Cognitive Disparities: The Impact of the Great Depression and Cumulative Inequality on Later-Life Cognitive Function

Critical period models (CPM) (cf. Kuh et al. 2003) take as their object of study how hardship is written into one’s physiology. There may be critical periods of development during which exposure to, for example, malnutrition or high stress causes irremediable damage to physiological systems (Hayward and Gorman 2004; Kajantie 2006; Lupien et al. 2009; McEwen and Gianaros 2011; Seeman et al. 2010). The effects of hardship are likely related to age, length, and extent of exposure (Manton et al. 1997).

As a type of CPM related to the prenatal period, Barker advanced through a series of studies (cf. 1990 and 1996) that fetal environment plays a role in the development of chronic diseases, independent of later-life risk factors—the fetal origins hypothesis. This prenatal period is understood as critical because it is a time of rapid growth and epigenetic plasticity (Barker 1990; Liu et al. 2010; Torche and Villarreal 2014). Prenatal exposure to a negative environment may actually cause the body to shift its phenotype (genetic expression) and/or cause functional or structural brain changes (McEwen 2009; Scholte et al. 2012).

Defining these critical periods is complex. Age of greatest neurological vulnerability varies by risk: for example, malnourishment, specific vitamin deficiencies, or stress (Benton 2010). More damage may result if adversity interrupts growth trajectories after they have begun than if there is hardship throughout the period. For example, evidence suggests if the first two trimesters in utero are characterized by higher nutritional inputs, which are then followed by deficiencies in the third trimester, the fetus may be at greater risk than if there were deficiencies through the entire fetal period (McEniry and Palloni 2010). Indeed, the third trimester to age 2 may be the most sensitive period for neurological development because of the pace of brain development at these ages (Benton 2010; Casey et al. 2000; Moceri et al. 2000). Between ages 3 to 6, cerebral volume reaches its maximum (Casey et al. 2000). This is a peak period of development in the frontal networks, critical to attention and planning (Thompson et al. 2000). Ages 7–10 appear to be critical for the development of the anterior corpus callosum, which is the region of the brain that declines the fastest in early stages of Alzheimer’s (Thompson et al. 2000). Reduced brain volume and development are already detectable in adversity-exposed adolescents (Lupien et al. 2009).

However, confounding factors, such as prolonged poverty, have made CPMs challenging to evaluate (George 2005). Researchers have managed selectivity in the context of CPMs by using natural experiments. Part of the underlying logic of natural experiments is that the age at which a cohort experiences social changes is important to its level of sensitivity (Yang 2008). Researchers have established associations between morbidity and mortality and age during the Spanish flu epidemic, the Dutch famine, the German food crisis, the Chernobyl disaster, the Food Stamp Program rollout, and exposure to natural disasters (Almond 2006; Almond and Currie 2011; Almond et al. 2009, 2011; Conley and Bennett 2000; Jürges 2013; Lindeboom et al. 2010; McEniry and Palloni 2010; Torche 2011). Specific to cognition, macroeconomic conditions at birth appear to be associated with later-life cognitive function for European elders born in the first half of the twentieth century (Doblhammer et al. 2013; Fritze et al. 2014).

The Great Depression is useful as an indicator of early-life conditions because of the breadth and the depth of the hardship experienced. Unemployment reached 27 % and even higher in some areas (Garcia 2014). With almost no federal assistance programs, approximately 60 % of the U.S. population met the federal guidelines for poverty, and one-fifth of all children in New York State and as high as 90 % of children in mining country were malnourished (Himmelgreen and Romero-Daza 2010). Because the Depression increased risk of malnutrition and allostatic load (physiological dysfunction related to stress) (cf. Geronimus et al. 2015; Juster et al. 2010; Seeman et al. 2010), it offers analytic leverage on whether early-life hardship is associated with later-life health. Although Cutler et al. (2007) found no significant long-term health effects of exposure to the Depression, they did not study cognitive function (further discussion in Online Resource 1).

Based on this body of research and using the Depression as an exogenous shock, I formulate a CPM hypothesis:

For example, the cohort born during the Depression will have the lowest cognitive function scores compared with earlier cohorts, and the cohort aged 0–2 will have the next lowest scores (Benton 2010; Casey et al. 2000). The earliest cohort (born in 1909–1918)—that is, the most neurocognitively developed at the onset of the Depression—will have the highest cognitive function scores. Evidence for the CPM hypothesis would be finding that intervening life-course factors do not significantly attenuate the association between early-life exposure to the Depression and cognitive function.

Nevertheless, deterministic arguments are seldom accurate in the social world. For instance, Elder (1998) found that by midlife, many Depression-era children were in good physical health and had overcome significant disadvantage. However, he also found that early adversity triggered a range of social transitions, including earlier marriages and antisocial behavior (Elder 1998). This is representative of a social trajectory model (STM) because he argued that the mechanism is sociobehavioral; Depression-era children possibly carry scars from that period that are not strictly physiological (Pol and Thomas 2002). In other words, it is not (necessarily²) that malnutrition or stress cause long-lasting physiological damage but rather that early hardship causes people to engage in riskier behaviors.

Do early-life experiences spur social trajectories that increase risk of cognitive dysfunction? The importance of distinguishing between physiological damage (CPM) and behavioral pathways (STM) relates to policy interventions. If cognitive function is associated with early-life hardship only through smoking behavior, for example, then investing in smoking-cessation programs may be the key intervention.

Evidence for the STM hypothesis would be finding that prenatal or early-childhood exposure to the Depression is associated with riskier behaviors (identified as risky within the CPM).

Although critical period models focus on measuring the independent effect of an exposure and the social trajectory model emphasizes behaviors, life-course theory requires us to consider “lives in motion,” (Elder 1998:7) examining entire life trajectories (Ben-Shlomo and Kuh 2002; Hayward and Gorman 2004). Most early-life conditions do not have direct, unalterable impacts on adulthood (Breteler 2001; Claussen et al. 2003; Mayer 2009).

Situated within life-course theory, cumulative inequality (Ferraro and Shippee 2009; Schafer et al. 2011, 2013) and cumulative disadvantage (DiPrete and Eirich 2006; Shuey and Willson 2008) theories posit that capital, including health, accumulates through the mechanisms of “structure, human agency, and chance” (Willson et al. 2007:1887). Regardless of the source of advantage, the advantaged have large initial gains that accumulate over the lifetime, resulting in positive outcomes (Elder 1998; O’Rand and Hamil-Luker 2005; Pearlin et al. 2007; Shuey and Willson 2008). The converse is also true: those who are disadvantaged become increasingly disadvantaged.

However, social factors are not solely detrimental. Resilience is important to cognitive abilities, just as with general health (Schafer et al. 2013). Among autopsied patients, 10 % to 40 % exhibit Alzheimer’s-like brain pathology but appear asymptomatic (Meng and D’Arcy 2012). Derived from medical reserve and neuroplasticity theories, cognitive reserve (CR) is the latent construct that represents this gap between brain pathology and cognitive dysfunction (Meng and D’Arcy 2012; Stern 2012). Researchers have operationalized CR using several social factors but have relied heavily on formal education as a proxy. Education is considered not as it relates to SES or poverty’s toll on health but rather in its direct effect on the brain. A meta-analysis of CR research shows a strong association between education and Alzheimer’s diagnosis (OR = 2.62; p < .001) (Meng and D’Arcy 2012; Siegel 2012). Higher occupational attainment may also be associated with lower Alzheimer’s risk through CR (Stern 2012). Reed et al. (2011) found that later-life leisure cognitive activities were the most predictive of CR, net of education and occupation. Insights from this research, however, are limited because CR studies often do not control for confounding variables.

These factors that may build CR should be considered in cumulative inequality models (CIM) along with four disadvantage-related elements: onset, duration, magnitude of exposure, and interpretation or perception of risk (Ferraro and Shippee 2009; Schafer et al. 2011). People who experience repeated and/or severe hardship at vulnerable ages are at higher risk than those who experience a shorter period or lesser degree of harm, and their perception of that harm³ influences the extent of damage (Case and Paxson 2010; Torche 2011; Willson et al. 2007).

This research motivates the CIM hypothesis:

Whereas CPMs focus narrowly on irremediable harm caused during a critical moment—to the exclusion of other periods—CIMs do not preclude resilience and imply that the link between early life and health will be affected by intervening factors. Evidence for the CIM hypothesis would be if those who have low SES in early, mid-, and later life have worse cognitive function than those who have a shorter period of disadvantage.

Data are from 11 waves of the HRS, a longitudinal, population-based, nationally representative biennial survey (n ≈ 30,000) covering a span of 20 years (1992–2012). The HRS is conducted by the University of Michigan (National Institute on Aging U01AG009740) (Health and Retirement Study 2014; RAND HRS Data 2014). In 1992, the HRS sampled two cohorts (born in 1890–1923 or 1931–1941) of noninstitutionalized individuals and their spouses (regardless of spouses’ ages). The HRS continues to add cohorts to replenish the sample and to better represent the contemporary middle-aged and older population. The HRS uses both face-to-face and telephone interviews to survey early-life SES, occupation, later-life wealth, lifestyle, health, and cognitive function. I incorporate historical data from the U.S. Census Bureau, including regional birth rates, infant mortality, geographic mobility, and state-level unemployment rates. Restricting the data set to respondents born between 1909 and 1934 (n = 14,481), who participated in or after 1996 (when the cognition measures included in this study became consistent) (n = 13,667) and who had at least one cognitive function score, results in a final estimation sample of 12,278.

The HRS adapted the Telephone Interview for Cognitive Status (TICS) to evaluate cognitive deficits and dementia-related declines (Ofstedal et al. 2002). The TICS has near-normal distribution and high test-retest reliability, is relatively insulated against ceiling effects, and effectively predicts cognitive impairment (Crimmins et al. 2011; Fong et al. 2009; Karlamangla et al. 2009). The University of Michigan Survey Research Center used a multivariate, regression-based procedure to impute missing, refusal, or not-applicable responses to cognition questions for nonproxied respondents who completed an interview that wave (for details, see Fisher et al. 2017).

From TICS, I distinguish fluid from crystallized intelligence because they are distinct aspects of cognition and show different life-course trajectories (Harvey et al. 2003; Horn 1982; van den Berg et al. 2010). Fluid intelligence is based in problem-solving ability, develops early, follows a normative decline, and is less correlated with education (Ghisletta et al. 2012). Crystallized intelligence is related to educational and cultural factors (e.g., vocabulary), develops later in life, and is more stable (Ghisletta and Lindenberger 2004; Ghisletta et al. 2012). This analysis focuses on fluid cognition because it is more reflective of neurophysiological health (Buckner 2004; Ghisletta et al. 2012).

Using TICS in Waves 3–11⁴ (1996–2012), I generate separate standardized scores for three items: (1) total word recall, (2) counting backward from 20, and (3) serial sevens (Cook et al. 2009). These measures are administered upon survey entry and then at every wave thereafter (biennially) unless the respondent requires a proxy. Total word recall is the sum of immediate and delayed word recall, wherein respondents recall immediately and after five minutes 10 words from one of four lists (RAND HRS Data 2014). Counting backward is denoted as 2 = correctly counts backward for 10 continuous numbers on the first try, 1 = successful on the second try, 0 = unsuccessful. Serial sevens is scored as 0–5, equaling the number of times the respondent correctly subtracts seven starting from 100 (e.g., 5 = five correct subtractions). Principal components analysis validates combining the three standardized scores into a single variable. I average the three standardized scores to generate fluid cognition.

To estimate Fluid Cognition testing the CPM and CIM hypotheses, I use growth curve models, which produce trajectories of cognitive change as a function of the predictors (Leahey and Guo 2001; Raudenbush and Bryk 2002). Individuals’ scores contribute to estimates even if there are few observations; thus, these models are more robust to rolling enrollment and attrition (Curran et al. 2010). Approximately 75 % of the sample has three or more observations of fluid cognition. These growth curve models are two-level hierarchical models with individuals at Level 2, individuals’ multiple fluid cognition scores at Level 1, and a random coefficient for Cohort:

Fluid cognition (Y _it ) is a function of the intercept α (the grand mean of all respondents’ cognitive tests), Age and Quadratic Age slopes (β₁ and β₂), Cohort (β₃), sets of time-invariant (w) predictors for person i and time-variant (x) predictors for person i at age t, the person-specific error term (μ), and the overall model residual (ε). The intercept parameter α can be interpreted as the mean standardized fluid cognition score at age 60.

To test the CPM hypothesis, I use a staged life-course approach to estimate nested models. For the STM hypothesis, I estimate multilevel binary and multinomial logistic and linear regression models to examine whether exposure to the Depression is associated with behavioral risk factors. To test the CIM hypothesis, I include interactions of life-course factors. I also predict marginal values of fluid cognition, displaying results graphically to aid in interpretation. I use clustered standard errors within persons in Stata 14. All results from analyses not displayed are available upon request.

I consider the CPM to investigate whether prenatal and/or early-childhood exposure to macroeconomic shock has an independent association with fluid cognition, regardless of other life-course factors. Across all four nested models, compared with the reference cohort (over age 10 at the onset of the Depression), I find that cohorts born during or closer to the Depression have statistically significantly lower fluid cognition (Table 2). Later life-course factors do not significantly attenuate the exposure coefficient—evidence for the CPM hypothesis.

In the full model, those hypothesized to be most vulnerable—that is, the Great Depression cohort—have the worst fluid cognition compared with the 1909–1918 cohort (b = –0.292, p < .001), net of all other factors. The fewer years of cognitive development a cohort had prior to the stock market crash, the worse their predicted fluid cognition compared with the reference cohort (aged 0–2, 3–6, and 7–10, respectively; b = –0.231, b = –0.202, b = –0.119, p < .001). The cohort coefficient magnitudes are comparable to or greater than those for education, which has been the most significant social predictor found to date (cf. Meng and D’Arcy 2012).

To further contextualize the cohorts’ cognitive disparities, at age 80, the Great Depression cohort’s fluid cognition is almost one-half standard deviation below the mean of 0, whereas the earliest cohort’s predicted score remains slightly below 0, net of other factors (see Fig. 1). By comparison, elders who have already been diagnosed with a memory disease have a mean fluid cognition of –0.9, only slightly lower than the Great Depression cohort’s fluid cognition at age 80.

Interactions between cohort and the age terms are nonsignificant. However, because they started with lower fluid cognition, the Great Depression cohort is predicted to reach any given cognitive impairment threshold 5 to 10 years prior to the earliest cohort (Fig. 1). The Great Depression cohort’s fluid cognition at age 75 is worse than the 1909–1918 cohort’s fluid cognition at age 82.

After education is controlled, sex is not significant. Although the associations between race/ethnicity and fluid cognition in the full model are attenuated by approximately one-half, the fluid cognition for people of color remains significantly lower than that of whites’ (black b = –0.379, Latinx b = –0.194, other b = –0.209, p < .001). The Latinx-white cognitive disparity is the most attenuated by other life-course factors (60 %). Childhood rurality is negatively associated with fluid cognition. Better early-life health, higher parents’ and respondents’ education, higher occupational attainment, greater later-life wealth, and socializing more frequently with neighbors are associated with higher fluid cognition. Widow(er)hood and separated/divorced/absent spouse are associated with higher cognitive function than partnered/married.¹⁴ Exercising more frequently, light alcohol use (reference = abstinent/rare), higher average BMI (reference = normal), lower emotional depression, and fewer comorbidities are associated with higher fluid cognition.

As noted, several behavioral factors are predictive of later-life cognitive function (Table 2). Is it possible that early-life harm increases the probability of engaging in “risky behaviors” over the life course, which subsequently increases the risk of later-life cognitive dysfunction? Here I examine whether behavioral factors¹⁵ are associated with exposure to macroeconomic shock, early-life poverty, and their interaction, controlling for age, sex, and race/ethnicity.

Contrary to expectation, the interactions between cohort and early-life SES are primarily nonsignificant (results available upon request). For example, respondents from the Great Depression cohort who had lower SES in childhood are no more likely to have lower occupational attainment than those who had lower SES in the 1909–1918 cohort.

I highlight several of the STM outcomes to demonstrate cohort differences in risky behaviors (see Table 3 and Fig. 2). Net of controls, those in the Great Depression cohort are significantly more likely than the three earliest cohorts (born before 1927) to have a high school diploma versus less than high school. Compared with the earliest cohort (born in 1909–1918), the Great Depression cohort was significantly more likely to attend college (vs. high school), to be wealthier, to work in professional/office occupations (vs. service/manual), to be light drinkers (vs. abstinent/rare), and to exercise more than once per week (vs. once/week or less). The CPM suggests that these factors are all associated with higher fluid cognition. However, the Great Depression cohort is significantly more likely than the earliest and neighboring cohort to consume three or more alcoholic drinks/day, which is associated with lower fluid cognition. There is not a clear, significant pattern by partnership status or socializing that would indicate the Great Depression cohort is at either higher or lower risk. Differences between the Great Depression cohort and the neighboring cohorts are primarily nonsignificant.

In sum, there is little evidence that the cohorts exposed to the Depression at earlier ages engage in riskier behaviors, net of controls. The STM hypothesis is not supported. The mechanism connecting early-life exposure to the Great Depression with fluid cognition does not appear to be behavioral.

To test whether life-course factors interact to affect fluid cognition and rate of cognitive decline, I estimate models interacting age, cohort, and early-life SES with other factors. In regard to the age interactions, few factors significantly predict rate of decline, so I do not display these results. That said, it is notable that blacks have a faster linear rate of fluid cognition decline than whites (race/ethnicity main effect: b = –0.27, interaction with age: b = –0.01, p < .001). Aligned with prior research (Early et al. 2013; Wilson et al. 2005), education and rate of decline are only weakly associated.

Similarly, little insight is gained from the interaction of cohort and life-course factors, although lower education does predict significantly worse fluid cognition for the two most recent cohorts compared with the 1909–1918 cohort (results available upon request). Contrary to my hypothesis, the interaction between cohort and early-life SES is nonsignificant. For instance, respondents in the Great Depression cohort who were from lower-SES families do not have multiplicatively worse fluid cognition, net of later-life factors (Table 4).

In contrast, later-life factors do affect the association between early-life SES and fluid cognition (Table 4). Fluid cognition is multiplicatively lower if an individual has lower early-life SES and lower occupational attainment or lower later-life wealth (Fig. 3). If an individual was in the lowest quartile of early-life SES, being in debt in later life predicts fluid cognition approximately three-quarters of a standard deviation lower than had s/he started in the highest quartile of early-life SES (fluid cognition = –0.74 vs. –0.14). The cognitive disparity by occupational attainment and wealth is narrower for those with higher early-life SES, net of other factors (Fig. 3). These results suggest that (1) higher early-life SES is protective against later-life risk factors for cognitive decline, and/or (2) midlife SES mitigates the negative effects of lower early-life SES.

I next predict fluid cognition for those who had a lifetime of low, mixed, or high SES, using estimates from the full CPM. Based on Table 2, the low-SES factors are Great Depression cohort, low parental education, poor childhood health, less than high school, in debt, and service occupations. The high-SES or mitigating factors are the most beneficial alternative category. I divide life into “early” and “late” after respondent’s education; control for age terms, proxy, death, partnership (married), and childhood region (Mid-Atlantic); and hold all other variables except race/ethnicity and sex at their means. Figure 4 displays predicted fluid cognition for blacks, Latinx, and white men and women at age 75.

Low life-course SES results in the lowest fluid cognition—more than 1 standard deviation below the higher-SES predictions (Fig. 4). We can interpret Fig. 4 to indicate that low early-life SES is worse for fluid cognition than low later-life SES. Notably, blacks and Latinx have significantly worse fluid cognition than whites across all groups. At age 75, blacks who have low SES throughout their lives have a predicted fluid cognition over 1 standard deviation below the mean, comparable with those who already have a memory disease diagnosis (–0.9). Although racial/ethnic differences are attenuated in the high life-course SES scenario, disparities remain.

I conduct multiple sensitivity analyses.¹⁶ The selection and attrition that are challenges for most longitudinal surveys are here exacerbated by the HRS’s rolling enrollment and aged sample. To account for selection into the sample, I restrict analyses to ages 76–83 or 72–83. Results are consistent (Online Resource 1, Fig. S1). In all models, I address the issue of selective mortality biasing cohort estimates by including indicators for whether respondents ever needed a proxy or died in 1992–2012.

As another robustness check, I vary cohort construction. I define the “most-exposed” cohort as under age 2 at the stock market crash or born during the Depression (born in 1927–1934; b = –0.26, p < .001), or conceived after the stock market crash (born in July 1930–1934; b = –0.29, p < .001). The cohort pattern and estimates for the other predictors are consistent. I also estimate models using single birth-year cohorts. Comparing neighboring birth-year cohorts minimizes problems with selective mortality and approaches a more fine-grained analysis of exposure. There is a significant dip in fluid cognition for those who were born in the year following the stock market crash relative to those born the previous or following year. Compared with the reference group born in 1918, birth year coefficients are b = –0.245 for 1929, b = –0.290 for 1930, and b = –0.264 for 1931 (p < .001). As additional sensitivity analyses for mortality selection, I conduct simulations (Online Resource 1, Figs. S2 and S3).

If cognitively impaired elders are more likely to misreport their early-life SES negatively, low early-life SES would inaccurately appear predictive of lower fluid cognition. This direction of misreporting may also bias the cohort estimates toward the null because earlier cohorts enter at older ages and average worse fluid cognition at first TICS. At their first eligible TICS test,¹⁷ 1,439 (11.7 %) and 517 (4.2 %) respondents (of n = 12,278) fall 1 or 2 standard deviations, respectively, below the fluid cognition mean. As a robustness check, I reestimate all the models but exclude individuals whose first test suggests impairment.¹⁸ Although the cohort and sex coefficients marginally increase, there is a modest decline in coefficients for some of the other factors, including race/ethnicity, family SES, and occupation. Nevertheless, estimates are consistent, and sex is the only variable that changes in significance (Online Resource 1, Tables S1 and S2).

Cardiovascular diseases are associated with early-life hardship (cf. Scholte et al. 2012); therefore, I test the subhypotheses that exposure to the Depression increases BMI, as well as risk of stroke, high blood pressure, heart disease, and diabetes. I find mixed results for these models, but overall cardiovascular health does not appear to be the mechanism connecting early-life exposure with fluid cognition (Online Resource 1, Table S3).

Finally, I use historical Census Bureau data to examine early-life contextual factors. I include childhood region’s average unemployment rate during the early 1930s to account for regional variation in severity of the Depression. Unemployment ranged from 20 % to 28 %. Fluid cognition is statistically significantly associated with, but not patterned by, regional unemployment. This suggests the variable is capturing other regional dissimilarities. I model fluid cognition as a function of total birth rate, infant mortality rate, and child mortality rate to assess the possibility of a culling effect (Myrskylä 2010). I find no indication that cohort estimates are downwardly biased (i.e., more conservative) because of selection in early-life birth and death rates (Online Resource 1, Fig. S4).

This study has limitations. Because TICS is a brief telephone-administered cognitive assessment, it is not as effective at evaluating cognitive function as clinically administered tests. Although TICS has been validated as a good indicator for pathological cognitive decline with high sensitivity and specificity (Crimmins et al. 2011; Crooks et al. 2005; Fong et al. 2009), this study cannot make strong claims that Alzheimer’s disease or other pathological cognitive decline caused low fluid cognition scores.¹⁹

An additional measurement limitation is that there are plausibly practice effects for the TICS. Respondents may perform better over time because they are familiar with the questions. This would downwardly bias estimates, making cognitive decline seem less steep. Vivot et al. (2016) found that associations between risk factors and cognitive decline are not significantly affected by practice effects.

Early-life factors are self-reported retrospectively, opening them to recall bias. However, respondents answer early-life questions at their first interview, when they are least likely to be suffering from cognitive impairment or comorbidities. Also, Alzheimer’s disease and other age-related cognitive impairment affect short-term memory and problem-solving skills much earlier than long-term memory (Buckner 2004). A respondent will thus be unable to answer the current questions well before their early-childhood memories are compromised (Buckner 2004; Greene and Hodges 1996), in which case the interviewer requests a proxy, and the respondent does not complete the TICS, excluding them from these analyses.

Nevertheless, I use two strategies to address possible recall bias. First, instead of operationalizing early-life SES with a single question (e.g., parents’ education), I include multiple domains of early life. This should be less sensitive to misreporting. The CPM encompasses six separate measures of early-life SES (see Table 2 and Fig. 4), and the CIM uses a composite measure of early-life SES (see Table 4 and Fig. 3). Second, in robustness checks, I exclude elders who are either 1 or 2 standard deviations below the fluid cognition mean at first TICS test. Although there is modest fluctuation in some coefficients (see the sensitivity analysis in Online Resource 1, and Tables S1 and S2), results are consistent. Measurement error in retrospective reports cannot be eliminated. In fact, missingness on covariates is associated with SES (b = –0.12, p < .001), so estimates for SES could be downwardly biased.

To manage selection into the sample and selectivity, I include indicators for proxy and mortality attrition, restrict analysis to those born 1909 to 1934, examine cohorts within the same age range, and use counterfactual simulations. Even so, bias may remain. If healthier individuals make up a greater share of earlier cohorts because they have survived to older ages, cohort estimates will be upwardly biased. On the other hand, if those most negatively affected by the Depression were selected out of the study by premature mortality, the bias would be toward the null, and estimates linking fluid cognition with exposure to the Depression or poverty would be conservative.