| HOME | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|---|
| ||||||||||||||||||||||||
RESEARCH ARTICLE |
1 Department of Human Development and Family Studies, Texas Tech University, Lubbock.
2 Leonard Davis School of Gerontology and 4 Department of Psychology, University of Southern California, Los Angeles.
3 Department of Sociology, California State University, Los Angeles.
Address correspondence to Dr. Du Feng, Department of Human Development and Family Studies, Texas Tech University, Box 41162, Lubbock, TX 79409-1162. E-Mail: du.feng{at}ttu.edu
 |
Abstract |
|---|
 TOP Abstract MethodsResultsDiscussionReferences |
|---|
Methods. The study sample was a subset of the Longitudinal Study of Generations and included data from 208 G1–G2 parent–child dyads and 538 G2–G3 dyads over 30 years. We used a latent growth–decline curve model based on full information maximum likelihood estimation in order to compare parents' and adult children's reports on older respondents' health and intergenerational solidarity by parents' attrition status.
Results. Results indicated that attrition due to mortality biased estimates of respondents' assessments of their functional health status over time, and parents' perceptions of the quality of the parent–child relationship deteriorated more rapidly among those who died by Time 7, but nonresponse did not seriously bias estimates of these measures. Using proxies, we found that functional impairment increased more rapidly when children reported about parents, especially in advanced old age.
Discussion. These results support the use of full information in estimating growth curves where mortality is present but raise concerns when using child proxies to evaluate parental health or the quality of intergenerational relationships.
ATTRITION of older respondents in longitudinal studies is almost universal, most often occurring due to death and incapacitation as well as weak commitment to research. However, little is known about the degree to which attrition biases estimation of trends in the health and family life of older people over time, and whether the use of proxy reports serves as a corrective to such biases. The purpose of our investigation was to examine (a) how growth–decline curves differ between older people with complete data and those with missing or incomplete data over time, and (b) whether proxy child-informants are reasonable substitutes for absent older respondents when estimating such growth–decline curves.
Using data collected by the Longitudinal Study of Generations (LSOG), a unique multiwave, multigenerational study, we examined two attributes of older adults that have received close scrutiny in gerontological studies: functional health and the quality of their intergenerational relationships. That the former is an attribute of the older individual (about which the informant may or may not have accurate knowledge), and the latter a joint social product (that both the respondent and the informant play a role in defining), provides the opportunity to examine attrition bias and the efficacy of proxy reports under two measurement conditions: where the unit of interest is the older individual and where it is the dyadic relationship.
In examining attrition patterns of the LSOG sample, our diagnostics (not shown) indicated that mortality was the dominant reason for sample loss among older respondents, with mental and physical incapacitation the major nonmortality reasons for sample exit. In addition, older parents with weaker emotional ties to their adult children were more at risk of exiting the sample than were those with stronger ties. Thus, in the presence of such systematic attrition, any analysis of health and family trends over the course of later life that relies only on complete data may underestimate the extent of health declines (due to attrition of the most frail) or overestimate growth in family cohesion (due to attrition of the least cohesive families). In this research we calculate the degree to which attrition in more than 30 years of longitudinal data has created statistical distortions in researchers' understanding of trajectories in the health and intergenerational relationships of older adults.
Methods for Handling Attrition
Researchers largely acknowledge classical approaches to handling systematic attrition based on listwise deletion or single imputation strategies as inadequate for reducing sample selection bias (Newman, 2003
). The most promising contemporary approaches to handling attrition include multiple imputation (Little, 1995) and maximum likelihood estimation (MLE). In this analysis, we focused exclusively on the MLE approach as a more general strategy to estimate growth–decline in later life when the reasons for sample-exit were age-specific.
MLE produces unbiased coefficients when it can be assumed that absent data are "missing at random." The missing-at-random assumption holds when the attrition process is systematic but measurable; that is, conditional on observed variables in the model (Little & Rubin, 1987). Notably, MLE under the missing-at-random assumption estimates unbiased parameters when all available data from respondents are analyzed regardless of the completeness of their longitudinal response patterns—a procedure known as full information MLE (Cnaan, Laird, & Slasor, 1997; McArdle & Hamagami, 1992, 2001).
Proxy Data From Adult Children
Although researchers often use data from informants in studies of elders who suffer from dementia (e.g., Beckett, 2000; Magaziner, 1997; Neumann, Araki, & Gutterman, 2000; Parker, Morgan, & Dewey, 1997; Stancliffe, 1999), it is generally not possible to compare these proxy data to similar data collected from the afflicted individuals. We take a more generalized approach to examining the utility of proxy data by relying on a full range of competency in the general older population captured in the LSOG sample.
The LSOG collected data from both elderly parents (primary respondents) and their adult children (informants), allowing us to examine how well children's reports about their parents' health and intergenerational relationships captured parents' reports about the same outcomes. In this research we were able to take advantage of the multigenerational design of the LSOG survey by examining the efficacy of proxy data. Because proxy data tend to be more accurate for objective information (Cummins, 1998), we note that the four variables we examined range along a continuum from more objective to more subjective: (a) Parents' functional health as measured by their ability to perform activities of daily living (ADLs), a fairly factual account of everyday competence but one that may be shaded by the parent's sense of autonomy and the child's vigilance toward the parent; frequency of (b) in-person contact and (c) phone contact between parents and children, behavioral dimensions of families that have a factual basis but are open to systematic distortion based on social desirability and expectations built into family roles; and (d) emotional closeness between parents and children, a subjective feeling that is most open to differences in perception and interpretation, with older generations tending to evaluate their intergenerational relationships more favorably than younger generations (Giarrusso, Stallings, & Bengtson, 1995). Clearly, all four variables are subject to some degree of cross-generation reporting bias. Although researchers have identified differences between adjacent generations in terms of absolute response levels, they have rarely investigated them with respect to response trajectories over time. Thus it is still possible that on a longitudinal basis, adult children's proxy data may be useful in adjusting for growth trends that are truncated by nonresponse in the older generation. Following on this discussion, we asked the following research questions:
|
METHODS |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
The survey was repeated in 1985, 1988, 1991, 1994, 1997, 2000, and 2004–2005 (T2–T8). The eligible sample in waves subsequent to T1 included all family members who were eligible at baseline (even if they did not respond at baseline) except for participants who (a) could not be located between T1 and T2 (n = 423), (b) had subsequently died or become mentally or physically incapacitated (n = 783), or (c) had divorced a G3 lineage member and had no G4 children (n = 33). The longitudinal response rate between T1 and T2 was 73%, with an average of 80% between waves since 1985. The sample is more than 90% Caucasian.
The subsample for the present study included 208 G1–G2 and 538 G2–G3 parent–child dyads that participated in any survey between T1 and T7. In terms of gender composition, 43% of G1–G2 dyads were mother–daughter, 15% mother–son, 31% father–daughter, and 11% father–son. Among G2–G3 dyads, 33% were mother–daughter, 23% mother–son, 23% father–daughter, and 21% father–son. G1 parents' mean age at baseline (T1) was 64.07 years (SD = 5.65), G2 parents' was 43.39 years (SD = 5.41). In our subsequent analyses, we pooled G1–G2 and G2–G3 dyads, resulting in 746 parent–child dyads with some fluctuation in sample size depending on the outcome variable of interest.
As in any long-term longitudinal study, particularly with older samples, very few respondents participated in every wave of measurement. Respondents may have exited the sample permanently due to death, or may have exited the sample in one wave with the potential to reenter it at a later wave. Therefore, the longitudinal patterns of nonresponse were quite complex. In order to differentiate nonresponse due to death from other types of nonresponse, and in order to create groups of sufficient sample size, we separated dyads into three groups based on the parent's participation status at T7: returned a survey, alive but did not respond, and died. Among G1–G2 dyads, 16% of the parents returned a survey, 9% were alive but did not respond, and 75% had died. Among G2–G3 dyads, 62% of the parents returned a survey, 17% were alive but did not respond, and 21% had died.
Measures
Dependent measures included (a) parents' self-reported functional health and children's reports of their parents' functional health; parents' and children's reports of the frequency of (b) in-person and (c) phone contact; and (d) the strength of the emotional bond they had with each other. Selected ADL and instrumental ADL (IADL) items measured parents' functional health from T2 to T7. Respondents answered questions on parent's level of difficulty in walking up and down stairs, walking more than one block, preparing meals, doing household chores, and taking care of own personal hygiene needs on a scale from 1 (no difficulty) to 4 (unable to do at all). Higher ADL and IADL scores indicated poorer functional health.
We measured intergenerational contact at each wave according to frequency of in-person and phone contact, with higher scores indicating more frequent contact (1 = not at all, 6 = daily). Researchers used Bengtson's Affectual Solidarity scale (Bengtson & Schrader, 1982) to measure emotional closeness at each wave. This 5-item 6-point (1 = not at all, 6 = very much/a great deal) scale includes questions on communication, getting along, closeness, and understanding between parents and children. Internal consistency reliability of this scale was between.85 and.92 across the seven waves.
Independent variables included parent's age (measured in years), parent's generation (–1 = G1, +1 = G2), parent's and child's gender (–1 = male, +1 = female), and parent's participation status at T7. By effect coding generation and gender such that the values equaled –1 and +1, the intercepts of Level 2 equations of the hierarchical linear modeling (HLM) analyses (see below) reflected the average of G1 and G2, and the average of men and women, on the respective dependent variables. As for parent's participation status, those who returned a survey at T7 served as the reference group, and we compared them to those who did not respond and to those who had died. Thus, we created two dummy variables, one indicating whether the parent was alive but did not respond at T7 and the other indicating whether the parent had died by T7. We centered parent's age around the grand mean (approximately 65 years) and divided by 10 (i.e., age per decade) in order to rescale the estimated fixed-effect coefficients obtained in the HLM analyses.
Analytic Strategy
Our analyses had two major goals: (a) In order to investigate the impact of attrition on estimated growth, we estimated growth–decline curves for parents under different participation statuses (complete data, incomplete data due to death, and incomplete data due to nonresponse). (b) In order to investigate whether reports by adult child informants recapitulated the growth trends of their parents and served as reasonable substitutes for them, we used proxy data from children to compare their growth–decline curves to those of their older parents.
The HLM program (Raudenbush, Bryk, Cheong, & Congdon, 2000) generated growth–decline trajectories using all available data to estimate latent growth curves. At Level 1 of the HLM analyses, we used the following equation to estimate linear and quadratic slopes in each of the dependent measures by age of parent (we centered age around the grand mean and rescaled it as age per decade):
|
|
At Level 2, we tested heterogeneity in effects of age on the dependent measures using parent's generation, parent's and child's gender, and the two dummy variables indicating parent's participation status. We wrote the Level 2 equations for each of the three components (i.e., intercept and linear and quadratic slopes) c = 1, 2, and 3, as
|
|
|
RESULTS |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
Table 1 presents fixed estimates predicting the three random effects (intercept and linear and quadratic slopes) for each of the dependent measures, assessed from the parent's and the adult child's perspectives. For parents' functional health reported from their own perspective, coefficients revealed that parents whose nonresponse at T7 was due to mortality had significantly more functional limitations at age 65 (the random intercept) than those who responded at T7 (B =.160, p <.01). However, they deteriorated with age less rapidly, as evidenced by their negative linear (B = –.022, p <.05) and quadratic effects (B = –.014, p <.05). Parents who responded at T7, although they had better health on average than those lost to mortality, exhibited more accelerated deterioration in their functional health as seen in the positive constant in the quadratic equation (B =.051, p <.01). The estimated growth curve for parents who did not respond but who were alive at T7 was indistinguishable from that for the reference group (B =.032, ns; B = –.011, ns; and B = –.005, ns, for the intercept, linear slope, and quadratic slope, respectively).
|
Turning to frequency of in-person and telephone contact between parents and their adult children, we see that growth trends were insensitive to parental attrition (none of the coefficients for effects of the parental participation status variables were significant). There was significant change in phone contact over time, from both parents' and children's perspectives, with the reference group showing a significant linear increase (B =.260, p <.01, from the parent; B =.293, p <.01, from the child) followed by a deceleration (B = –.142, p <.01, from the parent; B = –.094, p <.01, from the child). This suggests a rise and fall in phone contact with age. This curvilinear growth–decline pattern demonstrated with respect to the reference group of parents who participated at T7 was characteristic of the other two parental participation groups as well (as seen in Table 1, there were no significant group differences in the coefficients for the intercept or the slopes).
With regard to affectual solidarity, the results for participation status of parents from both parents' and children's perspectives revealed a stronger linear decline in affection when parents were lost to mortality (B = –.054, p <.01, for parent; and B = –.063, p <.01, for child). Overall, however, the decline accelerated only when children reported about the relationship (B = –.059, p <.05), but the rapidity of the decline did not vary by parent's participation status (B = –.021, ns, for parent alive but did not respond at T7; and B =.008, ns, for parent died by T7).
In order to more fully inspect the patterns suggested by estimates predicting parents' health in Table 1, we show in Figure 1 mean trajectories of parental health by participation status and the source of the report (parent or child). Comparing the three curves based on the parents' reports with the three curves based on children's reports, we can see that (a) both parents and children reported accelerating rates of impairment roughly after age 80, and (b) the increase in impairment appeared to be more rapid when taken from the child's perspective.
|
|
DISCUSSION |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
With regard to use of younger generations as proxies, we found that functional impairment tended to increase more rapidly when children reported about parents, especially in advanced old age. It is not possible to establish a true ADL or IADL score that forms a benchmark against which to evaluate participant responses; yet our results are consistent with research showing that older parents underestimate their dependency on adult children (in an effort to present themselves as autonomous) whereas adult children overestimate their parents' dependency on them (as a sign that they are fulfilling their filial duty; Marshall, 1995). Not surprisingly, affectual solidarity is stronger from the perspective of the parents' generation than from the perspective of children's generation, a finding consistent with the intergenerational stake phenomenon (Giarrusso, Feng, & Bengtson, 2005).
Overall, evidence based on this study tends to support the use of a full information MLE approach to handling mortality attrition. However, we found no biases with regard to nonresponse—a category dominated by cognitively impaired and physically frail older adults but one that also included those with low commitment to research. This suggests that earlier data obtained from sample dropouts who are still eligible for study inclusion are able to reproduce growth trajectories that are very similar to those with more complete data. There is evidence as well that proxies (i.e., adult children, in the current study) are better at estimating the quantity of parent–child interaction than they are at estimating the quality of their relationship. This confirms our expectation that informants are better at documenting more objective aspects of their intergenerational relationships.
In future research we plan to take advantage of the full-family design of LSOG by examining multiple child-informant dyads in the same families (i.e., siblings) in order to provide a more accurate consensus about the absent parent than relying on one child allows. It may also be possible to impute data for missing parents by adjusting proxy reports for known distortions based on generational position. But to better understand the metabolism of health and intergenerational relations over the last half of life and to minimize biases related to selective attrition, researchers must first acknowledge the mixture of groups represented in the older population in terms of their risk of, and reasons for, attrition.
|
Acknowledgments |
|---|
|
Footnotes |
|---|
Received for publication September 22, 2005. Accepted for publication February 1, 2006.
|
References |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
| ||||||||||||||||||||||||
| HOME | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|---|
