Early-Life State-of-Residence Characteristics and Later Life Hypertension, Diabetes, and Ischemic Heart Disease

Abstract

Objectives. We examined how state characteristics in early life are associated with individual chronic disease later in life.

Methods. We assessed early-life state of residence using the first 3 digits of social security numbers from blue- and white-collar workers from a US manufacturing company. Longitudinal data were available from 1997 to 2012, with 305 936 person-years of observation. Disease was assessed using medical claims. We modeled associations using pooled logistic regression with inverse probability of censoring weights.

Results. We found small but statistically significant associations between early-state-of-residence characteristics and later life hypertension, diabetes, and ischemic heart disease. The most consistent associations were with income inequality, percentage non-White, and education. These associations were similar after statistically controlling for individual socioeconomic and demographic characteristics and current state characteristics.

Conclusions. Characteristics of the state in which an individual lives early in life are associated with prevalence of chronic disease later in life, with a strength of association equivalent to genetic associations found for these same health outcomes.

Environments, whether at the level of country, state, county, or neighborhood, account for a meaningful amount of variance in chronic disease.^1–8 Although earlier work was entirely ecological in nature, thus making inferences to individual health outcomes problematic, more recent literature has avoided ecological bias by measuring individual-level outcomes and covariates.^4,9,10 In the United States, state is a level of geography that has varied social characteristics and policy:

It is one of the happy incidents of the federal system that a single courageous state may, if its citizens choose, serve as a laboratory; and try novel social and economic experiments without risk to the rest of the country.^11(p311)

The state is also methodologically advantageous as a level of analysis because there is less likely to be severe self-selection bias compared with smaller regions within metropolitan areas,¹² such as neighborhoods, in which wealth and race influence place of residence.¹⁰ Most typically, because of data constraints, only current context is examined, which is inconsistent with life-course theories of the etiology of chronic disease.^13–17 It is possible that the associations found with current measures of the social and economic environment are at least in part the result of impacts of earlier environments that are typically correlated with current context because many people remain geographically stable.¹⁸

Despite social and policy differences in the state context into which people are born and live, little is known about how characteristics may be associated with chronic disease. The majority of recent work has been at smaller levels of geography, with hypotheses focused on local physical and social environments as causal agents.^19–23 In contrast with this, at the state level there are broader macrocharacteristics that are affected by state-specific policies. Which specific characteristics of the state environment may be most correlated with prevalent health later in life is unknown. Also unknown is whether current state characteristics capture all of the same health outcome variance from early-life context and the extent to which individual-level factors may explain any early state context-associated differences.

We attempted to establish the relevant time period for measuring exposures in the state context and to determine the importance of the state context in explaining later-life patterns of chronic disease. Specifically, we examined how early-life state characteristics are associated with prevalent hypertension, type 2 diabetes, and ischemic heart disease. We were able to construct these environments through data linkage of individuals in an occupational cohort using the first 3 digits of social security numbers to determine state of early-life residence.²⁴ Thus, in our analysis, early life is operationalized as the time at which a social security card was issued. We examined the associations between early-life state characteristics and prevalent adult health outcomes. We also fit subsequent models to control for individual demographic data, census region, and current state characteristics. We also examined whether models including individual-level risk factors may in part explain the associations between early-life state context and chronic health conditions.

METHODS

The study population consisted of Alcoa employees from US manufacturing plants, with a total analytic sample size of 40 804 with 305 936 person-years of observation. Data were collected on employees beginning in 1996 through their date of retirement. Individuals were censored on December 31, 2012. Eligibility was restricted to individuals who were employed for at least 2 calendar years. The data are from 62 manufacturing plants, with at least 100 individuals per plant. We used the first 3 digits of employee social security numbers to determine the state in which their social security card was issued.²⁵ Less than 1% of our sample could not be linked to state of card issue.

Contextual Variables

We measured state context at 2 time points, early life and the baseline of observed cohort time between 1996 and 2012. The year of early life varied depending on age; we defined it by the year in which the individual was aged 15 years, which is an estimate of the average age at which individuals in this cohort began working and thus approximated when their social security card was issued. We chose the following state characteristics on the basis of previous findings in the literature as important indicators of social conditions that have also been consistently measured in the US census since 1940: percentage rural, percentage with less than a high school education, median income in 2000 dollars, unemployment rate, percentage non-White, and Gini coefficient of income inequality.

To facilitate comparison between factors and interpretation of odds ratios, we examined each contextual measure in the direction hypothesized to be associated with lower resources in the state, thus we reverse coded median income so that odds ratios refer to a 1 standard deviation lower median income. We obtained contextual measures from publicly available US census data from the 1940 to 2010 censuses. The distributions of these factors are shown in Figure A (available as a supplement to the online version of this article at http://www.ajph.org). For all analyses, state characteristics were Z scored (mean centered and divided by their standard deviation) to allow an easier comparison between different state characteristics. We calculated these Z scores separately for each early-life and current state measure. Figure B (available as a supplement to the online version of this article at http://www.ajph.org) presents a Pearson correlation plot showing how strongly measures are correlated, with more intense colors and larger dots indicating a stronger correlation. We also present these correlation plots separately for individuals (73%) staying in the same state (Figure C, available as a supplement to the online version of this article at http://www.ajph.org) and for individuals (27%) living in a different state later in life (Figure D, available as a supplement to the online version of this article at http://www.ajph.org).

Individual Variables

We used administrative records to obtain individual-level covariates and outcome data. These data include age, race, gender, whether the individual was an hourly or salaried worker, and a continuous measure of employment grade ranging from 1 to 76.²⁶ We also used an indicator of whether the plant was a smelter (10 out of 62) because these plants were not evenly distributed across states, and physical and chemical exposures at these plants might result in some worse health outcomes for workers in these plants or, conversely, select for positive health characteristics.²⁷

We defined prevalence of health outcomes as having at least 2 medical diagnoses for each of the respective health outcomes (hypertension, type 2 diabetes mellitus, and ischemic heart disease) over the time each individual was observed. The requirement of 2 claims increases the specificity of identifying clinical outcomes, with little decrease in sensitivity. Previous work has determined the accuracy of medical claims data for identifying health outcomes from administrative data.²⁸

Statistical Models

We fit all statistical models in a pooled logistic regression framework and used random effects to account for the clustering within both individuals (n = 40 804) and manufacturing plant (n = 62; a 3-level model). We fit 4 primary models for each of the 3 health outcomes, focusing on interpreting the estimates for state characteristics in early life. In addition to the 6 early-life state characteristics, the models contained the following covariates: model 1 included age, age-squared, race/ethnicity (Black, Latino, or other), and gender. Model 2 additionally included the smelter workplace indicator and early-life census region (West, Midwest, South, Northeast). Model 3 additionally included current state characteristics, whether the individual was an hourly worker, and individual-level employment grade. Model 4 also included individual risk factors.

Model 1 focused on how each early-life state characteristic might correlate with health as much as 50 years later; the intent was to establish that there was a sufficient signal between an exposure and outcome separated over a long period of time. Model 2 additionally controlled for early-life US census region and whether the plant was a smelter. We included these covariates because we were interested in estimates that are driven not by broad regional differences but rather by more specific state differences. Model 3 additionally included current state characteristics and socioeconomic measures. These factors occur temporally between early-life state and current health outcomes and so may be on the causal pathway, but it is nevertheless useful to characterize whether associations between early-life measures are driven in part by current individual state characteristics and individual socioeconomic measures.

The final model also included individual risk factors to explore whether current levels of these factors attenuated the relationship between early-life characteristics and later-life chronic disease. For hypertension prevalence, we included body mass index (weight in kilograms divided by height in meters squared) ²⁹; for diabetes, we included body mass index and pack-years smoked³⁰; and for ischemic heart disease, we included body mass index, pack-years smoked, high-density lipoprotein cholesterol, low-density lipoprotein cholesterol, triglycerides, systolic blood pressure, and diastolic blood pressure.³¹ We extracted these variables from employee medical records. Out of a total of 305 936 person-years observed for all of the primary demographic and health outcome variables, we had data on high-density lipoprotein cholesterol for 76 919 person-years, low-density lipoprotein cholesterol for 74 809 person-years, systolic blood pressure for 174 227 person-years, diastolic blood pressure for 174 067 person-years, body mass index for 167 294 person-years, and years smoked for 214 326 person-years.

Despite high amounts of missing risk factors resulting from the fact that records have not yet been collected from some plants, characteristics of the samples did not differ markedly (Table C, available as a supplement to the online version of this article at http://www.ajph.org). We thus fit models using multiple imputation analysis for missing data using PROC MI in SAS version 9.4 (SAS Institute, Cary, NC) with 5 imputed data sets. Multiple imputation equations included all outcome and predictor variables that were in model 3 as previously described, and we included no interaction terms.

Finally, we also performed 1 additional sensitivity analysis. We examined whether our estimates of association differed between individuals who had their social security card issued in the same state in which they worked later in life and individuals who had their card issued in a different state from where they worked later in life. The purpose was to identify whether selective migration had an impact on our findings and to identify associations based on changes in state characteristics over time among those who did not move.

We fit all models using inverse probability of censoring weights to address the bias that may exist if individuals who left the workforce earlier had different early-life state characteristics than those who remained at work. The weights were created by a prediction model for leaving the workforce by age 55 years (the dependent variable) as predicted by all of the contextual and individual characteristics previously described. Of the sample, 6% left work (i.e., were lost to follow-up) before age 55 years. The odds ratios (and 95% confidence intervals) for this model are shown in Table B (available as a supplement to the online version of this article at http://www.ajph.org). We created stabilized weights by dividing the marginal probability of censoring by the probability of leaving work. We used inverse probability weights for the models fit using SAS PROC MIXED. Individuals with a stabilized weight of greater than 20 (2.78% of the sample) were reassigned a weight of 20.

RESULTS

Table 1 shows the demographic, geographic, and health characteristics of the sample, which was more White, more male, and more working class than the general US population.

TABLE 1—

Demographic Characteristics of Individuals at First Year of Observation: Alcoa Cohort, United States, 1997–2012

Characteristic	Mean or % (n = 48 167)
Age, y, mean	47
Male, %	78
Race/ethnicity, %
Black	12
White/other	82
Latino	6
Hourly worker, %	67
Smelter worker, %	26
Current census region, %
West	8
Central	36
East	16
South	39
Early life census region, %
West	9
Central	35
East	18
South	38
Hypertension prevalence, %	24
Diabetes prevalence, %	8
Ischemic heart disease prevalence, %	5

Models of Early-Life State Characteristics

Figure 1 presents a Manhattan plot of –log(P) of the association between early-life state-of-residence characteristics and hypertension, diabetes, and ischemic heart disease. Dot shades distinguish the state characteristic, with models 1, 2, and 3 presented for each outcome. Points above the solid horizontal line are statistically significant at P < .05, and points above the dashed line are statistically significant at the Bonferroni adjusted value of P < .003 (.05/18 models per outcome).

FIGURE 1—

Manhattan plot of –log(P) of odds ratios for association of early-life state characteristics with later-life (a) hypertension, (b) diabetes, and (c) ischemic heart disease: Alcoa cohort, United States, 1997–2012.

Note. The solid horizontal line indicates the traditional P = .05 threshold. The dashed line indicates a Bonferroni-adjusted P value of .003 for each outcome. Within each indicator, the dots represent models 1–3, from left to right.

Table 2 presents the odds ratios for these models and also for model 4. Each set of 6 rows of odds ratios are from a single regression model; thus, this table shows results from the 4 different models for each of 3 outcomes (12 different regressions). For hypertension in model 1, the model of primary focus, a higher percentage of non-White, more rural, more without a high school degree, and a higher Gini were associated with higher levels of hypertension. Results were similar for model 2, but having less than a high school education was no longer significantly associated. Results from model 3, controlling for current individual socioeconomic position and state characteristics, were similar in direction and magnitude to those for model 2, although less than a high school education and higher unemployment now showed an association with lower levels of hypertension. Finally, compared with model 3, after controlling for risk factors in model 4, percentage less than high school was no longer associated, nor was the Gini index.

TABLE 2—

Association of Early-Life State Characteristics With Later-Life Health: Alcoa Cohort, United States, 1997–2012

Characteristic	Model 1, OR (95% CI)	Model 2, OR (95% CI)	Model 3, OR (95% CI)	Model 4, OR (95% CI)
Hypertension
% non-White	1.010 (1.006, 1.014)	1.010 (1.006, 1.014)	1.006 (1.001, 1.011)	1.006 (1.000, 1.011)
% rural	1.005 (1.001, 1.009)	1.004 (1.000, 1.009)	1.011 (1.005, 1.017)	1.011 (1.005, 1.016)
% < high school	1.022 (1.012, 1.032)	0.996 (0.984, 1.008)	0.987 (0.974, 0.999)	0.988 (0.975, 1.001)
% unemployed	0.999 (0.996, 1.003)	1.000 (0.996, 1.003)	0.995 (0.991, 0.999)	0.995 (0.991, 0.999)
Lower median income	0.991 (0.980, 1.001)	1.003 (0.991, 1.015)	1.002 (0.989, 1.015)	0.996 (0.984, 1.009)
Gini	1.022 (1.015, 1.029)	1.016 (1.009, 1.024)	1.013 (1.004, 1.022)	1.004 (0.995, 1.014)
Diabetes
% non-White	0.999 (0.996, 1.001)	0.998 (0.996, 1.001)	0.997 (0.994, 1.000)	0.997 (0.994, 1.000)
% rural	1.002 (1.000, 1.004)	1.001 (0.998, 1.004)	1.000 (0.997, 1.004)	1.001 (0.997, 1.004)
% < high school	1.004 (0.998, 1.010)	1.003 (0.996, 1.011)	1.003 (0.995, 1.011)	1.004 (0.996, 1.011)
% unemployed	1.002 (1.000, 1.004)	1.002 (1.000, 1.004)	1.002 (1.000, 1.004)	1.002 (1.000, 1.005)
Lower median income	0.999 (0.993, 1.006)	0.999 (0.992, 1.006)	1.000 (0.993, 1.008)	0.997 (0.989, 1.005)
Gini	1.004 (1.000, 1.009)	1.004 (0.999, 1.008)	1.005 (0.999, 1.010)	1.000 (0.995, 1.006)
Ischemic heart disease
% non-White	0.998 (0.996, 0.999)	0.998 (0.997, 1.000)	0.997 (0.995, 0.999)	0.997 (0.995, 0.999)
% rural	0.998 (0.996, 1.000)	0.998 (0.996, 0.999)	0.997 (0.995, 0.999)	0.997 (0.995, 1.000)
% < high school	1.012 (1.008, 1.016)	1.013 (1.008, 1.017)	1.013 (1.008, 1.017)	1.012 (1.007, 1.017)
% unemployed	1.001 (0.999, 1.002)	1.001 (0.999, 1.002)	1.000 (0.999, 1.002)	1.000 (0.999, 1.002)
Lower median income	0.995 (0.991, 0.999)	0.994 (0.990, 0.999)	0.993 (0.989, 0.998)	0.993 (0.988, 0.998)
Gini	1.002 (1.000, 1.005)	1.002 (0.999, 1.005)	1.002 (0.998, 1.005)	1.001 (0.997, 1.004)

Note. BMI = body mass index; CI = confidence interval; HDL = high-density lipoprotein; LDL = low-density lipoprotein; OR = odds ratio. Each set of 6 rows of ORs per outcome (and 95% CI) are from a single regression model; thus, this table shows results from 12 different regression models (4 models for each of 3 outcomes). All models include random effects for individual and plant location. In addition to the parameter estimates shown here, the models contain the following covariates: model 1 includes age, age-squared, race/ethnicity (Black, Latino, or other), and gender. Model 2 additionally includes smelter workplace indicator and the 4-category early-life census region. Model 3 additionally includes current state characteristics (the same 6 as the early-life characteristics) and whether the individual was an hourly worker and the employment grade. Model 4 additionally includes individual-level risk factors. For hypertension prevalence, we include BMI, for diabetes we include BMI and pack years smoked, and for ischemic heart disease we include BMI, pack years smoked, HDL cholesterol, LDL cholesterol, triglycerides, systolic blood pressure, and diastolic blood pressure.

For diabetes, percentage rural, percentage unemployed, and the Gini measure of income inequality were associated in the first model, and the odds ratios remained of similar magnitude in models 2, 3, and 4, although only the confidence interval for percentage unemployed did not cross 1.

For ischemic heart disease, percentage non-White, percentage with less than a high school education, and the Gini index were associated in model 1. After controlling for region, only percentage with less than a high school education remained associated, but lower median income became associated. After controlling for current state and individual socioeconomic characteristics (model 3), percentage non-White, percentage rural, percentage with less than a high school education, and lower median income were associated. In model 4, percentage non-White and percentage with less than a high school education were associated with ischemic heart disease.

Figure 2 presents the odds ratios from Table 2, model 3. This visualization shows that for hypertension, beginning with the Gini index and with subsequent factors in the clockwise direction, these factors are associated with slightly higher risk. Then for factors unemployment and income there are protective odds ratios. For diabetes, the odds ratio is closest to the null for all factors. For ischemic heart disease, lower median income was associated with lower risk, and less than a high school education was associated with an increased risk.

FIGURE 2—

Radar plot of odds ratios of association of early-life state characteristics with later-life health: Alcoa cohort, United States, 1997–2012.

Note. The figure presents the data from model 3 analyses. The location of each line along radii corresponds to odds ratios of association between the factor indicated on each radii and the health outcome as indicated in the legend to the right. The order of radii from 30° continuing clockwise to 330° is from increased to decreased risk of hypertension. The values on the left-hand side of the plot indicate the magnitude of odds ratio associated with each outcome. The range is from an odds ratio of 0.98 (center of circle) to 1.02 (fourth ring from the center), with labels of state characteristics on the outer ring.

Relationship Between Early-Life State and Current State

We determined whether odds ratios of association were similar among those who remained in the same state as in early life compared with those who moved states (Table 3).

TABLE 3—

Association of Early-Life State Characteristics With Later-Life Health, Stratified by Lifetime Mobility: Alcoa Cohort, United States, 1997–2012

Characteristic	Full Population, OR (95% CI)	Current State Same as Early Life, OR (95% CI)	Current State Different From Early Life, OR (95% CI)
Hypertension
% non-White	1.006 (1.001, 1.011)	0.983 (0.970, 0.996)	1.006 (0.999, 1.013)
% rural	1.011 (1.005, 1.017)	1.013 (0.989, 1.037)	1.010 (1.002, 1.018)
% < high school	0.987 (0.974, 0.999)	0.960 (0.943, 0.979)	1.010 (0.991, 1.030)
% unemployed	0.995 (0.991, 0.999)	0.991 (0.985, 0.997)	1.001 (0.994, 1.008)
Lower median income	1.002 (0.989, 1.015)	1.014 (0.996, 1.032)	0.972 (0.952, 0.993)
Gini	1.013 (1.004, 1.022)	1.014 (1.002, 1.027)	0.996 (0.981, 1.011)
Diabetes
% non-White	0.997 (0.994, 1.000)	0.997 (0.989, 1.005)	0.997 (0.993, 1.001)
% rural	1.000 (0.997, 1.004)	0.991 (0.977, 1.005)	1.004 (0.999, 1.009)
% < high school	1.003 (0.995, 1.011)	1.005 (0.994, 1.016)	1.005 (0.993, 1.017)
% unemployed	1.002 (1.000, 1.004)	1.004 (1.001, 1.008)	0.999 (0.994, 1.003)
Lower median income	1.000 (0.993, 1.008)	1.000 (0.990, 1.011)	0.996 (0.983, 1.009)
Gini	1.005 (0.999, 1.010)	1.004 (0.997, 1.011)	1.003 (0.993, 1.012)
Ischemic heart disease
% non-White	0.997 (0.995, 0.999)	0.993 (0.988, 0.998)	1.000 (0.997, 1.002)
% rural	0.997 (0.995, 0.999)	0.998 (0.990, 1.007)	0.999 (0.996, 1.002)
% < high school	1.013 (1.008, 1.017)	1.009 (1.002, 1.017)	1.015 (1.008, 1.023)
% unemployed	1.000 (0.999, 1.002)	1.001 (0.998, 1.003)	0.999 (0.996, 1.002)
Lower median income	0.993 (0.989, 0.998)	0.994 (0.987, 1.001)	0.990 (0.981, 0.998)
Gini	1.002 (0.998, 1.005)	1.002 (0.998, 1.007)	0.997 (0.991, 1.003)

Note. BMI = body mass index; CI = confidence interval; HDL = high-density lipoprotein; LDL = low-density lipoprotein; OR = odds ratio. Each set of 6 rows of ORs per outcome (and 95% CI) are from a single regression model; thus, this table shows results from 9 different regression models (3 models for each of 3 outcomes). All models include random effects for individual and plant location. In addition to the parameter estimates shown here, the models contain the following covariates: age, age-squared, race/ethnicity (Black, Latino, or other), gender, smelter workplace indicator, the 4-category early-life census region, current state characteristics (the same 6 as the early-life characteristics), whether the individual was an hourly worker, and employment grade (same covariates as model 3). The full population model is identical to that shown in model 3, and is repeated here to facilitate comparison of the 2 populations to the full population.

Focusing on our primary findings, the odds ratio for the Gini index and hypertension was similar in those with the same early-life and current state but closer to the null for those who moved states. By contrast, for percentage non-White and lower median income, the odds ratios were in different directions depending on whether individuals stayed in the same state or moved states.

DISCUSSION

Measures of early-life state social context were associated with chronic disease many years later. The strength of association was similar in models both with and without statistical control for individual sociodemographic characteristics and characteristics of current state context. The most consistent associations across models and health outcomes were for percentage non-White in early-life state and the Gini index of income inequality for hypertension and education for ischemic heart disease. Most measures of association did not change when including current levels of risk factors in the statistical models. The exception to this was the Gini, for which associations with hypertension and ischemic heart disease were substantially attenuated, suggesting that a potential pathway through which factors correlated with early-life state income inequality affect current chronic disease may be an individual risk factor, as has also been shown in other work.³² Finally, even without identifying mechanisms or specific causal factors, our findings, and the method of using the first 3 digits of social security numbers, can be used to inexpensively obtain information on factors that may act as confounders in studies of contemporaneous environments and chronic disease.

These findings for the association between Gini and hypertension are unlikely to be driven by selective migration because associations were the same when we examined them within the strata of the population who did not move states. In contrast, for the association of percentage of non-White and hypertension, there was a protective association for those living in the same state and an increased risk for those moving to a different state. In the case of this factor, results for those moving state are likely to be more valid for inference because the factors percentage non-White and percentage rural did not change much over time, as shown by the strong correlation in Figure C (available as a supplement to the online version of this article at http://www.ajph.org). Individuals did not seem to migrate to a state with a similar degree of rurality or percentage non-White because we found virtually no correlation between these early- and later-life characteristics among individuals who changed states. Although we did not focus inference on a comparison of current state characteristics to early-life characteristics, we present these odds ratios in Table D (available as a supplement to the online version of the article at http://www.ajph.org). We note that among our prominent findings current-state Gini was not associated with hypertension and current percentage of high school education was not associated with ischemic heart disease. Although the education findings can in part be explained by the fact that the variance in the percentage with less than a high school education is currently much smaller (Figure A), this is not true for the Gini.

In considering our findings, it is important to not interpret parameter estimates as indicating that the actual measure is causally associated with an outcome; instead, the factors are intended to capture aspects of the social environment in general. Second, these are ecological characteristics, and although certain states may have mean levels of characteristics associated with more resources, these resources are typically not distributed evenly across state residents. Finally, relationships between social factors and health are not constant over time; contextual characteristics associated with better health now may not have been associated with better health more than a half century ago.

We can also compare our findings with those of studies of individual-level early-life physiological differences and later-life chronic disease risks. Although the small magnitude of our findings diverges from those of some small studies, our findings are generally consistent with those of large, population-based studies and meta-analyses. For example, although some earlier studies showed large impacts of low birth weight on later-life hypertension,³³ meta-analyses of these associations showed very small correlations.^34,35

Limitations

It is important to consider our findings in light of several limitations of our study. Our sample is from an occupational cohort that is not a representative population of the United States. Even though the sample was racially and ethnically diverse and its socioeconomic position varied, our sample was more male, more White, and more working class than the general US population. Our sample was also located in 23 states that are distributed across census regions but are not representative. There are also nonrandom processes by which individuals move from the state in which they lived early in life to where they live during their working life. We address this limitation by showing that the majority of our results are robust to a stratified analysis of individuals who do not move states; thus, in this population, estimates of association were based on the change in characteristics of the states rather than the individual moving to a particular other state (Table 3). In fact, for hypertension, associations were generally stronger for those individuals who did not move states. Finally, our multiple imputation analysis for model 4 is valid only if data were missing at random. Although we cannot prove this, this approach seems reasonable enough because most of the missing data were the result of data being collected from only a subset of plants, not from differences in participation, and characteristics did not differ dramatically between those with and without data on risk factors.

Our study does not provide information on the early-life causes of these later-life chronic disease differences, nor on the mechanisms by which these differences occur. Although we presented specific associational estimates, these factors are themselves associated with myriad policy decisions and historical processes specific to states that correlate with the factors we examined. We presented 1 set of analyses in which we showed that a small selection of current health behaviors do not meaningfully change most of the estimates of association with early-life state context (Table 2, model 4), but these findings cannot be considered to be a rejection of the hypothesis that known risk factors entirely mediate the association between early-life environment and later-life chronic conditions, in part because these measures were only available during the observed study period. In addition, future work including formal mediation analysis will be necessary to better understand how current levels of risk factors and occupational exposures may be on the pathway between early-life environment and current health outcomes.³⁶ For example, exposure to fine particulate matter has been associated with incidence of ischemic heart disease in a subset of hourly workers in this population.^27,37 Part of the intervening process may be differential selection to workplace environments that have different levels of physical and social hazards. Both screening programs that were designed to avoid assigning higher risk individuals to higher risk jobs (e.g., high heat) as well as recent evidence of stronger healthy worker survivor bias³⁸ may have impacts.

Conclusions

Despite these limitations, our study offers important directions for future work. First, the size of the associations that we found suggests that, much like current thinking in genomewide associational studies used to identify single nucleotide polymorphisms associated with disease, the analysis of large population sizes either through the use of administrative databases or through meta-analytic combinations of cohorts is necessary for identifying whether there are meaningful and robust associations with early-life conditions. Despite the individually small strengths of association, we found that all 6 factors are associated with 1% to 2% lower odds of prevalent hypertension; strengths of association are at least as strong as recently documented genetic factors.³⁹ The social environment associations we document are also dramatic given the general exposure measures used in this study. Given the complex constellation of factors that may contribute to chronic disease across the life course, these studies are likely to provide useful information for the primary prevention of chronic disease. Future work should measure more specific state level exposures and analyze which specific policy changes could result in lower levels of chronic disease.