PERSPECTIVES: A Competing Risks Analysis of Older Americans' Poverty Entry and Exit Patterns in the Health and Retirement Study

Introduction

Several studies have used survey data to trace American adults' poverty entry and exit patterns over time using longitudinal, or panel, data sets (for example, Duncan 1984; Bane and Ellwood 1986).¹ Although such analyses are valuable for studying peoples' exposure to poverty over several periods, they are also subject to sample attrition. For research aiming to measure the likelihood of poverty entry and exit in later life, when respondents with low income and/or disability may be more likely than younger respondents to die or otherwise leave the panel, sample attrition could be problematic.

This article examines the effect of sample attrition because of death or for other (unknown) reasons, which we refer to as loss to follow-up (LTF), on the estimated variables associated with poverty entry and exit. We analyzed data reported in the University of Michigan's Health and Retirement Study (HRS), a nationally representative panel data set of Americans aged 50 or older, by 11,549 households who responded to the 2002 HRS biennial core survey. Those respondents were then invited to reinterview every 2 years through 2018.² Some respondents died or left the panel study because of LTF, and such attrition may be associated with factors affecting poverty transition status in later life. If people who remain in the panel are better off (if they have higher incomes, for example, or are healthier) than their attriting counterparts, that could lead to lower measured poverty rates among the survivors than for the population from which the initial sample was drawn.

Some analysts use proportional hazards models to evaluate the effects of various risk factors in panel surveys, yet those do not generally account for the possibility of nonrandom attrition over time. Accordingly, we examine both proportional hazards models and competing risks models that correct for nonrandom sample attrition. Our goal is to evaluate whether competing risks models produce different estimates of the variables associated with respondents' poverty transitions at older ages.³

In this article, we describe the methods we used to calculate older persons' poverty entry and exit rates. We then report our findings and discuss the results.

Methods

In this section, we describe the process used to collect HRS survey data, the reasons respondents may attrit from the survey, the statistical analysis we conducted to obtain the results, and the predictors and potential confounders we examined.

Data

The HRS survey protocol is detailed in HRS (2023b). HRS respondents are invited to be surveyed every 2 years either online or face-to-face, and all participants provide informed consent and receive small incentive payments. The data are housed at the University of Michigan's Survey Research Center and anonymized before analysts may access them.⁴ Our analysis uses the RAND HRS Longitudinal File, which is “a cleaned, easy-to-use, and streamlined data product containing information from Core and Exit Interviews of the HRS, with derived variables covering a large range of topics” (Bugliari and others 2024). From this file, we gathered information on 11,549 households surveyed in 2002, the first year for which the RAND file includes a poverty status indicator. We tracked which households entered and which ones exited poverty between consecutive waves, 2 years apart, in the period 2002–2018.

One advantage of using HRS data to examine poverty entry and exit patterns is the rich set of variables available in the database. The two main outcome variables for the older population we studied are (1) the hazard rate of entering poverty in wave t+1, having been nonpoor in wave t; and (2) the hazard rate of exiting poverty in wave t+1, having been poor in wave t. The RAND HRS file uses Census Bureau poverty thresholds and the family composition at the time of the interview to determine whether a household was in poverty in the previous year. Income comprises household labor earnings, pension benefits, Social Security income, disability program benefits, welfare benefits, withdrawals from deposit or individual retirement accounts, and income from self-employment, consulting, or any other source. Because the HRS is administered every 2 years, we can determine whether a household was poor or not at the time of each wave, but we cannot infer what the household's poverty status was between waves. Hence, the apparent poverty spells are actually consecutive reported occurrences across survey waves, which may or may not reflect continuous spells. This differs from many earlier studies on poverty, such as Card and Blank (2008), which used month-to-month Survey of Income and Program Participation data.

In addition to tracing respondents' poverty status over time, we also gathered 2002 HRS respondents' self-reported age, sex, race, ethnicity, education level, marital status, employment status, region of residence, number of marriages, and number of children younger than 18. We also included indicators for cohorts entering the sample at different times to assess whether cross-cohort patterns are statistically similar. We studied the Original HRS cohort (birth years 1931–1941), which was the first cohort surveyed, beginning in 1992; the Asset and Health Dynamics of the Oldest Old (AHEAD) cohort (birth years 1923 or earlier), first surveyed in 1993; and the Children of the Depression (CODA, birth years 1924–1930) and War Baby (birth years 1942–1947) cohorts, first surveyed in 1998.

Chart 1 compares 2002–2018 poverty rates for older Americans based on our HRS extract and the Current Population Survey (CPS) Annual Social and Economic Supplement. Poverty rates for older HRS and CPS respondents were generally consistent, ranging between about 7 percent and 16 percent. However, there were some differences. Among those aged 55–64, HRS respondents were slightly more likely to be poor (about 10–16 percent) than were CPS respondents (about 8–12 percent), although the HRS' relatively smaller sample sizes increase the standard error of its estimates and reduce the likelihood that estimated differences are statistically meaningful. In 2018, for persons aged 80 or older, the HRS poverty rate (9.5 percent) was lower than the CPS rate (13.2 percent); but again, the smaller HRS sample size suggests that this difference is not significant. Overall, we conclude that the two data sets provide relatively similar poverty status indicators by age in a given year. Yet, as we show in later sections, several other variables can influence poverty transitions.

Sample Attrition Definition

Respondents attrit from an HRS panel if they die or cannot be contacted by the survey team after substantial effort.⁵ The University of Michigan maintains an HRS Tracker file that records known deaths based on the dates reported by a respondent's spouse, partner, or other knowledgeable person; imputed dates based on respondents' last date known alive; or the dates on which HRS staff learned that respondents were deceased. HRS staff also periodically consult the Centers for Disease Control and Prevention's National Death Index (NDI) to confirm dates of death, if applicable, for those who were not interviewed.⁶ Table 1 shows the distribution of 2002–2018 HRS respondents by age, in each wave and overall. From 2002 to 2018, the share of respondents aged 80 or older increased from 19.4 percent to 46.2 percent. The median respondent age rose from about 60 in 2002 to about 70 in 2018 (not shown).

Statistical Analysis

We first offer tabular and graphic depictions of poverty entry and exit patterns over the study period. Next, we employ a multivariate hazards model to evaluate the influences of our predictors on the key outcomes. That approach assumes that attrition is independent of poverty exit or reentry (Schober and Vetter 2021). Then, we report results from competing risks regression models, which we used to control for effects of several independent variables on survival time without assuming that death, LTF, poverty entrance, and poverty exit are independent events. Our goal is to determine whether the variables associated with poverty hazards in later life differ when we use a model that assumes random attrition.

Predictors and Potential Confounders

The primary sociodemographic variables we examine as predictors of poverty entry and exit consist of age, sex, race, ethnicity, education level, marital status, employment status, region of residence, HRS cohort, marital history, and number of dependent children in household.⁷ Our first analysis measures those variables as of the 2002 HRS wave (which we call the “baseline”). In addition, because some of those control variables may vary over time, we provide an additional examination that considers the potential effects of such time-varying characteristics as changes in household headship, respondent and spouse health and employment status, marital status, and presence of dependent children.

Poverty Entry and Exit

Table 2 shows descriptive statistics for the HRS baseline respondents (those surveyed in 2002). The average age of all respondents was 69.8, similar to the average age of respondents in the subsamples that subsequently entered and exited poverty. The share of women among respondents entering poverty who were not poor at baseline was 54.1 percent, whereas the share of women among respondents exiting poverty was higher (73.4 percent). Among respondents entering poverty, 83.8 percent were White and 13.2 percent were Black, while 57.6 percent of those exiting poverty were White and 34.2 percent were Black. In addition, 6.5 percent of poverty entrants and 20.9 of those exiting poverty were Hispanic, while 93.5 percent of poverty entrants and 79.1 percent of those exiting poverty were non-Hispanic. Respondents with a high school education or higher constituted 75.9 percent of poverty entrants, but they constituted only 38.4 percent of the poverty exit subsample. Respondents working for pay accounted for 34.0 percent of the poverty entrant subsample, but they accounted for only 11.5 percent of those who exited poverty. Persons who were not working and reported a disability constituted 2.2 percent of poverty entrants, but they constituted 12.0 percent of those who exited poverty. Southerners made up 39.3 percent of those who entered poverty, but they accounted for 56.7 percent of those who exited poverty. The shares of each HRS cohort were similar among the entire sample and the two subsamples. The data suggest that respondents entering and exiting poverty were quite heterogeneous.

Table 3 documents poverty duration: how long, in terms of survey waves, HRS respondents' households remained in poverty. Almost 73 percent of the sample members were never in poverty in the period 2002–2018, and another 13.6 percent were poor in only a single wave. While the vast majority of the sample (86.5 percent) had no poverty experience or experienced poverty in only one wave, the remaining 13.5 percent of the sample experienced poverty in two or more waves, and 5.4 percent of the sample experienced poverty in four or more waves. Table 3 also shows the duration of respondents' longest poverty spells over the study period: 16.4 percent were poor for no longer than a single wave, yet 10.7 percent remained poor for two consecutive waves or more, a phenomenon referred to as “duration dependence” in the literature. The table shows heterogeneity in the chances that older Americans will enter and exit poverty, as well as duration dependence.

Modeling Poverty Entry and Exit Over Time

Chart 2 uses Kaplan-Meier survival curves to depict the time that passed until HRS respondents entered or exited poverty, conditional on remaining in the sample.⁸ Panel A shows the cumulative probability that a respondent who was not poor in one wave entered poverty in subsequent waves, and Panel B shows the cumulative probability that a respondent who was poor in one wave later exited poverty.⁹ A similar computation was conducted for each wave, and the cumulative distribution is shown by the red lines in Chart 2. Overall, the probability of entering poverty during 2002–2018 was far lower for older persons than their chances of exiting poverty once they had entered; this indicates the transitory nature of poverty for many older Americans.

Chart 2 also plots cumulative probabilities of poverty entry and exit, by wave, estimated using a competing risks model that allows for nonrandom attrition because of death and LTF. No other variables are included in these calculations. Interestingly, for poverty entry rates, the blue curve representing the competing risks model in Panel A lies below the Kaplan-Meier curve by 2 percentage points in 2004 and 12 percentage points in 2018. By contrast, poverty exit rates using the competing risks model were 17 percentage points lower than the Kaplan-Meier curve in 2004, but 32 percentage points lower in 2018. In sum, accounting for sample attrition in the longitudinal panel produces slightly lower poverty entry rates late in the study period, similarly lower poverty exit rates early in the period, and substantially lower poverty exit rates later in the period. This suggests that respondents who remain in the panel in the later part of the period are more likely to exit poverty than were those who attrited from the sample because of LTF or death, potentially indicating a change in the composition of the sample.

Multivariate Analysis Results

In contrast with simple survival curves, Table 4 presents multivariate model estimates that include baseline characteristics believed to be associated with poverty entry. Table 5 does the same for poverty exit. As we mentioned earlier, the variables include age, sex, race, ethnicity, education level, marital status, employment status, region of residence, HRS cohort, number of marriages, and number of dependent children in household. Both tables compare proportional hazards model results with results from a competing risks model. In addition, both tables show results for two separate specifications for each model: namely, with and without the effects of selected time-varying characteristics. The time-varying characteristics consist of the following occurrences between survey waves: a change in who is the head of household, poor health onset or an employment status change (own and spouse's), a marital status change, and the departure of dependent children. The tables show the adjusted hazard odds ratios for the proportional hazards model, the adjusted subhazard odds ratios for the competing risks model,¹⁰ and the standard deviations (95 percent confidence intervals) for both models for each variable; an estimated odds ratio greater than 1 indicates a greater likelihood of entering or exiting poverty, while an odds ratio less than 1 indicates a lower likelihood.

For the poverty entry determinants in Table 4, we found that few of the age effects in the hazards model are statistically significant at the 5 percent level or better; only persons aged 90 or older are significantly more likely than those aged 50–54 to enter poverty when the time-varying characteristics are included. Interestingly, once time-varying characteristics are controlled in our preferred competing risks model—which incorporates nonrandom attrition because of death and LTF—none of the age coefficients are significant at the 5 percent level or better. Accordingly, we see no evidence of differential age effects on the risk of poverty entry.

The likelihood of poverty entry by sex, race, and ethnicity are similar across the four models and specifications, with odds ratios of similar size and statistical significance. Men were 20 percent less likely than women to enter poverty, controlling for time-varying factors, whereas Black individuals were 2.5 times as likely as White individuals and Hispanic individuals were about 2.4 times as likely as Non-Hispanic individuals to enter poverty. The likelihood of poverty entrance increased with educational attainment, and again the odds ratios for each education level were similar in size and significance across the four models and specifications. Controlling for time-varying characteristics, married persons were 51–56 percent less likely to enter poverty than nonmarried individuals were. The presence of each additional dependent child was associated with a 14–16 percent higher risk of entering poverty across the models and specifications. Working for pay reduced peoples' poverty entry risk, with the estimated risks being slightly greater in the competing risks model specifications. Persons who reported having a disability were more than twice as likely as respondents who were unemployed, retired, or otherwise not in the labor force to enter poverty, with robust estimates in all four models and specifications. Residence in the South was consistently positively associated with poverty entry, with the risk estimated as 17–26 percent greater than that for residents of the West. Finally, there were few significant differences between HRS cohorts in the probability of entering poverty in later life, and no coefficient was significant at the 5 percent level in the competing risks framework.

In general, the hazard rate estimates prove to be comparable with those from the competing risks models. The key time-varying characteristics adding to the models' explanatory power include a change in the household head, divorce, worsening health of either the respondent or his or her spouse, a change in the respondent's employment status from unemployed to employed, and having children leave home. All of these factors are substantially and significantly correlated with a higher likelihood of poverty entry.

Table 5 shows poverty exit patterns. For all age groups older than 59, the coefficient estimates are statistically significant when time-varying characteristics are included in both the hazard and competing risks models. One key difference, however, is that the estimated magnitudes are often larger when attrition is not considered. For example, the risk of exiting poverty for an individual aged 85–89 was 2.63 times higher than that of an individual aged 50–54 in the hazards model, versus 1.83 times higher in the competing risks model. A person aged 90 or older would be 3.52 times more likely than someone aged 50–54 to exit poverty in the hazards model but only 1.80 times more likely in the competing risks framework. Therefore, the protective effect of older age is likely to be overstated in models that ignore endogenous attrition. Men were not differentially likely to exit poverty in later life, compared with women across both models.

Although the likelihood of poverty exit was similar by sex, other variables affected poverty exit patterns. For instance, the poverty exit odds ratios for Black and Hispanic individuals were about 27–33 percent lower than those of their respective reference groups (White and Non-Hispanic), and statistically significant, when controlling for time-varying characteristics. The likelihood of poverty exit declines as education level increases, with odds ratios of similar size and significance for each level across the four models and specifications. Among employment statuses, not working and reporting a disability was not a significant differentiator for exiting poverty at the 5 percent level or better. With time-varying characteristics controlled, people who worked for pay were about 1.9 times more likely to exit poverty than those who were unemployed, retired, or not in the labor force. In contrast with the poverty entry results, region of residence was not a statistically significant differentiator of poverty exit, and the results were similar across models and specifications. Finally, although Table 4 showed no significant difference between HRS cohorts, Table 5 shows that respondents from the three older cohorts (AHEAD, CODA, and HRS original) were less likely to exit poverty than the members of the War Baby cohort. The estimates with time-varying factor controls were similar in both models, as were those without the controls; but estimates under the two specifications differed from each other.

The estimated effects of the time-varying characteristics were not consistently and significantly different between the two models, including those of head of household changes, a marriage or divorce, children leaving home, or a spouse's employment transitions. When one's own health deteriorated, the likelihood of poverty exit declined; but somewhat surprisingly, the chances of poverty exit increased when a spouse's health deteriorated.

Discussion

Poverty rates among older Americans have declined by more than two-thirds in the last five decades (Li and Dalaker 2021), and the household poverty rate is lower among those headed by an individual aged 65 or older (9.0 percent) than among those with a head aged 18–64 (10.4 percent, Shrider and others 2021). Nevertheless, such static poverty measures reveal little about older persons' poverty exposure over time, and because most older Americans have stopped working, they may be increasingly vulnerable to poverty as they age.

This article evaluates older households' poverty entry and exit patterns using data from the HRS, a longitudinal survey of older Americans. Few researchers to date have explored poverty entry and exit patterns among the older population and the sociodemographic variables associated with those patterns,¹¹ although Larrimore, Mortensen, and Splinter (2020) used tax data to estimate household poverty transitions during a shorter period (2007–2018) than we examined. That analysis had the advantage of relying on administrative data to trace income changes, but the rich sociodemographic information available in the HRS permits us to examine additional variables, unavailable in tax records, that can illuminate movements into and out of poverty.

We have focused on whether and how variables associated with poverty entry and exit are influenced by nonrandom attrition over time, and whether modeling nonrandom attrition alters our interpretation of the factors associated with poverty transitions. We show that poverty entry models exhibit relatively less sample-attrition bias than poverty exit models do, and in many cases, the variables associated with poverty entry among older Americans are similar in proportional hazards and competing risks models. For example, differences in race and ethnicity effects are not statistically significant across models. Nevertheless, competing risks models indicate smaller effects of aging on poverty entry risk. After handling nonrandom attrition because of death and LTF and controlling for the time-varying characteristics, there are only moderate differential age effects on the risk of poverty entry, whereas the simpler hazards model estimates more significant age effects. Overall, though, we conclude that attrition bias is not highly problematic for analysts focusing on poverty entry.

There are a few more differences across models for the poverty exit data, particularly regarding the age effects. Specifically, older persons' chances of leaving poverty appear to be overstated in models that ignore endogenous attrition. When looking at race and ethnicity differences, Black people are 5 percentage points more likely than White individuals and Hispanic people are 6 percentage points more likely than non-Hispanic individuals to exit poverty in the competing risks model, compared with the simpler hazards model's estimates. Additionally, poverty exit was more likely for respondents who remained longer in the HRS panel than were those who attrited from the sample because of LTF or death. Such a finding implies a change in the representativeness of the sample over time.

Analysis of poverty transitions in the older population is becoming increasingly important because several government programs look at peoples' past resources when determining their eligibility for benefits. For instance, applicants for Medicaid nursing home and home care benefits must not only have little current income and assets; they also must have had limited financial resources over a recent period (usually the last 5 years). Medicare premiums for prescription drug and outpatient services coverage are likewise conditioned on participants' income in the last 2 years. The Temporary Assistance for Needy Families (TANF) program can provide food assistance to low-income grandparents who care for young children. However, TANF benefits are time-limited, and eligibility requirements do not account for the possibility that the aged may move in and out of poverty. For these reasons, our research confirms the importance of understanding attrition biases when examining which older Americans are particularly vulnerable to poverty transitions.

We also acknowledge caveats regarding our findings. The HRS income, wealth, and independent variables are self-reported and are subject to reporting error, potentially biasing estimates of the variables associated with poverty transitions (Bound, Brown, and Mathiowetz 2001).¹² We leave to future study an examination of that possibility. Several prior analyses (Meijer and Karoly 2017; Meijer, Karoly, and Michaud 2010; Sierminska, Michaud, and Rohwedder 2008) have compared income and wealth self-reports for HRS subsamples with other nationally representative surveys or administrative data from SSA, and they have generally concluded that HRS income and wealth measures suffer less from measurement error than measures from the other sources.¹³ Finally, we have focused on money income to define poverty, consistent with the Census Bureau's official poverty measure and a wide range of other poverty analyses. For this reason, future analysts could include in-kind benefits, such as health care, rent subsidies, and food stamps, in a broader analysis of financial vulnerability.¹⁴ Nevertheless, the official Census Bureau measure remains the most consistent measure of poverty used in the United States for the last half-century.

Our work has relied on data from the longest available survey panel to provide insights into the older population's poverty entry and exit patterns, and we have documented substantial heterogeneity, particularly in models allowing for competing risks of sample attrition. Policymakers concerned with programs designed to alleviate retirement insecurity may wish to consider a more dynamic perspective on financial insecurity at older ages and to take one that acknowledges the importance of attrition bias in the older population.