University of Southern California Dissertations and Theses

An analysis of nonresponse in a sample of Americans 70 years of age and older in the longitudinal study on aging 1984-1990

(USC Thesis Other)

An analysis of nonresponse in a sample of Americans 70 years of age and older in the longitudinal study on aging 1984-1990

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content AN ANALYSIS OF NONRESPONSE
IN A SAMPLE OF AMERICANS 70 YEARS OF AGE AND OLDER
IN THE LONGITUDINAL STUDY ON AGING 1984-1990
by
Adrienne H. Mihelic
A Thesis Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the Requirements for the Degree
MASTER OF SCIENCE
(Sociology — Applied Demography)
May 1995
Copyright 1995 Adrienne H. Mihelic
UNIVERSITY O F SO U TH ERN CALIFORNIA
T H E GRADUATE SCH O O L
U N IV ER SITY PARK
LO S AN GELES, CALI FORNIA 9 0 0 0 7
This thesis, written by
under the direction of h..f^.....Thesis Committee,
and approved by all its members, has been pre
sented to and accepted by the Dean of The
Graduate School, in partial fulfillment of the
requirements for the degree of
Master of Science, Sociology (Applied Demography)
Adrienne ft.' Mihelic
Dean
THESIS COMMITTEE
y, C h a i r / m a n _______
......
Acknowledgements
Many thanks to Prof. Eileen Crimmins, committee chair and mentor,
for her insightful comments and encouragement. Thanks also to Prof. David
Heer and Prof. Merril Silverstein for serving on my committee and for having
the patience to review this work in its several lives.
This research was supported by grant R03 HS08459 from the Agency
for Health Care Policy and Research.
Table of Contents
List of Tables and Figures.................................................................................. v
A b stract........................................................... vi
Introduction........................................................................................................... 1
Chapter I: Background ..................................................................................... 4
Types of Nonresponse............................................................................ 4
Individual Characteristics Affecting Nonresponse ............................ 6
Change in characteristics of non-respondents over time . . . 8
Effects of Survey Design and Sampling Characteristics.................... 8
Observation plan and changing rates of non-response over
tim e .................................................................................. 8
Other survey and sampling characteristics .......................... 9
Comparing Nonresponse in Other Panels .............................................. 10
Sum m ary................................................................................................ 11
Chapter II: Data ............................................................................................. 14
Longitudinal D e sig n ............................................................................. 14
National Death Index Match . ......................................................... 15
Distribution of Demographic Characteristics .................................. 16
Chapter III: Sample Dynamics ....................................................................... 19
Description of Sample Dynamics ...................................................... 16
Patterns of Nonresponse....................................................................... 23
Chapter IV: Modeling Loss To F ollow -up................................................... 27
Analytic A pproach............................................................................... 27
Model and M easures............................................................................. 29
Results ....................................................................................... 30
Multivariate analysis of loss to follow-up............................. 30
Characteristics affecting probability of nonresponse 34
Summary ........................................................................................ 36
Chapter V: Adjusting For Loss To Follow -up.............................................. 39
Background .......................................................................................... 39
Correcting for nonresponse ................................................... 39
Effect of nonresponse adjustm ents........................................ 42
Analytic Approach to Compensating for Nonresponse.................... 43
Constructing weights .................................................................. 43
Effect of weighting for nonresponse ..................................... 45
D isc u ssio n ................................................................................................... 47
Summary and Discussion........................................................................................ 49
References ............................................................................................................... 53
iv
List of Tables and Figures
Table 1: Comparison of Characteristics at Baseline for the
Total Sample and for the Sample Selected to
Include only Those Who Respond at Every Wave 17
Table 2: Descriptive Statistics for Nonresponse in 4 Waves
of the L S O A .................................................. 20
Figure 1: Transitions in Respondent Status Over 4 Waves
in the L S O A ............................................... 22
Table 3: Response Patterns for Three Waves of F ollow -up.................. 24
Table 4: Variable Definitions and Coding by Observation Plan . . . . 31
Table 5: Odds Ratios for Significant Predictors of Nonresponse . . . . 32
Table 6: Predicted Mean Probability of Becoming a
Nonrespondent for Specific Values of
Significant Independent V a ria b le s.................................... 35
Table 7: Comparison of Frequencies of ADLs and IADLs in
Each Wave Without Weights, With Only the
LSOA Sample Correction Weight, and with
Weights Constructed to Correct for Probability
of N onresponse....................................................................46
v
Abstract
Loss to follow-up is an important problem in longitudinal panels.
Although little analysis of nonresponse in panels of older persons has been
done, the literature suggests that such samples may be more vulnerable to
selection effects. This study employs an event history approach to ascertain
whether some groups are more likely than others to become nonrespondents and
whether the rate of attrition varies over time. The approach taken in this study
addresses several important shortcomings of previous analyses of nonresponse:
It investigates a variety of characteristics hypothesized to be associated with
nonresponse within a multivariate framework, and incorporates a dynamic
design to take advantage of time-varying covariates in the analysis of loss to
follow-up over multiple interviews. Persons of older ages, lower education,
who live alone, live in an institution, rent (not own), have poorer health status,
more functioning impairments, or have another sample person in the household
are more likely to become lost to follow-up. Estimated probabilities from this
analysis are used to construct weights to compensate for nonresponse. The
weighting approach differs from a simpler method using only initial response
characteristics which results in a loss of valuable information. Here, sample
members are retained in the sample until they become nonrespondents, and
cumulative weights adjust for the loss to follow-up in each interval. The effect
of these weights on sample totals is examined in the final section.
Introduction
Although the increasing availability of panel data has been a great boon
to researchers studying characteristics that change with age and over time,
analysis of panel data is complicated by the problem of sample attrition. All
studies using panel data must investigate the extent to which loss to follow-up
has occurred and whether the loss is likely to bias analytic results.
Surveys which collect panel data over time are especially vulnerable, as
accumulated effects are likely to introduce bias of greater magnitude.
Likewise, if rates of attrition or characteristics of attriters vary over time,
there will be implications (and complications) for both the quality and
magnitude of bias. Bias is introduced when loss to follow-up is not random
and is related to outcomes of interest because the selection of the sample may
result in confounded associations (Siegler & Botwinick, 1979; Schaie,
Labouvie & Barrett, 1973) and distorted prevalence estimates and transition
rates (Corder, Woodbury & Manton, 1992). For example, if an analysis of
functional health change among the elderly is undertaken using a sample in
which there is significant attrition that is related to functional health, the
sample is selected with respect to the characteristic of interest and analysis of
functional health will result in biased estimates of the correlates of health
change and the level of health of the sample.
1
Samples of older persons may be more vulnerable to selection effects
from non-response than younger samples because the proportion lost to follow-
up tends to increase with age (Rogers & Herzog, 1992) and because the
differences between participants and nonrespondents may be greater among
older persons than younger (Schaie et al., 1973). Attrition occurs because
respondents move or die and cannot be located or because they refuse to
participate. Loss may be more common among the old because they die more
frequently, have worse health, and more leisure time than other ages. Older
respondents who are no longer at the same address could have entered a
nursing home or other institution leaving no household members behind to
report their whereabouts. Those who refuse to respond may be physically
frail, cognitively impaired, or busy with social activities.
There has been limited consideration of how the causes, covariates and
rates o f non-response may vary over the duration of panels o f older samples,
differentially affecting each wave. In particular, previous studies of wave
nonresponse in older samples have suffered from two important shortcomings:
First, researchers have typically chosen to consider a limited number of
predictors, usually basic demographic characteristics and perhaps a few other
measures of interest to the researcher. And, although some use multivariate
techniques, most consist of uncontrolled comparisons. This may account for
the apparently inconsistent effects on loss to follow-up of some of the
characteristics reviewed in the next section. Second, for the most part these
studies consider only changes in the sample between two points in time. Most
panel surveys, however, consist of more than two interviews. Therefore it is
important to assess the effect of loss to follow-up over multiple interviews.
This research examines the level of nonresponse in a four wave sample
which is initially representative of the noninstitutionalized U.S. population 70
years of age and over. Following a review of previous research, a description
of how loss to nonresponse and death contribute to the decrement of the
sample over time is presented. Next, the covariates of nonresponse in this
sample are analyzed. The results of this analysis are used to propose a method
for adjusting for nonresponse. Finally, the effects of using data adjusted for
nonresponse are compared with the results using unadjusted data.
3
Chapter I: Background
Types of Nonresponse
Wave non-response is a special case of unit non-response: wave non
response occurs in the case when there is no data collected for a sample
element for any given wave in a panel survey on the condition that at least one
wave of data is obtained for that sample element (Lepkowski, 1989). The
term "attrition" refers to the case when a sample person is lost from the panel
for any reason and never returns.
Much research to date on the problem of non-response and attrition has
focused on the treatment of the sample to compensate for the loss. Fay (1989)
notes, however, that much of this analysis has focused exclusively on
predicting missing data based on observed data, and recommends that attention
i
be given to describing the process of non-response. Fay argues that if a
transition can be associated with non-response in the non-attrition case,
perhaps it can be inferred that the transition is also associated with attrition
non-response. The validity of this technique is predicated on the assumption
that attrition and non-attrition patterns are produced by the same causal
mechanisms. If true,. this would be a useful property to employ in replacing
missing data. However, others have noted that this assumption may be
unwarranted (Corder et al., 1992).
4
Comparing studies of non-response is difficult because the way in
which non-response is defined may differ considerably. For example, some
studies evaluate only attriters (those who stay lost after having been missing
for any interview) while others include those who do not respond at one (or
more) waves but then complete interviews at a later waves. Studies also differ
in whether they differentiate between types of loss to follow-up (outright
refusal vs. cause given), or reason for non-response (e.g. moved vs. poor
health). And, some studies consider attrition to include those lost to follow-up
because of death and those who are not self-respondents (i.e. those for whom
the survey is completed by proxy response). Further, non-respondents may be
defined as those within the sample frame who did not become participants,
those who dropped out only after the first wave was completed, or those who
did not respond to at least one wave at any point in the survey.
Refusal to respond has been explored as the extreme case of reluctance
to respond in the population 65 years of age and older (Adams et al., 1990).
Findings suggest that with increased age persons are more likely to be mildly
reluctant (as opposed to more reluctant, refused, and ready respondents).
Those who live alone.are more likely to be "ready respondents," whereas
those who live with others are more likely to respond by proxy or to be in a
nonrespondent category relative to being a self-respondent. Such a correlation
5
with living arrangements, however, may well be a selection effect
(Goldscheider, 1993).
The conceptualization of refusal to respond as an extreme case of
reluctance to respond may be flawed, however. There seems to be a clear
difference between refusal and other non-response. For example, Herzog and
Rogers (1992) review evidence that outright non-response declines at older
ages as the incidence of non-response with cause (health reasons) increases,
and that those who refuse with cause differ more clearly from respondents than
do outright refusers.
Individual Characteristics Affecting Nonresponse
There are several factors which are consistently implicated as causes or
covariates of loss to follow-up among older samples. Many studies have
reported that nonrespondents are likely to be in poorer physical health (Chyba
& Fitti, 1991; Manton, 1988; Rodgers & Herzog, 1992; Speare, Avery &
Lawton, 1991; Streib, 1982) or have lower levels of cognitive functioning
(Schaie et al., 1973; Siegler & Botwinick, 1979; Kalton, Lepkowski,
Montanari & Maligalig, 1990). In the Longitudinal Study of Aging those
more likely to become nonrespondents have been shown to have a greater
number of functional difficulties and worse self-reported health (Chyba and
Fitti, 1991; Speare et al., 1991). In the National Long Term Care Survey
6
nonrespondents to the initial survey subsequently had higher rates of mortality
and institutionalization than continuing respondents (Manton, 1988).
In general, those who are of lower socioeconomic status are also more
likely to be nonrespondents. Those with fewer years of education (Streib,
1982), lower income (Chyba & Fitti, 1991; Streib, 1982), and those who do
not own their homes (Kalton, 1990; Speare et al., 1991) are more likely to
become nonrespondents.
There are several other factors which are somewhat more ambiguous.
For example, household composition and family composition have not been
consistent in their effects nor in their level of significance. Not being married
is sometimes associated with nonresponse (Speare et al., 1991) and sometimes
not (Streib, 1982). Likewise, the response rate among blacks is sometimes
lower than whites (Foley et al., 1990), but sometimes not (Streib, 1982). Men
are sometimes found more likely to be non respondents (Jay, Liang, Lui &
Sugisawa, 1993; Streib, 1982; Kalton et al., 1990), although sex has also been
found to be insignificant (Adams et al., 1990). Although age is almost always
a predictor of likelihood of nonresponse (Rodgers & Herzog, 1992; Adams et
al., 1990), in a study of nonresponse in a Japanese sample, Jay et al. (1993)
find that its effect is nonlinear. The inconsistency in the effect of these
variables likely results from the fact that many are interrelated with other
7
variables which are sometimes included, and sometimes not included, in the
analysis or because of differences in the composition of the sample population.
Change in characteristics of non-respondents over time
When several waves of data are available, it becomes possible to
consider how the characteristics of attriters vary over time. In a study of non
response in a sample of individuals 65 years of age and older Norris (1985)
finds that over 5 waves, those who become lost to follow-up in earlier waves
are in better physical and psychological health than those who become non
respondents at later waves. And, although type of attrition is associated with
time at which the respondent first becomes lost to follow-up, an analysis of
type of attrition by time of attrition indicates that there is no effect of time
independent of attrition type. Although Norris’ data consists of 5 waves, the
survey tracks participants for a total of only 2 years. Also, Norris’ study
excludes those who do not remain lost to follow-up.
Effects of Survey Design and Sampling Characteristics
Observation plan and changing rates of non-response over time
In addition to individual characteristics that may influence a person’s
propensity to continue responding to the survey, there are also certain survey
attributes that may enhance the chances that a sample person will respond.
For example, there has been considerable debate over whether the length of
time between interviews influences response rates. Longer intervals decrease
8
respondent burden, and may allow more time to trace lost respondents (Kalton,
Kasprzyk & McMillen, 1989). Alternately, some researchers have speculated
that higher overall response rates result from more frequent contact with
sample members. However, a comparison of response rates in the National
Longitudinal Survey of Labor Markets and the Survey of Income and Program
Participation (SIPP) found that even though the intervals between waves were
different, that the rates in neither survey were found to decline over the years.
This suggests neither frequency of contact, nor length of survey lifetime affect
rates of nonresponse (Lillard, 1989). Although a common pattern of response
rates shows a dip at the first follow-up interview then a leveling off (Lillard,
1989; Streib, 1982), this initial flux is probably the weeding out of non-
interesteds and in general, regardless of the length of time between interviews,
most researchers have found that the rate of nonresponse does not vary as a
result of the passage of time (Lepkowski, 1989; Lillard, 1989; Norris, 1985).
Likewise, except for the accumulation of nonresponse over time, there is
apparently no effect of the total number of interviews.
Other survey and sampling characteristics
One o f the most potentially biasing sources of nonresponse may come
from lack of interest in a survey topic. Downes (1952), for example, found
that families that did not have a member with a chronic illness were less likely
to respond to a follow-up survey on chronic illness (in Strieb, 1982). Lillard
(1989) has commented that following multiple households belonging to the
same family in studies such as the PSID increases response rates.
Among older people, the institutionalized population may have even
higher rates of nonresponse than the community dwelling population. Lower
response rates may be expected among persons who are living in institutions
because they are more difficult to locate or because of resistance from the
institution or caregivers (Rodgers & Herzog, 1992).
Comparing Nonresponse in Other Panels
As noted above, nonresponse is an issue in all panel surveys. In the
first household interview of the National Medical Care Utilization and
Expenditure Survey (NMCUES 1980) 10% did not respond. By the fifth
interview conducted 15 months later, 22.1% of the sample were
non respondents (Wright, 1983). About 5% of the Survey of Income and
Program Participation sample did not respond to the first interview. By the
ninth wave, 36 months later, this rate had accumulated to 22.3% (Kasprzyk &
McMillen, 1987). Nonresponse to the annual interview of the PSID was high
in the first wave (24%), then leveled off to about 2% at each interview,
accumulating at about. 44% in the 20th wave (Hill, 1992).
While there are a number of reasons the loss to follow-up among older
persons might be expected to be higher than that in the rest o f the adult
population, loss to follow-up in surveys that are of only older persons
10
compares quite well with these general population surveys. Corder and
Manton (1991) compare rates of nonresponse in four longitudinal studies of
older persons and report low nonresponse rates of about 4% among those who
had previously completed a screening interview in the National Long Term
Care Survey, and a high of 25 % in the NHANES I follow-up which required a
physical exam. Both the LSOA and the National Nursing Home Survey fall
within the moderate range. The LSOA compares favorably with other surveys
of older persons and with general population surveys with an initial
nonresponse rate of about 9%, accumulating over the four interviews (six
years) to about 19%. The slightly younger male sample of the Retirement
History Survey, aged between 58 and 63 at baseline in 1969, was diminished
11% by the second interview 2 years later by death, institutionalization and
nonresponse. Forty-six percent had been lost by the seventh interview wave
(Burtless, 1987).
Summary
Regardless of the definition of non-response, type of sample, or
duration of the study, there are several factors which are consistently
implicated as prime suspects in causing or covarying with non-response.
In general, those who are of lower socio-economic status and in poorer
health are more likely to be non-respondents. Specific measures of these
constructs vary from study to study, but include race/ethnicity, educational
11
attainment, employment status, family income, types of dwelling structure, and
tenure type. Health status has been measured as mortality, institutionalization,
dependence on proxy, and by various measures of cognitive functioning and
morbidity, including self reported health status and measures of functional
health (usually ADL and IADL scales).
There are several other factors which are somewhat more ambiguous.
For example, household composition and family composition have not been
consistent in their effects nor in their level of significance. For example,
marital status is sometimes significant, and sometimes not. Men are usually
found more likely to be non-respondents, although sex has also been found to
be insignificant. Living with others sometimes has a positive effect on
nonresponse, sometimes the effect is negative. Although age is almost always
a predictor of likelihood o f non-response, it has been found to be insignificant
in some studies And the effects of age on the types of nonresponse are
suggestive but not conclusive. The inconsistency in effect of these variables
may likely be because they are either confounding or are confounded by some
other variable.
There has been some work done on changing rates of attrition in
longitudinal panel surveys, but none of the work thus far has considered
changes in rates of attrition in samples of older adults. The only exception to
this is a brief note which presents descriptive statistics on percentages o f non
response in the LSOA (Chyba & Fitti 1992), which indicates that the rate of
non-response increases over the three follow-up waves. This finding is in
contrast with the analysis on younger samples which found rates of attrition to
be constant.
Except for the work of Norris (1985) there has been no further
consideration of whether the characteristics of non-respondents vary over time.
And, although Norris’ work is an excellent starting point, the time frame for
the entire study is equivalent to one between-wave interval in the LSOA. This
observation begs the question of whether changes in non-respondents are a
function of historical time or survey time. While this cannot be addressed
conclusively in one or two studies, when research in this area begins to
accumulate, patterns may become evident.
13
Chapter II: Data
This study uses data from the Longitudinal Study of Aging (LSOA).
Detailed descriptions of this data set and the survey procedures are provided in
two NCHS publications (Fitti & Kovar, 1987; Kovar, Fitti & Chyba, 1992).
The sample for the LSOA is based on the Supplement on Aging to the
National Health Interview Survey. This was a multistage complex household
survey. The study began in 1984 with interviews of a sample representative of
the U.S. noninstitutionalized population 70 years of age and over. Three
follow-up interviews are conducted at approximately 2 year intervals, with the
last interview occurring in 1990. When a sample member is unable to respond
because of a physical or mental health problem proxy responses are accepted.
Longitudinal Design
As the sample is not replenished at each wave, but only original sample
members are followed, at each interview the age of the youngest sample
members increased so that by 1990 the sample is a group 76 years of age and
older. All four interview waves will be used in the analysis. For some parts
of the analysis the observations will be pooled over intervals to maximize the
number of transitions of interest.
At the original interview 7,541 persons representative of the community
dwelling population 70 years of age and older were interviewed. Because of
budgetary restrictions the sampling plan for the LSOA varied in 1986 so that
14
2,376 members of the original sample were not included in the first follow up
interview. For this reason, these 2,376 people are excluded from the analysis.
Thus, the working sample consists of 5,151 people for whom information
could be available at four dates with change over three intervals. Because
people die and are lost to follow-up, the number of people from whom data
are collected at each wave is less than the potential sample size after the
original date.
National Death Index Match
Of importance to this study is the fact that deaths for LSOA sample
members are tracked with the National Death Index as well as by using contact
persons whose names were solicited at the first interview (Fitti & Kovar,
1987). National Death Index (NDI) data is collected for sample members and
appended to survey files. This results in nearly complete knowledge of the
vital status o f panel members. A sample person is assumed to be dead if a
contact person has reported that person as deceased or if that person has a
"good match" (presumed deceased) with the National Death Index and did not
respond to any surveys subsequent to the date of death. The process of
matching the NDI records with the LSOA files has been completed using the
most recent update of the Index.
15
Distribution of Demographic Characteristics
The distribution of this sample by gender reflects the national
proportions for this age group. Approximately 60% of the sample is female.
The LSOA prioritized the inclusion of minority sample members. The initial
sample consisted of all individuals who were 55 years of age and older living
in a household where an NHIS survey was completed. In 1986 when
budgetary limitations required limiting the sample, all black or Hispanic
persons and their relatives over 70 years of age were retained in the sample.
Nevertheless, only 3% of the sample members claim hispanic origin and just
over 6% are black. Race will be included in all multivariate analyses.
However, due to the limited representation of other minorities, only the effects
of being black versus being white or other minority can be compared.
It should also be noted that the LSOA prioritized including the oldest-
old in its sample. Too much research on the elderly still inappropriately
groups older ages, includes primarily the young-old, or has a poor
representation of the very old. In the research proposed here, age will be
included as a continuous variable.
Table 1 shows the distribution of selected baseline characteristics for
the original representative sample and that sample after it has been selected by
nonresponse over the 6 year longitudinal survey period. This simple
presentation suggests that loss to nonresponse and death make the sample
16
Table 1: Comparison of Characteristics at Baseline for
the Total Sample and for the Sample Selected to Include
Only Those Who Respond at Every Wave
Full Sample Selected sample
n=5/15l n=2,431
F e m a le 64% 67%
B la c k 11% 11%
L i v e s a lo n e 37% 37%
Owns home 7 6 % 80%
H e a lth s t a t u s
P o o r 13% 7%
E x c e l l e n t 15% 19%
Number o f
f u n c t i o n i n g
d i f f i c u l t i e s
N one 60% 71%
One 13% 12%
s i x 2% 1%
T h ir t e e n 1 % 0%
o t h e r sa m p le p e r s o n 3 6% 37%
i n h o u s e h o ld
Mean a g e 78 77
Mean y e a r s o f 10 10
e d u c a t io n
Mean number o f 2 1
f u n c t i o n i n g
d i f f i c u l t i e s
17
younger, more likely to female, more likely to be a home-owner, and
healthier. And, although the selected sample includes more persons in
household which have multiple sample members, this selection does not affect
the sample composition by race, living arrangement or level of educational
attainment.
18
Chapter III: Sample Dynamics
All studies using panel data must investigate the extent to which loss to
follow-up has occurred and whether the loss is likely to bias analytic results.
This chapter is devoted to the sample dynamics in the LSOA, describing how
loss to nonresponse and death contribute to the decrement of the sample over
time. In the following chapter, covariates of nonresponse in this sample are
investigated and used to propose a method for adjusting for bias introduced by
nonresponse.
This section begins with a description of how loss to nonresponse and
death contribute to the decrement of the sample over time in the LSOA. First,
aggregate outcomes are presented for the entire sample for the four waves.
The same information is then presented in more detail in the form of
transitions among response categories and death. This gives a better idea of
the dynamics producing the aggregate outcomes. Individual patterns of
response over the four waves are then examined.
Description of Sample Dynamics
The number of respondents diminished at each successive wave (Table
2A.). Of the original 5,151 respondents, 80 percent were interviewed in
1986, 64 percent in 1988, and just over half of the sample was remaining at
the last wave in 1990. Two-thirds of this loss is due to death. By the second
wave 13 percent of the original sample had died; by the fourth wave more than
Table 2: Descriptive S t a t i s t i c s f o r Nonresponse i n 4 Waves of the LSOA
2A. Frequency and Percentages of Persons Not Interviewed at Each Wave by Death and Nonresponse
Wave 1 Wave 2 Wave 3 Wave 4
1984 1986 1988 1990
n % n % n % n %
Interviewed 5,151 100.0 4,114 79.9 3,314 64.3 2,676 52.0
Lost to Death
---- ----
645 12.5 1,291 25.0 1,857 36.1
Nonresponse
---- ----
392 7.7 546 10.5 618 12.0
Totat 5,151 100.0 5,151 100.0 5,151 100.0 5,151 100.0
2B. Percent of Those Alive on Previous Wave Not Interviewed on Next Wave Because of Death of Nonresponse
Wave 2 Wave 3 Wave 4
Lost to Death 12.5 14.3 14.7
Nonresponse ' 7.6 12.1 16.0
■ V
2C. Percent of Those Alive Who are Nonrespondents
Wave 2 Wave 3 Wave 4
8.7 14.1 IB.8
2D. Percent of Those Alive Who Become Nanrespondents for the First Time Among Those Who Were Not Previously Nonrespondents
to
O
Wave 2
8.7
Wave 3
9.7
Wave 4
11.1
one-third was dead. Interestingly, even in this older sample, the percentage
who died between interviews among those known to be alive at the preceding
interview was fairly constant across the three intervals (Table 2B). Between
the first and second interviews, twelve percent of the sample died; between the
third and fourth interviews this was about 15 percent.
The percent of the original sample not responding to the survey
increased with each wave. Among those not known to be dead at each wave
(Table 2C) the likelihood of nonresponse increased from almost 9 percent at
the second wave to 19 percent at the fourth wave. The percentage becoming
new nonrespondents (Table 2D) does not increase nearly as dramatically,
indicating that the likelihood of not responding to future waves is much higher
for those who are earlier nonrespondents.
While Table 2 indicates the overall level of nonresponse and death in
the sample, the likelihood of changing or retaining response status is better
shown in Figure 1 which indicates the outcomes on successive waves for those
who are interviewed and those who are nonrespondents at the earlier wave. At
each wave 75 to 80 percent of those interviewed at the previous wave were
interviewed at the next wave. At each successive wave about one-third of
those who were nonrespondents on the previous wave are interviewed, while
about 50 percent remain nonrespondents. While these percentages are fairly
constant over the waves, the numbers to which they are applied are changing
FjflUfg.1i, Transitions In Respondent Status oirer 4 Waves In the LSOA
Wave 1
Wave 2 :
Wave 3 Wave 4
1684
1988
1988
■ i -
1990
Interviewed
6161.
Interviewed
3175(77. 18*)
Interviewed
2501(75.47*)
Interviewed
4114
3314 2078
(79.9%) (73.6%)
(09.3%)
Nonresponse
Nonresponse Nonresponse
203(51.70*)
269(49.36*)
(12.1% )
to
ro
Dead Dead Dead
646
646
666
(12.6%) (14.3% ) (14.7% )
and this results in the build-up of nonrespondents and the decrease in
respondents at later waves.
Patterns of Nonresponse
There are 15 possible patterns of non-response (Table 3). Because
there are with 3 potential waves of response in this analysis, there are 4
general categories of response patterns to consider: response at all interviews
(XXXX), response and then attrited at some subsequent interview (XOOO),
present at the initial survey, lost, and then found (XOXX), or possibly lost
again (XOXO).
Consider the last two patterns (XOXX and XOXO). The final pattern
may simply be an elaboration on the lost and then found pattern, or it could be
a combination of that pattern and an attrition pattern. Likewise, the pattern
which is, for the purpose of the survey, considered attrition (XOOO), may
simply be a right censored lost and found pattern — that is to say, if the survey
continued, this individual might eventually transition to respondent status.
Thus far the discussion has implied the hypothesis that the different
patterns are associated with different processes. In practice difficulties arise in
differentiating the patterns, and ultimately this practice may be somewhat
arbitrary. It is not known whether these patterns are in fact distinct, and if so,
whether their differences can be detected within the LSOA, given the length of
time between intervals and the duration of the survey. Lepkowski (1989) has
23
Table 3; Response Patterns for Three Waves of Follow-up
I=interviewed N=nonresponse D=dead
pattern frequency percent
III 2431 47.2% 47.2% Completed all interviews
IID 442 8.6
ID. 597 11.6 32.7% Interviewed until death
D.. 645 12.5
IIN 302 5.9
INI 134 2.6 9.9% Two more interviews
Nil 70 1.4 still alive
INN 128 2.5
NIN 47 0.9 4.2% One more interview
NNI 41 0.8 still alive
IND 80 1.6 2.0% One more interview
NID 22 0.4 dead
ND. 50 1.0 1.4% Never interviewed again
NND 21 0.4 dead
NNN 141 2.7 2.7% Never interviewed again
_____ _____ presumed alive
100.1 100.1
24
asserted that the frequency distribution of patterns of wave nonresponse
depends on such factors as the survey topic and organization. In a comparison
of two surveys similar on both of these factors (the Income Survey
Development Program 1979 Research Panel (ISDP) and the 1984 Panel Survey
of Income and Program Participation (SIPP)), response at each wave was the
most common pattern, followed by patterns of attrition, and the least common
patterns were those non-attrition non-respondents (Lepkowski, 1989). The
distribution of the possible patterns observed by Lepkowski in those two
surveys is similar to the distribution of the nonresponse patterns observed in
the LSOA which has a rather different topic and observation plan than both the
ISDP and the SIPP. While this neither supports nor contradicts Lepkowski’s
theory regarding the determinants of the patterns of nonresponse, it also
permits speculation that perhaps the various response patterns are produced by
different subpopulations.
The pattern of response of individual cases presented in Table 3 gives
us a fuller picture of how sample nonresponse affects our ability to trace
changes in individual’s characteristics across interviews. Eighty percent of the
original respondents are interviewed at every wave before death. Another 10
percent of the sample is only missing at one interview and assumed to be still
alive. Only 4 percent of the sample are never interviewed again after the
initial interview, and one-third of these are dead before the fourth wave.
25
Thus, even with this level of attrition, the changes in individual characteristics
can be traced for most living members of the sample.
26
Chapter IV: Modeling Loss To Follow-up
In this chapter a multivariate analysis predicting non-response over the
four waves of the LSOA is used to identify characteristics of sample members
associated with loss to follow-up due to factors other than death.
Analytic Approach
The goal of this analysis is to account for the possibility that causes,
covariates, and rates of nonresponse may vary over time, differentially
affecting each wave. Therefore, a method that allows the incorporation of
time-varying covariates is required. And, because the event of non-response
takes place in the lifetime of the survey and has the possibility of occurring at
only 3 points in time, the analysis requires a method suitable for handling
coarsely measured discrete time intervals and multiple ties in the event of
interest.
A discrete time event history approach is chosen to estimate covariates
of non-response to take full advantage of information on timing of transitions
to non-response (Teachman, 1982). To accommodate the discrete-time
observation of transitions, a logistic regression model is chosen and adapted to
approximate a hazards rate (Allison 1984, Vuchinich, Teachman and Crosby
1991, Yamaguchi 1991). In this case the odds for the conditional
27
probabilities of the event of nonresponse occurring are modeled. This model
has the form:
ln{X(/,;X)/[l-X(/,;X)]}=a,+5:AXt
where X(r,;X) is the conditional probability of becoming a nonrespondent at
time tj for the covariate vector X (X =X /,...X J t ) and parameters b* (& =1,...,I)
and a* is the log-odds for the baseline group (af=ln{X0(ri)/[l-X0(4)]}). What is
different from a standard logistic regression, however, is the structure of the
input data set. In this approach a pooled data set is created in which each
observation gets a unique record for each interval (n = 10,758).
When the covariates are time constant this is a proportional odds
model, such that the odds of becoming a nonrespondent form a constant ratio
with respect to time among groups that are distinguished by the covariates.
However, time-varying covariates are included in most of the analyses.
Therefore, among the groups that are distinguished by the time-varying
covariates, the odds of becoming a nonrespondent forms a ratio which varies
with respect to time. The introduction of variables indicating each time period
allows us to determine whether there is an independent effect of time, and will
answer the question of whether the rate of nonresponse varies over time
(strictly speaking, it will answer the question of whether the odds o f becoming
a nonrespondent varies over time).
28
Model and Measures
For the purpose of this study, a transition to non-response will be
defined by the first record of non-interview status for any reason other than
death1. Because the sample frame for the LSOA consisted of respondents to
the 1984 HIS, there are no non-respondents for the first wave, thus eliminating
the problem o f left-censored cases. Individuals can become nonrespondents at
any wave after the baseline interview (1984). The event of nonresponse takes
place in the lifetime of the survey and has the possibility of occurring at three
points in time.
A model of loss-to-follow-up is posited based on previously discussed
research (refer to Chapter 1). It is expected that being older, male, black,
having fewer years of education, not owning a home, currently residing in an
institution, reporting worse overall health and a greater number of functioning
limitations are associated with higher odds of nonresponse. Indicators of the
interval of observation are also included in order to test whether there is an
effect of time on the probability of nonresponse, and an indicator for having
another sample person in the household to indicate family involvement in the
survey.
1 Predictors and rates of attrition from death have been considered
elsewhere. However, it should not be assumed that all loss to follow-up
conforms to the same patterns as mortality (Baltes 1968, Siegler and Botwinick
1982, Manton 1988).
29
To take advantage of fresh information collected at each interview, a
pooled data set is created in which each observation gets a unique record for
each interval (n = 10,756) and characteristics at the beginning of the interval
are related to response status at the end of the interval. This approach
assumes that the error terms for each observation (in this case each person-
interval) are not correlated.
The record for each person-interval contains two kinds of independent
variables (Table 4). The first are those variables which were measured only at
baseline and thus for analytic purposes must remain the same at each interval.
These variables typically remain constant over adult life: sex, race, and years
of education. The indicator for having another sample person in the household
and self-rated health status also remains constant. The second type of variable
was collected or can be inferred at each interview: age, home ownership,
whether the respondent lives in an institution, general health status, and
functioning abilities. An indicator for interval of observation is also attached
to each person-interval record.
Results
Multivariate analysis of loss to follow-up
A series of four nested logistic regression models were run so that the
effects of including groups of predictor variables covering different
characteristics could be compared (Table 5). This approach clarifies why
30
TabJejJj> Variab1e- Definitk>ns_and_Cod^^
Measures co llected only at baseline
Sex
Race
others)
Years of Education
0-18+)
Health Status
5=poor)
Other Sample Person in Household
Measures collected at each interview or change can be inferred
Age
Non-response
Living Alone
Owns H om e
Living in in s titu tio n a t time of interview
Number A D L * and IA D L b d iffic u ltie s
Duration of time in survey (in terv al flag)
(1=female, 0=men)
(1=black, 0=all
(in actual years:
( 1 = e x c e I L e n t,
(1=yes, 0=no)
(years reported)
(1=yes, 0=no)
(1=yes, 0=rvo)
(1=yes, 0=no)
(1=yes, 0=no)
(0-13)
(1=yes, 0=no)
* "A ctiv ities of Daily Living" is a commonly used scale to measure functioning a b ility in the
realm of basic personal care needs. The standard 7 item scale includes indicators of whether
i f , fo r health reasons, on individual has d iffic u lty with the following items: bathing or
showering; dressing; eating; g etting in or out of bed or chair; g ettin g outside; walking across
a-small room; and using the to ile t.
b "Instrum ental A ctiv itie s of Daily Living" is a commonly used scale to measure functioning
a b ility in the realm independent household maintenance. The standard 6 item scale includes
indicators of whether i f , for health reasons an individual has d iffic u lty with the following
items: preparing one's own meals; shopping for personal items; managing own money; using the
telephone; doing Light housework; and doing heavy housework.
31
Table 5: Odds Ratios for Significant Predictors of Nonresponse
n = 10.400
Intercept
Sex
Race
Age
Years education
Lives alone
Lives in institution
Owns home
Health status
Functioning
Other sample person
Interval.2
Interval 3
“2 Log L
Chi-sq
df
P
response
categories
* p < .05
~ p < .001
Model 1 Model 2 Model 3 Model 4
Odds Ratios Odds Ratios Odds Ratios Odds Ratios
0.014“ 0.025" 0.023“ 0.022”
1.189*
1.031"
0.945"
6148.962
covariates
V 78.627
4
.0001
1 = 930
0 = 9,270
1.027"
0.948"
2.108*
0.604"
6092.729
m2 v ml
56.233
3
.0001
1 = 930
0 = 9,270
1.023”
0.955"
1.180*
1.945*
0.616"
1.080*
1.030*
6075.865
m3 v m2
16.864
2
.001
1 = 930
0 = 9,270
1 . 022“
0.951"
1.383"
1.946*
0.607"
1.079*
1.032*
1.354
1.223*
6060.903
m4 v m3
14.963
3
.01
1 = 930
0 = 9,270
NJ
some studies find some variables are significant while others do not.
In the first model a basic set of socio-demographic variables are
included. Being female and being older both increase the odds of becoming a
nonrespondent, while higher levels of education reduce the odds that a person
will leave the sample.
In the second model when measures of living arrangements and home-
ownership are introduced, the effect of sex becomes insignificant, and the odds
for becoming a nonrespondent are twice as high for those who are currently
living in an institution compared to those who are not living in an institution,
and are lower for home owners than for renters. Household composition is
not significant.
In the third model two measures of health are included: self reported
health status, and the number of functioning difficulties. As would be
expected, those who have poorer self reported health status have higher odds
of being lost to follow-up, as do those who have difficulty with a greater
number of functioning activities. And, with health status controlled, living
alone becomes a significant predictor of nonresponse.
In the fourth model indicators of whether there are other sample
persons in the same household and the interval of observation are included.
As described above, the LSOA is a household sample. A total of 1,861 of the
5,151 sample members at the baseline interview are living in a household with
33
another sample member (921 households with multiple sample members). If a
sample person lives in such a household, regardless of the response status of
those other persons, his or her odds of nonresponse are over one third higher
than those who are the only sample person in a dwelling.
The coefficients of the time variables indicate whether (and how) the
odds of becoming a nonrespondent differ at the second and third intervals with
reference to the first. The significance of the dummy variable for the third
interval indicates that those persons responding at the third interview are more
likely to be lost by the fourth interval than respondents to the first interview
were to be lost by the second interview, all other characteristics constant.
Characteristics affecting probability of nonresponse
Having identified which characteristics are associated with nonresponse
in this sample, the next question is: How can variation in these characteristics
affect the actual probability of becoming a nonrespondent? To explore this
question the mean probability of becoming a nonrespondent is estimated for
selected categories of each of the significant predictor variables (Table 6a).
For example, if the entire sample were age 70, and each individual
retained all other observed characteristics, the mean probability of becoming a
nonrespondent would be 1.5%. If the whole sample were 90 years old the
probability of becoming a nonrespondent would increase approximately 3
percentage points. The effect of this 20 year increase in age is to increase the
Table 6: Predicted Mean Probability of Becoming a Nonrespondent for
a. Specific Values of Significant Independent Variables with All
Other Characteristics as Observed
b. Sample Members observed to have Specific Values of Significant
Independent Variables
Variable Value fa> Mean p Mean p
Age
Age 70
Age 90
,0755
.1112
.0819
.0950
Education
8 Years School
16 Years School
.0977
.0681
.1005
.0661
Living Arrangement
Lives with Others
Lives Alone
,0806
.1076
.0830
.1032
Tenure Type
Does Not Own
Owns Home
.1243
.0798
.1319
.0783
Living in an Institution
No
Yes
.0904
,1602
.0903
,1910
Self-Rated Health Status
Poor Health
Excellent Health
.1049
.0800
.1237
.0738
Functioning Health (ADL/IADL)
No Difficulties
Diff. w/ 6 functions
Diff. w/ 13 functions
.0866
.1025
.1551
.0778
.1252
.1553
Other Sample Persons in Household
NO .0840
Yes .1098
.0927
.0881
Respondent at 3rd Wave
No
Yes
.0867
.1037
.0862
.1056
35
probability of nonresponse by 47%. Similarly, having another sample person
in the household, regardless of whether that persons chose to response or not
results in a 30% increase in the probability of nonresponse to the subsequent
interview.
If the entire sample had difficulty with all thirteen of the ADL and
IADL functions, the mean probability would increase 80% from .0866 (having
no difficulty with any ADLs) to .1551. This is the largest contrast among the
conditions listed.
The second column (a) of Table 6 indicates that the predicted mean
probabilities of loss for sample individuals who actually have the observed
values of the characteristics listed differ more widely than the probabilities in
column a. This increased difference is caused by the fact that those who are
more or less likely to be non respondents in one characteristic are likely to also
have other characteristics leading to similar outcomes. The only exception to
this is in the confounded relationship between living arrangements and having
other nonrespondents in the household.
Summary
Some results of this analysis are consistent with previous research.
Those who are older, have lower educational attainment, are in worse health,
and rent their homes are more likely to become nonrespondents. However, no
effect of race, a characteristic that several previous studies found to predict
36
nonresponse, is found. Among those variables previously identified as
ambiguous in their relationship to nonresponse, sex becomes insignificant
when measures of living arrangement are included. And, when measures of
health status are included, living alone becomes a significant predictor of
nonresponse. The effect of currently living in an institution is also significant.
Some surveys collect information from multiple members of a single
household so that studies involving household behavior may be completed. In
this survey, having another sample person in the household increases the
probability of nonresponse considerably. Exactly why this occurs is unclear.
Those who have other sample persons in the household are more likely to be
older, perhaps they have been recently widowed, or are now providing care to
the other sample person. Some of these eventualities are controlled in the
estimation, although changes in the situation of the household and the choice to
not respond may occur after the last data were collected and thus will not be
accounted for in the analysis. Household response behavior is clearly an area
in which further research would be fruitful. The lesson from our finding is
not that survey designs including multiple sample members from a single
household should be avoided — this would result in a loss of valuable
information. Rather, awareness that such sample persons may be more prone
to loss to follow-up might call for special attention to retention of such sample
37
persons and possibly oversampling. Post-hoc analysis using such data may
include corrections for the disproportionate loss among these members.
Although studies of nonresponse in panels with younger samples find
that there is no effect of the interval on the odds of becoming a first time
nonrespondent, in this sample incidence of first time nonresponse becomes
more likely as time passes. Because most nonresponse interviews had reason
for noninterview recorded as "other" it was not possible analyze nonresponse
by reason. It is important that future survey collection efforts monitor reason
for nonresponse. If nonresponse is occurring because of physical difficulties
or anticipated respondent burden, as might be inferred from these results,
alternative or adapted interview mechanisms could be used.
These findings suggest that wave nonresponse in this sample of older
person is indeed not random. Characteristics found to be associated with
propensity to become a nonrespondent may be used to develop targeting
criteria for those interested in retention of sample persons. And, for those
using panel data with older samples, the process of identifying salient
characteristics and computing probabilities of becoming a nonrespondent may
be used to develop weights to compensate for decrement to the sample.
38
Chapter V: Adjusting For Loss To Follow-up
Background
Correctine for nonresponse
If loss to follow up occurs nonrandomly, that is, differentially selecting
persons with certain characteristics out of the sample, attention must be given
to correcting potential bias. There are three common approaches to correcting
for missing responses: discarding incomplete records, weighting, and
imputation (Little and Rubin, 1987).
The simplest approach is discarding incomplete data (for example, Mor
et al., 1989; Rogers, Rogers & Belanger, 1992). This approach is expedient
but should be accompanied by an analysis of the impact of excluding
observations. Some baseline analysis of characteristics related to the outcome
of interest may be used to determine whether excluding non respondents will
result in a seriously selected sample. One approach is to perform the analysis
both on the selected sample and on the sample including those who later
become nonrespondents (see for example Maddox & Clark, 1992). Similarly,
Corder, Woodbury and Manton (1992) apply the grade of membership
approach to 28 variables indicating health, functioning, and response status in
parallel analyses for both the total eligible sample and on a sample of only
those who later become nonrespondents. The resulting "pure type" profiles
are similar for both samples, and the authors conclude that no bias is
39
introduced by the nonrespondents. Presumably, then, these observations could
be excluded from the analysis with minimal consequence.
Weighting is perhaps the most commonly used approach to unit
nonresponse adjustment. Two methods are commonly employed, cell
adjustment and response probability weights. The first approach uses cross
tabulations of the characteristics affecting nonresponse to divide the sample
into subgroups. Each cell is then weighted with the inverse response
probability. Thus, cells representing subgroups of the sample population with
characteristics associated with higher rates of nonresponse will be "weighted
up" to compensate for the differential loss. This method is described in more
detail in Rowland and Forthofer (1993).
Response probability weights work on the same principle as weighting
class adjustments. However, the weights are constructed from multivariate
regressions predicting probability of nonresponse. Therefore, whereas
weighing class adjustments provide weights for coarse groupings of
subpopulations, response probability weights can be as unique as the
distribution of characteristics predicting nonresponse (Iannachione, Milne &
Folson, 1991). Class based weighting procedures and response probability
weights have been found to produces quite similar results (Lepkowski, Kalton
& Kaspryzk, 1989). While weighting is the most widely used approach to
correcting for nonresponse, and it will correct bias on observed variables, it
40
also tends to increase variances (lannachione et al., 1991; Kish, 1990; Little,
1986). It is, however, possible to adjust tests of significance for this increased
variance, just as it is done for sampling weights.
Imputation is similar to weighting, except that instead of giving
respondents additional weight to compensate for those lost with similar
characteristics, imputation actually replaces the missing data. The procedure
is to locate other observations with matching responses on items for which
information exists, and then allocate to the nonresponder the same values for
the missing items as the matched observations. Or, in the case where some
waves o f response are available and an assumption o f stability of
characteristics is justified, an individual’s response on earlier completed
interviews may be used for later missed interviews in a "carry-over"
imputation procedure (Kalton & Miller, 1986). Although weighting and
imputation produce similar sample means (Kalton, 1990; Kalton & Kish, 1981)
imputing has the advantage of controlling both bias and variance (Little, 1986).
Imputation is commonly employed to replace missing data in both the U.S.
Census and Current Population Surveys, And, although imputation procedures
are usually reserved for nonresponse to single survey items, exceptions exist.
For example, in cases where there is evidence of a household but no interview
data has been collected, the Census may impute records for entire households
(Shryock and Seigel, 1976). Also, it has been suggested that imputation might
41
be fruitfully employed in combination with weighting for multiple wave
surveys in cases when a sample member who was once a respondent and
becomes a nonrespondent responds again at a later interview (Kalton,
Lepkowski & Lin, 1985).
All of the approaches described above assume that those who are lost to
follow up are similar on unobserved characteristics to continuing respondents
that have the same observed characteristics. This assumption may be
untenable; If there are unobserved characteristics influencing nonresponse,
then a correction based on observed characteristics will not eliminate the bias.
Effect of nonresponse adjustments
When descriptive univariate or bivariate statistics including prevalence
and transition rates are being reported for a sample with nonrandom
nonresponse, correction may be required. However, when a multivariate
model is fully specified in a substantive analysis and there are no unobserved
characteristics contributing to the propensity to be lost to follow-up, then the
model itself should control for the distribution of characteristics in the sample.
The only way to determine how much difference an adjustment for
nonresponse will make is by making the adjustment and then comparing
corrected and uncorrected analyses. However, a number of studies have found
that correction has not altered the interpretation of the results (for example,
Kalton et a!., 1989; Rowland & Forthofer, 1993).
42
Analytic Approach to Compensating for Nonresponse
The probabilities of nonresponse calculated in the previous section are
now used to construct weights to compensate for nonresponse over the 3
survey intervals in the LSOA. The effect of this strategy is compared with an
uncorrected sample, and with the sample using only weights provided to
correct for design effects.
Constructing weights
Weights for the data are constructed based on the probabilities of
nonresponse for characteristics found to be salient predictors of nonresponse in
the previous analysis. The probabilities of nonresponse are estimated within
each interval of observation. While this approach is not new, response
probability weighting typically makes use of only first wave information
(Kalton & Miller, 1986). In other words, even if a sample member responds
to more than one interview, only their responses at the initial interview are
used to correct for their future nonresponse. Clearly this could result in
considerable waste of observed data. In this approach, each sample member is
retained as long as he or she has not missed an interview. Therefore, if a
*
persons does not become a nonrespondent until 1990, three waves of original
data are retained. The remaining sample is then weighted up to compensate
for this person’s loss in the final wave. This strategy allows us to include
considerably more information. This is particularly important in longitudinal
surveys that collect information on characteristics that are not expected to be
stable, and for which precision in estimating transitions is critical. Some loss
of information may be expected because reinterviews of persons who were
previously nonrespondents are not taken into account. However, because only
47 persons in this study have a pattern of response that would contribute in
this way (see Table 3, pattern NIN), the impact is probably quite small.
For each interval (e.g. 1986-1988) for all sample members responding
at the beginning of the interval, variables that were significant in the pooled
analysis are regressed on an indictor of response status. These probabilities
are then applied to the sample weight for design effect to correct for
nonresponse.
Because the LSOA is a multistage cluster sample, the LSOA sample
weight is provided to adjust by geographic area for baseline nonresponse and
the subsampling of certain populations. The weight adjusts the baseline
sample so that it is representative of the noninstitutionalized population that is
70 years of age and over in 1984 by sex, race and age. This sample design
weight is used for the baseline data, and then modified cumulatively at each
successive wave to compensate for nonresponse. Thus, each sample person
has a wave-specific weight that corrects for both the subsampling and that
person’s propensity to be lost to follow up. When these weights are applied,
the contribution of those persons who actually respond are inflated to also
44
represent those who did not respond, based on characteristics included in the
computation of the probabilities. Thus, the weights increase the sample size to
approximately what the sample would have been if the only source of attrition
had been death, and they alter the distribution of the characteristics so that
they take into account the propensity of some groups to become
nonrespondents more than others.
Effect of weighting for nonresponse
Table 7 compares the effect of weighting only with the normalized
LSOA weight to correct for sample design with the correction weight for
nonresponse on the distribution of the number of ADLs and IADLs across the
four interview waves. (In 1984 the nonresponse weight is equivalent to the
LSOA weight because there is no baseline nonresponse to account for). The
general pattern is that the LSOA weights alone make the sample a bit healthier
at each wave compared to the unweighted sample, adding a few more people
to the zero difficulties category. Alternately, since those in poor physical
functioning are less likely to respond, the nonresponse weights compensate for
that tendency, and thus move a small proportion out of the zero difficulties
category and into the.categories indicating difficulty with one or more
activities. Still, even though the nonresponse weights move the sample toward
worse health, they do not return to the levels of the unweighted data. In 1986
the effect of the nonresponse weights (compared to the data weighted with
45
TabU 7: ComparUon of Ftequanclaa (or AOL* and lADLa tn Each Wava Without Weight*, With Only tha LSOA Sample Conaotlon Walght, and with Weight* Con*t«uctad
! S ^ £ Q S £ L ! 2 L £ i S ^ S ^ ! l ! ! £ i S ! > £ i S S j l S S E S S 5 ^
ADI j a d l _
Nonresponse Nonresponse
1984
1986
1388
1990
O n
0
1
2
3
4
5
6
7
0
1
2
3
4
5
6
7
N
3k
N 3 k N . 3k E > 3k n 3k t ! *
3637 71.0 3770 73.6
. . . .
3341 65.3 3511 68.6
. . . .
526 10.3 507 9.9 829 16.2 800 15.6
319 6.2 291 5.7 308 6.0 271 5.3
199 3.9 176 3.4 i l.a. 170 3.3 150 2.9 n . a .
151 2.9 129 2.5 191 3.7 158 3.1
104 2.0 95 1.9 135 2.6 110 2.2
114 2.2 94 1.8
.. . .
141 2.8 116 2.3
. . . .
69 1.3 60 1,2
— — — —
5119 100X 5122 100X 5115 l o o i i 5119 100X
2367 58.7 2548 61.5 2740 61.2 1991 51.0 2159 53.8 2319 53.6
579 14.2 580 14.0 628 14.0 897 23.0 921 23.0 995 23.0
332 8.2 319 7.7 347 7.7 313 8.0 302 7.5 330 7.6
219 5.4 203 4.9 221 4.9 192 4.9 179 4.5 195 4.5
157 3.9 149 3.6 163 3.6 165 4.2 146 3.6 160 3.7
107 2.6 97 2.3 106 2.4 145 3.7 125 3.1 136 3.1
171 4.2 153 3.7 167 3.7 203 5.2 179 4.5 195 4.5
117 2,9 95 2.3 104
! • ?
— — — — — —
4069 100X 4143 100X 4475 1C 0X 3906 100X 4011. 100X 4326 100X
1637 56,0 " 2028 59.3 2260 58.6 1745 53.5 1897 55.7 2102 54.8
494 15.1 506 14.8 576 15.0 696 21.3 743 21.8 843 22.0
260 7.9 y 251 7.4 280 7.3 256 7.9 249 7.3 184 7.5
182 5.5 ^ 180 5.3 211 5.5 163 5.0 156 4.6 181 4.8
128 3.9 110 3.2 126 3.3 124 3.8 113 3.3 128 3.3
97 3.0 93 2.7 108 2.8 122 3.7 109 3.2 130 3.4
1 5 5 4.7 136 4.0 157 4.1 155 4.8 137 4.0 161 4.2
127 3.9 116 3.4 1?6 3,5
— — — — . . . . . .
3280 100X 3420 100X 3856 100X 3261 100X 3404 100X 3B36 100X
1398 53.1 1567 56.1 1812 55.3 1398 53.1 1544 55.4 1791 54.7
407 15.4 424 15.2 499 15.2 540 20.5 583 20.9 681 20.8
229 0.7 243 8.7 289 8.8 196 7.4 196 7.0 238 7.3
140 5.3 137 4.9 162 5.0 139 5.3 129 4.6 154 4.7
108 4.1 101 3.6 126 3.8 108 4.1 106 3.8 133 4.1
87 3.3 81 2.9 132 3.1 94 3.6 86 3.1 109 3.3
129 4.9 117 4.2 133 4.0 156 5.9 142 5.1 166 5.1
(17 5.2 123 4,4
.1 5t
* ,7
. . . . . . . . . . . . . . . . . .
2635 100X 2792 100X 1276 100X 2631 100X 2786 100X 3272 loox
LSOA sample weights) is to move about .2 percentage points of those in the
zero difficulties categories to a category of having difficulty with one or more
activity, for both ADL and IADL. In 1988 and 1990 the nonresponse weights
decrease the zero difficulties category by roughly .8 percentage points for both
ADL and IADL, and the effect is distributed among the other categories. In
1990 there is a subtle trend for the target of this redistribution to be among
categories of more disability. In fact, for IADL in 1990, the nonresponse
weights move people out of both the zero difficulties category and the one
difficulty category.
In this table it can be seen that by 1990 the actual sample responding to
ADL questions had diminished to 2,635. The nonresponse weights have
increased the sample responding to this question by 641 persons, making the
sample just 18 shy of what the total sample would have been in 1990 if the
only source of nonresponse had been death.
Discussion
For those using panel data with older samples, the process of
identifying salient characteristics and computing probabilities of becoming a
nonrespondent may be used to develop weights to compensate for decrement to
the sample. The approach suggested here allows for optimal use of each
sample members’ contribution to the survey by estimating response
probabilities within each interval and applying these to form a cumulative
47
weight. Such a method is a step toward restoring both sample size and
redistributing characteristics to maintain representativeness over time.
Summary and Discussion
All longitudinal panel studies are susceptible to selection because
sample persons either chose not to respond, become unable to respond, or
because sample persons relocate and cannot be found. This study addresses
the extent of the loss, the degree to which the loss is associated with certain
hypothesized characteristics, and it evaluates the effect of a weighting scheme
to compensate for nonrandom nonresponse.
Examination of the process of decrement to the sample in the
Longitudinal Study on Aging reveal considerable stability in the rates of loss to
both death and nonresponse. The assessment of sample dynamics suggests that
a critical component in the ultimate level of nonresponse in a panel is the
accumulation of nonresponders over time. And, although patterns of
nonresponse and subsequent response are the least common pattern, such
sequences make up a non-negligible 6 percent of the patterns. Thus, both
preventive measures to keep sample person from being lost to follow-up and
efforts to reinterview nonrespondents should be pursued. Some nonresponse is
inevitable, and efforts to implement rigorous field protocol to ascertain the
precise reason or circumstance for the nonresponse should be made.
An event history approach is employed to ascertain whether some
groups are more likely than others to become nonrespondents and whether the
rate of attrition varies over time. The approach taken in this study addresses
49
several important shortcomings of previous analyses of nonresponse: It
investigates a variety of characteristics hypothesized to be associated with
nonresponse within a multivariate framework, and incorporates a dynamic
design to take advantage of time-varying covariates in the analysis of loss to
follow-up over multiple interviews.
To some extent, who will become a nonrespondent can be predicted.
For example, persons who are older, have lower levels of education, are in
worse health, living in an institution, and those who have another sample
person in the household are more likely to become nonrespondents. This last
characteristic is a finding that has not been previously reported and has
important implications for survey design and protocol. Because the inclusion
of multiple sample persons from a given household is a feature that
significantly enrichens a data source, special attention must be paid to retaining
these survey participants. It is a matter of speculation, however, why such
persons are more likely to be lost. Although such factors as the relative older
age of couples selected from the same household and some aspects of failing
health have been controlled, it is possible that either increased frailty or
caregiving demands of one member for the other may reduce survey
participation. Similarly, even though this effect exists regardless of the
"other" sample person’s response status, a decision dynamic may exist in
50
households with multiple sample persons that works against continued
participation. This would be a fruitful area for future research.
Isolating specific characteristics such as these facilitates several
compensatory approaches. First, on the level of survey design and protocol, it
allows subgroups be targeted for either oversampling or for more intensive
strategies for retaining those persons in the sample. Second, for those using
panel data, this strategy for identifying characteristics associated with
nonresponse is especially useful because it maximizes the information from
time-varying covariates. This allows a more precise specification of those
characteristics that are associated with nonresponse. Importantly too, it has
permitted us to examine the effect of duration of time spent in the study on
propensity to become a nonrespondent, net of other factors. And, precise and
complete identification of characteristics associated with nonresponse is
important for constructing appropriate compensatory weights.
Most statistical evaluations of the selectivity effects of nonresponse do
find a significant nonrandom component. However, whether the selection will
introduce bias into the results can only be assessed by comparing corrected and
uncorrected results. In this study a strategy for compensating for nonresponse
is proposed that uses a cumulative weighting scheme. This approach is
different from common approaches that weight for wave nonresponse
occurring over the course of the survey based on characteristics of those later
non respondents at the inception of the survey. Such a strategy does not take
into account the possibility of change in characteristics between that first
interview and the time at which the sample member becomes a nonrespondent;
this constitutes a considerable waste of information. Because the approach
presented here retains sample persons in the analysis until the wave in which
they do not respond for the first time, nonresponders contribute the maximum
amount of information. Because this approach takes into account the most
recently collected information on the nonrespondents’ characteristics, the
weights are as precise as possible within the constraints of the data.
A comparison of the approach to weighting the data to compensate for
nonresponse indicates that the weights do correct prevalence estimates to adjust
for nonresponse on observed characteristics associated with nonresponse and
also restore the effective sample size to what it would have been if the only
source of decrement were death.
52
References
Adams, M. M. E., Scherr, P. A., Branch, L. G., Herbert, L. E., Cook, N.
R., Lane, A. M., Brock, D, B., Evans, D. A., & Taylor, J. O.
(1990). A comparison of elderly participants in a community survey
with nonparticipants. Public Health Reports. 105(6L 617-622.
Allison, P. D. (1984). Event History Analysis: Repression for Longitudinal
Event Data. Newbury Park: Sage.
Burtless, G. (1987). Occupational effects on health and income among the
elderly. In G. Burtless (Ed.), Work. Health and Income Among the
Elderly (pp. 103-150). Washington, D.C.: Brookings Institution.
Chyba, M. M., & Fitti, J. E. (1991). Response experience in a longitudinal
study of older persons. Proceedings of the American Statistical
Association: Section on Survey Research Methods. 135-140.
Corder, L. S., & Manton, K. G. (1991). National surveys and the health and
functioning of the elderly: The effects of design and content. Journal of
the American Statistical Association. 86(414), 513-525.
Corder, L. S., Woodbury, M., & Manton, K. G. (1992). Loss to follow-up
assessment: Application of grade of membership methods to aged
population longitudinal sample loss to follow-up in the National Long
Term Care Survey. Proceedings of American Statistical Association:
Section on Survey Research Methods. 357-361.
Downes, J. (1952). The longitudinal study of families as a method of research.
Milbank Memorial Foundation Quarterly Bulletin. 30. 101-118.
Fay, R.E. (1989). Estimating non-ignorable non-response in longitudinal
surveys through causal modeling. In D. Kaspryzyk, G. J. Duncan, G.
Kalton, and M. P. Singh (Eds.), Panel Surveys. New York: John Wiley
and Sons.
Fitti, J. E. & Kovar, M. G. (1987). A multi-mode longitudinal study of aging.
Proceedings of the American Statistical Association Section on Survey
Research Methods.
53
Foley, D. J., Branch, L. G., Madans, J. H., Brock, D. B., Guralnik, J. M.,
& Williams, T. F. (1990). Physical Function. In J. C. Coroni-Huntley,
R. R. Huntley, & J. J. Feldman (Eds.), Health Status and Well-being
of the Elderly: National Health and Nutrition Examination Survev-I
Epidemiologic Follow-up Study (pp. 221-236). New York: Oxford
University Press.
Goldscheider. F. (1993). Does it make sense to talk about changes in family
structure? Presentation at the Andrus Gerontology Center
Multidisciplinary Colloquium Series, January 1993.
Herzog, A.R. & Rodgers, W.L. (1992). The use of survey methods in
research on older Americans. In R. B. Wallace & R. F. Woolson
(Eds.), The Epidemiologic Study of the Elderly. New York: Oxford
University Press.
Hill, M. S. (1992). The Panel Study of Income Dynamics: A user’s guide.
Newbury Park: Sage.
Iannacchione, V. G., Milne, J. G., & Folson, R. E. (1991). Response
probability weight adjustments using logistic regression. Proceedings of
the American Statistical Association: Section on Survey Research
Methods. 637-642.
Jay, G. M., Liang, J., Liu, X., & Sugisawa, H. (1993). Patterns of
nonresponse in a national survey of elderly japanese. Journal of
Gerontology. 48(3), S143-S152.
Kalton, G. (1990). Comment on "Theory and issues in weighting survey data."
Proceedings of the American Statistical Association: Section on Survey
Research Methods. 147-149.
Kalton, G., Kasprzyk, D., & McMillen, D. B. (1989). Nonsampling errors in
panel surveys. In D. Kasprzyk, G. Duncan, G. Kalton, & M. P. Singh
(Eds.), Panel Surveys (pp. 249-270). New York: John Wiley & Sons.
Kalton, G. & Kish, L. (1981). Two efficient random imputation procedures.
Proceedings of the American Statistical Association: Section on Survey
Research Methods. 146-151,
Kalton, G., Lepkowski, J., & Lin, T.K. (1985). Compensating for wave
nonresponse in the 1979 ISDP Research Panel. Proceedings of the
American Statistical Association: Section on Survey Research Methods.
372-377.
Kalton, G., Lepkowski, J., Montanari, G. E., & Maligalig, D. (1990).
Characteristics of second wave nonrespondents in a panel survey.
Proceedings of the American Statistical Association: Section on Survey
Research Methods.. 462-467.
Kalton, G., & Miller, M. E. (1986). Effects of adjustments for wave
nonresponse on panel survey estimates. Proceedings of the American
Statistical Association: Section on Survey Research Methods. 194-199.
Kasprzyk, D. & McMillen, D. B. (1987). SIPP: Characteristics of the 1984
panel. Proceedings of the American Statistical Association: Social
Statistics Section. 181-186.
Kish, L. (1990). Weighting: Why, when, and how? Proceedings of the
American Statistical Association: Section on Survey Research Methods.
121-130.
Kovar, M. G. & Chyba, M. M. (1992). The Longitudinal Study of Aging:
1984-1990. Vital and Health Statistics. 1(28). Hyattsville: National
Center for Health Statistics.
Lepkowski, J. M. (1989). Treatment of wave nonresponse in panel surveys. In
D. Kasprzyk, G. Duncan, G. Kalton, & M. P. Singh (Eds.), Panel
Surveys (pp. 348-374). New York: John Wiley & Sons.
Lepkowski, J. M ., Kalton, G. & Kasprzyk, D. (1989). Weighting adjustments
for partial nonresponse in the 1984 SIPP panel. Proceedings of the
American Statistical Association: Section on Survey Research Methods.
296-301.
Lillard, L. A. (1989). Sample dynamics: Some behavioral issues. In D.
Kasprzyk, G. Duncan, G. Kalton, & M. P. Singh (Eds.), Panel
Surveys (pp. 497-511). New York: John Wiley & Sons.
Little, R. J. A. (1986). Survey nonresponse adjustments for estimates of
means. International Statistical Review. 54(2). 139-157.
55
Little, R. J. A. & Rubin, D. B. (1987). Statistical Analysis with Missing Data.
New York: Wiley.
Maddox, G. L. & Clark, D. O. (1992). Trajectories of functional impairment
in later life. Journal of Health and Social Behavior. 33. 114-125.
Manton, K. G. (1988). A longitudinal study of functional changes and
mortality in the United States. Journal of Gerontology. 43. S153-S161.
Mor, V., Murphy, J., Masterson-Allen, S., Willey, C., Razmpour, A.,
Jackson, M. E., Greer, D. & Katz, S. (1989). Risk of functional
decline among well elders. Journal of Clinical Epidemiology, 49(9),
895-904.
Norris, F. H. (1985). Characteristics of older nonrespondents over Five waves
of a panel study. Journal of Gerontology. 40(5). 627-636.
Rogers, R. G ., Rogers, A. & Belanger, A. (1992). Disability-free life among
the elderly in the United States. Journal of Aging and Health. 4(1), 19-
42.
Rogers, W. L., & Herzog, A. R. (1992). Collecting data about the oldest old:
Problems and procedures. In R. M. Suzman, D. P. Willis, & K. G.
Manton (Eds.), The Oldest Old (pp. 135-156). New York: Oxford
University Press.
Rowland, M. L., & Forthofer, R. N. (1993). Adjusting for nonresponse bias
in a health examination survey. Public Health Reports. 108(3).
380-386.
Schaie, K. W ., Labouvie, G. V., & Barrett, T. J. (1973). Selective attrition
effects in a fourteen-year study of adult intelligence. Journal of
Gerontology. 28(3). 328-334.
Shryock, H. S. & Siegel, J. S. (1976). The Methods and Materials of
Demography. New York: Academic Press.
Siegler, I. C., & Botwinick, J. (1979). A long-term longitudinal study of
intellectual ability of older adults: The matter of selective subject
attrition. Journal of Gerontology, 34(2). 242-245.
56
Speare, A. J., Avery, R., & Lawton, L. (1991). Disability, residential
mobility, and changes in living arrangements. Journal of Gerontology.
46(3), S133-S142.
Streib, G. F. (1982). Participants and drop-outs in a longitudinal study.
Journal of Gerontology. 21, 200-209.
Teachman, J. (1982). Methodological issues in the analysis of family
formation and dissolution. Journal of Marriage and the Family. 44,
1037-1052.
Vuchinich, S., Teachman, J. & Crosby, L. (1991). Families and hazard rates
that change over time: some methodological issues in analyzing
transitions. Journal of Marriage and the Family. 53(4). 898-912.
Wright, R. A. (1983). Survey design and administration of the National
Medical-Care Utilization and Expenditure Survey. Proceedings of the
19th National Meeting of the Public Health Conference on Records and
Statistics (DHHS-PHS84-1214), 177-180.
Yamaguchi, Kazuo (1991). Event History Analysis. Newbury Park: Sage.
57
INFORM ATION TO USERS
This manuscript has been reproduced from the microfilm master. UMI
films the text directly from the original or copy submitted. Thus, some
thesis and dissertation copies are in typewriter face, while others may
be from any type of computer printer.
The quality of this reproduction is dependent upon the quality of the
copy submitted. Broken or indistinct print, colored or poor quality
illustrations and photographs, print bleed through, substandard margins,
and improper alignment can adversely afreet reproduction.
In the unlikely event that the author did not send UMI a complete
manuscript and there are missing pages, these will be noted. Also, if
unauthorized copyright material had to be removed, a note will indicate
the deletion.
Oversize materials (e.g., maps, drawings, charts) are reproduced by
sectioning the original, beginning at the upper left-hand corner and
continuing from left to right in equal sections with small overlaps. Each
original is also photographed in one exposure and is included in
reduced form at the back of the book.
Photographs included in the original manuscript have been reproduced
xerographically in this copy. Higher quality 6" x 9" black and white
photographic prints are available for any photographs or illustrations
appearing in this copy for an additional charge. Contact UMI directly
to order.
A Bell & Howell Information Company
300 North Z eeb Road. Ann Arbor. M l 48106-1346 USA
313/761-4700 800/521-0600
UMI Number: 1 3 7 6 4 8 5
UMI Microform 1376485
Copyright 1995, by UMI Company. All rights reserved.
This microform edition is protected against unauthorized
copying under Title 17, United States Code.
UMI
300 North Zeeb Road
Ann Arbor, MI 48103

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Cross-Sectional And Longitudinal Effects Of Parental Divorce On Parent-Adult Child In-Person Contact

PDF

Determinants Of Neonatal And Postneonatal Mortality: The Case Of Malawi, 1992

PDF

Health And Well-Being By Marital Status: The Effect Of Age, Sex, Income And Social Support. Study Of 1985 Gss Data

PDF

Comparison of gene expression of SCG10 and Stathmin/p19 in aging rat brain: an in situ hybridization study

PDF

Survey of registered nurses' attitudes toward elderly patients

PDF

A comparative study of older and younger adults who have received job-specific classroom training to determine if there are significant differences as they relate to job placement

PDF

Complementarity problems over matrix cones in systems and control theory

PDF

Lateral variability in predation and taphonomic characteristics of turritelline gastropod assemblages from Middle Eocene - Lower Oligocene strata of the Gulf Coastal Plain, United States

PDF

Effects of age and gender on speed and accuracy of hand movements: and the refinements they suggest for Fitt's Law

PDF

The characterization of Huntington Beach and Newport Beach through Fourier grain-shape, grain-size, and longshore current analyses

PDF

Gender-biased education investment intention among Chinese rural parents: A glimpse into the intra-household resource allocation process

PDF

Filial expectations and social exchange patterns among older Taiwanese parents and their adult children

PDF

A qualitative study on the relationship of future orientation and daily occupations of adolescents in a psychiatric setting

PDF

Predictors of pretend play in Korean-American and Anglo-American preschool children

PDF

Occupational exposure to extremely low frequency electromagnetic fields as a potential risk factor for Alzheimer's disease

PDF

Intergenerational social support and the psychological well-being of older parents in China

PDF

Meta-analysis on the misattribution of arousal

PDF

Hand function in older adults: the relationship between performance on the Jebsen Test and ADL status

PDF

Processes of loss and recovery of physical abilities in a longitudinal study of Americans 70 years of age and older

PDF

A statical analysis and structural performance of commercial buildings in the Northridge earthquake

Asset Metadata

Creator Mihelic, Adrienne H. (author)

Core Title An analysis of nonresponse in a sample of Americans 70 years of age and older in the longitudinal study on aging 1984-1990

School Graduate School

Degree Master of Science

Degree Program Sociology - Applied Demography

Degree Conferral Date 1995-05

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

Tag gerontology,OAI-PMH Harvest,sociology, demography,sociology, theory and methods

Language English

Contributor Digitized by ProQuest (provenance)

Advisor Crimmins, Eileen M. (committee chair), Heer, David M. (committee member), Silverstein, Merril (committee member)

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c18-3224

Unique identifier UC11357810

Identifier 1376485.pdf (filename),usctheses-c18-3224 (legacy record id)

Legacy Identifier 1376485-0.pdf

Dmrecord 3224

Document Type Thesis

Rights Mihelic, Adrienne H.

Type texts

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

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

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus, Los Angeles, California 90089, USA