Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
Essays in empirical health economics
(USC Thesis Other)
Essays in empirical health economics
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
ESSAYS IN EMPIRICAL HEALTH ECONOMICS
by
Adam Kaufman
A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(ECONOMICS)
August 2009
Copyright 2009 Adam Kaufman
ii
Table of Contents
List of Tables iv
Abstract vi
Introduction 1
Chapter 1: A Quasi-Experimental Assessment of the Kids N Fitness Program for
the Reduction of Overweight and Obesity among Children 5
1.1 Introduction 5
1.2 Data 9
Figure 1.1: Sample Distribution >70
th
percentile BMI-z Scores for Paired Schools 17
1.3 Measuring the Intervention Effect at the School Level 19
1.4 Individual Level Model and Results 25
1.5 Impact of KnF Instructor Type 29
1.6 Conclusion 33
Chapter 1 References 35
Chapter 2: Conditional Hazard Modeling of Spells of Good and Poor Diabetes
Control in Children with Type 1 Diabetes 37
2.1 Introduction 37
2.2 Literature Review 38
2.3 Data 40
2.3.1 Spell Data 45
2.4 Econometric Model and Estimation 46
2.5 Results 51
2.6 Model with Unobserved Heterogeneity 56
2.7 Conclusions 59
Chapter 2 References 61
Chapter 3: Econometric Analysis of Elderly Drinking Behaviors, Changing
Economic Circumstances and Response to Spousal Death 63
3.1 Introduction 63
3.2 Literature Review 64
3.3 Data 69
3.3.1 Alcohol Consumption Dependent Variables 69
3.3.2 Spousal Death Variables: 73
3.3.3 Economic Variables: 77
3.3.4. Independent Covariates: 77
3.4 Linear Regression 81
3.5 Tobit Regression: 82
3.6 Results 89
3.7 Conclusion 95
Chapter 3 References 97
Bibliography 99
iii
Appendices 103
Appendix A: Detailed Information on the Kids N Fitness Program 103
Appendix B: Sensitivity Analysis – Results with A1C 7.0% and 9.0% 105
Appendix C: Example Determining Diabetes Control State and Spell Information 107
Appendix D: Wave 4 Example - Constructing the Became a widow(er) Variable 109
iv
List of Tables
Table 1.1: Summary Statistics of the Entire Sample Population 12
Table 1.2A: Child Enrollment Characteristics by Paired Schools & Test of 13
Equivalence for All Students
Table 1.2B: Child Enrollment Characteristics by Paired Schools & Test of 14
Equivalence for Students with Enrollment BMI >85th Percentile
Table 1.2C: Child Enrollment Characteristics by Paired Schools & Test of
Equivalence for Students with Enrollment BMI Percentile between 70th and 85
th
15
Table 1.3: Comparison of Observations and Missing Observations 19
Table 1.4A: Average Change in BMI-z score by School and Enrollment BMI
Percentile for All Students 20
Table 1.4B: Average Change in BMI-z score by School and Enrollment BMI
Percentile for Students with Enrollment BMI between 70th and 85th Percentile 20
Table 1.4C: Average Change in BMI-z score by School and Enrollment BMI
Percentile for Students with Enrollment BMI above the 85th Percentile 20
Table 1.5: Difference in Differences of BMI-z Score Intervention to Control
Schools 21
Table 1.6: Regression Analysis Difference in Intervention and Control BMI-z
Score Change 22
Table 1.7: Intervention and Control BMI Percentile Transition Probabilities 24
Table 1.8: Random Effects Regression Results at Individual Child Level (by BMI 27
Enroll Percentile)
Table 1.9: Regression Results for Underweight Sample Population (<15
Percentile) 29
Table 1.10: Regression Results Including Instructor Type (by Enrollment BMI
Percentile) 32
Table 2.1: Sample Population Characteristics 44
Table 2.2: Spell Information by Spell Type 45
Table 2.3: Duration Model Results – A1C Control Cutoff 8.0% 52
Table 2.4: Duration Model with Unobserved Heterogeneity Results – Control 58
Cutoff A1C 8.0%
v
Table 3.1A: Frequency Distribution of Drinking Behavior – Drinking Days Per
Week 70
Table 3.1B: Frequency Distribution of Drinking Behavior – Drinks per Day 71
Table 3.1C: Frequency Distribution of Drinking Behavior – Drinks Per Week 72
Table 3.2A: Descriptive Statistics of the Dependent and Independent Variables
(Male Population) 79
Table 3.2B: Descriptive Statistics of the Dependent and Independent Variables
(Female Population) 80
Table 3.3: Frequency of Observations on Sample of Individuals in the HRS Panel 81
Table 3.4A: Regression Results for Linear Regression and Tobit Models – Males
Days Drinking per Week 84
Table 3.4B: Regression Results for Linear Regression and Tobit Models – Males
Drinks per Day 85
Table 3.4C: Regression Results for Linear Regression and Tobit Models – Males
Drinks per Week 86
Table 3.4D: Regression Results for Linear Regression and Tobit Models –
Females Days Drinking per Week 87
Table 3.4E: Regression Results for Linear Regression and Tobit Models –
Females Drinks per Day 88
Table 3.4F: Regression Results for Linear Regression and Tobit Models –
Females Drinks per Week 89
Table 3.5: Elasticities of Demand – Male Drinking Behavior 91
Table 3.6: Elasticities of Demand – Female Drinking Behavior 94
Table B.1: Duration Model Conditional Hazard – Control Cutoff A1C 7.0% 105
Table B.2: Duration Model Conditional Hazard – Control Cutoff A1C 9.0% 106
Table C.1: A1C Values for Three Sample Patients 107
Table C.2: Example Transformation A1C to Diabetes Control States 107
Table D.1: Marital Status Change between Weave 4 and Wave 3 109
Table D.2: Number of Times Widowed between Wave 3 and Wave 4 109
Table D.3: Comparing Methods of Calculating Became a Widow(er) Variable 110
vi
Abstract
These essays examine econometric models of health behavior. The first essay
evaluates the Kids N Fitness program, an intervention to reduce childhood overweight
and obesity. The second essay uses a Childrens Hospital of Los Angeles private
dataset to examine determinants of diabetic glycemic control transitions. The third
essay analyzes drinking behavior and the response to spousal death among the elderly,
utilizing the Health and Retirement Study panel dataset. In each case, I analyze the
health behavior choice controlling for available individual characteristics including
economic variables. The models apply rigorous econometric techniques and are explicit
about the underlying assumptions.
I find that the Kids N Fitness program is effective at helping overweight children reduce
their BMI Z Score. In addition, the analysis points towards the conclusion that more
highly trained instructors do not yield improved program efficacy. I find that being a
teenager negatively impacts the probability that a child will return to good diabetes
control from a spell of poor control. Lastly, I find evidence that spousal death and being
a widow significantly impact elderly female drinking but not elderly male drinking.
The topics in these essays are generally approached from the medical literature and not
the economic literature. My analyses illustrate some of the potential value of applying
rigorous econometric and applied microeconomic techniques to these questions in
health and healthcare.
1
Introduction
In these essays I apply modern econometric models to data related to health behavior.
In these contexts, particular health related behaviors are seen as economic choices and
reduced form models are used to estimate the impact of characteristics on choices. The
essays demonstrate the breadth of application of econometric models to health behavior
including program evaluation, clinical practices and public health. Econometric modeling
is applicable in situations without random assignment and complicated questions of
endogeneity and dependence. While health behavior choices are becoming increasingly
important in health research, standard modeling in health literature often relies on
experimental design which cannot address these issues. Econometric techniques allow
for richer models which include, where possible, controls for economic variables and
rigorously address potential endogeneity issues.
These essays demonstrate the potential widespread value of applying economic
modeling to health topics. The breadth of the econometrics in these essays includes
comparing sample level estimates with individual level estimates, applying difference-in-
difference estimators, analyzing transition probabilities, comparing regression and Tobit
models, clustering standard errors, modeling duration probabilities and conditional
hazard functions, and controlling for unobserved heterogeneity. The essays address
prominent issues in health and help answer policy relevant questions. The essays
address three distinct health behaviors and show the value of applying micro-economic
modeling and econometric techniques to analysis of health behaviors.
2
The first essay evaluates the Kids N Fitness program, an intervention to reduce
childhood overweight and obesity. This essay evaluates data gathered from an
experiment of the KnF program with randomization at the school level. Using the
comparability of the school level results to support analyzing the intervention at the
individual child level, I make clear the assumptions and control for potential endogeneity
more rigorously. Ultimately, I find that the Kids N Fitness program is effective at helping
overweight children reduce their BMI-z score. Leveraging econometric techniques I am
able to control for the heteroskedacity in the error terms induced at the school level. And
importantly, given the assumptions, I can assess the impact of instructor type on
program success yielding policy relevant information. The analysis points towards the
conclusion that more highly trained instructors do not yield improved program efficacy.
The second essay uses a Childrens Hospital of Los Angeles private dataset to examine
determinants of diabetic glycemic control transitions. This data is collected as part of
regular clinical care for patients at CHLA and not part of a random experiment. Applying
the econometric technique of conditional hazard modeling I am able to assess the
impact of patient age and health insurance type on diabetes state transitions. While the
analysis is conditional based on observing patients in a particular state of control, the
modeling approach utilizes variations across patients to identify the impact. This
approach allows for explicit modeling of spell duration and does not require random
assignment. I find that being a teenager negatively impacts the probability that a child
will return to good diabetes control from a spell of poor control. I also find some
evidence that patients with government health insurance transition more quickly to states
of poor diabetes control. These results are of interest to practicing clinicians as they
provide input into characteristics that impact diabetes control.
3
The third essay analyzes drinking behavior and the response to spousal death among
the elderly. Alcohol consumption and its determinants is a well studied topic in the
health literature. However, generally the studies look at cross-sectional variations and
correlations. This essay considers the health behavior as reported in a large nationally
representative sample – The Health Retirement Study – similar to typical analysis but
uses the timing of spousal death to identify changes in alcohol consumption behavior.
The econometrics are rigorous in the required assumptions and their approach to
controlling for endogeneity. I find evidence that spousal death and being a widow
significantly impact elderly female drinking but not elderly male drinking. The analysis
shows that female alcohol consumption changes significantly as women’s circumstances
change while male drinking behavior seems to be relatively more stable.
The topics in these essays are generally approached from the medical literature and in
the economics literature from the point of view of costs. However, the modeling of
micro-economics and the techniques of modern econometrics have not been broadly
applied to the topics of health behavior choice. Taken together, the essays show the
breadth of the applicability of these techniques across health behaviors from energy
balance, diabetes and alcohol consumption choices. Additionally, the essays point to
the breadth of data from experimental outcomes, to clinical practice restults to survey
data which can be analyzed by these techniques. The issues of selection and
endogeneity inherent in economic choices are present in health behavior choice; the
techniques used in these essays can help control for observed and unobserved
heterogeneity, and provide a set of easily interpretable assumptions to better assess the
impact of characteristics on health behavior choices. Thus, questions of how individuals
4
make health behavior choices will not be answered entirely by randomized experiments.
My analyses illustrate some of the potential value of applying rigorous econometric and
applied microeconomic techniques to these questions in health and healthcare.
5
Chapter 1: A Quasi-Experimental Assessment of the Kids N Fitness Program for
the Reduction of Overweight and Obesity among Children
1.1 Introduction
The last decades have seen dramatic increases in the prevalence of elevated weights
and obesity in the population at large. This trend is particularly alarming among children
as the fraction that is overweight or obese has grown to be a substantial portion of the
entire childhood population. Wang, et al. (2002) described that the percentage of
children who are overweight and obese has grown substantially from 15.4% in the early
1970s to 25.6% in the early 1990s. The percentage now seems to have leveled off at an
alarming 31.9% (See Ogden et al. 2008 for a summary of recent obesity trends). The
health effects of elevated childhood weight are both immediate and long-term, and
include severe consequences such as significant increases in morbidity and mortality.
For example, Maffeis and Tato (2001) described that early onset obesity increased the
risk for diabetes, coronary heart disease, atherosclerosis, hip fracture and gout.
As the percentage of children overweight and obese has risen and the ill effects of
increased weight are better understood, the importance of policies aimed at decreasing
body weight has grown. On February 1
st
, 2007 President Bush made the following
statement, reinforcing the need to address childhood obesity. “Childhood obesity is a
costly problem for the country. It puts stress on American families. And we believe it is
necessary to come up with a coherent strategy to help folks all throughout our society
cope with the issue.
1
”
1
A transcript of the President’s statement is accessible at: http://georgewbush-
whitehouse.archives.gov/news/releases/2007/02/20070201-1.htmll
6
As a result of this increase in childhood obesity, numerous researchers have focused on
developing interventions that help decrease the population prevalence of overweight and
obese. They have worked to develop programs based on a variety of approaches
including broad environmental interventions and individually tailored approaches.
Programs can address an array of strategies (not mutually exclusive) ranging from
modifying the physical setting or built environment to make them more conducive for
healthy behavior, promoting physical activity, reducing TV watching, and improving
nutrition. (Dehgan et al. 2005)
The effectiveness of such interventions is an unsettled issue in the medical literature.
Surveying interventions aimed at decreasing childhood obesity Campbell et al. (2001)
concluded that “overall, the findings of the review suggest that currently there is limited
quality data on the effectiveness of obesity prevention programs and as such no
generalizable conclusions can be drawn.” Doak, et al. (2006) conducted another survey
of the literature with a focus solely on school-based interventions. They were able to
identify 25 such interventions and concluded that “the majority of overweight/obesity
prevention programs included in [the] review were effective.” Finally, a recent meta
analysis conducted by Kamath, et al (2008) concluded that “Pediatric obesity prevention
programs caused small changes in target behaviors and no significant effect on BMI as
compared to controls.”
My analysis adds to this literature by assessing the effectiveness of a particular
intervention aimed at helping overweight and obese children reduce their degree of
obesity as measured by their BMI-z score. The intervention, Kids N Fitness (KnF), was
7
developed at the Childrens Hospital Los Angeles (CHLA). It was developed in 2000 by a
multidisciplinary team
2
including endocrinologists, dietitians, nurses and physical
therapists as a 12 week clinic-based program (Monzavi et al. 2006). In 2006, CHLA
modified the program with the aim of altering the in-clinic program for delivery as an after
school program. The objective was to increase continued participation and reduce the
costs per patient while maintaining program effectiveness.
The KnF after school intervention involves both parents and children in the program. The
program supports improved physical activity and nutrition habits and aims to change
child and family behaviors to promote sustained improvements in child BMI. See
Appendix A for a more detailed description of the KnF program. From October 2006
through May 2007 CHLA conducted a trial of KnF at four pairs of schools. Final data
was collected from September to October 2007. From each of the four pairs of schools
one school was chosen to receive the treatment in a quasi random manner. In this
assignment process, school characteristics did not affect assignment. Thus, I am able to
assess the impact of KnF utilizing the quasi-experimental design.
My analysis addresses five questions. First, it assesses the effectiveness of the KnF
intervention in reducing BMI-z scores in those children at or above the 85
th
national
percentile for (age adjusted) BMI. Secondly, it extends the standard analysis in the
literature by analyzing how the KnF program affected children at-risk for becoming
overweight, i.e. those with a BMI between the 70
th
and 85
th
percentile in terms of
avoiding moving above the 85
th
percentile. Third, I compare results based on aggregated
data at the school level and results based on individual micro data. Most medical
2
See Valverde, et al. (1998) for a discussion of the elements of a multi-disciplinary school based
program.
8
studies examining school samples of this size would analyze only the micro data. The
analysis of the aggregated data requires fewer assumptions and therefore serves as a
robustness check. Fourth, I examine how a particular component, i.e. choice of
instructor type, affects the two groups described above. Since the sampling of children
is at school level, I apply modern econometric techniques to control for the sampling
process and allow for correlation among students at the school level through random
effects and at the pair level through a fixed effect. These individual and school effects
are time constant and effect BMI z score level. The econometric technique of
differencing removes these effects from the estimated equation. Lastly, time trends are
controlled for using a paired school fixed effect. Use of a paired school fixed effect,
allows me to control for systematic differences across pairs of schools, which if ignored,
could bias the evaluation of the intervention. While this later approach is common in
econometric analysis, it is not common in evaluations in the medical literature.
Summarizing the results, both the aggregate and micro analysis lead to the same
conclusion; KnF has a negative, statistically significant effect on the BMI-z scores of
those children who start between the 70th and 85
th
percentile, but has at best a minimal
effect on those at or above the 85
th
percentile. Further, I find that the intervention has a
significant negative effect on the number of children transitioning from the 70
th
-85
th
percentile to above the 85
th
percentile, while it has no effect on transitions from those
above the 85
th
percentile to the 70
th
-85
th
percentile. These results show the importance
of considering groups besides those above the 85
th
percentile, even if one’s interest is
only children above the 85
th
percentile. The results with respect to instructor qualification
are dramatic; the unlicensed instructors have a significant negative (good) effect on
those starting above the 85
th
percentile, while the credentialed teachers have no or a
9
slightly positive (bad) effect on these children. Both types of instructors have a desirable
effect on children starting in the 70
th
-85
th
percentile. This has important policy
implications, since the unlicensed after school instructors are less expensive and appear
to be more effective.
The chapter is organized as follows. Section 1.2 describes the KnF trial and trial data.
Section 1.3 presents results assessing the intervention at the school level. Section 1.4
discusses the observations based on the micro data. In both cases I look at the effect of
KnF at i) the end of intervention, ii) at a mean of two months after the end of the
intervention and iii) at a mean of eight months after the end of the intervention. Section
1.5 analyzes the impact on the KnF effectiveness when the program is delivered by a
credentialed elementary teacher or an unlicensed after school program instructors. I
present conclusions in section 1.6.
1.2 Data
The data comes from a trial conducted by CHLA on its program KnF. The KnF
intervention is an after-school program designed for both children and parents which
teaches about healthy nutrition and physical activity behaviors. The intervention is
designed for students in grades 3, 4 and 5. Both the child and a parent are expected to
attend weekly 90-120 minute sessions for six weeks. The KnF child classes cover topics
related to nutrition, physical activity and behavior change, while the parent classes cover
the topics of nutrition and behavior change support. In addition to the class, KnF
includes an organized physical activity for the children three times a week for 45
minutes. At the end of each class, a healthy meal with appropriate portion sizes is
served. Appendix A presents more detail on the KnF program.
10
My analysis uses data from a trial CHLA conducted on KnF. CHLA received funding
from the Johnson & Johnson Company to evaluate KnF using four pairs of California
schools. Two school pairs were selected from Northern California and two from
Southern California. Schools were invited to participate from within regions that had
predominately minority student populations, were able to devote 90 minutes per week for
the KNF classes, were willing to designate staff to be trained and teach KNF, and had
space for the physical activity component of KNF. The first school that enrolled in
each pair was assigned to the intervention, except in one pair where the first school
dropped after accepting but prior to enrolling students where a third school was recruited
- this school was assigned to be a control. Students enrolled in KnF and the school
administrative staff committed to providing instructors for KnF in a process where they
were blinded to school treatment assignment of intervention or control. Summarizing,
schools were paired on the basis of geographic and demographic information, and within
pairs one school was selected sequentially and approximately randomly to receive the
KnF intervention while the other school served as the control. Students, along with one
of their parents were asked to enroll in the study prior to the treatment assignment to the
schools. Students at all eight schools received initial information and health screening.
3
.
In total 323 students took part in the study. 160 students received the intervention, while
the remaining 163 were in control schools. The primary outcome of interest is the
change in the children’s BMI-z score. A child’s BMI-z score is calculated from the
3
It is relevant to note that the control group, like the intervention group, is a sample from the
population of motivated children and families, much in the same way that manpower training
programs are evaluated among those who volunteer for training. A child and parent in a treatment
school could decide not to participate, so we are measuring intent to treat. In addition, the
children and families understood that if their school was not chosen to receive the KnF
intervention during the trial, they would receive the intervention the following year.
11
Centers for Disease Control standard growth charts. It maps the accepted age and
gender adjusted child height and weight into the standard normal distribution. Since
children grow while aging, the BMI-z score is the accepted measure to compare relative
body mass. A child’s height and weight and calculated BMI-z score were measured at
four distinct times:
1. At enrollment;
2. At the conclusion of the KnF intervention or, for controls, six weeks after
enrollment; which is denoted as the 6 week period.
3. Three months after completion of the intervention which corresponded to the end
of the school year or semester, which I denote as the post measure.
4. Six to eight months after completion of the intervention which corresponded to
the beginning of the subsequent school year, which is denoted as the follow up
measure.
The study collected a very comprehensive dataset relating to measurements of the child
and parent over the intervention and the follow-up period. Demographic variables were
also collected at baseline but some of these data are missing for a number of children-
parent pairs. According to the CHLA staff, the missing data, including missing
information on the dependent variables, can be considered random. Therefore I make a
simplifying assumption and in all cases use only subjects with complete data on the
relevant fields. Thus for purposes of this analysis, I reduce the covariates to seven
individual characteristics with two specific to the child variables – child gender and age –
and five to the parent – parent BMI, parent age, parent working status, whether the
parent completed high school (or equivalent), and whether the parent was married at
time of the intervention. Note, that greater than 82% of the students are Hispanic so I
12
do not attempt to estimate treatment effects for Hispanics and Non-Hispanics. Table 1.1
presents the summary population data, including enrollment BMI-z score.
Table 1.1: Summary Statistics of the Entire Sample Population
Variable Obs Mean Std. Dev. Min Max
Child BMI Z Score 323 1.046 1.051 -1.93 2.88
Percent Male 323 0.455 0.499 0 1
Child Age 322 9.878 1.114 7.6 13.7
Parent BMI 315 31.443 6.254 15.5 56.4
Parent Age 319 35.154 6.466 25 56
Parent Working 321 0.421 0.494 0 1
Parent Completed High School 315 0.375 0.485 0 1
Parent Married 315 0.714 0.452 0 1
Given that the intervention was assigned approximately randomly at the school level, I
compare paired school populations to test whether they are comparable on the observed
characteristics. Table 1.2A presents the mean values for the entire school populations
and Table 1.2B below presents the values when the sample is reduced to overweight
and obese children. Table 1.2C presents the values for children between the 70
th
and
85
th
percentile of enrollment BMI. I test the equivalence of the means for the all
covariates between the pairs. The tables show where the equivalence is rejected.
13
Table 1.2A: Child Enrollment Characteristics by Paired Schools & Test of Equivalence
for All Students
KnF Control KnF Control KnF Control KnF Control
n 33 42 37 26 41 42 43 45
Child BMI Z Score
1.8492
(0.134)
1.8424
(0.087)
1.9485
(0.121)
1.9582
(0.103)
2.1068
(0.101)
1.6986
(0.097)
1.7480
(0.091)
1.7811
(0.093)
Percent Male
0.3571
(0.132)
0.4333
(0.092)
0.5000
(0.114)
0.7058
(0.113)
0.6000
(0.100)
0.3636
(0.104)
0.5769
(0.098)
0.4117
(0.123)
Child Age
9.8571
(0.230)
9.4300
(0.145)
9.8100
(0.178)
10.247
(0.274)
9.4960
(0.230)
9.6636
(0.195)
10.530
(0.288)
9.8588
(0.218)
Parent BMI
33.800
(1.413)
33.075
(1.085)
34.483
(1.888)
31.147
(1.135)
31.754
(1.000)
31.385
(1.312)
31.461
(1.259)
35.241
(1.523)
Parent Age
33.428
(1.759)
33.266
(1.308)
36.050
(0.927)
35.294
(1.462)
34.227
(1.287)
35.227
(1.559)
35.769
(1.332)
34.235
(1.041)
Parent Working
0.5000
(0.138)
0.3000
(0.085)
0.2500
(0.099)
0.6470
(0.119)
0.1666
(0.077)
0.3181
(0.101)
0.4000
(0.100)
0.7647
(0.106)
Parent Completed
High School
0.4285
(0.137)
0.1666
(0.069)
0.2500
(0.099)
0.4285
(0.137)
0.3750
(0.100)
0.3181
(0.101)
0.2400
(0.087)
0.6470
(0.119)
Parent Married
0.6428
(0.132)
0.7500
(0.083)
0.8000
(0.155)
0.6250
(0.125)
0.8750
(0.068)
0.5909
(0.107)
0.8400
(0.074)
0.7647
(0.106)
Std Errors in Parenthesis
Failed T-Test of Equivalence iin Pairs - * Significant 90% **Significant 95% ***Significant 99%
Pair 1 Pair 2 Pair 3 Pair 4
I find that for pair one the means of enrollment BMI-z score are not the same. However,
the equivalence is rejected only at the 10% level. All other pairs show similar enrollment
BMI-z scores. In one pair the percentage of male students is rejected as equivalent. I
cannot reject the children’s age are the same between any pair. For several pairs with
respect to four of the parent characteristics, I reject that the values are the same. In
general, the analysis shows that the pairs are not perfectly equivalent along the
observed characteristics.
The paired matching is better when considering the subset of children who are
overweight or obese, which is the subsample I ultimately analyze. The focus of the
CHLA staff is whether KnF helps those subjects classified as overweight and obese
reduce their BMI-z score. Standard definitions recognize children with a BMI-z score
above the 85
th
percentile as overweight or obese (James et al. 2001). As discussed, the
14
percentiles are based on accepted and standard height to weight charts and not on the
actual current population averages. In the study sample, 170 of the 323 subjects
(approximately 52.6%) had BMI-z scores that placed them above the 85
th
percentile.
The stated goal of the intervention was to reduce the fraction of children with such a
BMI-z score. Table 1.2B presents the schools’ population characteristics considering
only those children whose enrollment BMI was above the 85
th
percentile.
Table 1.2B: Child Enrollment Characteristics by Paired Schools & Test of Equivalence
for Students with Enrollment BMI >85
th
Percentile
KnF Control KnF Control KnF Control KnF Control
n 13 28 18 14 22 21 25 17
Child Enrollment BMI
Z Score
1.8492
(0.134)
1.8424
(0.087)
1.9485
(0.121)
1.9582
(0.103)
2.1068***
(0.101)
1.6986***
(0.097)
1.7480
(0.091)
1.7811
(0.093)
Percent Male
0.3571
(0.132)
0.4333
(0.092)
0.5000
(0.114)
0.7058
(0.113)
0.6000
(0.100)
0.3636
(0.104)
0.5769
(0.098)
0.4117
(0.123)
Child Age
9.8571
(0.230)
9.4300
(0.145)
9.8100
(0.178)
10.247
(0.274)
9.4960
(0.230)
9.6636
(0.195)
10.530*
(0.288)
9.8588*
(0.218)
Parent BMI
33.800
(1.413)
33.075
(1.085)
34.483
(1.888)
31.147
(1.135)
31.754
(1.000)
31.385
(1.312)
31.461*
(1.259)
35.241*
(1.523)
Parent Age
33.428
(1.759)
33.266
(1.308)
36.050
(0.927)
35.294
(1.462)
34.227
(1.287)
35.227
(1.559)
35.769
(1.332)
34.235
(1.041)
Parent Working
0.5000
(0.138)
0.3000
(0.085)
0.2500**
(0.099)
0.6470**
(0.119)
0.1666
(0.077)
0.3181
(0.101)
0.4000**
(0.100)
0.7647**
(0.106)
Parent Completed
High School
0.4285*
(0.137)
0.1666*
(0.069)
0.2500
(0.099)
0.4285
(0.137)
0.3750
(0.100)
0.3181
(0.101)
0.2400***
(0.087)
0.6470***
(0.119)
Parent Married
0.6428
(0.132)
0.7500
(0.083)
0.8000
(0.155)
0.6250
(0.125)
0.8750**
(0.068)
0.5909**
(0.107)
0.8400
(0.074)
0.7647
(0.106)
Std Errors in Parenthesis
Failed T-Test of Equivalence iin Pairs - * Significant 90% **Significant 95% ***Significant 99%
Pair 1 Pair 2 Pair 3 Pair 4
15
Table 1.2C: Child Enrollment Characteristics by Paired Schools & Test of Equivalence
for Students with Enrollment BMI Percentile between 70
th
and 85th
KnF Control KnF Control KnF Control KnF Control
n 7 8 10 5 9 7 5 8
Child Enrollment
BMI Z Score
0.8085
(0.062)
0.8012
(0.066)
0.733**
(0.044)
0.912**
(0.054)
0.8122***
(0.036)
0.6485***
(0.034)
0.826
(0.055)
0.8787
(0.073)
Percent Male
0.4285
(0.202)
0.125
(0.125)
0.4
(0.163)
0.2
(0.2)
0.5555
(0.175)
0.2857
(0.184)
0.6
(0.244)
0.875
(0.125)
Child Age
9.6857
(0.465)
9.5125
(0.297)
10.07
(0.244)
10.24
(0.282)
10.511
(0.398)
10.357
(0.414)
10.86
(0.400)
10.062
(0.521)
Parent BMI
29.2
(0.428)
30.857
(2.906)
28.41
(1.471)
25.98
(1.591)
31.555
(2.303)
31.928
(2.709)
29.86
(2.203)
35.175
(3.301)
Parent Age
33.285
(2.243)
34.375
(2.698)
33.4
(1.462)
36.2
(2.634)
38.888**
(2.306)
32.571**
(2.091)
34*
(1.264)
39.5*
(2.878)
Parent Working
0
(0)
0.25
(0.163)
0.4
(0.163)
0.8
(0.2)
0.5555
(0.175)
0.4285
(0.202)
0.4
(0.244)
0.75
(0.163)
Parent Completed
High School
0**
(0)
0.375**
(0.182)
0.5
(0.166)
0.5
(0.288)
0.3333
(0.166)
0.5
(0.223)
0.4
(0.244)
0.75
(0.163)
Parent Married
0.4285*
(0.202)
0.875*
(0.125)
0.8888
(0.309)
0.4
(0.244)
0.75
(0.163)
0.6666
(0.210)
1
(0)
1
(0)
Std Errors in Parenthesis
Failed T-Test of Equivalence iin Pairs - * Significant 90% **Significant 95% ***Significant 99%
Pair 1 Pair 2 Pair 3 Pair 4
When examining only those children within the 70
th
to 85
th
or those above the 85
th
percentile enrollment BMI, I reject fewer null hypothesis tests of equivalence between
the pairs’ means than for the whole population. However, the data rejects the null
hypothesis in several pairs of equivalence of the enrollment BMI-z scores in pair 3 at the
1% level for both the 70-85
th
and the above 85the percentile groups, and at the 5% level
for pair 2 in the population of children 70
th
-85
th
percentile. Additionally, three of the eight
characteristics are rejected as equivalent for the above 85
th
percentile group for pair 4
and pair 1 shows differences for parent characteristics of completed high school and
percent married. While these results may raise concerns about the appropriateness of
the analysis, I proceed under the assumption that the pairs represent appropriate
matching. This assumption along with the assumption of quasi-random intervention
assignment allows me to identify the intervention effects.
16
As a final check on the pairing, I examine the sample distribution of children above the
70
th
percentile BMI-z score. I want to confirm that the observed transition probabilities or
population changes are not simply a feature of the data. This would occur if certain
schools exhibit a substantial mass of children close to the percentile thresholds of
interest as compared to other schools. While I do not formally test the equivalence of
these distributions given the small number of students in the subsamples, I conclude that
the distribution of students is similar among the pairs. The figures in Figure 1.1 show the
relative sample frequencies for each school in each pair.
17
Figure 1.1: Sample Distribution >70
th
percentile BMI-z Scores for Paired Schools
D is tribution O verweight Students Schools 1 and 4
0
2
4
6
8
10
12
14
16
18
70-
72.5
72.5-
75
75-
77.5
77.5-
80
80-
82.5
82.5-
85
85-
87.5
87.5-
90
90-
92.5
92.5-
95
95-
97.5
97.5-
100
P ercentile R ange
Count of Students
S chool 1
S chool 4
D is tribution Overweight Students Schools 2 and 3
0
1
2
3
4
5
6
7
8
9
10
70-
72.5
72.5-
75
75-
77.5
77.5-
80
80-
82.5
82.5-
85
85-
87.5
87.5-
90
90-
92.5
92.5-
95
95-
97.5
97.5-
100
Percentile R ange
Count of S tudents
S chool 2
S chool 3
D is tribution of Overweight Students Schools 5 and 6
0
2
4
6
8
10
12
14
16
18
70-
72.5
72.5-
75
75-
77.5
77.5-
80
80-
82.5
82.5-
85
85-
87.5
87.5-
90
90-
92.5
92.5-
95
95-
97.5
97.5-
100
Percentile R ange
Count of Students
School 5
School 6
18
Figure 1.1: Continued
D is tribution Overweight Students Schools 7 and 8
0
1
2
3
4
5
6
7
8
9
10
70-
72.5
72.5-
75
75-
77.5
77.5-
80
80-
82.5
82.5-
85
85-
87.5
87.5-
90
90-
92.5
92.5-
95
95-
97.5
97.5-
100
Percentile R anges
Count of Students
School 7
School 8
As the measure of KnF effectiveness is based on comparisons of pre-intervention BMI z
scores and after intervention measures at different points, sample attrition between
measures represents an additional potential source of bias. Attrition is of particular
concern if it is correlated with enrollment BMI z scores. Examining the 155 children with
initial BMI above the 85
th
percentile and complete data on the seven characteristics,
there are 6 week measures for 72% of the children (111 children), post measures on
71% (110 children), and the follow up measure on 59% (92 children). This is a relatively
high attrition rate and cause for some concern. Similar to all missing data, the Center
staff believes that attrition occurred randomly and therefore would not be correlated with
outcomes. I cannot check the difference in BMI z score change for those children
missing after intervention measures, but I do check to see if those with after intervention
measures differ significantly from those without after intervention measures by their
enrollment BMI z score. Table 1.3 shows the mean enrollment BMI z scores for both
intervention and control groups for those with after intervention measures and those
missing the after intervention measures.
19
Table 1.3: Comparison of Observations and Missing Observations
Observations Missing Observations Missing
6 Week Measure
1.887
(0.070)
n=55
1.931
(0.107)
n=21
1.771
(0.063)
n=56
1.920
(0.073)
n=23
Post Measure
1.893
(0.064)
n=62
1.927
(0.147)
n=14
1.742*
(0.065)
n=48
1.927*
(0.075)
n=31
Follow Up Measure
1.930
(0.072)
n=51
1.836
(0.098)
n=25
1.790
(0.069)
n=41
1.840
(0.073)
n=38
Std Errors in Parenthesis
Failed T-Test of Equivalence iin Pairs - * Significant 90% **Significant 95% ***Significant 99%
Intervention Control
As Table 1.3 illustrates for all but one of the period observations, the enrollment BMI z
score of children with after interventions measures is not statistically different than for
children with missing observations. This holds true both for the children receiving the
intervention and those in control. The only case where I reject that the enrollment BMI
was the same for observed and missing observations was for the post measure among
the controls. Here the equivalence is rejected at only the 10% level. Thus, I conclude
that sample attrition is not a relevant source of potential bias and do not control for
attrition in my analysis.
1.3 Measuring the Intervention Effect at the School Level
I first assess the impact of the KnF intervention by examining the differences in BMI-z
score change between the control and the intervention schools. Given that the KnF
evaluation was designed with approximately random assignment at the school level, the
natural assessment is to review changes in the mean BMI-z scores between the pairs.
This assessment requires the weakest assumption. However, due to the fact that there
20
are only four observations at the pair level, I cannot control for child/parent
characteristics in the school level analysis. Table 1.4 shows the BMI-z score change for
each of the schools with Table 1.4A the change for the entire population, 1.4B for
children 70-85
th
percentile and 1.4C for children greater than 85
th
percentile.
Table 1.4A: Average Change in BMI-z score by School and Enrollment BMI Percentile
for All Students
School # Obs mean Std Error # Obs mean Std Error # Obs mean Std Error
1:KnF 21 -0.1452 0.2945 25 -0.154 0.311328 14 -0.0357 0.2008
4: Control 32 0.0494 0.1170 27 -0.063333 0.155638 21 0.0205 0.2210
2: KnF 31 0.0235 0.2760 29 -0.061035 0.251713 28 -0.0411 0.2414
3: Control 24 0.0488 0.1521 15 -0.072 0.14915 9 0.0733 0.2997
5: KnF 30 0.0180 0.1470 37 -0.107838 0.235645 29 -0.0472 0.3675
6: Control 24 0.0175 0.1793 29 -0.134483 0.252722 29 -0.0155 0.3341
7: KnF 33 -0.0882 0.1368 36 -0.2175 0.26514 28 -0.0807 0.2851
8: Control 43 0.0305 0.1991 44 -0.005 0.299329 38 0.0066 0.3899
Enroll to Follow Up BMI z Score Enroll to 6 Week BMI z Score Enroll to Post BMI z Score
Table 1.4B: Average Change in BMI-z score by School and Enrollment BMI Percentile
for Students with Enrollment BMI between 70
th
and 85
th
Percentile
School # Obs mean Std Error # Obs mean Std Error # Obs mean Std Error
1:KnF 21 -0.1452 0.2945 25 -0.154 0.311328 14 -0.0357 0.2008
4: Control 32 0.0494 0.1170 27 -0.063333 0.155638 21 0.0205 0.2210
2: KnF 31 0.0235 0.2760 29 -0.061035 0.251713 28 -0.0411 0.2414
3: Control 24 0.0488 0.1521 15 -0.072 0.14915 9 0.0733 0.2997
5: KnF 30 0.0180 0.1470 37 -0.107838 0.235645 29 -0.0472 0.3675
6: Control 24 0.0175 0.1793 29 -0.134483 0.252722 29 -0.0155 0.3341
7: KnF 33 -0.0882 0.1368 36 -0.2175 0.26514 28 -0.0807 0.2851
8: Control 43 0.0305 0.1991 44 -0.005 0.299329 38 0.0066 0.3899
Enroll to Follow Up BMI z Score Enroll to 6 Week BMI z Score Enroll to Post BMI z Score
Table 1.4C: Average Change in BMI-z score by School and Enrollment BMI Percentile
for Students with Enrollment BMI above the 85
th
Percentile
School # Obs mean Std Error # Obs mean Std Error # Obs mean Std Error
1:KnF 7 -0.2114 0.1060 10 -0.208 0.076388 7 -0.0500 0.0644
4: Control 20 0.0280 0.0250 17 0.01 0.031178 12 0.0617 0.0517
2: KnF 15 -0.0027 0.0294 18 -0.052222 0.037273 16 -0.0206 0.0390
3: Control 7 -0.2114 0.1060 10 -0.208 0.076388 7 -0.0500 0.0644
5: KnF 19 -0.0089 0.0179 21 -0.069524 0.028136 16 -0.0369 0.0572
6: Control 12 -0.0258 0.0374 12 -0.1725 0.050649 13 -0.1054 0.0698
7: KnF 20 -0.0545 0.0248 21 -0.092381 0.02603 18 -0.0078 0.0407
8: Control 16 -0.0163 0.0150 17 -0.076471 0.033099 14 -0.1343 0.0775
Enroll to Post BMI z Score Enroll to Follow Up BMI z Score Enroll to 6 Week BMI z Score
21
Table 1.5 shows the difference in the difference between the intervention schools and
control schools. The results are shown by school pair.
Table 1.5: Difference in Differences of BMI-z Score Intervention to Control Schools
Percentile All 70-85th 85-100th All 70-85th 85-100th All 70-85th 85-100th
Pair 1 -0.1946 -0.2667 -0.2394 -0.0907 -0.0333 -0.2180 -0.0562 -0.0755 -0.1117
Pair 2 -0.0252 -0.0356 0.2088 0.0110 -0.1883 0.1558 -0.1144 -0.1150 0.0294
Pair 3 0.0005 -0.2413 0.0169 0.0266 -0.0388 0.1030 -0.0317 -0.1642 0.0685
Pair 4 -0.1186 -0.1083 -0.0383 -0.2125 -0.3858 -0.0159 -0.0873 -0.0546 0.1265
Difference in Enroll to 6 Week
BMI Z Score Change
Difference in Enroll to Post BMI
Z Score Change
Difference in Enroll to Follow
Up BMI Z Score Change
Assessing the intervention’s impact at the aggregate school level, I assume that
differences in outcomes can be explained by the intervention, pair effects and school
effects. It is illustrative to view this assumption within a regression framework. I assume
that there is a school fixed effect in levels and a pair effect which differs by period.
Formally, let i index the school, j index the school pair, and t represent the period of
observation, where t=0 represents the enrollment measure; t=1 the 6 week measure; t=2
the follow up measure; and t=3 the post measure. Thus the model in levels, with i, i’
belonging to the same pair is given by:
t i t jt i t i
it t jt j t i it
D Y
D Int Y
' ' '
ε γ α
ε γ δ α
+ + =
+ + + =
(1)
Differencing Y
it
-Y
i0
t=1,2,3 removes the school fixed effect. Taking the difference in
difference between paired schools for each t removes the pair effect. This yields the
standard difference in difference estimator:
) ( ) (
) ( ) (
0 ' ' 0 '
' 0 ' ' 0
i t i i it to ii
to ii j t i t i i it
Int Y Y Y Y
ε ε ε ε μ
μ δ
− − − =
+ = − − −
(2)
22
Equation (2) can be estimated with Ordinary Least Squares. Analysis of (2) estimates
the impact of KnF by changes in the population means between intervention and control
schools between different observation periods. The results of this regression analysis
are in Table 1.6.
Table 1.6: Regression Analysis Difference in Intervention and Control BMI-z Score
Change
Percentile All 70-85th 85-100th All 70-85th 85-100th All 70-85th 85-100th
Intervention
-0.084**
(0.042)
-0.162***
(0.046)
-0.013
(0.071)
-0.066
(0.042)
-0.161**
(0.077)
0.0062
(0.060)
-0.072***
(0.021)
-0.102***
(0.034)
0.0281
(0.044)
Standard Errors in Parenthesis
* Significant 90% ** Significant 95% *** Significant 99%
Difference in Enroll to 6 Week
BMI z Score Change
Difference in Enroll to Post BMI
z Score Change
Difference in Enroll to Follow
Up BMI z Score Change
The results of the analysis point to KnF’s differential impact by enrollment BMI
percentile. KnF significantly impacts children’s BMI-z score for the entire population and
for children in the 70-85
th
initial percentile. Within these populations, the intervention
works to decrease BMI-z score. The results are statistically significant for the 70-85
th
percentile at all measures after the intervention and for the entire population at the
immediate post intervention and the follow up. However, the intervention does not seem
to significantly impact children with a BMI percentile greater than the 85
th
.
To illustrate the magnitude of the impact of KnF for children between the 70
th
and 85
th
percentile, I calculate the implied difference in a child’s weight at enrollment and six
weeks after that would be attributable to the intervention. I calculate the change for the
median child whose enrollment BMI is between the 70
th
and 85 percentile. The mean
child in this BMI range was 10.1 years old and had BMI-z score of 0.93, approximately
the 82
nd
percentile, at enrollment. She weighed 34.55 kg and was 135.76 cm tall. Over
23
the six week course, she aged 0.1 years and grew 0.94 cm. If her BMI-z score had
stayed the same over the 6 week course, her weight would have been 35.15kg at the 6
week measure (gain of 0.6kg). The estimated intervention impact for this child between
enrollment and 6 week’s would be a decrease in BMI-z score of 0.162. This implies that
the child’s post intervention BMI-z score would be 0.77. A 10.2 years old, 135.76 cm tall
girl with a BMI-z score of 0.77 would weigh 34.2kg (loss of 0.35kg). Thus for the median
child, the KnF intervention works to reduce the child’s weight over the 6 week period by
0.95kg (35.15-34.2) and reduce her BMI percentile from the 82
nd
to the 78
th
. Note that
the child would still be expected to gain weight but that the weight gain is decreased by
the intervention.
At the aggregate school level, I examine the sample transition probabilities and the
impact of KnF on transitions to a BMI percentile above or below the 85
th
percentile. I
categorize three states for children based on their BMI percentile – less than 70
th
percentile, 70-85
th
percentile, greater than 85
th
percentile. Table 1.7 shows the percent
of children whose enrollment BMI put them in the categories 70-85
th
percentile and
greater than 85
th
that transition into each of the categories at the three intervention
measures. The transition probabilities are shown separately for children in the control
schools and children in the intervention schools.
24
Table 1.7: Intervention and Control BMI Percentile Transition Probabilities
Transition Probabilities 6 Week Outcomes <70th %ile 70-85th %ile >85th %ile
2 13 6
9.5% 61.9% 28.6%
5 16 2
21.7% 69.6% 8.7%
0 4 58
0.0% 6.5% 93.5%
0 4 57
0.0% 6.6% 93.4%
Transition Probabilities Post Outcomes <70th %ile 70-85th %ile >85th %ile
6 9 5
30.0% 45.0% 25.0%
10 15 0
40.0% 60.0% 0.0%
0 4 50
0.0% 7.4% 92.6%
0 7 63
0.0% 10.0% 90.0%
Transition Probabilities Follow Up Outcomes <70th %ile 70-85th %ile >85th %ile
5 6 6
29.4% 35.3% 35.3%
8 11 3
36.4% 50.0% 13.6%
0 4 40
0.0% 9.1% 90.9%
0 3 54
0.0% 5.3% 94.7%
Control 70th to 85th %ile Initial BMI
(n=28)
Intervention 70th to 85th %ile Initial BMI
(n=31)
Control >85th %ile Initial BMI
(n=85)
Intervention >85th %ile Initial BMI
(n=85)
Control 70th to 85th %ile Initial BMI
(n=28)
Intervention 70th to 85th %ile Initial BMI
(n=31)
Control >85th %ile Initial BMI
(n=85)
Intervention >85th %ile Initial BMI
(n=85)
Control 70th to 85th %ile Initial BMI
(n=28)
Intervention 70th to 85th %ile Initial BMI
(n=31)
Control >85th %ile Initial BMI
(n=85)
Intervention >85th %ile Initial BMI
(n=85)
As Table 1.7 illustrates, KnF did not have a dramatic impact on the transition
probabilities of children who started with a BMI percentile greater than the 85
th
percentile. The percent of children who transitioned from greater than the 85
th
percentile
to the lower percentile groupings is statistically the same between the intervention and
control groups. However, it does appear that KnF had a substantial impact on the
transition probabilities of children who began in the 70-85
th
percentile group. At each of
the after intervention measures, dramatically fewer children who began in the 70-85
th
percentile group and received the intervention moved into the greater than 85
th
percentile group. At six weeks the difference is between 28.6% in the control group and
25
8.7% in the intervention; in the post measure the difference is between 25% control and
0% intervention; at follow up the difference is 35.3% control to 13.6% intervention.
Thus, when examining KnF at the school level from both regression and transition
probability framework, I find that KnF has a significant impact on children whose initial
BMI percentiles are between the 70
th
and 85
th
percentile. The intervention does not
appear to significantly impact children who begin with a BMI percentile greater than the
85
th
percentile. This implies that the intervention with respect to the population
prevalence of obesity, works primarily by reducing the number of children whose BMI
percentile would increase to cross to above the 85
th
percentile and serves to prevent
normal weight children from becoming overweight or obese.
1.4 Individual Level Model and Results
I examine the impact of KnF at the individual child level. Analysis at the individual level
enables me to control for differences in the observables among the sample students at
different schools. I control for arbitrary heteroskedacity and correlation across students
at the same school by clustering at the school level. Lastly, by examining the
intervention at the micro-level, I am able to examine the impact of using after-school
instructors as KnF instructors versus using credentialed teachers as instructors.
Since the study design was based on quasi randomization mechanism among paired
schools, several assumptions are required to act as-if the randomization was at the child
level. I allow for a fixed effect for each child, but assume that paired schools have
identical linear and higher order trends. In addition, to preserve degrees of freedom, I
make the standard assumption that the intervention effect does not depend on any of the
26
individual child/parent characteristics. I model a child’s BMI-z score at a given time t
it
Y as follows:
,
it t jt i i t t i it
D Int X Y ε γ α δ β + + + + = (3)
Where in (3), X
i
is a vector of child and parent characteristics. Inclusion of X
i
is
motivated by the findings in Table 1.2 that even after attempting to match schools as
similarly as possible, observable and most likely unobservable differences remains. Int
i
is a dummy variable=1 if the child received the KnF intervention,
t
D is a dummy for
each of four periods (i.e. t=1,2,3,4), and ε
it
is an independent across child and time error.
Further, α
i
is a random effect; γ
jt
is a time-varying school pair trend coefficient; and
t
β is
a time varying vector of coefficients. I normalize γ
j0
and
0
β to zero. For each t=1,2,3, I
subtract off
0 i
Y to obtain
io it t jt i t t i i it
D Int B X Y Y ε ε γ δ − + + + = −
0
(4)
Since I have assumed no time-varying individual or school effects correlated with
i
Int ,
(i.e quasi-random assignment), OLS will yield consistent estimates for equation (4).
Table 1.8 shows the regression estimates of (4) .
27
Table 1.8: Random Effects Regression Results at Individual Child Level (by BMI Enroll
Percentile)
Percentile 70-85 85-100 70-85 85-100 70-85 85-100
n 42 111 41 110 36 92
Intervention
-0.161***
(0.058)
-0.068**
(0.028)
-0.279***
(0.096)
-0.057
(0.047)
-0.201*
(0.110)
0.0254
(0.047)
Male
-0.0055
(0.059)
-0.010
(0.024)
0.0547
(0.088)
-0.045
(0.030)
-0.076
(0.109)
0.0661
(0.045)
Child Age
-0.035
(0.029)
-0.0027
(0.011)
-0.050
(0.041)
-0.037***
(0.014)
-0.029
(0.047)
-0.018
(0.021)
Parent BMI
0.0042
(0.004)
-0.0020
(0.002)
0.0046
(0.006)
0.0018
(0.002)
0.0192***
(0.007)
0.0014
(0.003)
Parent Age
0.0028
(0.004)
-0.00001
(0.002)
-0.0013
(0.007)
0.0058**
(0.002)
0.0115
(0.008)
0.0027
(0.004)
Parent Working
-0.024
(0.076)
-0.019
(0.027)
-0.080
(0.119)
-0.028
(0.032)
-0.056
(0.127)
0.0083
(0.048)
Parent Completed
High School
-0.052
(0.067)
0.0303
(0.027)
-0.060
(0.103)
0.0351
(0.032)
0.0118
(0.116)
-0.0008
(0.048)
Parent Married
0.0812
(0.067)
-0.0017
(0.029)
0.0002
(0.084)
0.0311
(0.031)
-0.168*
(0.092)
0.0177
(0.046)
Std Errors in Parenthesis, Random Effects Grouping on School
Coefficients not reported: Intercept, Pair Dummies
* Significant 90% ** Significant 95% *** Significant 99%
Enroll to 6 Week BMI z
Score Change
Enroll to Post BMI z
Score Change
Enroll to Follow Up
BMI z Score Change
The regression results at the individual level are quite similar to the results at the school
level. The impact of the intervention at six weeks is quite similar with the two separate
estimates for both populations, supporting the validity of the assumptions underlying the
individual model. For the 70-85
th
percentile group, the magnitude of the intervention
impact is greater at the individual estimates for the post and follow up measures;
however, the result is not significant at the follow up. Note also that the impact of the
intervention is statistically significant for the above 85
th
percentile group in the 6 week
measure. In most cases the individual characteristics do not have a significant impact
supporting the contention of quasi-random assignment. However, in the case of the
children with enrollment percentiles in the 85-100
th
range, it appears that older children
28
show significant reduction in BMI-z score at both the post and follow up. In both these
cases, the magnitude of the impact is small. There are also several specifications where
the characteristics related to the parents impact significantly. In summary, both
individual and school level analysis point to the significant impact of KnF on reducing
BMI-z score in the short run and with some persistence for children whose initial BMI is
in the 70
th
-85
th
percentile. The individual analysis reflects that KnF also impacts in the
short run the BMI-z score of children with enrollment percentiles above the 85
th
.
However this impact is not shown in the aggregate analysis.
While KnF seems effective at helping overweight and at-risk for overweight children
reduce BMI-z score, the CHLA staff also wanted to ensure that the program does not
promote weight loss among those students who are already underweight. As Doak, et al
(2006) discussed, there is concern that population-delivered interventions designed to
help children reduce body weight may have that effect on those children with low BMI
causing potential health risks associated with underweight. This effect may work, for
example, if the intervention were to create a stigma against high weight which impacted
underweight as well as overweight and obese.
Within the complete sample, only 15 students had enrollment BMI-z scores below the
15
th
percentile. Eight of these received KnF while seven were in the control group.
Table 1.9 shows the regression estimates of the effect of the KnF intervention on those
children with initially low BMI scores. Since the sample size is small, I include only child
age, parent BMI and pair dummies in the regression as these were the covariates found
to be significant in at least two of the specifications for the overweight children.
29
Table 1.9: Regression Results for Underweight Sample Population (<15 Percentile)
Enroll to 6 Week BMI z
Score Change
Enroll to Post BMI z
Score Change
Enroll to Follow Up BMI z
Score Change
n 11 10 7
Intervention
0.1757
(0.172)
0.0205
(0.278)
-0.834***
(0.133)
Child Age
-0.204**
(0.084)
-0.255**
(0.128)
-0.144*
(0.087)
Parent BMI
0.0214
(0.013)
0.0302
(0.020)
0.0419***
(0.009)
Std Errors in Parenthesis, Random Effects Grouping on School
Coefficients not reported: Intercept
* Significant 90% ** Significant 95% *** Significant 99%
The results support the contention that KnF does not cause underweight children to
further decrease their BMI-z score. The intervention is found to significantly increase
BMI-z score for underweight children at the immediate after intervention measure. As
KnF is designed to teach healthy nutrition and physical activity habits, it seems
appropriate that KnF should impact low weight children to improve their BMI as shown
by the analysis.
1.5 Impact of KnF Instructor Type
Modeling the intervention assignment as-if random at the individual level, conditional on
the pair effects, allows me to assess the impact of the KnF instructor type on the
effectiveness of the intervention. The KnF intervention was delivered in schools with
school staff serving as the program’s instructors. In the study, subjects received
instruction from teachers with two different levels of teacher schooling. Two of the
schools had instructors who were after school instructors and two had credentialed
teachers as instructors.
30
Examining the impact of instructor type on KnF outcomes is important because instructor
costs represent a very significant portion of the total KnF costs. After receiving initial
training from CHLA staff, each instructor was expected to deliver the complete KnF
program with only administrative support from CHLA. The KnF curriculum called for
each instructor to spend approximately 4.5 hours per week in class preparation and
execution. Including the 5 hour initial training, each instructor spent an estimated 32
hours per KnF course over the six weeks. These staff members (and the school) were
compensated for the time they spent to deliver the program. The program cost varied
dramatically between those schools with after school staff instructors and those with
credentialed teachers as instructors. After school instructors were reimbursed at a rate
of $15/hour, and credentialed teachers at a rate of $40/hour. Total non-instructor costs
for a KnF program were approximately $2,500 per course. Thus, using credentialed
teachers represented an increase in total course costs of nearly 27%.
With an understanding of instructor type cost and benefit, KnF program staff can make
an informed policy decision about the benefits of employing more highly trained and
more costly instructors. Theoretically, higher skilled instructors should illicit better
results. This implies that students of credentialed teachers should decrease their BMI-z
scores by a greater amount than students of KnF instructors with less training, i.e. the
after school instructors. This increased benefit in the form of greater BMI-z score
reduction could then be compared with the additional cost to determine if future
implementations warrant the more highly trained and more costly instructors.
In the study, instructor type was determined by school. Both of the Southern California
schools provided instruction from after school instructors, while the Northern California
31
schools provided instruction from credentialed teachers. Although schools and therefore
students were not randomized to the type of KnF instructor, extending the assumptions
that allowed me to treat intervention assignment as random, I can estimate the impact of
instructor type. Under these same assumptions, I am able to treat KnF instructor type as
uncorrelated with the error term in the model represented by (4) and can estimate the
marginal impact of having credentialed teachers as compared to after school staff
instructors. The required assumption is analogous to our existing assumption that there
are no unique school-intervention-instructor type interactions and trends beyond the pair
dummies. This assumption is stronger than the previous assumptions as it implies that
children in Southern California and Northern California would respond to the KnF
instructor type the same after controlling for the trend through the pair effect. I do allow
the intervention to interact with the instructor type by estimating an equation that treats
the interventions as separate. I estimate the following differenced equation:
io it t jt i t i t t i i it
D IntAfter IntCred B X Y Y ε ε γ ϕ ϑ − + + + + = −
0
(5)
Where IntCred is a dummy variable=1 if the child received KnF from a credentialed
teacher and IntAfter is a dummy variable=1 if the child received KnF from an after school
instructor. Table 1.10 presents the results of OLS of (5) with clustered standard errors.
32
Table 1.10: Regression Results Including Instructor Type (by Enrollment BMI Percentile)
Percentile 70-85 85-100 70-85 85-100 70-85 85-100
n 42 111 41 110 36 92
Intervention w/
Credentialed
-0.130
(0.090)
-0.014
(0.033)
-0.329***
(0.122)
0.0508
(0.038)
-0.203
(0.141)
0.0963
(0.063)
Intervention w/ After
School
-0.183**
(0.077)
-0.134***
(0.038)
-0.210
(0.141)
-0.208***
(0.048)
-0.199
(0.170)
-0.077
(0.077)
Male
-0.0023
(0.059)
-0.025
(0.024)
0.0416
(0.090)
-0.066**
(0.029)
-0.077
(0.116)
0.0415
(0.047)
Child Age
-0.039
(0.030)
-0.0072
(0.011)
-0.044
(0.041)
-0.041***
(0.013)
-0.029
(0.051)
-0.022
(0.021)
Parent BMI
0.0049
(0.004)
-0.00098
(0.002)
0.0037
(0.006)
0.0030
(0.002)
0.0192**
(0.007)
0.0028
(0.003)
Parent Age
0.0022
(0.004)
-0.00020
(0.002)
-0.00079
(0.007)
0.0059**
(0.002)
0.0116
(0.008)
0.0019
(0.004)
Parent Working
-0.026
(0.076)
-0.018
(0.025)
-0.067
(0.120)
-0.026
(0.031)
-0.057
(0.128)
0.0255
(0.049)
Parent Completed
High School
-0.054
(0.067)
0.0301
(0.027)
-0.065
(0.103)
0.0425
(0.031)
0.0119
(0.117)
0.0052
(0.048)
Parent Married
0.0803
(0.067)
-0.0017
(0.028)
0.0078
(0.085)
0.0468
(0.030)
-0.168*
(0.092)
0.0210
(0.045)
Std Errors in Parenthesis, Random Effects Grouping on School
Coefficients not reported: Intercept
* Significant 90% ** Significant 95% *** Significant 99%
Enroll to 6 Week BMI z
Score Change
Enroll to Post BMI zScore
Change
Enroll to Follow Up BMI z
Score Change
Both instructor types are found to significantly impact children with enrollment BMI
between the 70-85
th
percentile, and these estimates are in line with the estimates of the
other specifications. However, when analyzing the impact for the students whose initial
BMI is above the 85
th
percentile, the results are striking and dramatic. The intervention
with after school instructors is found to significantly decrease BMI-z score. The
intervention with credentialed teachers has either a non-significant impact or actually
significantly increases BMI-z score as shown in the post measure. Thus, the analysis
seems to suggest that more highly trained instructors in the form of credentialed
teachers actually illicit poorer responses from these heavier students as measured by
BMI-z score change.
33
This result is highly surprising given the hypothesis that more highly trained instructors
should be more effective at teaching the material in the KnF intervention. My prior
hypothesis would be that the effect of having a credentialed teacher would be highly
negative. While surprising and certainly warranting of further investigation, these results
are in no way conclusive of instructor impact given the necessary assumptions. Since
the random design of the study was not constructed to estimate the effects of instructor
type, these effects could also be interpreted as a Southern California, Northern
California intervention interaction.
There is a potential health education theory explanation for the results that less trained
after school instructors may have been more effective at eliciting behavior change. The
improved intervention effect with after school instructors may result from their greater
cultural sensitivity in delivering the intervention. The vast majority of the students in the
trial were Hispanic with Spanish spoken at home. The after school instructors were also
Hispanic while the credentialed teachers were not. Therefore, a plausible alternative
explanation for the differential impact is that the after school instructors delivered the
KnF program in a more culturally sensitive way. It is relatively accepted in the health
promotion literature that programs are more effective when they are culturally sensitive
(See for example Yancey et al. 2006 on the impact of cultural sensitivity on physical
activity promotion interventions.)
1.6 Conclusion
This paper assesses the results of a randomized trial at the paired school level of the
KnF intervention. KnF aims to reduce the prevalence of childhood overweight and
obesity. CHLA intends to offer more courses of KnF and is interested in its effectiveness
34
and efficient design. CHLA conducted the evaluation to measure KnF delivered in a
school context, to study the efficacy of the program and to better understand the
effectiveness of particular elements of the intervention. My analysis aims to rigorously
assess KnF efficacy based on a variety of assumptions. Additionally, the Center staff is
interested in assessing the impact of elements of the KnF intervention so that it can
optimize the program design in terms of costs and impacts. My analysis measures the
impact of KnF instructor type on outcomes as instructor costs represent a significant
portion of the entire program costs.
I conclude that the KnF intervention has a significant impact on the BMI-z score of
children whose initial BMI is greater than the 70
th
percentile at enrollment. This impact
persists for those students who are between the 70
th
and 85
th
percentile to both the post
and follow up measures. Additionally, I find that KnF has no negative impact on
students who enter the program underweight. Summarizing, the program is found to
effectively help children at risk for becoming overweight reduce their BMI-z score and
remain at a healthy weight. This effect seems to persist in the population whose initial
BMI is between the 70
th
and 85
th
percentile at least six months post intervention. The
intervention also seems to impact the BMI-z score of children whose enrollment BMI is
above the 85
th
percentile. However this effect is not strong and does not seem to
persist. Importantly, the intervention reduces the population prevalence of overweight
and obesity. The program seems to work on the population level by reducing the
transition of children into the overweight category. Lastly, I conclude that instructors with
formal teaching credentials and thus more training do not appear to improve the
outcomes of the program and in this analysis actually reduce the program’s
effectiveness.
35
Chapter 1 References
Campbell, K and Waters, E and O’Meara, S and Summerbell, C (2001) “Interventions for
preventing obesity in childhood. A systematic review” Obesity Reviews 2:149-
157
Davis, K and Christoffel, K.K. (1994) “Obesity in Preschool and School-age Children,
Treatment Early and Often May Be Best” Archives Pediatric Adolescent Medicine
148:1256-1261
Dehghan, Manshid and Akhtar-Danesh, Noori and Merchant, Anwar (2005) “Childhood
obesity, prevalence and prevention” Nutrition Journal 4:24-32
Doak, C.M. and Visscher, T.L.S. and Renders, C.M. and Seidell, J.C. (2006) “The
prevention of overweight and obesity in children and adolescents: a review of
interventions and programmes” Obesity Review 7: 111-136
Edmundson, Elizabeth and Parcel, Guy and Feldman, Henry and Elder, John and Perry,
Cheryl and Johnson, Carolyn and Williston, B.J. and Stone, Elaine and Yang,
Mingua and Lytle, Leslie and Webber, Larry (1996) “The Effects of the Child and
Adolescent Trial for Cardiovascular Health upon Psychosocial Determinants of
Diet and Physical Activity Behavior” Preventive Medicine 25: 442-454
Flodmark, Carl-Erik and Ohlsson, Torsten and Ryden, Olof and Sveger, Tomas (1993)
“Prevention of Progression to Severe Obesity in a Group of Obese
Schoolchildren Treated with Family Therapy” Pediatrics 91:880-884.
Hedley, Allison and Ogden, Cynthia and Johnson, Clifford and Carroll, Margaret and
Curtin, Lester and Flegal, Katherine (2004) “Prevalence of Overweight and
Obesity Among US Children, Adolescents and Adults, 1999-2002” Journal of the
American Medical Association 291:2847-2850
James, Philip and Leach, Rachel and Kalamara, Eleni and Shayeghi, Maryam (2001)
“Worldwide Obesity Epidemic” Obesity Research 9:228S-233S
Kamath, CC and Vickers, KS and Eherlich, A and McGoern, L and Johnson, J and
Singhal, V and Paulo, R and Hettinger, A and Erwin, PJ and Montori, VM (2008)
“Clinical review: behavioral interventions to prevent childhood obesity. A
systematic review and meta-analysis of randomized trials” Journal of Clinical
Endocrinology & Metabolism 10.1210: 2006-2411
Maffeis, Clauio and Tato, Luciano (2001) “Long-Term Effects of Childhood Obesity on
Morbidity and Mortality” Hormone Research 55: 42-45
Monzavi, Roshanak and Dreimane, Daina and Geffner, Mitchell and Braun, Sharon and
Conrad, Barry and Klier, Mary and Kaufman, Francine R. (2006) “Improvement in
Risk Factors for Metabolic Syndrome and Insulin Resistance in Overweight
Youth Who Are Treated With Lifestyle Intervention” Pediatrics 117: 1111-1118
36
Ogden, Cynthia and Carroll, Margaret and Flegal, Katherine (2008) “High Body Mass
Index for Age Among US Children and Adolescents, 2003-2006” Journal
American Medical Association 299.20: 2401-2405
Valverde, M.A. and Patin, R. V. and Olveira, F. L. C. and Vitolo, M. R. (1998) “Outcomes
of obese children and adolescents enrolled in a multidisciplinary health program”
International Journal of Obesity 22:513-519
Wang, Youfu and Monteiro, Carlos and Popkin, Barry (2002) “Trends of obesity and
underweight in older children and adolescents in the United States, Brazil, China
and Russia” American Journal Clinical Nutrition 75: 971-977
Yancey, Atronette and Ory, Marcia and Davis, Sally (2006) “Dissemination of Physical
Activity Promotion Interventions in Underserved Populations” American Journal
of Preventive Medicine 31(4): 82-91
37
Chapter 2: Conditional Hazard Modeling of Spells of Good and Poor Diabetes
Control in Children with Type 1 Diabetes
2.1 Introduction
This chapter examines child and family characteristics that impact patients’ proper
control of their diabetes. Diabetes is a potentially debilitating disease that can lead to
both short-term and long-term complications and requires constant management by
patients to maintain the disease in proper control. The seminal, NIH funded Diabetes
Complications and Control Trial (“DCCT”) showed that better diabetes control by
patients, in particular better control of blood sugar levels improves long-term outcomes
in people with type 1 diabetes (The Diabetes Control and Complications Trial Research
Group, 1993). Since the DCCT concluded clinicians have been urging their patients to
achieve better control (American Diabetes Association, 2003). This analysis examines
the impact of child/family characteristics on the duration of spells of good and poor
glycemic control. While progress has been made in understanding the individual patient
determinants of being in a state of good or bad diabetes control (see for example
Hanberger et al. 2008 for a recent analysis of the determinants of glycemic control in a
representative sample of Swedish children), little work has been done to model the
determinants of transitions between these states. This analysis estimates the impact of
various characteristics and of spell duration on diabetes control state transition
probabilities.
I estimate separate conditional hazard models for probability of returning to good
diabetes control and for transitioning to poor diabetes control. The model explicitly
includes both spell duration dependence and patient characteristics. I use a dataset of
38
patients from the Childrens Hospital Los Angeles Comprehensive Diabetes Center
(“Center”) to estimate the model. The results of the study are of particular interest to the
practicing clinicians at the Center as a possible guide for treating patients with diabetes.
The chapter is structured as follows. Section 2.2 contains a literature review. Section
2.3 describes the data and section 2.4 the econometric model and specification. The
results are presented in section 2.5. I conclude in section 2.6.
2.2 Literature Review
Substantial work has been done studying the correlation between treatment specifics,
individual patient characteristics and glycemic control. This literature aims to establish
both correlation and causation of patient and family characteristics, socio-economic
characteristics and treatment protocols with glycemic control. For example Chalew et al.
(2000) “evaluated the impact of multiple factors including a special multidisciplinary
management program on glycosylated hemoglobin in children with Type 1 diabetes.”
The study was particularly interested in the impact race played in moderating successful
diabetes treatments. Analyses of this type are primarily from the medical literature and
are primarily interested in examining clinical treatment approaches and public health
interventions. For example Ellis, et al. (2005) examined the effect of various insulin
treatment regimens and Kaufman, et al. (1999) examined the relationship between office
visits and diabetes control.
In addition to studies relating diabetes control to various clinical practices, the literature
also addresses specific patient characteristics that may affect patient control. In
particular, clinicians have recognized the relationship between patients’ age and their
39
diabetes control. Vanelli, et al (1999) showed a correlation between a patient’s gender
and age and diabetes control; girls were in worse control than boys and adolescents had
less optimal control than younger patients. Additionally, several researchers have
shown a relationship between a patient’s socio-economic status and diabetes control.
Bihan, et al (2005) concluded “deprivation status is associated with poor metabolic
control.” In their study Hanberger et al. (2008) found that older children and those with
longer durations of diagnosed diabetes had poorer control.
Two strands in the economics literature are related with my analysis of diabetes. First,
there are a number of studies that estimate the costs of diabetes. The review article
from Songer et al (2004) found that “The economic cost of diabetes is estimated to be
as much as $US100 billion per year in the US alone (1997 values). This estimated cost
has increased notably over time.” A second relevant strand is that which examines
duration models and estimates the parameters on conditional hazards. Conditional
hazard modeling represents the probability of an event as an empirical likelihood that is
made a function of observable characteristics. Observations of event occurrence
identify the impacts of these various characteristics on the event’s probability. This
approach has not been widely used in studies of health behaviors and has never been
applied to diabetes management and to the probability that patients will be in various
states of diabetes control. The economics literature has applied duration conditional
hazard modeling to a diverse set of phenomenon with particular use in labor economics.
See for example Ham and Rea (1987) for an application of these models to
unemployment.
40
My analysis builds on the spirit of the medical literature in attempting to assess the
determinants of diabetic control while applying the econometric techniques of duration
conditional hazard modeling. My analysis differs from the more standard treatment of
the health literature in that I study spell transition probabilities and not population
correlations. I apply the econometric techniques used in labor economics to study spells
to diabetes control.
2.3 Data
To assess the impact of various child/family characteristics on diabetes control transition
probabilities, I examine a longitudinal dataset containing multiple observations on
individual patients with type 1 diabetes. Within the data, I am able to observe patients
who transition from states of good to poor diabetes control as well as transition from
poor control to good diabetes control. The dataset comes from the Childrens Hospital
Los Angeles’ Comprehensive Diabetes Center (“Center”) and contains information on
2334 patients with type 1 diabetes who received diabetes care through the Center from
2001 through the end of 2005.
The data contains a limited set of demographic information on each patient. Information
on a patient’s age, ethnicity and time since diabetes diagnosis are included as
independent variables in the analysis. The patient’s health insurance type is included
and coded either as private insurance covering HMO, PPO or POS, or government
insurance covering CCS, MediCal and Healthy Families. Due to a flaw in the database
design, the Center only maintained the value of each demographic data element
collected at the most recent visit. This does not impact time-constant variables;
41
however, for the potentially time changing variable, insurance type, I assume that it did
not change from 2001 through 2005. Due to legal confidentiality regulations the Center
did not collect and therefore the dataset does not contain any additional patient
demographic information such as economic status, family education levels and parental
marital status.
The dataset does contain patient address in a confidential section. I applied for and
received CHLA Human Subjects Research Board approval to access the confidential
address information and match it with the patient’s census track. With each address
matched to the corresponding census track, I associated census track level socio-
economic variables to each of the patient observations. Note that only the most recent
address was maintained and thus I have to assume that the address was constant
through the period of observation. The matching with the patients’ census track allows
me to add the time-changing variables of census track level poverty rates and median
income. These economic variables are intended to help control for the unobserved
characteristics of each child/family, particularly those correlated with insurance type.
The dataset contains information on the patients’ A1C levels measured each half year
from 2001 through 2005. While instantaneous diabetes control is measured by a
patient’s current blood sugar level, the laboratory A1C test of a patient with diabetes
provides sufficient information to determine his/her average control state over the
previous months. The A1C measures a biological marker of the patient’s mean blood
glucose level over the preceding three months. The measured A1C level increases
monotonically as the patient’s blood glucose shows worse control on average. The A1C
variable is a continuous variable and ranges from four to 25. In practice it is collected on
42
each patient with diabetes periodically and used to assess average control over the
preceding time period.
I transform each A1C measure to a state variable coding for whether the patient was in
good control over the preceding period or in poor control. As described below, I model
the control process as a discrete state transition model. The American Diabetes
Association guidelines for proper diabetes care state that patients with an A1C level
below 8.0% are considered to have been in “Good Control” while those with levels equal
to or above 8.0% are considered to have been in “Poor Control”
4
. Each patient’s 10, half
year A1C values are transformed into a discrete variables coding for the patient’s control
state in each half year period. As my data is biannual, I make the assumption that the
six month A1C level is an applicable measure for control over the entire six months.
I define four spell types – New Spells Poor Control, New Spells Good Control,
Interrupted Spells Poor Control, and Interrupted Spells Good Control. A patient is
determined to have transitioned from good control to poor control and start a new spell
of poor control if there is an observation of good control followed by an observation of
poor control, and vice versa for transitions from poor control to good control and spells of
good control. Interrupted spells are based on the patient’s control state when the
observation period first begins. Thus if the patient is first observed in good control (poor
control) she/he has an interrupted spell of good control (interrupted spell of poor control).
The spell, either new or interrupted, continues through each of the following half year
periods if the patient’s control state does not change in the following observations.
4
I conduct a sensitivity analysis to the A1C cutoff for control. Appendix B contains the results of
all analysis using an A1C cutoff of 7.0% and 9.0%.
43
Explicitly a spell of poor control is determined to end when the patient returns to control
as shown by the first observation of an A1C level below 8.0% after the initial transition to
poor control. A spell of good control is determined to end when the patient transitions to
poor control as measured by the first observation of an A1C level equal to or above
8.0%. Thus spell duration is measured in half year increments based on the six month
A1C observations.
Spells where the patient does not transition to another state to conclude the spell are
deemed censored spells. Censored spells occur in the data for two reasons. First, a
number of spells are ongoing when the observation period concludes in 2005. Second,
a number of observations on A1C are missing throughout the dataset. If the
observations on the patient contain a missing A1C value before the patient ends a spell,
that spell is treated as censored at that point. Appendix C illustrates the determination of
spell starting, spell length and spell termination type for three sample patients.
The original dataset contains information on 2334 patients with type 1 diabetes;
however, the analysis is ultimately conducted on a subset of the sample population.
First, the analysis is restricted to only those patients that had complete information on
ethnicity, age, length of time with diabetes, healthcare payer and address. This leaves
739 children in the complete sample. Second, for each spell type, analysis is conducted
on only observed spells. Thus, only those children who are observed with a particular
spell type are included in each of the four separate analyses. Table 2.1 shows the
characteristics of the sample population and those patients who are included in each of
the four separate analyses. For age and duration with diabetes the values in 2003 are
shown.
44
Table 2.1: Sample Population Characteristics
All
Patients
New Spells of
Good Control
Interrupted Spells
of Good Control
New Spells of
Poor Control
Interrupted Spells
of Poor Control
Number of Children 739 332 248 298 340
Age
16.11
(0.177)
15.29
(0.259)
15.68
(0.338)
15.39
(0.265)
16.84
(0.237)
Duration with
Diabetes
8.675
(0.159)
8.216
(0.238)
8.633
(0.272)
8.254
(0.232)
9.336
(0.225)
Ethnicity=White
0.592
(0.018)
0.612
(0.026)
0.631
(0.030)
0.595
(0.028)
0.523
(0.027)
Government
Insurance
0.343
(0.017)
0.353
(0.026)
0.246
(0.027)
0.350
(0.027)
0.438
(0.026)
Census Track
Poverty Rate
12.78
(0.369)
12.04
(0.521)
11.71
(0.616)
12.68
(0.577)
14.10
(0.556)
Census Track
Median Income
54957
(935.5)
56070
(1442)
59042
(1764)
54799
(1498)
50324
(1217)
Standard Errors in parenthesis
As Table 2.1 illustrates there is only complete information on 31.6% of the patients and
of those 33% to 46% enter spells of the various types. The reduced sample causes
concern about selection bias of two types; (1) selection of those patients for whom the
data was complete and (2) selection of those patients who enter into a particular spell
type. The first type of potential selection bias arises from the fact that when patient data
was missing information on personal characteristics, these observations were not
included in the analysis. The Center staff asserted that missing data was random. The
data analyzed here comes from a secondary clinical database which the Center did not
actively maintain and missing data resulted. Thus, I assume that the missing individual
data occurred randomly.
The second type of potential selection bias, the selection of individual patients who enter
particular spell types, cannot be considered a random event. In fact, the object of my
transition analysis is precisely to examine the differential impact of characteristics on
patients in various control states. Therefore, my analysis should be considered a
conditional analysis that is conditional on the fact that the patient has been observed in a
45
certain type of spell. Ham and LaLonde (1996) show that this initial sample selection
bias may not be trivial when the selection is codetermined with an independent variable.
However, here the analysis does not assess the effect of a particular treatment or
program. Rather the estimation is concerned with estimating determinants of
duration/transition for various diabetes control states.
2.3.1 Spell Data
For each spell type there were both completed and censored spells. Additionally, for the
new spells, there are occasions where individual patients had multiple spells. Table 2.2
shows the information on total number and completed spells as well as spell duration for
each of the four spell types.
Table 2.2: Spell Information by Spell Type
New Spells of
Good Control
Interrupted Spells
of Good Control
New Spells of
Poor Control
Interrupted Spells
of Poor Control
Number of Children 332 248 298 340
Number of Spells 472 248 409 340
Percent Completed Spells 52.8% 44.0% 53.8% 50.6%
Complete Spell Duration
1.755
(0.085)
3.220
(0.202)
1.522
(0.066)
2.139
(0.126)
Censored Spell Duration
3.058
(0.159)
5.079
(0.284)
2.802
(0.168)
4.035
(0.232)
Standard Errors in parenthesis
As Table 2.2 illustrates for each spell type roughly half of the observed spells are
censored. Censoring resulted both at the end of the observation period and from
missing data on the dependent variable – diabetes control state. Similar to the missing
information on the independent variables, the missing information on the dependent
variable creates a source of potential bias. This potential source of bias is worrisome if
46
missing observations are somehow correlated with the A1C and diabetes control state.
This would be the case if missing observations derived from the fact that patients with
different control states differentially showed up for office visits, or if some other similar
process resulted in the missing data. During interviews, the Center staff reiterated that
missing data, including on the dependent variable was a random occurrence and
therefore downplayed any possible selection bias. In my analysis, I treat the data as
missing randomly.
Table 2.2 also illustrates that censored spells lasted significantly longer than the
completed spells. This held true for all spell types. This will impact estimates of spell
duration dependence as we observe fewer spells that end with longer durations. In
particular, there are some censored spells that last the entire observation period
implying that the patient stayed in one control state through the observation period of
2001 through 2005. To address this data limitation in the econometric work, I use
duration dummies and do not treat duration dependence as linear.
2.4 Econometric Model and Estimation
I model the process of diabetes control state transition as an individual child/family
choice between the costs and benefits of good control. Type 1 diabetes affects a
patient’s ability to produce the hormone insulin which naturally regulates the levels of
sugar in the blood. Diabetes treatment requires a patient to expend both considerable
time and energy in order to keep their blood glucose (glycemic levels) within the normal
range. For example, patients must test their blood sugar multiple times a day and often
must take multiple daily injections of insulin. The costs of compliance in money, time
and mental anguish are significant. The benefits to maintaining proper glycemic control
47
are potential short-term improvements in how a patient feels and the well-documented
reduction in long-term complications. Diabetes is one of the leading causes of
blindness, amputations, renal failure and heart disease and the likelihood of all of these
complications increase with poor glycemic control (American Diabetes Association,
2003). However, the true extent of the possible complications and their relation to poor
glycemic control is not understood equally by all patients. A patient faces a tradeoff
between the immediate perceived costs of maintaining control and the discounted
benefits. Therefore it is not surprising that a significant proportion of time we observe
patients in poor diabetes control.
I model the glycemic control decision as a decision regarding the current period’s
expected costs to achieve control as compared to the discounted perceived benefits. A
child/family determines the “optimal” control state and acts to achieve that state. Thus
the model focuses on diabetes control state and not the control level. The model
focuses on state transitions in the context of spells of particular states. While the data
contains information on levels as represented by the A1C value, I model the process as
a duration model for several reasons. A duration model allows me to specify the
transition from good control to poor control separately from the transition from poor
control to good control. Duration models of state transition explicitly take into account
spell duration and the state of the individual in previous periods through explicitly
including spell duration as one of the variables. Thirdly, the data limitations described
above make a duration model more appealing than a continuous dynamic model. The
data does not contain any exogenous time-changing variables making it impossible to
find an appropriate instrument for lagged control level. Lastly, while several studies in
the medical literature examine A1C in levels, there are numerous studies that examine
48
diabetes control using an A1C cutoff level. For example Ellis et al. (2005) described
their study as follows, “A randomized controlled trial
was conducted with 127
adolescents with type 1 diabetes and
chronically poor metabolic control (HbA
1c
[A1C]
8% for the past
year) who received their diabetes care in a children’s
hospital located in a
major Midwestern city”. Thus an analysis of control states is consistent with the medical
literature. I test the sensitivity of my results to the chosen control state cutoff by
performing the same analysis with A1C of 7.0% and A1C of 9.0% as control cutoffs.
Note that A1C levels of 7.0% and 9.0% also represent clinically relevant cutoff values.
Appendix B presents these results.
The model is a discrete state duration model where the transition to a state of diabetes
control is modeled as conditional on having been in the other control state for a period of
time. The decision as modeled allows for child/family demographic variables as well as
spell duration to potentially influence the decision. The model allows me to explicitly
define the state transition as depending on child age, duration of time with diabetes,
ethnicity, insurance type and census track poverty levels and census track median
income. These variables affect the control decision by either affecting the relative cost
of maintaining control and the assessment of the discounted future benefits of control.
My analysis offers an empirical reference point to understand the impact of these
variables. Additionally, the duration model allows for estimation of the effect of the
duration of the spell on the conditional probabilities. The model is completed by adding
a random “care shock” which models the fact that given a particular set of control
behaviors, the actual control outcome has a random component. Child/family control
decisions are based on expectations but patients following the exact same treatment
regimens experience different results. The DCCT and numerous subsequent random
49
control trials have shown that despite tight vigilance by medical experts, individual
responses to treatment vary. See for example, Weintrob, et al (2003) for the variance in
type 1 diabetes patients’ response to type 1 diabetes treatments.
For each spell (denoted s) beginning at time t
os
, and where the individual has remained
in the current state for t-1 periods, denote the conditional probability of transitioning
states at time t
os
+t as λ
s
. The conditional probability will depend explicitly on the
child/family characteristics and the spell duration as well as the care shock. The
conditional probability function is given by ( )
ts ts os s s
X t t θ λ λ , ; , = where X
ts
represents
the values of the characteristics at time t
os
+t, and θ
ts
represents the control shock which
is assumed independent across spells and time. Following the literature, for example
Ham and Rea (1987), I model λ
s
as a logit probability model and estimate the reduced
form equation shown in (1).
( ) ( ) ( ) [ ] ) , ( exp 1 / 1 , ; ,
'
0
t t h X c X t t
os ts ts ts s
+ + − + = β θ λ (1)
Where c represents a constant, X
ts
the vector of child/family characteristics and h(.) a
function of the duration of the spell. Given that the spell durations of the censored spells
are significantly longer than the completed spells I represent h(.) as two dummy
variables. The empirical specification will include a dummy variable if the period
represents spell duration=1 and a second dummy variable if the period represents spell
duration=2.
The estimation proceeds similar to Heckman and Singer (1984) and Ham and Rea
(1987) and maximizes the likelihood of observing the pattern of spell transitions in the
data conditional on the specified hazard function (1). For completed spells, the spell’s
50
starting period t
os
and end period t
s
*
are observed and thus the spell ends in the semi-
annual period t
os
+ t
s
*
. The probability of observing that particular completed spell is
given by:
( ) [ ] ) , ( ) , ( 1 ,
*
0 0
1
1
*
0
*
s s s
t
t
s s
t t t t t t g
s
λ λ
− =
Π
−
=
(2)
Where the conditioning on X
ts
and θ
ts
are still present but dropped for notational
simplicity. For censored spells, the spell ending period is not observed. The information
on these spells ends at some s t
_
after which no information is observed. The probability
of observing such a spell is given by the survivor function:
[ ] ) , ( 1 , 1
0
1
_
0
_
t t t t G
s
t
t
s
s
s
λ − =
−
Π
=
(3)
Thus denoting the observations with completed spells by C and those with censored
spells as IN, the likelihood function can be written as follows:
− =
Π Π
) , ( 1 ) , (
_
0
*
0
s
s
IN i
s s
C i
t t G t t g L
ε ε
(4)
As in Ham and Rea (1987), I maximize (4) with respect to c, β, α
1
, and α
2
. Where α
1
,
and α
2
represent the coefficients on the duration time dummies. As discussed above, for
new spells of good control and new spells of poor control, I observe occasions of
multiple spells per individual. However, even in the case of multiple spells per individual,
the model represented by (4) treats each spell as independent. For each spell type, I
51
estimate two separate specifications which differ in how they treat the child’s age. In the
first specification, both age and age squared are entered as continuous variables in the
hazard. The second specification uses a dummy variable to code if the child is a
teenager at the observation. Median income and age squared variables are transformed
by subtracting the mean and dividing by the standard deviation. Thus, the parameter
estimates represent the impact of change of one standard deviation.
2.5 Results
Given the nature of the model in (4) I use the STATA logit command to estimate the
coefficients. Each period of each spell is coded as a separate observation with the logit
command estimating on the period where the spell is completed. Table 2.3 shows the
results of maximizing (4).
52
Table 2.3: Duration Model Results – A1C Control Cutoff 8.0%
Age
-0.119
(0.064)*
-0.399
(0.082)***
-0.015
(0.056)
0.109
(0.090)
Age Squared
0.254
(0.245)
1.270
(0.319)***
0.010
(0.243)
-0.421
(0.543)
Teen
-0.419
(0.198)**
-1.000
(0.213)***
0.031
(0.179)
0.269
(0.227)
Duration with
Diabetes
-0.000
(0.026)
-0.015
(0.022)
-0.036
(0.031)
-0.032
(0.027)
0.021
(0.023)
0.011
(0.018)
-0.047
(0.035)
-0.037
(0.029)
Ethnicity=White
-0.296
(0.191)
-0.298
(0.190)
0.155
(0.221)
0.129
(0.219)
-0.218
(0.174)
-0.217
(0.174)
-0.143
(0.244)
-0.145
(0.244)
Government
Insurance
-0.190
(0.208)
-0.179
(0.207)
-0.169
(0.234)
-0.152
(0.232)
0.189
(0.191)
0.185
(0.190)
0.545
(0.264)**
0.510
(0.263)*
Census Track
Poverty Rate
-0.018
(0.013)
-0.018
(0.013)
-0.026
(0.014)*
-0.023
(0.014)
-0.004
(0.012)
-0.004
(0.012)
-0.020
(0.016)
-0.020
(0.016)
Census Track
Median Income
0.079
(0.119)
0.072
(0.119)
-0.130
(0.135)
-0.074
(0.131)
-0.101
(0.117)
-0.104
(0.116)
-0.149
(0.170)
-0.148
(0.171)
Duration=1
1.580
(0.229)***
1.571
(0.228)***
0.796
(0.209)***
0.877
(0.207)***
1.423
(0.189)***
1.427
(0.189)***
0.202
(0.258)
0.188
(0.255)
Duration=2
0.825
(0.277)***
0.811
(0.277)***
0.114
(0.265)
0.150
(0.264)
0.591
(0.239)**
0.594
(0.239)**
0.351
(0.267)
0.341
(0.266)
Constant
0.223
(1.058)
-1.182
(0.377)***
5.011
(1.375)***
-0.690
(0.406)*
-1.875
(0.869)**
-2.051
(0.318)***
-3.282
(1.397)**
-1.926
(0.390)***
Standard Errors in parenthesis - *** significant at 1%, ** significant at 5%, * significant at 10%
Spells Poor Control
Transition to Good Control
Spells Good Control
Transition to Poor Control
New Spells Interrupted Spells New Spells Interrupted Spells
Table 2.3 reveals an interesting pattern of effects on the probability that a patient in a
spell of a particular diabetes control state transitions to the other control state. A child’s
age seems to significantly impact the probability of ending a spell of poor control and
transitioning to good control. Age does not impact transitions in the opposite direction
from good control to poor control. The dummy variable on teen is significant and
negative for both new and interrupted spells of poor control. This implies that teenagers
are significantly less likely to return to good control than are younger patients. To assess
the magnitude of the impact, I look at the estimates from the interrupted spells of poor
control and evaluate the likelihood of a child returning to good control at the average
values for the other variables. I find that a child who is not a teenager has a 26.0%
likelihood of transitioning to good control. A teenager has only an 11.4% likelihood of
53
transitioning. It is also of note that the impact of being a teen is asymmetric in that
teenagers appear no more likely to transition to poor control; they seem to persist longer
in the state of poor control. As Hanberger et al. (2008) showed, it is known in the
medical literature that teenagers have generally poorer diabetes control than younger
patients. These results seem to suggest that this steady state difference results from
longer spells of poor control and not necessarily more spells of poor control.
A patient’s duration of time with diabetes is not found to significantly impact any of the
transition probabilities. This result is not consistent with the medical literature. For
example, Hanberger et al (2008) found that children with A1C levels less than 7.0% had
an average duration of diabetes of 5.0 years, those with A1C 7.0-7.9% had an average
diabetes duration of 5.9 years, those with A1C 8.0-9.9% had an average duration of 6.9
years, and children with A1C levels 10% or above, had an average duration of diabetes
of 7.9 years. My analysis implies that additional experience with the disease does not
improve the likelihood of ending spells of poor control or maintaining spells of good
control.
The patient ethnicity and socio-economic variables are also not significant in the
transition probabilities. White patients are not significantly more likely to transition either
to good or poor control. The census track poverty and median income also do not affect
transition probabilities in this analysis. Note, however, that these results may be
misleading for three reasons. First, since the duration models consider spell duration
and the analysis is conditional on observing a spell of a particular state, these variables
may impact the likelihood of a patient starting a particular spell but not concluding the
spell. Second, I do not observe the economic variables over time as the family’s
54
situation changes as the data only contains one address. Lastly, the data is based on
census track information which is, at best, a proxy for the individual family.
The government insurance variable seems to only significantly impact the transition of
interrupted spells of good control. Patients who begin the observation in good control
with government insurance are significantly more likely to transition to poor control than
those with private insurance. Evaluating the hazard probability for interrupted spells of
good control in the specification with the teen dummy at the means of the
characteristics, a child without government insurance has 9.0% probability of
transitioning to poor control. If that same child had government insurance the probability
would increase 14.1%. Note however, that these effects are statistically significant at
approximately the 5% level. Additionally, given that the economic variables are
controlled for at the census track level, the government insurance variable may also
proxy for unobserved socio-economic characteristics of the family that are correlated
with having government insurance.
Lastly, it seems that duration of a spell negatively impacts transition probabilities. In all
the specifications except the interrupted spells or good control, patients are most likely to
transition in the first period and often more likely in the second period of a spell. This
implies that the longer a patient persists in a particular spell the less likely they are to
transition to the other state. This phenomenon was clear in the data as shown in Table
2.2 by the greater average length of censored spells. It is not possible to distinguish true
state dependence from unobserved heterogeneity in the current model. Unfortunately in
these results, it is not clear if some spells simply persist in a given state longer or if
patient behavior actually changes as the spell continues.
55
As described, I conducted a sensitivity analysis to the choice of A1C cutoff. I conducted
the same analysis for the four spell types using the clinical relevant A1C cutoffs of 7.0%
and 9.0% as comparisons. The results using A1C 7.0% confirm the majority of the
conclusions above. Appendix B shows the tables with the results using these alternative
control cutoffs. For A1C 7.0%, the age affects are no longer significant. In all cases but
the interrupted spells of poor control duration with diabetes, ethnicity, and insurance type
are also not significant. Note that lowering the A1C cutoff has the impact of creating
substantially more interrupted spells of poor control as many patients start the
observation period above the control cutoff. I observe that duration with diabetes has a
significant but small impact to decrease transitioning to good control. White patients are
now more likely to transition to control and patients with government insurance less likely
in the interrupted spells. In all cases, the economic variables continue to be non-
significant and longer durations continue to decrease transition likelihood.
Likewise, the results using A1C 9.0% as the control cutoff confirm the general results.
We see that teenagers are less likely to transition to good control. However, here the
effect of age does not appear as asymmetric where in the interrupted spells of good
control, teenagers are significantly more likely to transition to poor control. Duration with
diabetes and ethnicity are not seen to significantly impact the transition probabilities.
The results for government insurance are significant for more of the spell type transitions
and to a greater statistical level. I observe that patients with government insurance are
significantly more likely to end a spell of good control, both new and interrupted, and
transition to poor control. Again the economic variables are not significant and I observe
negative duration dependence.
56
2.6 Model with Unobserved Heterogeneity
I extend the model to allow for unobserved heterogeneity at the spell level. The results
and examination of the data leave open the question if some spells are different from
others along unobserved characteristics. In particular, the data exhibits a clear
distinction between censored and non-censored spells as shown in Table 2.2. A t-test of
the equivalence of spell length rejects that censored spells have the same duration as
non-censored spells at the 1% level. The estimated results of negative duration
dependence might represent a selection bias on the types of spells that persist and true
duration dependence.
To examine these questions, I extend the model to incorporate unobserved
heterogeneity at the spell level. Formally, the intercept term in the logit hazard
probability will take either one of two discrete types. The model changes so that the
probability λ
s
becomes a function of the unobserved type k which enters through the
constant c.
( ) ( ) ( ) [ ] ) , ( exp 1 / 1 , , ; ,
'
0
t t h X c k X t t
os ts k ts ts s k
+ + − + = β θ λ (5)
Where c
k
now represents a type specific constant, and the rest of the function (5) is as
specified in (1). The probability of observing a particular completed spell for a particular
type k is given by.
( ) [ ] ) , ( ) , ( 1 ,
*
0 0
1
1
*
0
*
s s k s k
t
t
s s k
t t t t t t g
s
λ λ
− =
Π
−
=
(6)
57
The probability of observing a particular spell for a particular individual is given by the
expected value of g
k
with respect to the probability distribution of the unobserved
heterogeneity k. Given that there are two types of k the expected value, given in the
following sum, represents the likelihood of observing a completed spell.
( ) ( )
∑
=
=
2
1
*
0
*
0
) ( ; , ,
k
s s k i i
k prob k t t g t t g (7)
Similarly the probability of observing a particular censored spell for type k is given by:
[ ] ) ; , ( 1 , 1
0
1
_
0
_
k t t t t G
s k
t
t
s
s k
s
λ − =
−
Π
=
(8)
The likelihood of observing the censored spell is given by:
∑
=
− =
−
2
1
_
0
_
0
) ( ; , 1 , 1
k
s
s k
i
i
k prob k t t G t t G (9)
Thus, we are left with the sample likelihood given by the same expression as in (4). In
the case now with unobserved heterogeneity the maximization is with respect to the
coefficients β, α
1
, and α
2
of the original model and the heterogeneity parameters c
1
, c
2
and the probability(k=1). As the probabilities of the unobserved types must sum to one,
the maximization is with respect to just one type probability. Table 2.4 shows the results
of estimating the coefficients in (8)
5
.
5
Given the unobserved heterogeneity in the model, I cannot use any preprogrammed statistical
algorithms. I have programmed the optimization in Fortran using the GQOPT optimization library.
Code is available upon request.
58
Table 2.4: Duration Model with Unobserved Heterogeneity Results – Control Cutoff A1C
8.0%
New Interrupted New Interrupted
Teen
-0.799
(0.201)***
-1.744
(0.206)***
-0.219
(0.166)
0.169
(0.201)
Duration with
Diabetes
0.023
(0.018)
-0.092
(0.021)***
0.066
(0.017)***
-0.013
(0.021)
Ethnicity=White
-0.270
(0.170)
0.391
(0.176)**
-0.564
(0.159)***
-0.187
(0.218)
Government
Insurance
-0.123
(0.177)
-0.133
(0.185)
0.490
(0.171)***
0.657
(0.247)***
Census Track
Poverty Rate
-0.030
(0.012)**
-0.051
(0.011)***
-0.008
(0.010)
-0.027
(0.014)*
Census Track
Median Income
0.094
(0.135)
-0.229
(0.114)**
-0.079
(0.095)
-0.106
(0.14)
Duration=1
0.467
(0.260)*
-1.335
(0.204)***
0.250
(0.221)
-0.36
(0.25)
Duration=2
0.003
(0.313)
-1.330
(0.315)***
-0.228
(0.292)
-0.117
(0.330)
Constant 1
0.503
(1.584)
4.170
(1.430)***
-0.204
(1.144)
-0.968
(3.431)
Constant 2
-29.80
(2.052)***
0.059
(0.408)
-4.131
(0.789)***
-30.35
(4.625)***
Gamma
-1.072
(2.658)
0.286
(0.616)
-0.518
(1.248)
-0.723
(6.786)
Probability 1
74.5% 42.9% 62.7% 67.3%
Standard Errors in parenthesis - *** significant at 1%, ** significant at 5%, * significant at 10%
Spells Poor Control
Transition to Good Control
Spells Good Control
Transition to Poor Control
The results shown in Table 2.4 suggest that the observed negative duration dependence
was a function of unobserved heterogeneity. In no cases are the coefficients on the
dummies for duration=1 or duration=2 positive and significant. In fact, in two of the spell
types the coefficients are negative with the particularly interesting result that the
coefficients on the duration dummies in the interrupted spell of good control are now
negative and significant at the 1% level.
The asymmetric impact of being a teenager remains and its magnitude increases after
allowing for unobserved heterogeneity. The coefficients on the transition to good control
59
in both new and interrupted spells became more negative and are significant at the 1%
level. The coefficients on the teen dummy for the transition to poor control remain non-
significant. This reinforces the assessment of the impact of being a teen that comes
from the results with no heterogeneity.
The significant positive impact of government insurance on transitions to poor control
remains with the unobserved heterogeneity. The coefficient is of a similar magnitude for
the interrupted spells. For the new spells, the coefficient on government insurance is
significant, positive and of a significant magnitude. Additionally, after controlling for the
unobserved heterogeneity I see some evidence that poverty rate in the census track
decreases the transition into good control. However, the magnitude of these effects are
quite small.
2.7 Conclusions
This analysis represents an extension of the understanding of the relationship between
patient characteristics and their diabetes control. The chapter takes a different modeling
approach than common in the literature and looks at diabetes control spells and state
transitions. I chose this modeling approach both to overcome data limitations and to
focus my research particularly on control states. In this context, I explored the impact of
duration in a particular spell on the likelihood of transition.
The results suggest that teenagers are significantly less likely to end spells of poor
control and transition to good control. There does not seem to be a similar impact for
transitions to poor control. This result was relatively robust both to different cutoffs of the
A1C level for control and to the introduction of unobserved heterogeneity.
60
When significant, government insurance reduces the likelihood that a patient is in good
control. However, the results are not conclusive across specifications. The census track
level economic variables are not significant. Variables at the track level may not
adequately control for the socio-economic factors that are correlated with government
insurance, further reducing the significance of the government insurance results.
Interestingly, the models without unobserved heterogeneity seem to suggest negative
duration dependence. In the case of this model, negative duration dependence would
imply that as patients spent longer in a certain control state they were less likely to
transition to other state. Negative duration dependence would imply that through some
process, such as becoming discouraged or learning by doing, patient state changed less
and became more stable. The alternative explanation for the result that spells are more
likely to end early or to continue as they last longer, is that there is some form of
unobserved heterogeneity that makes some spells simply more likely to last longer. This
would imply that there is no fundamental change as spells progress. The analysis
without controlling for unobserved heterogeneity suggests that patients seem more likely
to end any particular spell in the first and/or second observational period than at
subsequent times. However, the analysis including unobserved heterogeneity at the
spell level reversed these findings. This seems to suggest further research is needed to
determine if duration in a particular diabetes state impacts transition probabilities.
61
Chapter 2 References
American Diabetes Association (2003), “Implications of the Diabetes Control and
Complications Trial,” Diabetes Care 26: S25-S27
Bihan, Helene and Laurent, Silvana and Sass, Catherine and Nguyen, Gerard and Huot,
Caroline and Moulin, Jean Jacques and Guegen, Rene and Le Toumeline,
Philippe and Le Clesiau, Harve and La Rosa, Emelio and Reach, Gerard and
Cohen, Regis (2005) “Association Among Individual Deprivation, Glycemic
Control, and Diabetes Complications” Diabetes Care 28:2680-2685
Chalew, Stuart and Gomwz, Ricardo and Butler, Ashley and Hempe, James and
Comptom, Terry and Rao, Jayashree and Vargas, Alfonso (2000) “Predictors of
glycemic control in children with Type 1 diabetes: The importance of race”
Journal of Diabetes and Its Complications 14: 71-77
Diabetes Control and Complications Trial Research Group (2003) “The Effect of
Intensive Treatment of Diabetes on the Development and Progression of Long-
Term Complications in Insulin-Dependent Diabetes Mellitus” New England
Journal of Medicine 329: 977-986
Ellis, Deborah and Frey, Maureen and Naar-King, Sylvie and Templin, Thomas and
Cunningham, Phillippe and Cakan, Nedim (2005), “Use of Multisystemic Therapy
to Improve Regimen Adherence Among Adolescents With Type 1 Diabetes in
Chronic Poor Metabolic Control” Diabetes Care 28:7 1604-1610
Ham, John and LaLonde, Robert (1996) “The Effect of Sample Selection and Initial
Conditions in Duration Models: Evidence from Experimental Data on Training”
Econometrica 64:1 175-205
Ham, John and Rea, Samuel (1987) “Unemployment Insurance and Male
Unemployment Duration in Canada” Journal of Labor Economics 5:3 325-353
Hanberger, Lena and Lindblad, Bengt and Samuelsson, Ulf and Ludvigsson, Johnny
(2008) “A1C in Children and Adolescents With Diabetes in Relation to Certain
Clinical Parameters” Diabetes Care 35:5 (927-929)
Heckman, James and Singer, Burton (1984) “Econometric Duration Analysis” Journal of
Econometrics 24: 63-132
Kaufman, Francine Ratner and Halvorson, Mary and Carpenter, Sue (1999) “Association
Between Diabetes Control and Visits to a Multidisciplinary Pediatric Diabetes
Clinic” Pediatrics 103:948
Songer, Thomas and Zhang, Ping and Engelgau, Michael (2004) “Cost-of-Illness Studies
in Diabetes Mellitus: Review Article” Pharmacoeconomics. 22(3):149-164
62
Vanelli M and Chiarelli F and Chiari G and Tumini S and Costi G and di Ricco L and
Zanasi P and Catino M and Capuano C and Porcelli C and Adinolfi B and Cieri F
and Giacalone T and Casni A (1999) “Metabolic Control in Children and
Adolescents: Experience of Two Italian Regions” Journal of Pediatric
Endocrinology and Metabolism 12:403
Weintrob, Naomi and Benzaquen, Hadassa and Galatzer, Avinoam and Shalitin, Shlomit
and Lazar, Liora and Fayman, Gila and Lilos, Pearl and Dickerman, Zvi and
Philiip, Moshe (2003) “Comparison of Continuous Subcutaneous Insulin Infusion
and Multiple Daily Injection Regimens in Children with Type 1 Diabetes: A
Randomized Open Crossover Trial” Journal of the American Academy of
Pediatrics 112.3: (559-564)
63
Chapter 3: Econometric Analysis of Elderly Drinking Behaviors, Changing
Economic Circumstances and Response to Spousal Death
3.1 Introduction
This analysis aims to study the impact of the death of a spouse on an elderly individual’s
drinking behavior. The analysis controls for individual characteristics as well as
changing macroeconomic circumstances. Understanding the determinants of drinking
behavior is important for developing appropriate prevention and treatment policies and
has been a question of interest in both the medical/epidemiological and economics
literature. While there is general consensus that birth cohort, age and other non-time-
changing characteristics impact drinking behaviors, there is still debate about how
individual specific and economy wide time-changing variables impact drinking. (Breslow
and Smothers 2004)
My analysis complements the current literature to look at how spousal death impacts
drinking behaviors using a longitudinal dataset and controlling for a broader range of
time-changing and economic variables. I utilize the Health and Retirement Study (HRS)
as my primary dataset which, due to its panel nature, represents a valuable dataset to
study the question of drinking behavior. HRS has been used previously in a number of
studies of drinking behavior primarily looking at the impact of drinking on health (see for
example Ostermann and Sloan 2001 or Perreria and Sloan 2002).
The chapter is organized as follows. Section 3.2 presents a literature review of topics
related to alcohol consumption in both the epidemiological/medical literature and the
64
economics literature. Section 3.3 describes the data. Section 3.4 presents the linear
regression model and 3.5 presents the Tobit regression model. Section 3.6 details the
results. Section 3.7 provides conclusions.
3.2 Literature Review
The medical and public health literature on drinking among the elderly is interested in
two questions. First, the literature addresses the question of what characteristics
determine likelihood for an individual to drink. Second, the literature examines the
impact of drinking on other outcomes. The literature often focuses on the discrete
variable of heavy drinking and note the level of drinking. It looks independently at both
the number of days a person drinks in a given time frame and the number of drinks the
individual consumes per day when she/he drinks.
One strand of the literature examines primarily non-time varying characteristics of
individuals that are correlated with higher drinking levels. These studies primarily
examine the impact of age, birth cohort and demographics on drinking. The literature
has grown as researchers examine new datasets and including recent attention to
longitudinal datasets. For example Breslow & Smothers (2004) used data from the
National Health Interview Survey (1997-2001) to “estimate the gender and age-specific
quantity and frequency of drinking among Americans age 60 years and older.”
Karlamanga, et al (2006) used NHANES data and Merrick, et al (2008) use data from
the Care File of the Medicare Current Beneficiary Survey. This literature primarily uses
cross-sectional data and often discretizes the dependent variable into heavy drinking
and non-heavy drinking. The results are presented either as correlation tables or the
results of regression analysis.
65
Interest in the correlates of heavy drinking has resulted in the breadth of survey data
reviewed and an emerging consensus on the impact of several characteristics and
inconclusive results on others. Breslow, Faden and Smothers (2003) used data from
three national surveys to “estimate gender-specific prevalence of alcohol consumption in
elderly Americans by 5-year age groups.” The paper presents tables of prevalence
estimates adjusted for demographics and lifestyle. They find a prevalence of heavy
drinking of 10.1% among men and 2.2% among women, with heavy drinking decreasing
among men with age and holding stable for women with age. Kerr et al. (2009) studied
data from six cross-sectional National Alcohol Surveys from 1979 to 2005. The
dependent variables in the study were both number of monthly drinks and number of
days where five or more drinks were consumed. Independent variables included age,
birth cohort, real income, ethnicity, martial status and education. They conclude that
their “models indicate the importance of income, ethnicity, education and marital status
in determining these alcohol measures.” All articles I reviewed in this literature al treat
the all individual variables as exogenous. Summarizing my understanding of a
consensus in the literature, male and female drinking patterns have been found to be
different in all analyses with men drinking more than women. When their impact is
controlled for, better education and higher wealth were correlated with greater drinking
behavior. These results treated education and wealth as exogenous to current drinking.
Results for age and birth cohort have been inconclusive.
A second strand of the literature attempts to look directly at causes of drinking by
examining the impact of “stressors” on drinking. This literature follows from theories of
how individuals perceive the benefits of alcohol consumption. In particular, the theory of
66
tension-reduction posits that people drink as a result of stressful events in an attempt to
mitigate the impact of those events. This literature was developed primarily in the 80s
and 90s and while it continues to receive attention, the evidence seems to point that, at
best, tension-reduction is only one of many elements that reinforces drinking (Veenstra
2006). In a 2006 Veenstra et al. reviewed all the articles in the literature from 1990
through 2005 and concluded that “evidence points towards a relationship between the
occurrence of life-events and alcohol use in the general population. The direction of that
effect is, however, not unequivocal”.
A third strand of the medical literature takes drinking as the independent variable and
attempts to estimate its impact on other outcomes. These studies looked at longitudinal
datasets where drinking in previous periods is considered the independent variable and
the timing of follow-on events identifies its impact. A study by Perreira and Sloan (2002)
used the same dataset that I use in my analysis. They studied the “association of
problem drinking history and alcohol consumption with the onset of several health
conditions and death.” They found that heavy drinking at baseline significantly impacted
the likelihood of developing functional impairments, psychiatric problems and memory
problems. Heavy drinking, however, was not associated with adverse health events
(such as cancer, diabetes, stroke, etc.) Again, this analysis examines correlations over
time with baseline behavior that, if interpreted causally, treats alcohol consumption in the
baseline period as exogenous.
The Veenstra et al. (2206) review examined 16 studies, 12 of which were cross-sectional
and four of which were longitudinal looking at the impact of life stressors on alcohol
consumption. These studies examined four different types of life events or stressors –
67
(1) negative life events such as death of a spouse, (2) chronic conditions such as
ongoing job stress, (3) personal emotional distress such as depression, and (4) minor
daily irritations. The results of these studies do not paint a clear picture of the impact of
stressors as some events were found to have different impacts in different studies. For
example “divorce and financial problems were found to be related with both higher and
lower alcohol use.” (Veenstra 2006)
The economics literature on drinking primarily follows the work of Becker and Murphy
(1988) and examines alcohol consumption in the context of rational addiction. This
literature is concerned with how individuals choose an optimal alcohol consumption path.
In this context, changes in prices provide the exogenous variation needed to assess how
current consumption is impacted by past and future consumption. Grossman,
Chaloupka, and Sirtalan (1995) described that “intertemporal complementarity – a key
aspect of the rational addiction model – arises because increases in past or future
consumption (caused by reductions in past or future prices) cause current consumption
to rise.” They find, using panel data on adolescents and young adults that past and
future consumption of alcohol positively impact current consumption. They conclude
that alcohol consumption fits the model of rational addiction. While not the primary
interest of their analysis, their estimates of the impact of race, gender, education and
income were generally consistent with the findings from the epidemiological/medical
literature (Grossman et al. 1995). Since Grossman, Chaloupka, and Sirtalan (1995),
many studies, looking at different populations have tested the rational addiction model of
alcohol consumption. See Wagenaar, Salois, and Komro (2009) for a review of 112
studies examining the impact of prices and taxes on alcohol consumption.
68
While most studies, even in the economics literature, estimated reduced form models or
first order conditions, there have been studies that developed structural models of
drinking in a dynamic context. Arcidiacono, Sieg, and Sloan (2007) modeled individuals
as jointly determining smoking and drinking behavior. They used the same dataset I
used for my analysis and focused only on males. They discretized the state-space into a
heavy/not heavy drinkers, heavy/not heavy smokers and looked at a dynamic discrete
planning model. They were primarily interested in estimating the degree to which
individuals were forward-looking when they decided how much to drink and smoke.
They provided a model in which different levels of forward-looking behavior were nested
which allowed them to test for the degree of forward-looking behavior that best fit the
data. They concluded that their “analysis thus provides strong evidence for that
hypothesis that older individuals are forward-looking and take future risks associated
with smoking and heavy drinking into consideration when determining their choices.”
My analysis fits within the literature analyzing the impact of various characteristics on
individual drinking behavior. I follow the literature and estimate a reduced form model
with drinking behavior as the dependent variable. The analysis focuses primarily on
spousal death and its impact on elderly drinking behavior and in this sense fits into the
literature on life-stressors. I extend the standard medical/epidemiological literature to
look at the impact of economic variables and control more rigorously for the potential
endogeneity of marital status. Data limitations, particularly with respect to available price
information, limit my ability to follow the economics literature and estimate the impact of
lagged and future consumption in my models.
69
3.3 Data
The primary data come from the Health and Retirement Study (HRS) which is a
longitudinal household survey dataset designed to follow the behavior and health of the
elderly in the United States. HRS conducts surveys every two years with the first survey
conducted in 1992. In particular, my analysis uses the HRS data cleaned and
processed by RAND (HRS-RAND) with funding from the National Institutes of Aging.
HRS-RAND contains data through the eighth wave of HRS in 2006.
3.3.1 Alcohol Consumption Dependent Variables – In each wave of the HRS,
respondents were asked about their alcohol consumption. The HRS changed the
structure of the questions regarding alcohol consumption between the second and third
wave. As the questions are dissimilar, I include only the data from waves 3-8 to ensure
consistency. The analysis includes lagged variables and I construct some variables
using the difference in observations between successive waves. Therefore wave 3 data
will not be used as a separate observation. Thus, the dataset has consistent alcohol
consumption information on five distinct waves - HRS waves 4-8.
The HRS survey, starting with wave three, asks three separate questions about alcohol
consumption behavior. First an individual is asked if she/he ever drinks any alcohol
including beer, wine or liquor. If the individual responds yes, she/he is then asked “In
the last three months, on average, how many days per week have you had any alcohol
to drink?” An individual is then asked, “In the last three months, on days you drink,
about how many drinks do you have?” Days with drinking per week is commonly
70
referred to as frequency in the literature. Drinks per drinking episode is referred to as
quantity. HRS-RAND data contains two variables related to drinking:
1. drinkd - Number of days/week drinks
2. drinkn - Number of drinks/day when drinks
I calculate a third drinking variable which is the product of drinkd and drinkn and known
as the composite measure in the literature. This variable represents the number of
drinks per week consumed by the individual.
3. drink_week – Number of drinks/week (drinkd * drinkn)
Table 3.1 shows the frequency distribution of the responses for each of the dependent
variables. The table includes the observations in each of the five waves used in the
analysis. As the literature has shown a distinction between male and female alcohol
consumption behavior (see for example Breslow et al 2003 for differential drinking
patterns among men and women), I present the drinking behavior data separately for
men and women.
Table 3.1A: Frequency Distribution of Drinking Behavior – Drinking Days Per Week
Number Obs. % of Obs Number Obs. % of Obs
0 days 21,153 60.9% 39,113 77.6%
1 day 3,565 10.3% 3,841 7.6%
2 days 2,272 6.5% 2,110 4.2%
3 days 1,858 5.4% 1,462 2.9%
4 days 913 2.6% 662 1.3%
5 days 910 2.6% 637 1.3%
6 days 440 1.3% 253 0.5%
7 days 3,608 10.4% 2,310 4.6%
Males Females
71
Table 3.1B: Frequency Distribution of Drinking Behavior – Drinks per Day
Number Obs. % of Obs Number Obs. % of Obs
0 drinks 21,196 61.1% 39,172 77.7%
1 drink 5,162 14.9% 6,379 12.7%
2 drinks 4,464 12.9% 3,481 6.9%
3 drinks 2,006 5.8% 909 1.8%
4 drinks 805 2.3% 230 0.5%
5 drinks 307 0.9% 65 0.1%
6 drinks 471 1.4% 89 0.2%
7 drinks 57 0.2% 15 0.0%
8 drinks 85 0.2% 14 0.0%
9 drinks 9 0.0% 1 0.0%
10 drinks 65 0.2% 9 0.0%
11 drinks 3 0.0% 2 0.0%
12 drinks 59 0.2% 19 0.0%
13 drinks 0 0.0% 2 0.0%
14 drinks 2 0.0% 0 0.0%
15 drinks 9 0.0% 1 0.0%
16 drinks 7 0.0% 0 0.0%
> 16 drinks 12 0.0% 0 0.0%
Males Females
72
Table 3.1C: Frequency Distribution of Drinking Behavior – Drinks Per Week
Number Obs. % of Obs Number Obs. % of Obs
0 drinks 21,207 61.1% 39,191 77.8%
1 drink 1,874 5.4% 2,537 5.0%
2 drinks 1,742 5.0% 1,974 3.9%
3 drinks 944 2.7% 1,021 2.0%
4 drinks 1,286 3.7% 1,199 2.4%
5 drinks 376 1.1% 379 0.8%
6 drinks 1,181 3.4% 782 1.6%
7 drinks 1,174 3.4% 1,113 2.2%
8 drinks 477 1.4% 261 0.5%
9 drinks 335 1.0% 150 0.3%
10 drinks 416 1.2% 219 0.4%
11 drinks 1 0.0% 1 0.0%
12 drinks 560 1.6% 199 0.4%
13 drinks 0 0.0% 0 0.0%
14 drinks 1,152 3.3% 852 1.7%
15 drinks 188 0.5% 64 0.1%
16 drinks 78 0.2% 20 0.0%
17 drinks 0 0.0% 0 0.0%
18 drinks 144 0.4% 30 0.1%
19 drinks 0 0.0% 0 0.0%
20 drinks 81 0.2% 15 0.0%
21 drinks 640 1.8% 233 0.5%
22 drinks 0 0.0% 0 0.0%
23 drinks 0 0.0% 0 0.0%
24 drinks 84 0.2% 11 0.0%
25 drinks 25 0.1% 5 0.0%
26 drinks 0 0.0% 0 0.0%
27 drinks 1 0.0% 0 0.0%
28 drinks 264 0.8% 60 0.1%
29 drinks 0 0.0% 0 0.0%
30 drinks 37 0.1% 3 0.0%
> 30 drinks 452 1.3% 69 0.1%
Males Females
Table 3.1 illustrates several features of the data. First, a large percentage of both men
and women abstain from drinking altogether. Almost 61% of men and 78% of women
abstain from drinking. Second, while the likelihood of observing more drinks per day or
more drinks per week generally declines as the number of drinks increases, this is not
the case for number of days drinking. Reporting 7 days of drinking is more common for
men than any frequency other than abstinence. For women 7 days of drinking is more
73
common than any frequency other than zero or one days of drinking. Lastly, the pattern
of observed behavior in the data supports that drinking behavior is significantly different
between men and women. I the test the proportion of men and women who report the
same number of days drinking per week and rejected that it is the same for all days.
Based on these results and the propensity in the literature, I proceed to conduct my
analysis separately for men and for women.
3.3.2 Spousal Death Variables: HRS contains a relatively rich set of information on the
individuals’ current and past marital status. For each wave, the data contains
information on the respondents’ current marital status, their total number of marriages
and the total number of times they have been widowed. Using the marital status
information I can determine if an individual’s spouse has died between survey waves.
I construct a dummy variable, Became a widow(er), which codes for an individual whose
spouse dies between two waves. This variable represents the primary time-changing
variable of interest in my analysis. It takes the value 1 if the respondent became a
widow in the period and 0 otherwise.
The Became a widow(er) dummy variable is constructed as follows
6
. First, I compare
each respondent’s marital status in the current wave with that in the previous wave. If
the current period response is that the individual is a widow(er) while the previous period
response is that the individual was not a widow(er), I set the value of Became a
widow(er) equal to 1. In addition to information on individuals’ marital status, the data
contains information on the number of times an individual has been widowed. If in
6
Appendix D illustrates an example of coding the Became a widow(er) variable for the wave 4
responses.
74
successive waves the data shows an increase in the number of times the individual has
been widowed this reveals that the individual was widowed in the time between the
waves. In these cases I set the value of Became a widow(er) equal to 1.
Thus, the data contains two sources of information on whether the individual was a
widow(er) for each survey period. In most cases the information gathered from
examining the number of times widowed results in the same coding of the Became a
widow(er) variable as examining the information on change in the widowed marital
status. However, in a small minority of the cases the reported marital status and the
count of times widowed did not agree. In those cases, I include all values of Became a
widow(er) equal to 1 even if the two data are not in alignment. This yields the most
expansive definition of Became a widow(er) available in the data.
There are two endogeneity issues related with marital status and spousal death that
could potentially bias the results of my analysis. First, it is unclear whether current
marital status should be treated as exogenous from current drinking behavior. In
particular, drinking activity may impact whether individuals get married and/or stay
married. To address this form of endogeneity I conduct the analysis on two different
samples of individuals as described below. Second, there is a potential issue about
whether spousal death is exogenous from current drinking behavior; I proceed with the
analysis under the assumption that it is exogenous. I assume that a spouse dying
between given waves of the HRS is not correlated with the individual’s drinking behavior.
This excludes the direct causal pathway whereby an individual increases the likelihood
of his/her spouse dying in a method someway connected to drinking, for example
through neglect while drinking or while drunk driving.
75
Part of the exogeneity of marital status is a subtle issue that spousal death requires an
individual to be married at the time his/her spouse passes away. This means the
individual has to both have become married and remained married. The source of this
potential endogeneity is then the selection bias of those who become and remain
married to a point where their spouse could die. If the population of people who are in a
state of marriage such that their spouse could die so that they could become a widow is
somehow different than the general population, the estimates of the impact of spousal
death may be biased. I address this potential selection bias by assessing the impact of
spousal death on two different sample populations. Estimation with these two sample
populations require two different exogeneity assumptions and comparison of the
estimates between them provides a robustness check on the results.
First I assess the impact of spousal death on the entire sample population and assume
that there is no selection bias. This does allow for a correlation on the observable
marital status but not with unobservable characteristics. I include a dummy variable
Married equal to 1 if the individual is married at the wave interview and a dummy
variable Widowed if the individual is a widow at the time of the interview. Being married
or widowed impacts behavior, but conditional on marital status in the particular period,
the spousal death is uncorrelated with the error terms.
Second, I conduct the analysis on a subset of the sample that by construction excludes
the potential source of bias in the correlation with martial status and unobservable
characteristics. For this sub sample, I assume that any potential endogeneity derives
from the possible selection bias of those who are married and in a position to become
76
widowed. I conduct the analysis on the subset of individuals who are married at Wave 3
and whose marriage either continues for the length of the survey, ends in the spouse
dying, or who drop out of the sample. This is a subsample of only those individuals who
could become a widow(er) and thus excludes the potential bias by construction. Slightly
greater than 45% of the survey sample meets these criteria.
In the analysis of both samples I include a dummy variable if the individual is widowed at
the observation. An individual who becomes a widow(er) in a given period will have the
Widowed dummy variable equal to one and have the dummy variable Became a
Widow(er) equal to one. Identification of the effect of spousal death comes from the
assumption that the Widowed variable controls for the impact that the individual is not
married and is widowed. Thus, if being a widow(er) impacts drinking through the a
causal pathway related to change in behaviors associated with no longer being married,
for example socializing more or less frequently, this effect would appear in the coefficient
of the Widowed variable. This implies that the Became a Widow(er) variable will identify
only those impacts of the spouse dying in that period. This assumption implies the
strong assumption that there is no adjustment period to the new state of being a
widow(er) and that the transition occurs fully within one period. If there were a transition
and over time adjustment process longer than one period, than the Became a Widow(er)
variable would not identify the impact of the spousal death alone but also this transition
process. Given that the period length between observations is two years, I believe it is a
plausible assumption that all adjustment to being a widow occurs within this period and
the variable Became a Widow(er) will estimate the impact of spousal death.
77
3.3.3 Economic Variables: I analyze how individuals’ drinking behavior responds to
changes in their income and to macroeconomic changes in the business cycle and
alcohol prices. The HRS-RAND data contains reliable information on household per
capita income in each wave. I assume that lagged income can be treated as exogenous
from current drinking behavior and use lagged real income in the various specifications.
The HRS-RAND contains information on individuals’ census division of residence at the
date of survey response. Using data from the Bureau of Labor Statistics, I obtain the
monthly CPI and use this data to transform the HRS-RAND income data to real terms.
CPI data is reported with a value of 100 as the base value and therefore real income is
reported as nominal income divided by the CPI.
Additionally, using data from the Bureau of Labor Statistics I associate with each
observation, for each individual, the monthly, seasonally adjusted, unemployment rate
and the CPI alcohol price. Both the general CPI and alcohol CPI are reported on the
same scale and therefore the analysis of alcohol price is conducted on the ratio of
alcohol prices to general prices. Note that this data is available at the census division
level monthly and can be matched to the specific month the survey was conducted for
each individual
7
. I assume that individual employment status is not exogenous from
current drinking and thus use macroeconomic variables as a measure of general
business cycle effects.
3.3.4. Independent Covariates: As previous literature has shown, an individual’s age,
birth cohort, ethnicity and education impact drinking behavior. HRS-RAND includes
7
Note that in the publically available files, HRS does not contain more precise information
beyond census division on individuals’ location of residence.
78
consistent information across all the waves on these variables. They are included as
independent variables in the analysis.
To construct the final dataset I remove observations with missing data. While the HRS-
RAND contains entries for up to 30,405 individuals per wave data is missing for some
individuals in some of the waves. This is mainly due to the fact that some individuals
were added to the HRS sample in later waves. I assume data to be missing randomly
and exclude the individual wave observations with any missing data from this analysis.
This leaves a substantial dataset with more than 24,310 individuals included and more
than 15,000 individual observations in each wave. Table 3.2 presents the descriptive
statistics of the dependent and independent variables.
79
Table 3.2A: Descriptive Statistics of the Dependent and Independent Variables (Male
Population)
Wave (survey year)
Wave 4
(1998)
Wave 5
(2000)
Wave 6
(2002)
Wave 7
(2004)
Wave 8
(2006)
Number of Individuals 6314 7686 6869 6406 6933
Days Drinking
(Frequency)
1.427
(0.03)
1.357
(0.026)
1.414
(0.028)
1.489
(0.03)
1.483
(0.028)
Number Drinks
(Quantity)
0.868
(0.02)
0.841
(0.018)
0.882
(0.018)
0.897
(0.021)
0.945
(0.019)
Drinks per Week
(Composite
3.434
(0.103)
3.222
(0.088)
3.386
(0.092)
3.557
(0.111)
3.59
(0.092)
Widowed
0.095
(0.004)
0.09
(0.003)
0.1
(0.004)
0.103
(0.004)
0.086
(0.003)
Became a Widow(er)
0.027
(0.002)
0.023
(0.002)
0.025
(0.002)
0.021
(0.002)
0.02
(0.002)
Age
68.362
(0.121)
67.816
(0.111)
69.273
(0.113)
70.388
(0.112)
68.608
(0.123)
Percent White
0.849
(0.005)
0.848
(0.004)
0.846
(0.004)
0.843
(0.005)
0.83
(0.005)
Years of Education
11.894
(0.046)
12.181
(0.041)
12.277
(0.043)
12.332
(0.044)
12.613
(0.041)
Real Household Per
Capita Income
165.91
(4.744)
175.175
(5.354)
165.601
(3.396)
168.708
(4.371)
185.547
(11.129)
Unemployment Rate
4.439
(0.009)
3.915
(0.007)
5.754
(0.007)
5.488
(0.007)
4.601
(0.006)
Real Alcohol Prices
1.013
(2.269E-4)
1.011
(2.149E-4)
1.019
(2.19E-4)
1.015
(2.77E-4)
0.993
(3.318E-4)
Standard Errors in Parenthesis
80
Table 3.2B: Descriptive Statistics of the Dependent and Independent Variables (Female
Population)
Wave (survey year)
Wave 4
(1998)
Wave 5
(2000)
Wave 6
(2002)
Wave 7
(2004)
Wave 8
(2006)
Number of Individuals 9177 10997 10002 9455 9932
Days Drinking
(Frequency)
0.659
(0.017)
0.641
(0.016)
0.715
(0.017)
0.746
(0.018)
0.818
(0.018)
Number Drinks
(Quantity)
0.339
(0.008)
0.316
(0.007)
0.366
(0.008)
0.389
(0.009)
0.428
(0.009)
Drinks per Week
(Composite
1.123
(0.036)
1.076
(0.034)
1.193
(0.035)
1.282
(0.039)
1.423
(0.042)
Widowed
0.309
(0.005)
0.297
(0.004)
0.311
(0.005)
0.32
(0.005)
0.282
(0.005)
Became a Widow(er)
0.054
(0.002)
0.044
(0.002)
0.048
(0.002)
0.044
(0.002)
0.042
(0.002)
Age
67.952
(0.122)
67.64
(0.109)
68.888
(0.111)
69.929
(0.11)
68.129
(0.113)
Percent White
0.815
(0.004)
0.815
(0.004)
0.817
(0.004)
0.816
(0.004)
0.798
(0.004)
Years of Education
11.678
(0.034)
11.921
(0.03)
12.035
(0.032)
12.079
(0.032)
12.338
(0.031)
Real Household Per
Capita Income
135.008
(2.091)
140.993
(2.358)
142.093
(4.687)
138.745
(2.35)
155.144
(7.548)
Unemployment Rate
4.449
(0.008)
3.911
(0.006)
5.752
(0.006)
5.483
(0.006)
4.6
(0.005)
Real Alcohol Prices
1.012
(1.864E-4)
1.01
(1.79E-4)
1.018
(1.821E-4)
1.014
(2.292E-4)
0.991
(2.781E-4)
Standard Errors in Parenthesis
As Table 3.2 illustrates, in each wave more than 8.5% of men and 28% of women are
widow(er)s. In each wave more than 2% of men and 4% of women became widowed.
Male drinking behavior is relatively similar between waves, while women tended to drink
more often and in greater quantities in later waves. Note that the drinking behavior, on
average, is quite different between men and women. T-tests that the mean drinking
behaviors between the genders are the same are rejected; this further supports
separating the analysis by gender.
81
Across the five waves there are 34,208 observations on men and 49,563 observations
on women. There are from one to five observations per individual. Table 3.3 shows the
distribution of observations per individual within the data. As the table illustrates, there
are 24,310 individuals in the sample with 10,172 men and 14,138 women. Most
commonly there are observations on each individual for each of the five waves, but for
more than 70% of the sample there are less than five observations and there is only a
single observation for 21.5% of the men and 19.2% of the women.
Table 3.3: Frequency of Observations on Sample of Individuals in the HRS Panel
Number of
Observations
Males in
Sample
% of
Males
Females in
Sample
% of
Females
1 Observation 2183 21.5% 2711 19.2%
2 Observations 1184 11.6% 1478 10.5%
3 Observations 1090 10.7% 1389 9.8%
4 Observations 2188 21.5% 3069 21.7%
5 Observations 3527 34.7% 5491 38.8%
Totals 10172 14138
3.4 Linear Regression
The linear regression specification assumes that the drinking behaviors, days per week,
drinks per day and drinks per week, are generated by a process that can be
approximated by a linear regression model. In this specification, observed values of
zero days drinking, seven days drinking, zero drinks per day, and zero drinks per week
are assumed to be the optimal choice of the individual and not corner solution results of
model where the individual has been constrained from selecting a negative quantity.
Additionally, in this specficiation, the drinking variables are assumed continuous. I
estimate the following equation for each of the six dependent variables:
it i it it
X y μ δ β β + + + =
1 0
(1)
82
Where X
it
is a vector of individual explanatory variables, δ
i
an individual i specific effect
and μ
it
is an individual specific contemporaneous shock. X
it
will contain only variables
assumed exogenous from the individual specific error term. (1) represents the random
effects model in which δ
i
is uncorrelated with the explanatory variables. In (1), δ
i
can be
ignored in terms of its impact on the parameter estimates and the equation can be
estimated with the pooled OLS estimator. However, given that there are multiple
observations on a number of individuals within the sample, δ
i
will introduce
heteroskedacity into the model. Therefore, in the linear regression model I cluster the
standard errors by individual.
X
it
contains variables, as discussed above, shown to impact drinking behavior and
assumed exogenous from the individual time specific shock. X
it
contains the
demographic information of birth year cohort, age, squared age, ethnicity and education.
X
it
contains the economic variables of lagged real per capita household income, census
division unemployment rate at time of survey, and census division real alcohol prices at
time of survey. X
it
will contain a dummy variable for Widowed and Became a Widow(er)
in both samples and an additional dummy variable for Married when the complete
sample is analyzed.
3.5 Tobit Regression:
As discussed, for each of the six dependent variables, a significant portion of the
observed values are equal to zero. For men approximately 61% of the observations
were reported as zero and for women more than 77% of the observations were reported
as zero. Also, as described above, in the case of days drinking per week, the data
83
shows a non-linearity at 7 days. Thus the linear regression model, which treats
observed zero values for each of the drinking behaviors and observed 7 days per week
as resulting from the individual optimization, may bias the estimated coefficients.
To address this potential bias, I estimate a model where the observed drinking behavior
is modeled as a latent variable model. y
it
represents the observed drinking behavior and
*
it
y represents the latent variable. For days drinking per week the latent variable is
modeled as left censored at zero and right censored at seven days. This yields the
following relationship:
≥
< >
≤
=
+ + + =
0 7
7 0
0 0
*
* * *
*
1 0
*
it
it it it
it
it
it i it it
y if
y and y if y
y if
y
X y μ δ β β
(2)
For drinks per day and drinks per week the latent variable is modeled as left censored at
zero. This yields the following relationship:
≤
=
+ + + =
otherwise y
y if
y
X y
it
it
it
it i it it
*
*
1 0
*
0 0
μ δ β β
(3)
I make the same exogeneity assumptions regarding the appropriate covariates as in the
linear regression model and estimate (2) and (3) using Tobit regression. (1), (2) and (3)
are estimated separately for men and women. The equations are also estimated
separately for both the complete sample and the sub-sample of those married at wave
three whose marriage does not end or ends with spousal death. Table 3.4 presents the
84
Linear Regression and Tobit Results. The estimated partial effects are shown for the
Tobit results.
Table 3.4A: Regression Results for Linear Regression and Tobit Models – Males Days
Drinking per Week
Linear
Regression
Tobit
(partial effects)
Linear
Regression
Tobit
(partial effects)
Age
0.02
(0.036)
0.011
(0.034)
0.032
(0.054)
0.02
(0.053)
Age Squared
-1.649E-4
(2.598E-4)
-1.345E-4
(2.502E-4)
-2.061E-4
(3.805E-4)
-1.52E-4
(2.76E-4)
Ethnicity = White
0.452
(0.049)***
0.478
(0.058)***
0.531
(0.072)***
0.624
(0.095)***
Years of Education
0.097
(0.007)***
0.109
(0.007)***
0.097
(0.009)***
0.115
(0.009)***
Married
-0.061
(0.067)
-0.088
(0.062)
Widowed
-0.018
(0.101)
-0.022
(0.1)
-0.144
(0.135)
-0.09
(0.145)
Became a Widow(er)
-0.102
(0.095)
-0.141
(0.101)
0.005
(0.12)
-0.09
(0.134)
Lagged Real Household
Per Capita Income
3.08E-4
(2.051E-4)
2.2E-4
(1.478E-4)
8.3E-4
(1.343E-4)***
5.52E-4
(9.661E-5)***
Unemployment Rate
0.025
(0.018)
0.033
(0.017)*
0.007
(0.025)
0.013
(0.023)
Real Alcohol Prices
0.494
(0.944)
0.502
(0.865)
-0.416
(1.341)
-0.322
(1.255)
Not reported: birth cohorts, constant
Standard Errors in parenthesis, clustered by individual - *** significant at 1%, ** significant at 5%, * significant at 10%
Complete Sample Married at Wave 3
85
Table 3.4B: Regression Results for Linear Regression and Tobit Models – Males Drinks
per Day
Linear
Regression
Tobit
(partial effects)
Linear
Regression
Tobit
(partial effects)
Age
-0.057
(0.029)**
-0.026
(0.025)
-0.018
(0.028)
-0.006
(0.03)
Age Squared
2.39E-4
(1.979E-4)
5.698E-5
(1.774E-4)
2.94E-7
(1.907E-4)
-4.153E-5
(2.136E-4)
Ethnicity = White
0.209
(0.039)***
0.27
(0.043)***
0.261
(0.052)***
0.34
(0.063)***
Years of Education
0.016
(0.004)***
0.051
(0.005)***
0.024
(0.005)***
0.053
(0.006)***
Married
-0.315
(0.057)***
-0.2
(0.046)***
Widowed
-0.204
(0.07)***
-0.114
(0.068)*
0.069
(0.079)
0.024
(0.088)
Became a Widow(er)
-0.01
(0.061)
-0.069
(0.067)
-0.039
(0.062)
-0.077
(0.078)
Lagged Real Household
Per Capita Income
5.74E-5
(6.44E-5)
9.986E-5
(7.785E-5)
2.006E-4
(5.61E-5)***
2.645E-4
(4.451E-5)***
Unemployment Rate
0.032
(0.012)***
0.032
(0.011)***
0.022
(0.014)
0.018
(0.014)
Real Alcohol Prices
-0.491
(0.594)
-0.171
(0.575)
-1.205
(0.76)
-0.712
(0.735)
Not reported: birth cohorts, constant
Standard Errors in parenthesis, clustered by individual - *** significant at 1%, ** significant at 5%, * significant at 10%
Complete Sample Married at Wave 3
86
Table 3.4C: Regression Results for Linear Regression and Tobit Models – Males Drinks
per Week
Linear
Regression Tobit
Linear
Regression Tobit
Age
-0.117
(0.148)
-0.059
(0.125)
0.063
(0.153)
0.04
(0.153)
Age Squared
3.295E-4
(0.001)
-4.605E-5
(7.798E-4)
-8.253E-4
(0.001)
-7.528E-4
(0.001)
Ethnicity = White
1.184
(0.182)***
1.474
(0.206)***
1.277
(0.248)***
1.752
(0.315)***
Years of Education
0.091
(0.02)***
0.263
(0.023)***
0.119
(0.024)***
0.272
(0.028)***
Married
-1.199
(0.291)***
-0.824
(0.229)***
Widowed
-0.821
(0.343)**
-0.482
(0.33)
0.232
(0.406)
0.065
(0.434)
Became a Widow(er)
0.242
(0.341)
-0.16
(0.336)
0.174
(0.332)
-0.162
(0.384)
Lagged Real Household
Per Capita Income
3.725E-4
(3.318E-4)
3.899E-4
(3.899E-4)
0.001
(3.459E-4)***
0.002
(2.484E-4)***
Unemployment Rate
0.119
(0.062)*
0.142
(0.056)**
0.098
(0.079)
0.088
(0.071)
Real Alcohol Prices
-1.128
(3.057)
-0.03
(2.837)
-4.885
(3.963)
-2.922
(3.683)
Not reported: birth cohorts, constant
Standard Errors in parenthesis, clustered by individual - *** significant at 1%, ** significant at 5%, * significant at 10%
Complete Sample Married at Wave 3
87
Table 3.4D: Regression Results for Linear Regression and Tobit Models – Females
Days Drinking per Week
Linear
Regression
Tobit
(partial effects)
Linear
Regression
Tobit
(partial effects)
Age
0.05
(0.016)***
0.043
(0.016)***
0.088
(0.029)***
0.07
(0.03)**
Age Squared
-3.377E-4
(1.182E-4)***
-3.579E-4
(1.247E-4)***
-6.171E-4
(2.214E-4)***
-5.709E-4
(1.903E-4)**
Ethnicity = White
0.396
(0.025)***
0.512
(0.036)***
0.484
(0.039)***
0.661
(0.07)***
Years of Education
0.086
(0.005)***
0.111
(0.006)***
0.091
(0.007)***
0.117
(0.008)***
Married
0.124
(0.039)***
0.098
(0.035)***
Widowed
-0.138
(0.041)***
-0.137
(0.043)***
-0.272
(0.054)***
-0.206
(0.061)***
Became a Widow(er)
0.124
(0.037)***
0.112
(0.041)***
0.128
(0.048)***
0.06
(0.054)
Lagged Real Household
Per Capita Income
4.266E-4
(2.148E-4)**
1.79E-4
(1.207E-4)*
8.976E-4
(1.276E-4)***
5.709E-4
(7.437E-5)***
Unemployment Rate
0.045
(0.011)***
0.043
(0.01)***
0.057
(0.018)***
0.053
(0.017)***
Real Alcohol Prices
-1.069
(0.596)*
-0.499
(0.522)
-2.134
(1.003)**
-1.215
(0.894)
Not reported: birth cohorts, constant
Standard Errors in parenthesis, clustered by individual - *** significant at 1%, ** significant at 5%, * significant at 10%
Complete Sample Married at Wave 3
88
Table 3.4E: Regression Results for Linear Regression and Tobit Models – Females
Drinks per Day
Linear
Regression
Tobit
(partial effects)
Linear
Regression
Tobit
(partial effects)
Age
-0.003
(0.01)
0.012
(0.009)
-0.008
(0.015)
0.01
(0.014)
Age Squared
-1.54E-5
(7.12E-5)
-1.141E-4
(6.327E-5)*
3.05E-5
(1.077E-4)
-9.606E-5
(1.016E-4)
Ethnicity = White
0.149
(0.016)***
0.224
(0.019)***
0.172
(0.026)***
0.265
(0.034)***
Years of Education
0.029
(0.002)***
0.049
(0.002)***
0.029
(0.003)***
0.047
(0.003)***
Married
0.005
(0.019)
0.027
(0.017)
Widowed
-0.05
(0.02)**
-0.056
(0.021)***
-0.052
(0.023)**
-0.056
(0.027)**
Became a Widow(er)
0.045
(0.021)**
0.046
(0.021)**
0.027
(0.023)
0.007
(0.025)
Lagged Real Household
Per Capita Income
1.725E-4
(8.81E-5)*
1.03E-4
(5.416E-5)*
3.583E-4
(5.58E-5)***
2.414E-4
(3.191E-5)***
Unemployment Rate
0.023
(0.005)***
0.023
(0.005)***
0.025
(0.008)***
0.024
(0.007)***
Real Alcohol Prices
-0.156
(0.276)
-0.099
(0.252)
-0.142
(0.42)
-0.187
(0.392)
Not reported: birth cohorts, constant
Standard Errors in parenthesis, clustered by individual - *** significant at 1%, ** significant at 5%, * significant at 10%
Complete Sample Married at Wave 3
89
Table 3.4F: Regression Results for Linear Regression and Tobit Models – Females
Drinks per Week
Linear
Regression
Tobit
(partial effects)
Linear
Regression
Tobit
(partial effects)
Age
0.043
(0.043)
0.069
(0.035)*
0.045
(0.067)
0.075
(0.058)
Age Squared
-3.583E-4
(2.96E-4)
-6.706E-4
(2.235E-4)**
-3.72E-4
(4.797E-4)
-7.232E-4
(4.822E-4)
Ethnicity = White
0.63
(0.061)***
0.966
(0.077)***
0.783
(0.094)***
1.205
(0.139)***
Years of Education
0.115
(0.01)***
0.203
(0.011)***
0.116
(0.013)***
0.199
(0.014)***
Married
0.113
(0.084)
0.149
(0.071)**
Widowed
-0.271
(0.084)***
-0.27
(0.086)***
-0.396
(0.098)***
-0.323
(0.113)***
Became a Widow(er)
0.321
(0.098)***
0.258
(0.087)***
0.299
(0.118)**
0.122
(0.109)
Lagged Real Household
Per Capita Income
7.782E-4
(3.963E-4)*
4.471E-4
(2.235E-4)*
0.002
(2.81E-4)***
9.643E-4
(1.438E-4)***
Unemployment Rate
0.082
(0.023)***
0.09
(0.02)***
0.107
(0.034)***
0.107
(0.031)***
Real Alcohol Prices
-1.905
(1.229)
-0.88
(1.044)
-3.218
(1.895)*
-1.85
(1.67)
Not reported: birth cohorts, constant
Standard Errors in parenthesis, clustered by individual - *** significant at 1%, ** significant at 5%, * significant at 10%
Complete Sample Married at Wave 3
3.6 Results
Examining first the results for elderly men, I find that whites drink both more often and
more drinks per episode. As is often the case, the parameter estimates from the linear
regression are similar in magnitude and significance to the Tobit estimates. With respect
to ethnicity, analysis on the complete population and the sub-population yield similar
results. The estimates show that whites drink more than one drink more per week
compared to other ethnicities as the baseline. Likewise more highly educated
individuals drink significantly more frequently and more per day than less educated.
Following the Tobit estimates for the whole population, an additional four years of school
would imply that an individual would consume approximately one more drink per week.
90
These results are consistent with the literature on drinking behavior that consistently
finds whites and higher educated drink more (Kerr et al. 2009). An individual’s age does
not appear to impact male drinking behavior.
The analysis on the effect of the business cycle as measured by the unemployment rate
offers inconclusive evidence that elderly men drink more when the economy is weaker.
The results of the analysis on the whole population, particularly for number of drinks per
day, seem significantly impacted by the unemployment rate. Higher unemployment
rates in the census division implies greater drinking behavior. However, while the sign of
the effect is still positive, the significance of these results do not hold up in the sub-
sample analysis. Lagged real income seems to positively impact drinking behavior.
However the effect is significant only in the analysis on the married sub-sample. For
elderly male behavior, alcohol prices at the census CPI level of aggregation do not seem
to significantly impact behavior.
Examining the impact of spousal death, I find that, for elderly men, becoming a widow in
the particular period does not impact drinking behavior significantly. In fact, for some of
the specifications, the linear regression estimate is positive while the Tobit estimate is
negative and neither is significant. Examining the whole population, I find that both
married men and widowed men drink less than their single counterparts. However,
looking at just those men married at wave 3, I find no significant impact on drinking
behavior for men who become widowers.
To understand the magnitude of the estimated impacts, I calculate the elasticity of
demand of each drinking behavior with respect to each of the independent variables.
91
The elasticities are calculated at the mean values of the independent variables and with
respect to the linear regression estimates. The elasticity measures the expected
percentage change in the drinking behavior for a one percent change in the independent
variable. Table 3.5 shows the mean values of the variables and the elasticity
calculations for elderly men for each of the three dependent variables.
Table 3.5: Elasticities of Demand – Male Drinking Behavior
Mean x
Value
Male Drink
Days
Male Drink
Number
Male Drink
Week
Age 68.7431
0.972
(1.736)
-4.418
(2.24)**
-2.348
(2.964)
Age Squared 4818.1
-0.553
(0.871)
1.295
(1.071)
0.462
(1.421)
Ethnicity = White 0.843427
0.265
(0.029)***
0.198
(0.037)***
0.29
(0.044)***
Years of Education 12.2755
0.827
(0.06)***
0.224
(0.058)***
0.326
(0.071)***
Married 0.804508
-0.034
(0.038)
-0.285
(0.051)***
-0.28
(0.067)***
Widowed 0.093053
-0.001
(0.007)
-0.021
(0.007)***
-0.022
(0.009)**
Became a Widow(er) 0.023166
-0.002
(0.002)
-2.663E-4
(0.002)
0.002
(0.002)
Lagged Real Household Per Capita
Income
173.336
0.037
(0.025)
0.011
(0.013)
0.019
(0.017)
Unemployment Rate 4.82115
0.085
(0.061)
0.176
(0.063)***
0.166
(0.086)*
Real Alcohol Prices 1.00957
0.347
(0.663)
-0.558
(0.675)
-0.331
(0.897)
Not reported: birth cohorts, constant
Standard Errors in parenthesis, clustered by individual - *** significant at 1%, ** significant at 5%, * significant at 10%
The elasticities illustrate the magnitude of the impact of the particular independent
variable on the drinking behavior. Other than the elasticity with respect to age for
number of drinks per day, all of the significant elasticities are less than 0.35 showing that
while male drinking behavior seems to respond significantly to a number of independent
92
variables it does so relatively inelastically. In particular, the elasticity with respect to the
marital status variables are significant for both the measures of the number of drinks.
For the Married dummy the elasticity is significant at the 99% level and equal to -0.285
for drinks per day and significant at the 99% level and equal to -0.28 for drinks per week.
Interpreting the elasticity of a dummy variable is based on the non-plausible event that
the individual were 1% more married than the mean. This increase would result in a
0.28% decrease in the quantity of drinks consumed per day and quantity of drinks
consumed per week.
Examining female drinking behavior reveals that it depends significantly both on non-
time-changing and time changing characteristics. Similar to men, women who are white
and more highly educated drink significantly more often and more per day than
minorities and individuals who are less educated. The estimated Tobit partial effect of
being white on drinks per week is just less than one additional drink per week. However,
in contrast to elderly men, age seems to significantly impact the number of days women
drink. The effect appears to be non-linear. Using the complete sample linear regression
estimates, women continue to drink more days per week until approximately the age of
72 and then begin to decrease the number of days they drink
8
.
The analysis shows that women are impacted by the economic environment and drink
more on average as unemployment rises. Women drink significantly more days per
week and drinks per day as unemployment rises. This result is consistent across all the
specifications. This contrasts with men who seem to only be impacted by the
unemployment in the number of days they drink. Similar to men, women’s drinking
8
Note that if drinking impacts death and therefore heavy drinkers exit the sample before attaining
older ages, this estimated non-linearity may result from this non-random sample attrition.
93
behavior does not appear to be significantly impacted by prices
9
. Women do seem to be
impacted by household income with lagged income positively impacting both days
drinking and drinks per day. The impact of lagged income is more significant when the
analysis is conducted on those women who were married at wave 3 as the results are
significant only to the 10% level in the complete population.
Most interesting for my analysis, women appear to be significantly impacted by being a
widow and by becoming a widow in a particular period. Recall that the dummy variable
Widowed equals 1 when the women is a widow in a particular period. The dummy
variable Became a widow will equal 1 in the period the woman becomes a widow. The
results reveal an interesting relationship. Widowed women tend to drink both less
frequently and consume less when they drink than non-widowed women, a result which
holds in both the whole population and among those who were married at wave 3.
However, in all the specifications with the complete sample and for all the linear
regression analysis on those married at wave three, women drink more in the year they
become a widow. For example, looking at the results for those married at wave three, I
find that Widowed women drink -0.272 days less, but those that Became a Widow drank
0.128 days more. For each of the three behaviors and for both the linear regression and
Tobit models within the complete population these impacts are significant and
consistent. For the married population looking at the regression models for both days
drinking and drinks per week the effects are consistent and significant. In all cases, the
estimated decrease in drinking from being a widow is larger than the increase in the
period where the woman becomes a widow.
9
Note that the price levels used in the analysis are the real alcohol CPI prices at the census
division level. The fact that these do not seem to significantly impact drinking behavior may
reflect the high level of both product and geographic aggregation and should not necessarily be
interpreted that prices do not significantly impact frequency and quantity of drinking.
94
To examine the magnitude and economic significance of these estimates, I calculate the
elasticities with respect to the independent variables. The calculations are conducted at
the mean levels for each of the variables. Results are presented in table 3.6.
Table 3.6: Elasticities of Demand – Female Drinking Behavior
Mean x
Value
Female Drink
Days
Female Drink
Number
Female Drink
Week
Age 68.3586
4.749
(1.538)***
-0.547
(1.932)
2.381
(2.409)
Age Squared 4799.44
-2.253
(0.789)***
-0.2
(0.925)
-1.406
(1.161)
Ethnicity = White 0.812377
0.447
(0.026)***
0.327
(0.035)***
0.419
(0.039)***
Years of Education 12.0261
1.435
(0.085)***
0.958
(0.073)***
1.127
(0.098)***
Married 0.55939
0.096
(0.03)***
0.007
(0.029)
0.052
(0.039)
Widowed 0.300193
-0.058
(0.017)***
-0.041
(0.016)**
-0.066
(0.02)***
Became a Widow(er) 0.045736
0.008
(0.002)***
0.006
(0.003)**
0.012
(0.004)***
Lagged Real Household Per Capita
Income
144.013
0.085
(0.043)**
0.067
(0.034)*
0.092
(0.047)**
Unemployment Rate 4.8287
0.301
(0.075)***
0.303
(0.069)***
0.323
(0.091)***
Real Alcohol Prices 1.00921
-1.5
(0.835)*
-0.427
(0.753)
-1.572
(1.016)
Not reported: birth cohorts, constant
Standard Errors in parenthesis, clustered by individual - *** significant at 1%, ** significant at 5%, * significant at 10%
The elasticity calculations reveal that, at the mean, women’s drinking behaviors are
significantly and negatively impacted by being a widow. Their drinking behavior was
positively and significantly impacted by becoming a widow. However in both cases the
effects are of relatively small magnitude as female drinking behavior seems to respond
relatively inelastically to changes in marital status and to spousal death. The elasticity at
the mean for Widow is between -0.041 and -0.066. The elasticity for becoming a widow
95
is of an order of magnitude smaller and varies from 0.006 to 0.012. The analysis does
show a significant and positive income elasticity, although the magnitude is small. The
elasticity with respect to the unemployment rate is significant at the 1% level for each of
the drinking behaviors and approximately 0.3 in each case. While still inelastic this is a
relatively larger response as compared to other covariates.
3.7 Conclusion
This analysis examines determinants of drinking behaviors among elderly men and
women. In particularly, it examines how spousal death impacts drinking behavior. A
longitudinal dataset – the Health and Retirement Study – is used to construct a variable
that codes for the individual’s spouse dying between survey observations. Using two
different assumptions on the exogeneity of spousal death the analysis examines its
impact on drinking frequency and quantity. The analysis controls for economic variables
at the census division level and includes individual demographics.
The results support several accepted findings in the literature. First, male and female
drinking behaviors differ significantly. Second, individuals who are white and more
highly educated tend to drink both more frequently and consume more on days that they
drink. When income impacts drinking behavior higher income individuals drink more.
Age is found to only significantly impact the number of days that a woman drinks and
does so in a non-linear fashion with individuals drinking more until about 72 and then
drinking less. Age does not significantly impact the quantity of alcohol women consume
nor does age significantly impact male drinking behavior.
96
Interestingly, macro-economic variables, as measured by census division
unemployment, positively and significantly impact elderly female drinking. The estimates
for the impact of unemployment rates on men are positive but not significant, indicating
that female drinking behavior is more sensitive to changes. Likewise, women seem to
be impacted significantly by family income where men are not. However the elasticities
of these effects are small.
Women are significantly impacted by their marital status and by their spouse’s death
while men are note. For each of the three different drinking behaviors, being a widowed
significantly and negatively impacted female drinking levels. For women, becoming a
widow in the between current survey waves significantly and positively impacted drinking
behavior. However, the magnitudes of these effects, as measured by the elasticities at
the mean values was not large.
A natural extension of this research is to explore the impact of changes in marital status
and economic variables in a dynamic context. The HRS dataset lends itself to this
analysis due to its panel nature. Unfortunately, the publically available data provides
information only on the individuals’ census division of residence. The economic
information associated with census division does not provide for powerful enough
instruments to control for the endogeneity of lagged drinking variables. Gathering the
HRS private information and the associated more precise economic data will be a topic
of future research.
97
Chapter 3 References
Arcidiacono, Peter and Sieg, Holger and Sloan, Frank (2007) “Living Rationally under
the Volcano? An Empirical Analysis of Heavy Drinking and Smoking” International
Economic Review 48(1): 37-65
Becker, Gary and Murphy, Kevin (1988) “A Theory of Rational Addiction” Journal of
Political Economy 96 #4: 675-700
Brennan, Penny and Moos, Rudolf (1990) “Life Stressors, Social Resources, and Late-
Life Problem Drinking” Psychology and Aging 5 #4: 491-501
Breslow, Rosalind and Faden, Vivian and Smothers, Barbara (2003) “Alcohol
Consumption by Elderly Americans” Journal of Studies on Alcohol 64: 884-892
Breslow, Rosalind and Smothers, Barbara (2004) “Drinking Patterns of Older Americans:
National Health Interview Surveys 1997-2001” Journal of Studies on Alcohol 65(2):
232-240
Gallo, William and Bradley, Elizabeth and Siegel, Michele and Kasl, Stanislav (2001)
“The Impact of Involuntary Job Loss on Subsequent Alcohol Consumption by Older
Workers: Findings from the Health and Retirement Survey” Journal of Gerontology
56B #1: S3-S9
Glass, Thomas and Prigerson, Holly and Kasl, Stanislav, Mendes de Leon, Carlos
(1995) “The Effects of Negative Life Events on Alcohol Consumption Among Older
Men and Women” Journal of Gerontology 50B #4: S205-S216
Grossman, Michael and Chaloupka, Frank and Sirtalan, Ismail (1995) “An Empirical
Analysis of Alcohol Addiction: Results from the Monitoring the Future Panels” NBER
Working Paper No. 5200
Karalamangla, Arun and Zhou, Kefei and Ruben, David and Moore, Alison (2006)
“Longitudinal trajectories of heavy drinking in adults in the United States of America”
Addiction 101: 91-99
Karalamangla, Arun and Sarkisian, Catherine and Kado, Deborah and Dedes, Howard
and Liao, Diana and Kim, Sungjin and Reuben, David and Greendale, Gail and
Moore, Alison (2009) “Light to Moderate Alcohol Consumption and Disability:
Variable Benefits by Health Status” American Journal of Epidemiology 169: 96-104
Kerr, William and Greenfield, Thomas and Bond, Jason and Rehm, Jurgen (2009) “Age-
period-cohort modeling of alcohol volume and heavy drinking days in US National
Alcohol Surveys: divergence in younger and older adult trends” Addiction 104: 27-37
Merrick, Elizabeth and Horgan, Constance and Hodgkin, Dominic and Garnick, Deborah
and Houghton, Susan and Panas, Lee and Saitz, Richard and Blow, Frederic (2008)
“Unhealthy Drinking Patterns in Older Adults: Prevalence and Associated
Characteristics” Journal of the American Geriatric Society 56: 214-223
98
Moore, Alison and Gould, Robert and Reuben, David and Greendale, Gail and Carter,
Kallin and Zhou, Kefei and Karlamangla, Arun (2005) “Longitudinal Patterns and
Predictors of Alcohol Consumption in the United States” American Journal of Public
Health 95(3): 458-464
Ostermann, Jan and Sloan, Frank (2001), “Effects of Alcohol Consumption on Disability
among the Near Elderly: A Longitudinal Analysis” Milbank Quarterly 79 #4: 487-515
Perreria, Krista and Sloan, Frank (2002) “Excess alcohol consumption and health
outcomes: a 6-year follow-up of men over age 50 from the health and retirement
study” Addiction 97: 301-310
Veenstra, Marja and Lemmens, Paul and Friesema, Ingrid and Garretsen, Henk and
Knottnerus, Andre and Zweitering, Paul (2006) “A literature overview of the
relationship between life-events and alcohol use in the general population” Alcohol &
Alcoholism 41 #4: 455-463
Wagenaar, Alexander and Salois, Matthew and Komro, Kelli (2009), “Effects of beverage
alcohol price and tax levels on drinking: a meta analysis of 1003 estimates from 112
studies” Addiction 104: 179-190
Welte, John and Mirand, Amy (1993) “Drinking, Problem Drinking and Life Stressors in
the Elderly General Population” Journal of Studies in Alcohol Jan 1995: 67-73
99
Bibliography
American Diabetes Association (2003), “Implications of the Diabetes Control and
Complications Trial,” Diabetes Care 26: S25-S27
Arcidiacono, Peter and Sieg, Holger and Sloan, Frank (2007) “Living Rationally under
the Volcano? An Empirical Analysis of Heavy Drinking and Smoking” International
Economic Review 48(1): 37-65
Becker, Gary and Murphy, Kevin (1988) “A Theory of Rational Addiction” Journal of
Political Economy 96 #4: 675-700
Bihan, Helene and Laurent, Silvana and Sass, Catherine and Nguyen, Gerard and Huot,
Caroline and Moulin, Jean Jacques and Guegen, Rene and Le Toumeline,
Philippe and Le Clesiau, Harve and La Rosa, Emelio and Reach, Gerard and
Cohen, Regis (2005) “Association Among Individual Deprivation, Glycemic
Control, and Diabetes Complications” Diabetes Care 28:2680-2685
Brennan, Penny and Moos, Rudolf (1990) “Life Stressors, Social Resources, and Late-
Life Problem Drinking” Psychology and Aging 5 #4: 491-501
Breslow, Rosalind and Faden, Vivian and Smothers, Barbara (2003) “Alcohol
Consumption by Elderly Americans” Journal of Studies on Alcohol 64: 884-892
Breslow, Rosalind and Smothers, Barbara (2004) “Drinking Patterns of Older Americans:
National Health Interview Surveys 1997-2001” Journal of Studies on Alcohol 65(2):
232-240
Campbell, K and Waters, E and O’Meara, S and Summerbell, C (2001) “Interventions for
preventing obesity in childhood. A systematic review” Obesity Reviews 2:149-
157
Chalew, Stuart and Gomwz, Ricardo and Butler, Ashley and Hempe, James and
Comptom, Terry and Rao, Jayashree and Vargas, Alfonso (2000) “Predictors of
glycemic control in children with Type 1 diabetes: The importance of race”
Journal of Diabetes and Its Complications 14: 71-77
Davis, K and Christoffel, K.K. (1994) “Obesity in Preschool and School-age Children,
Treatment Early and Often May Be Best” Archives Pediatric Adolescent Medicine
148:1256-1261
Dehghan, Manshid and Akhtar-Danesh, Noori and Merchant, Anwar (2005) “Childhood
obesity, prevalence and prevention” Nutrition Journal 4:24-32
Diabetes Control and Complications Trial Research Group (2003) “The Effect of
Intensive Treatment of Diabetes on the Development and Progression of Long-
Term Complications in Insulin-Dependent Diabetes Mellitus” New England
Journal of Medicine 329: 977-986
100
Doak, C.M. and Visscher, T.L.S. and Renders, C.M. and Seidell, J.C. (2006) “The
prevention of overweight and obesity in children and adolescents: a review of
interventions and programmes” Obesity Review 7: 111-136
Edmundson, Elizabeth and Parcel, Guy and Feldman, Henry and Elder, John and Perry,
Cheryl and Johnson, Carolyn and Williston, B.J. and Stone, Elaine and Yang,
Mingua and Lytle, Leslie and Webber, Larry (1996) “The Effects of the Child and
Adolescent Trial for Cardiovascular Health upon Psychosocial Determinants of
Diet and Physical Activity Behavior” Preventive Medicine 25: 442-454
Ellis, Deborah and Frey, Maureen and Naar-King, Sylvie and Templin, Thomas and
Cunningham, Phillippe and Cakan, Nedim (2005), “Use of Multisystemic Therapy
to Improve Regimen Adherence Among Adolescents With Type 1 Diabetes in
Chronic Poor Metabolic Control” Diabetes Care 28:7 1604-1610
Flodmark, Carl-Erik and Ohlsson, Torsten and Ryden, Olof and Sveger, Tomas (1993)
“Prevention of Progression to Severe Obesity in a Group of Obese
Schoolchildren Treated with Family Therapy” Pediatrics 91:880-884.
Gallo, William and Bradley, Elizabeth and Siegel, Michele and Kasl, Stanislav (2001)
“The Impact of Involuntary Job Loss on Subsequent Alcohol Consumption by Older
Workers: Findings from the Health and Retirement Survey” Journal of Gerontology
56B #1: S3-S9
Glass, Thomas and Prigerson, Holly and Kasl, Stanislav, Mendes de Leon, Carlos
(1995) “The Effects of Negative Life Events on Alcohol Consumption Among Older
Men and Women” Journal of Gerontology 50B #4: S205-S216
Grossman, Michael and Chaloupka, Frank and Sirtalan, Ismail (1995) “An Empirical
Analysis of Alcohol Addiction: Results from the Monitoring the Future Panels” NBER
Working Paper No. 5200
Ham, John and LaLonde, Robert (1996) “The Effect of Sample Selection and Initial
Conditions in Duration Models: Evidence from Experimental Data on Training”
Econometrica 64:1 175-205
Ham, John and Rea, Samuel (1987) “Unemployment Insurance and Male
Unemployment Duration in Canada” Journal of Labor Economics 5:3 325-353
Hanberger, Lena and Lindblad, Bengt and Samuelsson, Ulf and Ludvigsson, Johnny
(2008) “A1C in Children and Adolescents With Diabetes in Relation to Certain
Clinical Parameters” Diabetes Care 35:5 (927-929)
Heckman, James and Singer, Burton (1984) “Econometric Duration Analysis” Journal of
Econometrics 24: 63-132
101
Hedley, Allison and Ogden, Cynthia and Johnson, Clifford and Carroll, Margaret and
Curtin, Lester and Flegal, Katherine (2004) “Prevalence of Overweight and
Obesity Among US Children, Adolescents and Adults, 1999-2002” Journal of the
American Medical Association 291:2847-2850
James, Philip and Leach, Rachel and Kalamara, Eleni and Shayeghi, Maryam (2001)
“Worldwide Obesity Epidemic” Obesity Research 9:228S-233S
Kamath, CC and Vickers, KS and Eherlich, A and McGoern, L and Johnson, J and
Singhal, V and Paulo, R and Hettinger, A and Erwin, PJ and Montori, VM (2008)
“Clinical review: behavioral interventions to prevent childhood obesity. A
systematic review and meta-analysis of randomized trials” Journal of Clinical
Endocrinology & Metabolism 10.1210: 2006-2411
Karalamangla, Arun and Zhou, Kefei and Ruben, David and Moore, Alison (2006)
“Longitudinal trajectories of heavy drinking in adults in the United States of America”
Addiction 101: 91-99
Karalamangla, Arun and Sarkisian, Catherine and Kado, Deborah and Dedes, Howard
and Liao, Diana and Kim, Sungjin and Reuben, David and Greendale, Gail and
Moore, Alison (2009) “Light to Moderate Alcohol Consumption and Disability:
Variable Benefits by Health Status” American Journal of Epidemiology 169: 96-104
Kaufman, Francine Ratner and Halvorson, Mary and Carpenter, Sue (1999) “Association
Between Diabetes Control and Visits to a Multidisciplinary Pediatric Diabetes
Clinic” Pediatrics 103:948
Kerr, William and Greenfield, Thomas and Bond, Jason and Rehm, Jurgen (2009) “Age-
period-cohort modeling of alcohol volume and heavy drinking days in US National
Alcohol Surveys: divergence in younger and older adult trends” Addiction 104: 27-37
Maffeis, Clauio and Tato, Luciano (2001) “Long-Term Effects of Childhood Obesity on
Morbidity and Mortality” Hormone Research 55: 42-45
Merrick, Elizabeth and Horgan, Constance and Hodgkin, Dominic and Garnick, Deborah
and Houghton, Susan and Panas, Lee and Saitz, Richard and Blow, Frederic (2008)
“Unhealthy Drinking Patterns in Older Adults: Prevalence and Associated
Characteristics” Journal of the American Geriatric Society 56: 214-223
Moore, Alison and Gould, Robert and Reuben, David and Greendale, Gail and Carter,
Kallin and Zhou, Kefei and Karlamangla, Arun (2005) “Longitudinal Patterns and
Predictors of Alcohol Consumption in the United States” American Journal of Public
Health 95(3): 458-464
Monzavi, Roshanak and Dreimane, Daina and Geffner, Mitchell and Braun, Sharon and
Conrad, Barry and Klier, Mary and Kaufman, Francine R. (2006) “Improvement in
Risk Factors for Metabolic Syndrome and Insulin Resistance in Overweight
Youth Who Are Treated With Lifestyle Intervention” Pediatrics 117: 1111-1118
102
Ogden, Cynthia and Carroll, Margaret and Flegal, Katherine (2008) “High Body Mass
Index for Age Among US Children and Adolescents, 2003-2006” Journal
American Medical Association 299.20: 2401-2405
Ostermann, Jan and Sloan, Frank (2001), “Effects of Alcohol Consumption on Disability
among the Near Elderly: A Longitudinal Analysis” Milbank Quarterly 79 #4: 487-515
Perreria, Krista and Sloan, Frank (2002) “Excess alcohol consumption and health
outcomes: a 6-year follow-up of men over age 50 from the health and retirement
study” Addiction 97: 301-310
Songer, Thomas and Zhang, Ping and Engelgau, Michael (2004) “Cost-of-Illness Studies
in Diabetes Mellitus: Review Article” Pharmacoeconomics. 22(3):149-164
Valverde, M.A. and Patin, R. V. and Olveira, F. L. C. and Vitolo, M. R. (1998) “Outcomes
of obese children and adolescents enrolled in a multidisciplinary health program”
International Journal of Obesity 22:513-519
Vanelli M and Chiarelli F and Chiari G and Tumini S and Costi G and di Ricco L and
Zanasi P and Catino M and Capuano C and Porcelli C and Adinolfi B and Cieri F
and Giacalone T and Casni A (1999) “Metabolic Control in Children and
Adolescents: Experience of Two Italian Regions” Journal of Pediatric
Endocrinology and Metabolism 12:403
Veenstra, Marja and Lemmens, Paul and Friesema, Ingrid and Garretsen, Henk and
Knottnerus, Andre and Zweitering, Paul (2006) “A literature overview of the
relationship between life-events and alcohol use in the general population” Alcohol &
Alcoholism 41 #4: 455-463
Wagenaar, Alexander and Salois, Matthew and Komro, Kelli (2009), “Effects of beverage
alcohol price and tax levels on drinking: a meta analysis of 1003 estimates from 112
studies” Addiction 104: 179-190
Wang, Youfu and Monteiro, Carlos and Popkin, Barry (2002) “Trends of obesity and
underweight in older children and adolescents in the United States, Brazil, China
and Russia” American Journal Clinical Nutrition 75: 971-977
Weintrob, Naomi and Benzaquen, Hadassa and Galatzer, Avinoam and Shalitin, Shlomit
and Lazar, Liora and Fayman, Gila and Lilos, Pearl and Dickerman, Zvi and
Philiip, Moshe (2003) “Comparison of Continuous Subcutaneous Insulin Infusion
and Multiple Daily Injection Regimens in Children with Type 1 Diabetes: A
Randomized Open Crossover Trial” Journal of the American Academy of
Pediatrics 112.3: (559-564)
Welte, John and Mirand, Amy (1993) “Drinking, Problem Drinking and Life Stressors in
the Elderly General Population” Journal of Studies in Alcohol Jan 1995: 67-73
Yancey, Atronette and Ory, Marcia and Davis, Sally (2006) “Dissemination of Physical
Activity Promotion Interventions in Underserved Populations” American Journal
of Preventive Medicine 31(4): 82-91
103
Appendices
Appendix A: Detailed Information on the Kids N Fitness Program
The following text comes directly from the KnF CHLA Institutional Review Board filing.
The filing forms part of the informed consent that each child and family signed upon
entering study. The document was intended to describe what would happen and
therefore is in the future tense.
Elementary school children in the 3rd grade and up, who are enrolled in after school
programs at selected elementary schools will be invited to participate in the KnF
program. Individuals must have medical clearance from a physician or nurse practitioner
due to the exercise requirement of the program prior to their enrollment in KnF.
Individuals will be excluded if they are disinterested, non-ambulatory, are unable to read
and write, or due other medical reasons noted by their physician, or allied health team
staff.
The six session program includes nutrition education, exercise, and behavior change
components lasting 90 to 120 minutes per session. Parents/guardians participate in a
parent support and nutrition class. Participants will be strongly encouraged to participate
in all 6 sessions. Students will engage in exercise led by the after school teacher for at
least 45 minutes 3 times a week on the days that the nutrition program is not offered.
The class ends with dinner where families are served a healthy meal with appropriate
portion sizes.
At enrollment, child subjects will complete the KnF (CATCH) knowledge questionnaire,
and will answer a series of questions based on their activities and eating behavior of the
past 24 hours (24 hour Activity and Eating recall- SPAN). A parent/caregiver will be
asked to complete sections entitled Global Health, Your Child’s Physical Activities, Your
Child’s Everyday Activities, Well-Being, Self Esteem, and Facts about you) of the Child
Health Questionnaire. Since parents are actively involved in the program, it is our desire
to determine whether or not their participation in the program has any effect on their own
weight. Therefore, we will perform pre and post measurements of their height, weight,
and percentage of body fat on the parents. Right before starting the program and during
the last session, we will take height, weight, body fat percentage, blood pressure, and
waist circumference of the children. Students will work on weekly group fitness goals
that they will work on outside of the program during the course of the intervention.
Subjects will be instructed to complete the CHLA KnF logbook daily for the entire six
weeks to learn more about their routine eating and exercise habits. Parents will be asked
to sign off each day data is recorded to ensure their involvement and promote accurate
responses. They will also work on an individual and group fitness goal every week. They
will rate their success at achieving their goals every week.
At week six, height, weight, body fat percentage, BMR, Fat Mass, and Fat Free Mass,
and complete the written measures described below. Parent’s height, weight, and body
fat percentage will also be recorded.
104
At the end of the pilot, (three 6 week interventions have been completed during a school
year), all the children will again complete a 24 activity and eating recall, do repeat blood
work, and complete the CATCH health behavior questionnaire, and CHQ.
Specific classes will for the six week study program are described below:
1. Pyramid Power: The first class consists of name games and fun activities designed to
allow participants get to know each other as well as to promote exercise. In the nutrition
education portion of the class, the Food Guide Pyramid and the food groups are
introduced. These concepts are reinforced with group activities, including Food Guide
Pyramid bingo. The participants are given the KnF Logbook as homework to keep track
of their eating and exercise habits for the succeeding seven days. The logbook is
designed to take no more than five minutes a day to complete.
2. Portion-size Power- During the exercise class, kids rotate through different exercises
stations in small groups. Subjects are introduced to the food label and serving size. They
learn by measuring actual food samples onto serving plates/containers in effort to
emulate portions normally used at home, then find and measure the actual serving size.
A coordinating demonstration will be given showing advised food portion-sizes.
3. Fat counts! – The exercise for this session is resistance training with a theraband.
Participants are instructed how to use a theraband in several exercises to increase
muscle strength and tone. They will then learn how to utilize the Nutrition Food Label
with a focus on total fat compared with total calories per serving for various food items. A
taste-test of low-fat foods will follow.
4. Sugar Savvy- Participants warm up with various soccer drills to help them learn and/or
practice the skills needed to play soccer. Then a game of indoor soccer is played with
multiple balls to help keep everyone involved in the game and build exercise endurance.
. Subjects will then continue product label reading exercises with a focus on sugar
content in foods.
5. Healthy Cooking- Relay races and “bombing the cones” are the athletic activities for
week 5. Parents learn healthy cooking techniques and children learn about the
importance of fiber in their diet and the eating of fruits and vegetables.
6. Dining out- Subjects will play indoor dodgeball followed by stretching exercises
designed to improve flexibility. They will then be taught techniques for eating out.
Strategies will be exchanged, as well as an opportunity to practice ordering mock foods
using local restaurant menus.
Follow-up: At 3 to 6 months post pilot, heights, weights, BMI, Body fat %, BMR, Fat
Mass, and Fat Free Mass, and waist circumference will be taken of the children. Heights,
weights, BMI, Body fat %, BMR, Fat Mass, and Fat Free Mass will be taken of the
parents.
105
Appendix B: Sensitivity Analysis – Results with A1C 7.0% and 9.0%
Table B.1: Duration Model Conditional Hazard – Control Cutoff A1C 7.0%
Number of Spells
Number Individuals
Age
-0.008
(0.072)
-0.122
(0.083)
0.050
(0.080)
-0.058
(0.106)
Age Squared
-0.025
(0.315)
0.458
(0.369)
-0.178
(0.415)
0.387
(0.621)
Teen
0.032
(0.258)
-0.204
(0.222)
0.365
(0.228)
-0.376
(0.347)
Duration with
Diabetes
0.034
(0.033)
0.021
(0.023)
-0.092
(0.035)**
-0.092
(0.032)***
-0.031
(0.029)
-0.035
(0.023)
0.012
(0.055)
0.045
(0.038)
Ethnicity=White
0.210
(0.275)
0.214
(0.271)
0.598
(0.265)**
0.594
(0.265)**
0.073
(0.249)
0.059
(0.249)
0.164
(0.351)
0.173
(0.349)
Government
Insurance
-0.089
(0.296)
-0.091
(0.295)
-0.813
(0.302)***
-0.805
(0.301)***
0.099
(0.281)
0.107
(0.282)
0.377
(0.368)
0.332
(0.361)
Census Track
Poverty Rate
-0.007
(0.017)
-0.007
(0.017)
0.023
(0.016)
0.023
(0.016)
0.017
(0.017)
0.017
(0.017)
-0.038
(0.025)
-0.038
(0.025)
Census Track
Median Income
-0.028
(0.151)
-0.032
(0.149)
0.088
(0.125)
0.093
(0.125)
0.004
(0.135)
0.017
(0.135)
-0.705
(0.332)**
-0.741
(0.327)**
Duration=1
1.118
(0.254)***
1.122
(0.254)***
0.422
(0.239)*
0.449
(0.238)*
0.988
(0.255)***
0.995
(0.256)***
1.096
(0.348)***
1.135
(0.339)***
Duration=2
0.324
(0.329)
0.331
(0.329)
-0.053
(0.297)
-0.029
(0.296)
0.187
(0.317)
0.200
(0.317)
0.560
(0.433)
0.591
(0.427)
Constant
-2.326
(1.138)**
-2.392
(0.460)***
-1.025
(1.370)
-2.756
(0.439)***
-2.066
(1.263)
-1.544
(0.419)***
-0.717
(1.746)
-1.603
(0.527)***
Standard Errors in parenthesis - *** significant at 1%, ** significant at 5%, * significant at 10%
195 469 194 86
249 469 268 86
Spells Poor Control
Transition to Good Control
Spells Good Control
Transition to Poor Control
New Spells Interrupted Spells New Spells Interrupted Spells
106
Table B.2: Duration Model Conditional Hazard – Control Cutoff A1C 9.0%
Number of Spells
Number Individuals
Age
-0.428
(0.215)**
-0.046
(0.179)
0.338
(0.155)**
1.149
(0.281)***
Age Squared
1.127
(0.728)
-0.338
(0.680)
-1.465
(0.653)**
-6.070
(1.616)***
Teen
-0.714
(0.343)**
-0.891
(0.323)***
0.368
(0.302)
1.222
(0.287)***
Duration with
Diabetes
0.029
(0.034)
0.010
(0.031)
-0.031
(0.040)
-0.068
(0.036)*
0.058
(0.029)**
0.033
(0.024)
-0.013
(0.033)
-0.028
(0.028)
Ethnicity=White
-0.087
(0.271)
-0.111
(0.269)
-0.017
(0.298)
-0.104
(0.295)
0.176
(0.249)
0.163
(0.250)
0.491
(0.266)*
0.519
(0.263)**
Government
Insurance
-0.139
(0.281)
-0.116
(0.279)
-0.048
(0.302)
0.035
(0.302)
0.709
(0.249)***
0.711
(0.249)***
0.799
(0.273)***
0.817
(0.271)***
Census Track
Poverty Rate
-0.007
(0.017)
-0.003
(0.017)
-0.007
(0.020)
-0.010
(0.019)
0.001
(0.015)
0.002
(0.015)
0.011
(0.015)
0.010
(0.015)
Census Track
Median Income
0.312
(0.187)*
0.348
(0.186)*
0.051
(0.194)
0.089
(0.186)
-0.097
(0.165)
-0.089
(0.166)
-0.001
(0.167)
-0.043
(0.164)
Duration=1
1.411
(0.354)***
1.486
(0.351)***
0.505
(0.308)
0.605
(0.303)**
1.280
(0.241)***
1.284
(0.241)***
0.400
(0.274)
0.346
(0.273)
Duration=2
0.929
(0.407)**
0.981
(0.404)**
0.024
(0.369)
0.101
(0.364)
0.673
(0.292)**
0.672
(0.291)**
0.344
(0.295)
0.305
(0.295)
Constant
5.116
(3.605)
-1.285
(0.613)**
0.022
(3.023)
0.339
(0.594)
-8.382
(2.410)***
-3.437
(0.474)***
-21.11
(4.275)***
-4.637
(0.457)***
Standard Errors in parenthesis - *** significant at 1%, ** significant at 5%, * significant at 10%
195 166 224 390
252 166 296 390
Spells Poor Control
Transition to Good Control
Spells Good Control
Transition to Poor Control
New Spells Interrupted Spells New Spells Interrupted Spells
107
Appendix C: Example Determining Diabetes Control State and Spell Information
The following table shows an example of the A1C measures for three patients in the
sample. This illustrates the raw data on the dependent variable in the dataset. Blank
values in the table illustrate the missing data from the original dataset
Table C.1: A1C Values for Three Sample Patients
AIC
2001-1
AIC
2001-2
AIC
2002-1
AIC
2002-2
AIC
2003-1
AIC
2003-2
AIC
2004-1
AIC
2004-2
AIC
2005-1
AIC
2005-2
Subject 1 7.10 7.20 7.25 7.20 8.30 9.95 10.43 8.35 8.30 8.00
Subject 2 7.80 7.70 10.00 11.60 8.05 11.90
Subject 3 7.65 8.10 7.80 9.30 8.55 8.00 8.20 9.40 10.00 12.60
I transform the A1C values into diabetes control states. The transformation depends on
the A1C cutoff value used. The primary analysis in the paper uses the accepted level of
A1C 8.0% as the cutoff. Appendix B shows the results using cutoffs of 7.0% and 9.0%.
The table below shows the diabetes control states using the 8.0% cutoff.
Table C.2: Example Transformation A1C to Diabetes Control States
State
2001-1
State
2001-2
State
2002-1
State
2002-2
State
2003-1
State
2003-2
State
2004-1
State
2004-2
State
2005-1
State
2005-2
Subject 1 Good Good Good Good Poor Poor Poor Poor Poor Poor
Subject 2 Good Good Poor Poor Poor Miss Poor Miss Miss Miss
Subject 3 Good Poor Good Poor Poor Poor Poor Poor Poor Poor
Based on the pattern of control states I determine the diabetes control spells. Subject 1
has two spells. The subject has an interrupted spell of good control that lasts for four
periods. In 2003-1 the subject begins a new spell of poor control that continues for six
periods and ends with a censored observation. Subject 2 also two spells. The subject
has an interrupted spell of good control that lasts for two periods. In 2002-1 the subject
108
begins a new spell of poor control that continues for three periods and ends with a
censored observation. Subject 3 has four spells. The subject has an interrupted spell of
good control that lasts one period. The subject then begins, in period 2001-2, a new
spell of poor control that lasts one period. In period 2002-1 the subject begins a new
spell of good control that lasts one period. In period 2002-2 the subject begins a new
spell of poor control that lasts seven periods and ends with a censored observation.
109
Appendix D: Wave 4 Example - Constructing the Became a widow(er) Variable
The following table illustrates the changes in marital status between wave 4 and wave 3.
The information that a particular respondent has the status Widowed in wave 4 but not
wave 3 is used to identify the fact that the respondent became a widow(er).
Table D.1: Marital Status Change between Weave 4 and Wave 3
Wave 3 Status Married / Partnered Separated / Divorced Widowed Never Married Totals
Married / Partnered 10,334 133 527 3 10,997
Separated / Divorced 70 1,125 122 11 1,328
Widowed 56 81 2,802 13 2,952
Never Married 6 14 21 380 421
Totals 10,466 1,353 3,472 407 15,698
Wave 4 Marital Status
From examining the change in marital status, I observe that 670 (527+122+21)
individuals are newly widow(ed) between the time of the wave 3 interview and the wave
4 interview. For these 670 individuals the value of Became a widow(er) is set to one for
their wave 4 observations.
The following table illustrates how the information on number of times widowed changed
between Wave 3 and Wave 4. If the number of times widowed increased between the
waves, I conclude the individual was widowed in the time between Wave 3 and Wave 4.
Table D.2: Number of Times Widowed between Wave 3 and Wave 4
Wave 3 Times Widowed 0 Times 1 Time 2 Times 3 Times Totals
0 Times 11,765 579 0 0 12,344
1 Time 0 3,348 14 0 3,362
2 Times 0 0 52 1 53
3 Times 0 0 0 2 2
Total 11,765 3,927 66 3 15,761
Wave 4 Times Widowed
110
Examining the data on times widowed reveals that 594 (579+14+1) individuals reported
a higher number of times becoming a widow in Wave 4 versus Wave 3.
Information on changing widow marital status does not always align with information on
changes in the number of times an individual has been widowed. The following table
illustrates the information for changes between Wave 3 and Wave 4.
Table D.3: Comparing Methods of Calculating Became a Widow(er) Variable
Marital Status Change to Widow 0 1 Totals
0 29,726 9 29,735
1 85 585 670
Totals 29,811 594 30,405
Increase Count of Times Widowed
In 85 cases the marital status change is not accounted for in the count of times widowed.
There are 9 cases where the count of times widowed codes for an individual becoming a
widow that are not accounted for in the Marital Status.
Therefore, for purposes of this analysis, I conclude that 679 individuals became a widow
between Wave 3 and Wave 4.
Abstract (if available)
Abstract
These essays examine econometric models of health behavior. The first essay evaluates the Kids N Fitness program, an intervention to reduce childhood overweight and obesity. The second essay uses a Childrens Hospital of Los Angeles private dataset to examine determinants of diabetic glycemic control transitions. The third essay analyzes drinking behavior and the response to spousal death among the elderly, utilizing the Health and Retirement Study panel dataset. In each case, I analyze the health behavior choice controlling for available individual characteristics including economic variables. The models apply rigorous econometric techniques and are explicit about the underlying assumptions.
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Essays on economic modeling: spatial-temporal extensions and verification
PDF
Essays on health and well-being
PDF
Essays on human capital accumulation -- health and education
PDF
Empirical studies of monetary economics
PDF
Essays on the equitable distribution of healthcare
PDF
Essays in political economics
PDF
Essays on the economics of subjective well-being in transition countries
PDF
Three essays on health economics
PDF
New approaches using probabilistic graphical models in health economics and outcomes research
PDF
Essays on macroeconomics of health and labor
PDF
Essays on health and aging with focus on the spillover of human capital
PDF
Three essays on health & aging
PDF
Economic aspects of obesity
PDF
Essays on the econometrics of program evaluation
PDF
Essays on labor and development economics
PDF
Essays on health economics
PDF
Large N, T asymptotic analysis of panel data models with incidental parameters
PDF
Essays in health economics: evidence from Medicare
PDF
Empirical analysis of factors driving stock options grants and firms volatility
PDF
Three essays on the credit growth and banking structure of central and eastern European countries
Asset Metadata
Creator
Kaufman, Adam
(author)
Core Title
Essays in empirical health economics
School
College of Letters, Arts and Sciences
Degree
Doctor of Philosophy
Degree Program
Economics
Publication Date
08/03/2009
Defense Date
06/18/2009
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
econometric models of health,Health Economics,health program review,OAI-PMH Harvest
Language
English
Contributor
Electronically uploaded by the author
(provenance)
Advisor
Ham, John C. (
committee chair
), Goeree, Michelle (
committee member
), Mittelman, Steve (
committee member
)
Creator Email
abkaufma@usc.edu,akaufman@dpshealth.com
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-m2457
Unique identifier
UC174395
Identifier
etd-Kaufman-3121 (filename),usctheses-m40 (legacy collection record id),usctheses-c127-172959 (legacy record id),usctheses-m2457 (legacy record id)
Legacy Identifier
etd-Kaufman-3121.pdf
Dmrecord
172959
Document Type
Dissertation
Rights
Kaufman, Adam
Type
texts
Source
University of Southern California
(contributing entity),
University of Southern California Dissertations and Theses
(collection)
Repository Name
Libraries, University of Southern California
Repository Location
Los Angeles, California
Repository Email
cisadmin@lib.usc.edu
Tags
econometric models of health
health program review