interaction test: Linear regression of ln(healthcare costs) on gestational age and its interactions
with group characteristics.
Character
istics
Coe
ff.
[SE
]
Ch.
Coe
ff.
[SE]
Ch.
Coe
ff.
[SE]
Ch.
Coeff.
[SE]
Ch.
Co
eff.
[SE
]
Ch.
Coe
ff.
[SE]
P1 (vs.
full-term)
0.1
39* P1*AF
DC
0.02
2
Other
respira
tory
(OR)
0.81
* P1*Mo
tor dis.
0.29
P2*DL
M
-
0.1
7 Hernias
1.18
7*
[0.0
6]
[0.0
4]
[0.0
1]
[0.3]
[0.
11]
[0.0
7]
P2(vs. full-
term)
0.3
10*
P2*AF
DC
0.58
6* P1*O
R
-
0.07
8
P2*Mo
tor dis.
0.15 GI
function
al dis.
0.7
5* P1*Herni
as
0.12
2
[0.0
8]
[0.0
7]
[0.0
8]
[0.15]
[0.
02]
[0.4
0]
Male
0.063
0*
Eligibili
ty status
missing
-
1.47
1*
P2*O
R
0.17
*
Seizur
e dis.
1.41*
P1*GI
fn. Dis.
-
0.0
5
P2*Herni
as
-
0.50
1*

[0.006
]
[0.0
2]
[0.0
8]
[0.08]

[0.
12]
[0.2
0]
Year of
birth 2003
(vs. 2002)
0.0
65* P1*elig.
miss.
0.03
2
LRT
0.82
*
P1*Sei
zure
dis.
-0.08
P2*GI
fn. Dis.
-
0.0
8
Miscella
neous
infection
s
0.96
3*
[0.0
1]
[0.0
9]
[0.0
2]
[0.31]
[0.
11]
[0.0
1]
African-
American
(vs.
Hispanic)
-
0.3
5*
P2*elig.
miss.
-
0.69
*
P1*L
RT
0.03
5
P2*Sei
zure
dis.
-0.17
GI
defects
1.2
9* P1*Misc.
infx.
-
0.12
8
[0.0
1]
[0.1
6]
[0.0
9]
[0.27]
[0.
07]
[0.0
8]
White
-
0.2
2* Age
-
0.12
*
P2*L
RT
-
0.33
0* NDD
1.16*
P1*GI
defects
-
0.3
2
P2*
Misc.
infx
-
0.40
0*
[0.0
1]
[0.0
01]
[0.1
0]
[0.03]
[0.
32]
[0.1
0]
Other
Race
(Asian/Am
. Indian)
-0.06
P1*Age
-0.01
URT
1.76*
P1*N
DD
-0.15
P2*GI
defects
-0.38
F (48, 476903) =
15.59
p-value < 0.0001
[0.04
]
[0.01]
[0.00
8]
[
0
.
1
2
]
[0.33
]
Race
missing
-
0.4
1*
P2*Age
0.01
P1*U
RT
-
0.05
72
P2*N
DD
-0.38*
Endocrin
e/
Metaboli
c.
1.2
0*
P1 = 33-36 weeks
of GA;
P2 = < 33 weeks of
GA; [0.0 [0.0 [0.0 [0.08] [0.
197

1] 1] 4] 03] CHD = Congenital
heart disease
SSI = Social
security income
AFDC = Aid for
families with
dependent children
NDD =
Neurodevelopment
al delay
DLM = delayed
milestones
GI =
Gastrointestinal
*p<0.05

CHD
(Yes/No)
0.9
9*
No. of
follow-
up
periods
0.09
* P2*U
RT
-
0.46
*
Vision
disorde
rs
1.04*
P1*Endo
/ Metab.
-
0.0
5
[0.0
4]
[0.0
01]
[0.0
6]
[0.04]
[0.
17]
SSI (vs.
expansion)
1.51*
Hospital
admissio
n.
t-1

0.19*
CNS
0.82*
P1*Visi
on dis.
-0.02
P2*Endo/
Metab.
-
0.43
*
[0.03
]
[0.02] [0.09]
[
0
.
1
8
]
[0.17
]
P1*SSI
-
0.2
9*
Asthma/
Bronchi
tis
1.11
* P1*C
NS
0.45
P2*Vis
ion dis.
-0.25* Nutrition
al
disorders
0.7
6*
[0.1
4]
[0.0
1]
[0.4
4]
[0.09]
[0.
03]
P2*SSI
-
0.6
6*
P1*Ast
hma_
bronch
-
0.00
5
P2*C
NS
0.03
Delaye
d
milesto
nes
0.93*
P1*Nutri
_dis.
0.2
0
[0.0
7]
[0.0
6]
[0.2
1]
[0.04]
[0.
20]
AFDC
(vs.
expansion)
0.1
1*
P2*Ast
hma_
bronch
0.13 Motor
disord
ers
1.04
* P1*DL
M
0.02
P2*Nutri
_dis.
-
0.4
8*
[0.0
1]
[0.0
7]
[0.0
6]
[0.17]
[0.
20]
198

Appendix 7. Attrition rates by wave (cross-sectional) [Note: Attrition does not include deaths]

0%
10%
20%
30%
40%
50%
0 6 18 36 60
Attrition rate by wave (cross-sectional)
Full-term
33-36 weeks of GA
<33 weeks of GA
199

Appendix 8. Comparison of unweighted and weighted fixed effects OLS estimates [Attrition weights
calculated for each wave cross-sectionally].
Dependent
variable =
ln(overall
healthcare
costs)
(A) (B) (C) (D) (E) (F) (G) (H)
(I)
Controls 33-36 weeks of GA <33 weeks of GA
Unweigh
ted
Coeff.
[SE]
Weight
ed
Coeff.
[SE]
%
Differen
ce in
coefficie
nts (UW
- W)
Unweigh
ted
Coeff.
[SE]
Weight
ed
Coeff.
[SE]
%
Differen
ce in
coefficie
nts (UW
- W)
Unweigh
ted
Coeff.
[SE]
Weight
ed
Coeff.
[SE]
%
Differen
ce in
coefficie
nts (UW
- W)
N 455665 455665 14320 14320 6987 6987
R-sq 0.445 0.436 0.483 0.475 0.584 0.580
adj. R-sq 0.325 0.315 0.367 0.357 0.502 0.497
Constant
6.28*** 6.07***
23%
6.44*** 6.28***
17%
7.58*** 7.42***
17%
[0.013] [0.014] [0.076] [0.079] [0.11] [0.12]
SSI
t-1
(vs.
expansions)
1.30*** 1.30***
0%
1.14*** 1.16***
-2%
0.97*** 0.91***
6%
[0.074] [0.085] [0.33] [0.33] [0.14] [0.15]
AFDC
t-1

0.042**
0.065**
*
-2%
0.10 0.12
-2%
0.38** 0.35**
3%
[0.013] [0.015] [0.074] [0.078] [0.12] [0.13]
Elig. cat.
Missing
t-1

-0.41***
-
0.27***
-13%
-0.34* -0.15
-17%
-1.16***
-
1.01***
-14%
[0.026] [0.028] [0.16] [0.17] [0.28] [0.28]
CHD
t

1.21*** 1.34***
-12%
1.32*** 1.38***
-6%
0.92*** 0.95***
-3%
[0.052] [0.058] [0.24] [0.24] [0.24] [0.23]
Hospitalizati
on
t-1

-0.097***
-
0.099**
*
0%
-0.16 -0.14
-2%
-0.012
-
0.00029
-1%
[0.019] [0.021] [0.094] [0.096] [0.098] [0.10]
12-18 (vs. 6-
12 mos)
-0.80***
-
0.74***
-6%
-0.97***
-
0.96***
-1%
-1.20***
-
1.22***
2%
[0.011] [0.012] [0.063] [0.067] [0.100] [0.11]
18-24
months
-1.23***
-
1.22***
-1%
-1.12***
-
1.14***
2%
-1.59***
-
1.61***
2%
[0.014] [0.016] [0.080] [0.084] [0.12] [0.12]
24-30
months
-1.68***
-
1.61***
-7%
-1.89***
-
1.88***
-1%
-2.32***
-
2.31***
-1%
[0.015] [0.017] [0.086] [0.089] [0.13] [0.13]
30-36
months
-2.44***
-
2.34***
-10%
-2.82***
-
2.74***
-8%
-3.01***
-
2.96***
-5%
[0.016] [0.017] [0.094] [0.097] [0.13] [0.14]
36-42
months
-2.29***
-
2.22***
-7%
-2.69***
-
2.66***
-3%
-2.92***
-
2.94***
2%
[0.017] [0.019] [0.11] [0.11] [0.14] [0.14]
42-48 -2.73*** -
-7%
-2.65*** -
-5%
-3.16*** -
-7%
200

Dependent
variable =
ln(overall
healthcare
costs)
(A) (B) (C) (D) (E) (F) (G) (H)
(I)
Controls 33-36 weeks of GA <33 weeks of GA
Unweigh
ted
Coeff.
[SE]
Weight
ed
Coeff.
[SE]
%
Differen
ce in
coefficie
nts (UW
- W)
Unweigh
ted
Coeff.
[SE]
Weight
ed
Coeff.
[SE]
%
Differen
ce in
coefficie
nts (UW
- W)
Unweigh
ted
Coeff.
[SE]
Weight
ed
Coeff.
[SE]
%
Differen
ce in
coefficie
nts (UW
- W)
N 455665 455665 14320 14320 6987 6987
R-sq 0.445 0.436 0.483 0.475 0.584 0.580
adj. R-sq 0.325 0.315 0.367 0.357 0.502 0.497
months 2.66*** 2.60*** 3.09***
[0.018] [0.020] [0.11] [0.11] [0.14] [0.15]
48-54
months
-2.51***
-
2.42***
-9%
-2.44***
-
2.39***
-5%
-2.90***
-
2.83***
-7%
[0.019] [0.020] [0.11] [0.11] [0.15] [0.15]
54-60
months
-2.81***
-
2.74***
-7%
-2.71***
-
2.66***
-5%
-3.02***
-
3.01***
-1%
[0.019] [0.020] [0.11] [0.12] [0.15] [0.16]
***p<0.001; **p<0.01; *p<0.05; t = current 6-month period t-1 = previous 6-month period; SSI: Social security income AFDC: Aid for families with dependent
children CHD: Congenital heart disease SE = standard error

201

Abstract (if available)

Abstract Ch2 Objective: This study evaluated the cost-effectiveness of a 100% human milk-based diet composed of mother’s milk fortified with a donor human milk-based human milk fortifier (HMF) versus mother’s milk fortified with bovine milk-based HMF to initiate enteral nutrition among extremely premature infants in the neonatal intensive care unit (NICU). Methods: A net expected costs calculator was developed to compare the total NICU costs among extremely premature infants who were fed either a bovine milk-based HMF-fortified diet or a 100% human milk-based diet, based on the previously observed risks of overall necrotizing enterocolitis (NEC) and surgical NEC in a randomized controlled study that compared outcomes of these two feeding strategies among 207 very low birth weight infants. The average NICU costs for an extremely premature infant without NEC and the incremental costs due to medical and surgical NEC were derived from a separate analysis of hospital discharges in the state of California in 2007. The sensitivity of cost-effectiveness results to the risks and costs of NEC and to prices of milk supplements was studied. Results: The adjusted incremental costs of medical NEC and surgical NEC over and above the average costs incurred for extremely premature infants without NEC, in 2011 US$, were $74,004 (95% confidence interval [CI], $47,051–$100,957) and $198,040 (95% CI, $159,261–$236,819) per infant, respectively. Extremely premature infants fed with 100% human milk-based products had lower expected NICU length of stay and total expected costs of hospitalization, resulting in net direct savings of 3.9 NICU days and $8,167.17 (95% confidence interval, $4,405–$11,930) per extremely premature infant (p<0.0001). Costs savings from the donor HMF strategy were sensitive to price and quantity of donor HMF, percentage reduction in risk of overall NEC and surgical NEC achieved, and incremental costs of surgical NEC. Conclusions: Compared with feeding extremely premature infants with mother’s milk fortified with bovine milk-based supplements, a 100% human milk-based diet that includes mother’s milk fortified with donor human milk-based HMF may result in potential net savings on medical care resources by preventing NEC. ❧ Ch3 Background: Infants who survive advanced necrotizing enterocolitis (NEC) at the time of birth are at increased risk of having poor long term physiological and neurodevelopmental growth. The economic implications of the long term morbidity in these children have not been studied to date. This paper compares the long term healthcare costs beyond the initial hospitalization period incurred by medical and surgical NEC survivors with that of matched controls without a diagnosis of NEC during birth hospitalization. Methods: The longitudinal healthcare utilization claim files of infants born between January 2002 and December 2003 and enrolled in the Texas Medicaid fee-for-service program were used for this research. Propensity scoring was used to match infants diagnosed with NEC during birth hospitalization to infants without a diagnosis of NEC on the basis of gender, race, prematurity, extremely low birth weight status and presence of any major birth defects. The Medicaid paid all-inclusive healthcare costs for the period from 6 months to 3 years of age among children in the medical NEC, surgical NEC and matched control groups were evaluated descriptively, and in a generalized linear regression framework in order to model the impact of NEC over time and by birth weight. Results: Two hundred fifty NEC survivors (73 with surgical NEC) and 2,909 matched controls were available for follow-up. Medical NEC infants incurred significantly higher healthcare costs than matched controls between 6–12 months of age (mean incremental cost = US $5,112 per infant). No significant difference in healthcare costs between medical NEC infants and matched controls was seen after 12 months. Surgical NEC survivors incurred healthcare costs that were consistently higher than that of matched controls through 36 months of age. The mean incremental healthcare costs of surgical NEC infants compared to matched controls between 6–12, 12–24 and 24–36 months of age were US $18,274, $14,067 (p<0.01) and $8,501 (p=0.06) per infant per six month period, respectively. These incremental costs were found to vary between subgroups of infants born with birth weight <1,000 grams versus ≥ 1,000 grams (p<0.05). Conclusions: The all-inclusive healthcare costs of surgical NEC survivors continued to be substantially higher than that of matched controls through the early childhood development period. These results can have important treatment and policy implications. Further research in this topic is needed. ❧ Ch4 Background: The long-term healthcare cost among preterm survivors, classified by gestational age, has not been studied well. Previous studies have largely ignored the role of unobserved heterogeneity, leading to biased inferences regarding the treatment effect of preterm births. Objectives: To evaluate the incremental healthcare costs and drivers of costs, among preterm survivors, within a US public payer population from 6 months to 5 years of age. Study Design & Methods: Children born moderate to late preterm (33-36 weeks of gestational age), very to extremely preterm (< 33 weeks of GA) or full-term and enrolled in the Texas Medicaid fee-for-service program were identified from their birth hospitalization claims. An unbalanced panel structure was used to summarize data on time-varying covariates and outcomes within each period. Two-part model, with fixed effects in each step, was used to estimate logarithm of total healthcare costs, separately, among preterm and control groups. The incremental effect of preterm births was obtained using Blinder-Oaxaca type decompositions and standard errors were estimated by bootstrapping. Attrition weighting was used to account for non-random attrition. Cumulative costs per child were estimated after applying a discount rate of 3% to costs incurred beyond 12 months of age. Results: The incremental impact of moderate to late preterm birth, per se, was not significant beyond 6 months, after controlling for unobserved child specific factors. When combined with an additional risk factor, birth weight < 2,500 grams, the average incremental cost of M/L preterm birth was US$ 560 per child for the total period between 6 and 60 months of age (p < 0.05). Over the same period, the average incremental cost of preterm birth at < 33 weeks of GA was $4,800 per child (p < 0.001). Neurodevelopmental delay, asthma/bronchial disorders, respiratory symptoms and refractory vision disorders were top drivers of the healthcare cost difference between preterm and full-term survivors. Weighting for attrition had a significant impact on the treatment effect estimates. Conclusions: Controlling for unobserved heterogeneity while evaluating the long-term consequences of preterm birth should be considered. M/L preterm with birth weight < 2,500 grams and V/E preterm born children bear significant financial impact to Texas Medicaid program through 5 years of age. These findings have important policy implications.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Impact of pharmacy-based transitional care on healthcare utilization and costs

PDF

Evaluating approaches to reduce inappropriate antibiotic use in the United States

Asset Metadata

Creator Ganapathy, Vaidyanathan (author)

Core Title Selected papers on the evaluation of healthcare costs of prematurity and necrotizing enterocolitis using large retrospective databases

School School of Pharmacy

Degree Doctor of Philosophy

Degree Program Pharmaceutical Economics and Policy

Publication Date 04/28/2015

Defense Date 03/16/2015

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag Blinder-Oaxaca,costs,fixed effects,generalized linear models,gestational age,necrotizing enterocolitis,OAI-PMH Harvest,prematurity,propensity match

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Hay, Joel W. (committee chair), Ingles, Sue Ann (committee member), Romley, John A. (committee member)

Creator Email muraleevaidya@gmail.com,vaidyang@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c3-563464

Unique identifier UC11300633

Identifier etd-GanapathyV-3412.pdf (filename),usctheses-c3-563464 (legacy record id)

Legacy Identifier etd-GanapathyV-3412.pdf

Dmrecord 563464

Document Type Dissertation

Format application/pdf (imt)

Rights Ganapathy, Vaidyanathan

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA