Uses of Panel Study of Income Dynamics Data in Research on Aging

Late-life consequences of early-life exposures

There is a growing recognition that conditions and contexts of early life shape later health, well-being, and mortality outcomes. This is partly due to accumulating evidence produced using data from the PSID and other longitudinal data sources. One of the markers of items in this literature is an allusion to the “long arm of childhood” (e.g., Hayward and Gorman 2004; Latham 2015). The contexts considered in this literature include the household, the neighborhood, and the place of residence (i.e., city, county, or even state); they can also be traced to social institutions (e.g., the military) and to interaction between individual-level behaviors and contextual factors. While much contextual information can be collected in, or inferred from, responses to survey items such as those used in the PSID, the existence of geographic identifiers (such as census tract or county-of-residence codes) in the PSID opens up a vast potential for linkage of individual longitudinal records to area-specific information.

Several PSID-based studies adopt some form of this long-time-horizon research design. For example, Karraker, Schoeni, and Cornman (2015) select a sample of male PSID respondents aged 20 to 40 in 1972 (their baseline year), tracking their survivorship and mortality over a 35-year follow-up period. Although their principal focus is on the association between several psychological traits and longevity, they do control for one early-life household-contextual factor, namely, living in a household that contains smokers. Having lived in a household containing a smoker significantly increases the chances of early death, whether in a baseline-only model or a model that includes measures of marital status and income over the life cycle. However, if the respondent’s own later smoking behavior is added to the model, the effect of the earlier-life exposure to smoking loses significance.

The own-household context depicted in the Karraker, Schoeni, and Cornman (2015) study is located near the “micro” level end of a continuum for which one’s country of residence might represent the most “macro” of levels. Lillard et al. (2015) adopt the latter focus, analyzing the effects on self-reported health status over the adult life cycle of exposure to income inequality prior to adulthood. Their measure of income inequality is constructed for the United States as a whole. Using 25 years of PSID data, they find a consistent and robust health-reducing effect of early life exposure to income inequality, even after controlling for additional early-life exposures (such as childhood health and poverty status), parents’ education, and one’s own lifetime income.

To introduce neighborhood-, city-, or county-level contextual factors into a PSID data analysis, it is necessary to link external information using geo-coded identifiers. An example of this approach is the study by Spring (2017), who uses PSID sample members age 45 or older in 1999, following them through 2013, at which time the surviving individuals are 59 to 104 years old. Spring focuses on the built environment, using indicators of the per-capita density of supportive services (seven types of establishments) and of commercial decline (using three categories of establishments). Counts of both categories of establishments come from the economic census. Quartiles of the two density measures are then determined, and each neighborhood (zip code) is categorized, using judgmental coding, into one of five built-environment types (e.g., “low density,” “service dense,” “commercially declined,” and so on). Spring examines the association between short-term (one-year) exposure and long term (a cumulative measure, averaging over all survey waves) exposure to these environmental types, and finds significant differences in self-rated health across selected categories of exposure.

Another example of this category of research design is found in Johnson and Schoeni (2011), who use data on nearly 4,400 children who were in the initial (1968) PSID sample and followed up to 2007. Johnson and Schoeni investigate the distinct effects of weight at birth, childhood poverty, and other family-background characteristics, along with local-area factors measured at a very fine level of geographic detail—census tracts—on the incidence of chronic diseases and conditions during adulthood. They find significant effects of childhood poverty on the onset of hypertension, diabetes, and cardiovascular diseases, all of which are well-established factors in late-life disability (Freedman and Martin 2000).

While Spring’s (2017) results reveal the power of local-area contextual factors to explain individual-level health outcomes, they also serve to illustrate some of the challenges posed by the linkage of longitudinal survey data to area-specific contextual measures. For example, there are multiple levels of relevant context, and these distinct contexts may not fall into neatly nested hierarchies. Whereas multiple zip codes—widely used in applied research to represent “neighborhood”—may align neatly with county boundaries, there are many examples of local labor markets that encompass more than one county, and may even cross state lines (for example, W. Connecticut–New York City–Northern New Jersey; or Philadelphia–Camden and adjacent areas). Moreover, not only may features of differing spatial aggregates within a hierarchy exert influences on individual or household outcomes, there may be additional influences that can be traced to adjacent contexts as well (e.g., housing prices in San Francisco influence housing and travel behavior many miles away). State-level policies or programs may influence individual-level outcomes throughout a state, but they may, as well, be modified at the county level within the state. Finally, data-reduction techniques—such as those employed in Spring’s study, which maps the population densities of ten types of establishments into a single five-level categorical variable—are an understandable response to a need to conserve degrees of freedom when there may be a large number of contextual dimensions. However, the resulting summary measures might obscure which of the several underling observable attributes are the most influential. These difficulties are further amplified when we consider multiyear sequences of contextual exposures, whether experienced over a child’s or an adult’s lifetime: adverse exposures may matter more at some ages than others; they may matter only if sustained over some well-defined set of years; their consequences may fade over time, or may be overcome by a favorable change in environment; and so on. All such questions remain as additional research opportunities, presenting potential users of the PSID with essentially limitless possibilities.

As a final point concerning the usefulness of area-specific measures, it is worth noting that linking individual-level survey data to local-area contextual measures can serve the purpose of addressing omitted variable bias in evaluation or outcomes research. One example (although not one based on PSID data) is the use of geographic distance to the nearest specialty hospital as an instrumental variable, in a study of the effectiveness, with respect to later survival, of catheterization and revascularization in the treatment of acute myocardial infarction (Newhouse and McClellan 1998). Many possible variations on this approach can be envisioned.

Patterns of functioning and disability

Health, functioning, and disability status are among the leading topics of interest in social gerontology. The PSID lends itself to a wide variety of study designs in this area. One such design is the use of longitudinal data on health or disability transitions to calculate active life expectancy (ALE); this can be viewed as a type of “trajectory” model (see [3] above). Even though the PSID did not begin to routinely ask about activities of daily living (ADL)–type disability until 1992, well into its existence, the temporal span of the resulting sequence of disability measures compares favorably with other leading data sources used for this type of research (in particular, the Health and Retirement Study or HRS).

Moreover, compared to other data sources used in ALE research, the PSID allows for an “origin” point much earlier than age 65 or 70 that is often used. Laditka and Laditka (2016) use a sample of nearly nine thousand PSID respondents age 40 and older in 1999 and followed until 2011. They emphasize differences in ALE by race (white and black, only), gender, and level of education, finding (as have others working in this area) that less-educated people live shorter lives, on average, than more-educated people; they also spend a higher percentage of their lives after age 40 living with a disability than do more-educated people.

One question that is gaining increasing attention among researchers, practitioners, and advocates is the differences between people “aging with” disability—that is, spending most or all their adult years with a disability—versus “aging into” disability—that is, making what is generally assumed to be a first transition into having a disability late in life. In particular, one issue on which almost no information can be found is how people’s late-life circumstances differ (with respect to material resources, health status, living and care arrangements, use of services, and so on) according to whether their period of late-life disability began late in life or was preceded by a possibly lengthy period of adult life spent with disability. The PSID, with its long span of coverage, would seem to be an obvious source of data with which to investigate the “with versus into” question. However, I have not located any such study. Clarke and Latham (2014) cover one of the two periods—the “aging with” period—but not the other. Their sample consists of PSID respondents ages 20 to 34 in 1979, and followed until 2009; thus, they obtain good coverage of the adult part of the life cycle. They classify people as “living with” disability if they report having a physical or nervous condition that limits the kind of work they can do—the work-disability question mentioned before—on four or more occasions during the 1981 to 1994 period. Persons so classified are found to report poorer health, to work fewer hours, to have lower probabilities of employment, and to have lower household income than those not aging with disability.

Shuey and Willson (2017), who use a slightly longer sequence of PSID years (1981–2013) and a very different approach to classifying disability patterns compared to Clarke and Latham (2014), address patterns of work disability. Shuey and Willson apply a latent class estimator to their 33-year sequences of work-disability indicators, concluding that a five-class solution provides the best fit; interestingly, they find that about two-thirds of their sample falls into a “never disabled” group, while 4 percent have a high probability of living with a work disability throughout adulthood (ages 22–32 through 54–64). A surprisingly high percentage of the sample (more than 24 percent) arrives at the threshold of “old age” (age 65) with a high chance of having a work disability. And, using probabilistically assigned class-membership indicators as predictors, they show that both the early- and midlife-onset (of disability) classes are less likely to own a home, more likely to be in poverty, more likely to have fewer assets, and less likely to have pension coverage, in 2013 (when they are 54–64), compared to the never-disabled group.

Given the findings from both the preceding studies, it is reasonable to anticipate that the late-life circumstances of those with disabilities during their adult lives will be less favorable than those who have comparable levels of disability that came late in life. This is an interesting and important question, and has implications for policy and practice, yet it is one that remains largely in the not-yet-addressed realm. The PSID also presents opportunities for modeling trajectories in other domains, such as body weight change and cognition, both of which have been shown to exhibit distinctive patterns of downturn in the years immediately preceding death (see Alley et al. [2010] and Sliwinski et al. [2006], respectively). With its long time span and detailed coverage of adult-life circumstances, the PSID seems like a promising basis for detecting early markers of these end-of-life phenomena.

Finally, the topic of late-life disability can be linked to the contextual-factors issue discussed above. Montez, Hayward, and Wolf (2017), for example, demonstrate an association between disability and selected features of one’s state-level policy context (cigarette tax rate and supplementation of the Earned Income Tax Credit). However, these findings are based on cross-sectional data from the American Community Survey and reflect one’s current state of residence only. Using the PSID, it would be possible to account for a much longer history of contextual influences on disability in various policy domains.

Intrafamily and intergenerational financial support and caregiving

An extensive body of research has been directed toward determining whether resource flows within families reflect altruistic motives—the idea that person A’s generosity toward person B increases A’s well-being—or are a form of exchange between self-interested parties—the creation or discharge of a debt for which repayment, in some form and at some time, is expected. This distinction matters, because altruistic resource flows can be seen as a manifestation of “family solidarity” (Bengston et al. 2002); however, altruistically motivated transfers may be “crowded out” by public transfer programs. Moreover, different motivations for intergenerational support may lead to different patterns of inequality within families in sharing the “burden” of parent care. However, despite the many theoretical models that have been developed to explain observed resource flows, there remains a “dominant feeling of disillusion in the [economics] profession about the explanatory power of these models” (Arrondel and Masson 2006, 978).

Apart from the lack of discriminatory power among these various theoretical models, there is an empirical problem: it is difficult to establish whether an observed intrafamily flow of resources is part of a self-interested exchange transaction when the data cover only a brief period in the family members’ lives. A parent may transfer resources to a young-adult child (helping, for example, with a home purchase) in anticipation of repayment in the form of coresidence or personal-care services decades later. Much of the available data on intrafamily support, however, are temporally limited. The PSID’s 1988 intergenerational transfers module, for example, provides measures of personal-care support and financial support between parents and children during only one year—1987—of the lives lived in either of those generations. Those data have been used in many studies (Altonji, Kayashi, and Kotlikoff 2000; Couch, Daly, and Wolf 1999; Furstenburg, Hoffman, and Shrestha 1995; Schoeni 1996), but they cannot provide much resolution to the altruism-exchange issue. The more recent 2013 version of PSID intergenerational transfers data includes measures of support over much longer time periods, suggesting that a more powerful test of the altruism-versus-exchange issue could be undertaken.

In addition to the newer 2013 transfer data, the PSID presents even greater opportunities for research on lifetime patterns of intrafamily support. Researchers in this area often refer to the three “currencies” of familial support—shared housing space (i.e., coresidence), money, and time devoted to personal-care service (Soldo and Hill 1993). Regarding the time horizon for these various forms of support, a recent paper showed that among a surprisingly large percentage of mother-child pairs, and during the part of the life cycle beginning when the mother is 58 years old, the child has never left the parental home (Wiemers et al. 2017). If this shared living arrangement were interpreted as one governed by the terms of an implicit exchange-motivated “contract,” then the time period over which the exchange is conducted is evidently quite long. And, as noted earlier, in addition to the 50-year coresidence histories produced in the PSID, there are equally long sequences of annual information on income received from, and financial help provided to, family members outside the household. Despite the lack of specificity about kin relationships in those latter measures, they clearly represent a form of intrafamily resource flows, and would seem to provide a resource with which to investigate possible exchange-motivated behaviors.

Migration and family geography

Proximity to one’s grown children during late life has been characterized as a resource for family support and long-term care service use (Wolf 1994). Proximity to family members is, of course, a consequence of (and is altered by) any move one makes, as well as any moves made by other family members, leading to a complex set of interacting move- and proximity-variables sequences.

The PSID has throughout its history included self-reported indicators of whether a move has happened (since the last interview) as well as whether the respondent has nearby relatives. Much more detailed information can, however, be developed through the linkage of survey data to geographic locations (i.e., named spatial entities such as tracts, zip codes, or counties or polar coordinates). Sequences of individual locational data can be used to compute the (approximate) ⁸ distance traversed in a residential relocation. Richer still is the more detailed kin-proximity data that can be derived by calculating the (approximate) distances between two different PSID respondents who are linked by known familial relationships. Migration and location choice are heavily influenced by contextual factors (besides the presence of kin in a given location, itself a contextual factor), so the sorts of linkages to external data discussed earlier are again relevant in migration studies.

One example of a PSID-based study that entails most of these features is a new paper by Spring et al. (2017). Their central interest is on the effect of proximity to nonresident kin on whether and to where people move, although they limit their sample to individuals who remain within the same metropolitan area throughout the panel period studied (1980–2013). Using linked records for individuals that trace back to the same 1968 households, they are able to compute the distance from each “ego’s” location to the locations of other PSID individuals to whom they are related. The analysis is conducted at the census tract level, which allows for a very detailed depiction of geography. Moreover, using the tract identifiers they are able to incorporate area characteristics—the number and average value of housing units, and racial composition—into their model of destination choice. Among their findings are that movers tend to move toward aging parents; among whites and higher-income individuals, there is a tendency to move closer to parents and children; and among lower-income individuals, there is a tendency to move toward extended-family kin. This is a very richly formulated analysis and one that makes use of many of the distinctive features of the PSID. Yet (as is often the case) despite this richness, it leaves open many questions, in particular the long-distance moves, and the role of proximity to kin in such moves, that were removed, by design, from the Spring et al. analysis.

Trade-offs in breadth and temporal scope of analysis

Carriers of the PSID “gene” include members of the original 1968 sample families and their descendants. A sample member’s age in 1968 is, as well, an indicator of how many years of presample experience is unrecorded in the data. ⁹ And even though 50 years is a very long time, it is less than the entire lifetime of most people, placing bounds on the portion of sample members’ lifetimes that can be studied. These design features create various trade-offs regarding study design that all users of the PSID must confront. These issues are quite self-evident; nevertheless, it is useful to review them.

Figure 1 shows “lifelines” for individuals from a handful of birth cohorts whose lives overlap with the PSID study period through 2018; the horizontal axis represents years and the vertical axis represents age (0 through 90, although many PSID sample members are observed to even older ages). Each cohort is characterized by its year of birth and by the ages covered by the PSID study period. For example, the earliest cohort shown consists of people born in 1923, and observed at ages 45 (in 1968) to 95 (in 2018). Successively later cohorts—for example, those born in 1938 or 1953—are observed throughout most of their adulthood—ages 30 to 80 and 15 to 65, respectively. As fewer of the years lived after turning 65 are observed, more of those lived prior to adulthood are observed. There are, however, limits on the extent to which childhood, the work life, and postretirement experience can all be directly observed. People born in 1968, the first year of PSID data collection, can be traced throughout their childhood, but their adult years are censored at age 50 (for now—as the PSID continues to age, this cohort’s lifetimes will become increasingly visible).

OPEN IN VIEWER

Figure 1

Coverage by Age for Selected Birth Cohorts in PSID Sample

The trade-offs confronting researchers are evident: (1) people observed late in life (after, say, age 65) have adult lives that are incompletely observed and childhoods that are completely unobserved; (2) for a few cohorts (e.g., those born in 1953), their entire adult lives are observed but little or none of either childhood or old age is observed; and (3) those whose childhoods are fully observed (e.g., born in or after 1968) have a range (from very little to the majority) of their adult lives directly recorded. Assuming that the PSID continues to operate and follow these more recent cohorts’ lives, or course, these limitations will gradually be overcome. The limitations associated with the earlier cohorts cannot, however, be eliminated.

As a consequence of these unavoidable design features, comparisons over time in the circumstances of those over age 65, or alternatively those under 18, are limited to those in a well-defined and somewhat narrow set of birth cohorts. And studies that rely on joint observation of parents and children face even more stringent limits. It is important to recall, of course, that however restrictive any such limitations might be, they are much less restrictive for PSID users than for users of any other longitudinal data source (for the United States). Moreover, with each additional round of data collection, these constraints are incrementally relaxed.

Completeness of family-network representation

The PSID “gene pool” is limited to members of households included in the original 1968 sample and their descendants. For the many study designs that rely on linking the records of parents and their children, this limitation has consequences for the completeness of family networks studied: any children of original PSID respondents who had already “split off” from their parent’s household before 1968 are excluded. Given the strong role of age in the split (or “emancipation”) process, there are additional study-design limitations that must be superimposed on the cohort-representativeness diagram shown in Figure 1.

We can get a sense of the family-representation problem using public-use microdata from the 1970 U.S. Census (Ruggles et al. 2017), which is reasonably close to 1968 for purposes of illustrating the overlap of families and households. ¹⁰ I use the census data to address two questions. First, what is the age profile of coresidence with one’s parents? Second, what is the age profile of not living with all of one’s children (or siblings)? These two questions are just two different ways of addressing the problem of family members who are excluded from the PSID gene pool by design.

An indicator of coresidence with parents is directly provided in the census records. For the analysis of “missing” children (or siblings), I use just two data elements: the number of children ever born (to women) and the number of own children in one’s household, both of which are available in the 1970 Census. I classify a woman’s household as one with “missing children” if the number of children ever born exceeds the number present in the household. ¹¹

Figure 2 shows the proportions of individuals (not distinguished by gender) living with and without parents present. Having neither parent present is rare up to about age 16 and rises sharply at age 18 and thereafter. This suggests, in turn, that among the original 1968 PSID households, nearly all children younger than 18 are present. However, Figure 3 shows that the proportion of young children whose household fails to contain one or more of their (presumably older) siblings rises steadily, from about 0.1 even among the very youngest children to over 50 percent by age 16. Figure 3 also shows that for very young mothers (e.g., 18 or younger), a substantial proportion have had children who are not current coresidents; these missing children may have died (we cannot tell, using these data), or might be living with a grandparent or adoptive parent. Even among mothers in their 20s, 10 percent or more are in households for which at least one child is “missing” (as defined here).

OPEN IN VIEWER

Figure 2

Proportion of Children Coresiding with Parent(s), by Age, 1970 Census

OPEN IN VIEWER

Figure 3

Proportion of Mothers (Children) with One or More Children (Siblings) Not in Household, 1970 Census

Together, these graphs illustrate a problem facing users of linked parent-child PSID records: the earlier the range of birth cohorts included among the children’s generation, the more likely it is that some family members will be missed. Even if a study limits its inclusion of the offspring generation to those born after 1968, there is at least a small chance that some older siblings are irrevocably lost from the sample. Fortunately, the PSID contains many data elements with which the existence, if not the detailed circumstances, of others in the sample members’ nuclear-family kin networks can be determined. Using these items, researchers could assess potential selection biases or external-validity limitations of their analytic samples.

The issues reviewed above should be viewed as factors that limit, but do not preclude, broad-coverage life-cycle studies of individual trajectories and of their association with the life-cycle circumstances of family members to which they can be linked. Care must be taken not only to select the ages and years for which both background factors and outcomes are to be measured, but to recognize, and convey to readers, the built-in limits on the generalizability of findings from these designs.

In addition to the issue of completeness of family-network coverage just described, it bears mentioning that step-relationships are not fully captured over time: given that one can acquire the PSID “gene” only through birth or adoption, children born to the spouse of a PSID gene carrier prior to that marriage, or after the dissolution of the marriage (and consequent split-off of the ex-spouse’s family), are not considered to be PSID sample members. Therefore, some step- and half-siblings of PSID sample members are excluded from the sample by design.

Survival bias in retrospective reports of early circumstances

The 2014 module of questions on childhood circumstances, the planned linkage to 1940 Census records, and other retrospectively reported characteristics (such as military service) are all examples of data whose distributions can be affected by a type of survival bias: if an earlier-life circumstance (for example, living in poverty during childhood) is correlated with survival to a later age, then those who remain alive to tell us about that earlier-life circumstance are not representative of the population that experienced that earlier-life circumstance. This problem is not always fully recognized, and even when it is recognized, it presents challenges.

To illustrate just one of the relevant variants of this problem in the PSID context, we can use a 1940 birth cohort life table. ¹² This birth cohort will be the youngest, among all PSID individuals, for whom a 1940 Census data linkage is possible. They became 72 (and therefore met the condition of the “age 72 rule”) in 2012. They will be 77 in 2017 and 79 in 2019, which are likely times at which they will be asked to provide any data items that would be used in the linkage effort.

The cohort life table (for all races and both sexes) tells us that already by 1968, the first year of PSID data collection, modest selectivity was present: about 8 percent of the birth cohort had died by 1968. Among survivors to 1968, only about 61 percent will survive to age 77, while 56 percent will survive to age 79. This degree of sample loss through mortality could pose problems for the users of the retrospectively reported information.

There are a number of potential solutions to this selective-retrospective reporting problem, such as the use of external data sources to model the selection process, joint modeling of the reporting and survivorship processes, and possibly others. In general, these approaches are technically complicated, narrow in their applicability, and not widely used. The availability of rich data sources such as the PSID, coupled with a growing recognition of the methodological issues suggested by these data structures, can be expected to help spur further development and applications of appropriate models and methods.