Understanding Geographic Disparities in Mortality

Abstract

A rich literature shows that early-life conditions shape later-life outcomes, including health and migration events. However, analyses of geographic disparities in mortality outcomes focus almost exclusively on contemporaneously measured geographic place (e.g., state of residence at death), thereby potentially conflating the role of early-life conditions, migration patterns, and effects of destinations. We employ the newly available Mortality Disparities in American Communities data set, which links respondents in the 2008 American Community Survey to official death records, and estimate consequential differences based on the method of aggregation we use: the unweighted mean absolute deviation of the difference in life expectancy at age 50 measured by state of birth versus state of residence is 0.58 years for men and 0.40 years for women. These differences are also spatially clustered, and we show that regional inequality in life expectancy is higher based on life expectancies by state of birth, implying that interstate migration mitigates baseline geographic inequality in mortality outcomes. Finally, we assess how state-specific features of in-migration, out-migration, and nonmigration together shape measures of mortality disparities by state (of residence), further demonstrating the difficulty of clearly interpreting these widely used measures.

Introduction

For decades, the United States has lagged comparably high-income nations and some middle-income countries on major population health indicators (Kulkarni et al. 2011; Woolf and Aron 2013), including life expectancy (Schwandt et al. 2021). Recently, life expectancy in the United States declined (Case and Deaton 2015), while mortality inequalities increased (Chetty et al. 2016; Currie and Schwandt 2014; Dwyer-Lindgren et al. 2016; Ezzati et al. 2008; Montez and Zajacova 2013; Murray et al. 2006; Wang et al. 2013). Understanding longevity disparities across subpopulations is a critical step in addressing America's growing health disadvantages.

Geography—encompassing physical, social, and policy environments—is a key axis of mortality disparities (Black et al. 2015; Chetty et al. 2016; Dwyer-Lindgren 2016; Ezzati et al. 2008; Murray et al. 2006; Wang et al. 2013; Woolf and Schoomaker 2019). Places have a causal impact on mortality among older adults (Deryugina and Molitor 2020; Finkelstein et al. 2021), and early-life exposures have long-term impacts on subsequent health and longevity (Galobardes et al. 2006; Haas 2008; Hayward and Gorman 2004; Palloni 2006; Schwandt and von Wachter 2020; Warner and Hayward 2006). Linking these processes between place effects in early life and place effects later in life is migration. Not surprisingly, when faced with circumstances that threaten lives and livelihoods (e.g., Boustan et al. 2020; Deryugina and Molitor 2020; Hornbeck 2012; Hornbeck and Naidu 2014), or when hoping for better prospects (e.g., Kennan and Walker 2011), individuals often migrate. Indeed, more than 30% of U.S.-born adults in recent cohorts leave their state of birth (Molloy et al. 2011), raising the possibility of consequential differences in geographic mortality disparities calculations using alternative measures of “place.”

Nonetheless, most research on spatial disparities in mortality implicitly or explicitly aggregates death outcomes by individuals' place of residence at death (e.g., Dwyer-Lindgren et al. 2016) or at some point during late adulthood (Chetty et al. 2016; Finkelstein et al. 2021). Conceptually, the resulting measurements of spatial disparities are difficult to interpret, as they are a mixture of persistent early-life health differences among nonmigrants, life course migration patterns (which are shaped by early-life health) of both in-migrants and out-migrants, and later-life health environments (e.g., quality of medical care) of residents. This mixture of factors suggests major hurdles for both learning about the causes of measured spatial disparities and suggesting remedies for the disparities.

At the same time, a rapidly advancing literature is beginning to tie together (contemporaneously measured) state-level factors with geographic disparities in mortality in order to understand key sources of health differences at a point in time and across decades (e.g., Montez et al. 2020; Montez and Farina 2021; Montez, Hayward et al. 2019; Montez, Zajacova et al. 2019). For example, Montez et al. (2020) showed that individuals residing in a U.S. state with more liberal policies live more than two years longer than those living in less liberal states. Data limitations often do not allow these analyses to examine life course exposures to state policies and thus face similar interpretation issues as the larger literature on geographic disparities in (mostly) old-age outcomes such as mortality. Specifically, it is difficult to estimate whether the policies causally affect health in old age, since policy environments are chosen through in- and out-migration, and it is also difficult to examine the cumulative and dynamic effects of exposures to policies in earlier life on old-age mortality without information on where individuals lived in early life.

Indeed, limited work has attempted to decompose these factors and no work has contrasted alternative measures of spatial disparities based on place of birth and place of death. A recent exception is Xu et al. (2020), which showed that state of birth explains a similar amount of variation in late-life mortality as state of residence, suggesting that these two alternative measures of geographic disparities could differ, but it did not otherwise directly estimate these measurements nor have large enough samples to fully decompose state-specific migration experiences to understand the sources of these differences. We expand on the empirical findings of Xu et al. (2020) by providing the first quantification of the extent that mortality disparities differ when measured by state of birth versus state of residence and disentangling the role of in-migrants and out-migrants in explaining the differences across life expectancy measures.

Using the Mortality Disparities in American Communities data set, we find important differences between measures of life expectancy by state of residence and state of birth. Overall, we find that the method of aggregating individuals by state of residence in later life underestimates the extent of geographic inequality in mortality outcomes compared with the method that aggregates individuals by state of birth.

We then proceed by decomposing the difference in life expectancy by state of residence and state of birth into the difference in the life expectancy of in-migrants relative to stayers, the difference in the life expectancy of out-migrants relative to stayers, and in- and out-migration rates. Surprisingly, we find that state in- and out-migration rates are largely uncorrelated with the life expectancy of stayers, which is inconsistent with a simple story that migrants select destinations on the basis of health environments. Instead, we show that states both lose healthy out-migrants and gain healthy in-migrants and that the net effect of these flows both differs widely across states but also is clustered by region. For example, we find that the mortality risk of in-migrants is substantially lower than the mortality risk of nonmigrants in many southern states, while in many states in the Northeast and Midwest, the mortality risk of these two subgroups is similar.

Finally, we explore several counterfactual simulations in order to decompose the roles of selective migration and potential “place effects” that aggregate to produce life expectancy differences by state of birth and state of residence. We find evidence that the nonrandom sorting of migrants to destinations based on state of birth and unobserved mortality risk plays an important role in explaining why life expectancies by state of residence are significantly different than life expectancies by state of birth for certain states. “Place effects” also contribute to the patterns in the data. We also present an assessment of our mortality models and robustness checks to validate that our results are not sensitive to alternative assumptions

Data

The analysis of geographic heterogeneities in mortality patterns requires a considerable amount of mortality data across different locations. We make use of the newly available Mortality Disparities in American Communities (MDAC) restricted data set to perform our analysis. The MDAC data set links respondents in the 2008 American Community Survey (ACS) to official death records from the National Death Index. The currently available mortality follow-up period extends until December 31, 2015.

The MDAC data set contains approximately 4.5 million individuals who were surveyed as part of the original 2008 ACS. More than 300,000 of these individuals died over the following seven years. We restrict our analysis to individuals who were born in one of the 50 U.S. states or Washington, D.C., and who were 50 or older in 2008.¹ We further drop individuals who did not provide valid personal information that allows them to be matched to official death records (∼0.8% of the sample). In total, our sample has close to 1.5 million individuals.

In Table 1 we provide descriptive statistics of our sample by gender, age group, and mortality status by 2015. Since very few prior papers have used the MDAC data set to analyze mortality (e.g., Miller et al. 2019), the table also presents comparable statistics from the National Vital Statistics System (NVSS) over a roughly similar follow-up period. Further details about the validation of the MDAC and NVSS samples is available in online appendix A.

Methods: Primary Calculations

Because of the restricted nature of the data, we use “cell counts” as opposed to individual-level data. We construct aggregated death rates from the linked 2008 ACS respondents and official death records from 2008 to 2015 in two ways. We aggregate individuals by (1) their state of residence at the time of the 2008 interview or (2) their reported state of birth (we treat Washington, D.C., as a state in the analysis). We further stratify the sample by five-year age group and gender to calculate the raw probability of surviving throughout the 7+-year follow-up period by gender, five-year age group, and either state of residence or state of birth. Using these cells as inputs, we compute period life expectancies at ages 50 and 65 by state of birth and then by state of residence.

The computation of period life expectancies involves two steps. We first need to obtain traditional one-year mortality rates as a function of age from the disclosed mortality probabilities by age group in the follow-up period. To do this, we rewrite the probability of surviving throughout the follow-up period for a given age group in terms of one-year mortality rates as a function of age. We further assume that mortality rates grow exponentially with age, which leads us to expressing the probability of surviving throughout the follow-up period across age groups as a nonlinear equation with two parameters.² We use weighted nonlinear least squares (NLLS) to fit these two parameters for each state of birth and gender and for each state of residence and gender. In total, we run 204 (2 genders × 51 states × 2 aggregation methods) different regressions. After age 90, we impute sex-specific mortality rates that are equal to the observed mortality rates at the national level following the methodology presented in Chetty et al. (2016) (in online appendix C, we show that the results are robust to other age cutoff values).

Next, we apply standard life table formulas to compute life expectancies at ages 50 and 65 from the age schedules of mortality estimated in the first step (Olshansky and Carnes 1997; Preston et al. 2000). We obtain standard errors for the 204 life expectancy estimates using parametric bootstrap (Chetty et al. 2016). We validate the magnitude of our standard errors by comparing our estimates to estimates that follow the classical approach introduced by Chiang (1984). Further technical details regarding these computations can be found in online appendix B.

Life Expectancies by State of Residence and State of Birth

Figure 1 displays life expectancy at age 50 for each state (aggregating (1) those born in the state and (2) those who reside in it at the time of their death) using the mortality events that occurred in the period 2008–2015. Panel a compares the two state-based measures for men and panel b does the same for women. A state would lie on the dashed 45-degree line if the two measures of life expectancy were identical. This would happen, for example, if there was no migration into and out of the state, or if the mortality patterns of in-migrants and out-migrants were the same. States that are located to the right of the 45-degree line are those that have a higher life expectancy when mortality is aggregated by state of birth than by state of residence. Ohio is one such example, for both men and women. The estimated male life expectancy at age 50 by state of birth in Ohio is 30.5 years (standard error (SE) = 0.20), while this measure by state of residence is 29.6 years (SE = 0.15). In contrast, states located to the left of the 45-degree line have a higher life expectancy when individuals are aggregated by location of adult residence than by birth. For example, the estimated life expectancy for males born in Florida is 29.5 years (SE = 0.17), while this measure for men residing in Florida is 30.7 years (SE = 0.26). These differences of close to one year in life expectancy are significant in both a substantive and a statistical sense.

Comparing across the Figure 1 panels, the link between the two life expectancy measures is weaker for men than for women. For men, the unweighted (weighted) correlation coefficient between the two measures is equal to .77 (.82).³ In contrast, the unweighted (weighted) correlation coefficient for women is higher and equal to .83 (.91).⁴ Being less susceptible to the presence of outliers, the unweighted (weighted) mean absolute deviation of the difference in the two measures across states is equal to 0.58 (0.50) years for men and 0.40 (0.29) years for women. The difference between genders is close to being statistically significant (p = .06).

Previous literature has shown that the American South has the lowest levels of life expectancy by state of residence (Chetty et al. 2016; Murray et al. 2006; Wang et al. 2013, among many other papers). We confirm this pattern in panel a of Figure 2, where we show male life expectancies at age 50 by state of residence. In panel b we show male life expectancies at age 50 by state of birth. To ease the comparison of differences between these maps, panel c displays the difference between the two life expectancy measures (state of residence – state of birth) for each state. This map demonstrates that the sign of the difference is geographically clustered and varies substantially by census division. All states in the East North Central, West North Central, and Middle Atlantic divisions except for Minnesota have higher life expectancy point estimates by state of birth than by state of residence. The opposite is true for almost all states in the South Atlantic and East South Central divisions, which already had the lowest life expectancies by state of residence.

Thus, the extent of inequality in mortality outcomes across divisions is higher if we measure life expectancy by state of birth than by the typically used state of residence. In panel d of Figure 2 we highlight the states where the difference in life expectancy measures is statistically significant at a 10% level. Most of the states that have statistically significant higher life expectancies by state of residence are in the South Atlantic division: Florida, Georgia, Maryland, South Carolina, and Virginia. Those where life expectancy is higher when calculated by place of birth are mainly in the East North Central and Middle Atlantic divisions: Illinois, Indiana, New York, Ohio, and Pennsylvania.

In Figure 3 we present life expectancies at age 50 for women. The patterns are similar to those documented for men, but differences are slightly smaller in magnitude. One potential explanation for this gender difference is that the overall migration rate is higher for men than for women. We can directly assess this by analyzing the IPUMS version of the 2008 ACS sample. By the time of the 2008 ACS interview, 41.7% of women and 42.1% of men aged 50 or older are not residing in their state of birth. Although this difference is statistically significant, it is unlikely to be the driver of the weaker relationship of the two life expectancy measures, as it represents only a 0.4-percentage-point increase in the baseline migration probability for women. Instead, the relationship between health status and migration decisions might differ by gender. For example, Halliday and Kimmitt (2008) found that a lower reported health status is associated with a lower propensity to migrate for men younger than 60 but not for women. Thus, different migration motives across genders might help explain why the discrepancies in life expectancy by state of birth and by state of residence are lower for women than for men.

Because of disclosure rules for minimum sample sizes in each “cell” for the MDAC data set, we were unable to consider separately mortality rates for racial minority groups or mortality rates at younger ages. To indirectly assess the former, we repeat the foregoing analysis with a subsample that excludes Black and Latino individuals and find similar patterns (see online appendix Figure A3).⁵ To consider the possibility of differences in results based on age, we also calculate life expectancies at age 65 for men and women and find very similar patterns at this older age. The relationship between life expectancies at age 65 by state of residence and by state of birth is portrayed in online appendix Figure A4. Figure A5 shows that the disparities across male life expectancies at age 65 are closely linked to geographic regions. Figure A6 shows the geographic patterns in life expectancies at age 65 for women. Overall, the results are evidence that the empirical patterns in the discrepancies between life expectancy by state of residence and by state of birth are not driven by any specific age group or racial/ethnic group.

Analysis of Subpopulations: Stayers, In-Migrants, and Out-Migrants

Three broad groupings of individuals are considered in the construction of life expectancies by state of residence and state of birth for a given state $s$⁠: individuals who were born in $s$ and are residing in $s$ by the time of the ACS interview (“stayers”),⁶ individuals who were born in $s$ but are observed in a different state (“out-migrants”), and individuals who were not born in $s$ but who are observed in $s$ (“in-migrants”). For each state, we estimate life expectancies for the three different subgroups. The calculation of life expectancies by state of birth assigns positive weight only to the first two subgroups, while life expectancies by state of residence consider only stayers and in-migrants. The weights assigned to each subgroup are closely related to the state in-migration and out-migration rates. We pay particular attention to the differences in male life expectancy of out-migrants and in-migrants relative to stayers across states.

We first compute the life expectancy of stayers using the 2008 ACS matched with official death records. As in the previous analysis, we stratify the stayer population by five-year age group and gender. Because of disclosure requirements, we are unable to report match rates and numbers of stayers in cells that have fewer than 20 deaths. To obtain reliable life expectancy measures at age 50, we drop states that have missing information in three or more of the eight age groups.⁷ We calculate life expectancies at age 50 for stayers using the same two-step approach as before. First, we use a weighted NLLS model based on the Gompertz mortality model to estimate age-specific mortality rates. Then, we follow standard life table procedures to calculate life expectancies in a second step. Online appendix B presents the details.

To calculate mortality rates and cell sizes of in-migrants and out-migrants, we combine the information about mortality rates and number of stayers in each cell with our previous data on natives and residents.⁸ Then, we calculate life expectancies for in-migrants and out-migrants with our two-step estimation strategy.

where $Dj$ corresponds to the number of deaths of individuals of subpopulation $j$ and $Nj$ is the total number of individuals from that subpopulation.

Thus, differences in mortality rates by state of residence and by state of birth can be decomposed into two additive terms. The first term considers the difference in mortality rates of in-migrants relative to stayers $mIn−mStay$⁠, while the second term considers the difference in mortality rates of out-migrants relative to stayers $mOut−mStay$⁠. The terms are weighted by the in-migration (⁠$rIn$⁠) and out-migration rate (⁠$rOut$⁠) in that specific age group, respectively.

This equation shows that the difference in life expectancy by state of residence and by state of birth can be (approximately) decomposed into two additive terms: (1) the difference in life expectancy between in-migrants and stayers and (2) the difference in life expectancy between out-migrants and stayers. The combination of these two additive terms explains the discrepancy for any given state between life expectancy calculated for residents and life expectancy calculated for those born in the state.¹⁰

As a first step to determine the contribution of these factors to variation in differences in male life expectancy at age 50 by state of residence and by state of birth, Figure 4 plots the relationship between the two factors at the state level,¹¹ showing no significant relationship between the relative mortality advantage of in-migrants and out-migrants. The slope of the fitted line is equal to 0.05 (SE = 0.09), suggesting that states with higher in-migrant life expectancy are not systematically experiencing lower out-migrant life expectancy.

A second takeaway from Figure 4 is that the cross-state standard deviation in the in-migrant mortality advantage is higher (0.92) than the cross-state standard deviation in the out-migrant mortality advantage (0.67), suggesting that a larger component of the difference between life expectancy by state of residence compared with state of birth is differential state gains from in-migrants rather than differential state losses from out-migrants (see online appendix Table A1). Importantly, there is a difference in the right tails of these two distributions. In seven states, the difference in life expectancy of in-migrants and stayers is greater than two years. Six of these seven states are in the South region. In contrast, only Louisiana and Mississippi have an out-migrant mortality advantage that is higher than two years. Of of the 43 states in the analysis, 33 (31) have an out-migrant (in-migrant) mortality advantage. In online appendix Figure A6, we show the smoothed distribution of both components.

Two different mechanisms can be behind the differences in mortality advantage of in-migrants across states—place-based selection and causation.¹² The first mechanism is ex ante health advantage of in-migrants relative to stayers, where the difference in mortality risk between in-migrants and the stayer subpopulation at the destination at the time of the move is different across states. The second one is the presence of differences in causal “place effects” across locations (see also Deryugina and Molitor 2020 and Finkelstein et al. 2021). The interaction between the two can also be relevant. For example, detrimental place effects might have a bigger effect on the mortality outcomes of more vulnerable subgroups. Thus, place effects might exacerbate or mitigate ex ante differences in mortality outcomes between subgroups. While we do not attempt to formally disentangle these two different channels, we provide suggestive evidence in the following that the selection channel is playing an important role in explaining the cross-state variation in the mortality advantage of in-migrants.

The relationship between the levels of life expectancies of stayers, in-migrants, and out-migrants is also informative. We find that the correlation between the life expectancy of stayers and out-migrants is high and equal to .75 for men. In contrast, the correlation between the life expectancy of stayers and in-migrants is lower and equal to .55 (see Table 2).¹³

In Figure 5 we present the baseline levels of the life expectancy of in-migrants, out-migrants, and stayers for the 13 states in which the difference between the male life expectancy at age 50 by state of residence and by state of birth is statistically significant at a 10% significance level. In all selected states except Colorado, the life expectancy of stayers is lower than that of out-migrants. This suggests that the “healthy migrant hypothesis” (Palloni and Morenoff 2001) holds in our data, even for the states in the South, where the difference between the life expectancy by state of residence and by state of birth is positive. However, we do not have information about the health of out-migrants at the time of migration to formally test this hypothesis.

Finally, Figure 5 shows that the life expectancy of stayers is on average lower in the South Atlantic and East South Central states than in the Middle Atlantic and East North Central states. This suggests that the underlying causal place effects of states in the South could be more detrimental than in the Midwest and Northeast.¹⁴ However, the life expectancy of in-migrants relative to stayers is substantially higher in this set of southern states. Thus, in-migrants shift up the life expectancy by state of residence in these states. This does not happen in the Middle Atlantic and East North Central states, where the life expectancy of in-migrants and stayers is similar in magnitude.

We are unable to formally quantify the cross-state difference in the relative ex ante health selection of in-migrants relative to stayers at the time of migration. As mentioned before, the ex ante differences can be exacerbated if detrimental place effects have a bigger effect on individuals with already vulnerable health status.¹⁵ The empirical patterns that we highlighted from Figures 4 and 5 are consistent with a higher ex ante health selection of in-migrants in southern states than in states in the Midwest and Northeast. If the role of “place effects” was substantial, we would expect to observe a positive correlation between the mortality advantage of in-migrants and out-migrants.¹⁶ However, we showed in Figure 4 that these two variables are virtually uncorrelated.

Assessing the Role of Migration Flows

In the previous section, we abstracted away from the role of migration in-flows and out-flows and focused entirely on the cross-state variation in mortality differences between in-migrants, out-migrants, and stayers. In this section we further investigate which aspects of migration are descriptively important in explaining the differences between life expectancy by state of residence and by state of birth across states.

Individuals who are born in locations with detrimental “place effects” on health outcomes might be able to mitigate the adverse effects of their place of birth on health by migrating to healthier locations. For example, some research has shown that one important way in which individuals respond to natural disasters like the American Dust Bowl or Hurricane Katrina is by migrating to unaffected locations (Boustan et al. 2020; Deryugina and Molitor 2020; Hornbeck 2012). A broader literature has documented that most migration is motivated by employment, education, and family considerations (Cooke 2011; Kaplan and Schulhofer-Wohl 2017; Wolf and Longingo 2005), though whether these migration processes would be linked to geographic differences in life expectancy has not been examined.

where $ysg$ is a migration outcome of interest in state $s$ and the explanatory variable $LEStayers,s$ is the life expectancy of the subpopulation of stayers in state $s$⁠.

Migration patterns might be substantially different for working-age populations and retirees. For example, return migration might be more prevalent after retirement. We primarily focus on the population that is between 50 and 64 years old by the time of the 2008 ACS interview to mitigate survivorship bias as well as classification error from return migrants. Our outcomes are state out-migration and in-migration rates.

We use the life expectancy of stayers as a proxy for the place effects on health of different states. As has been highlighted in the previous section, life expectancies by state of residence and by state of birth are a combination of many different factors, including the life expectancy of stayers, in-migrants, and out-migrants and migration rates. In panel a of Figure 6, we show the linear relationship between out-migration rates and life expectancy of stayers at age 50 for men. We weigh each state by the inverse variance of the life expectancy of stayers to address the concern that our dependent variable is subject to measurement error and that it is estimated with a different level of precision across states. Finally, we omit the eight states in which the population of stayers is too small to obtain reliable estimates of life expectancy for the stayer subpopulation.

As can be seen in panel a, the relationship between out-migration rates and the life expectancy of stayers is virtually flat.¹⁷ The estimated slope is equal to 0.01 (SE = 0.01). Overall, there is no evidence that out-migration rates are higher in locations where the life expectancy of stayers is lower than average.

In panel b of Figure 6, we show the relationship between the in-migration rates and life expectancy of stayers. As before, the relationship is flat. The estimated coefficient is equal to −0.02 (SE = 0.02), and it is not statistically significant. States where stayers have a higher life expectancy at age 50 do not appear to be attracting relatively more immigrants than states with potentially more detrimental place effects. It is worth noting that our data cannot explore migration below the state level (e.g., county); Figure A8 in the online appendix instead explores further aggregated units (interregional migration) and shows that the relationships are similar to our results in Figure 6. Appendix Figure A9 reports findings for older men (65–79) to explore differences between retirees and shows the same relationships as our main results. In unreported results, we verify that this pattern is driven by the White non-Hispanic subpopulation.

Although the results do not consider individual characteristics that might help explain self-selection into migration, the aggregate patterns of in- and out-migration are not consistent with a narrative in which net migration flows primarily from locations with lower life expectancy to locations with higher life expectancy. A potential explanation is that individuals give a higher priority to wage differentials and other work amenities over health considerations when deciding where to settle (see Kennan and Walker 2011). This pattern is consistent with findings from the literature on the rural–urban migration process.

Understanding Differences in Place-Based Life Expectancy Measures

Even though the magnitudes of the migration flows across states do not appear to be systematically associated with the life expectancy of stayers, migration can still make the interpretation of life expectancies by state of residence difficult. In particular, the representative state of origin and the health composition of who moves in and out of a state relative to the health of the population of stayers may vary substantially across states.

We first describe differences in the states of origin of in-migrants across states by showing the relationship between the life expectancy of male stayers in a given state $s$ and a summary measure of the states of origin of its in-migrants in terms of life expectancy. Specifically, for a given state $s$⁠, we weigh the life expectancy of stayers in all other states by the share of in-migrants who come from each state. This measurement does not allow a “healthy migrant” effect at the individual level but focuses only on the composition of in-migrants from “healthy” or “unhealthy” states of origin. We obtain these shares from the IPUMS ACS 2008.¹⁸

Panel a of Figure 7 presents the relationship between the life expectancy of stayers and this summary measure of the states of origin of in-migrants. The positive relationship indicates that on average, states with a higher life expectancy of stayers receive in-migrants who come from states of origin where stayers also have high levels of life expectancy. Instead, if destination choice was independent of state of origin, we would expect to observe a flat relationship between the life expectancy of stayers and the representative state of origin of in-migrants.

However, the slope of this relationship is also significantly lower than one (slope = 0.28; SE = 0.28). This implies that in states where life expectancy is lower (higher) than average, in-migrants come on average from states of origin that are healthier (unhealthier) relative to the destination. For example, the estimated life expectancy gap between the states of origin of in-migrants and stayers is equal to 1.32 years in Georgia, where life expectancy of stayers is low, and is −1.37 years in the state of Washington, where life expectancy of stayers is high. These differences across states illustrate the way in which migration mitigates state-of-birth-based disparities—migration induces convergence in state life expectancies. The least (most) healthy locations receive in-migrants who on average come from more (less) healthy locations.¹⁹

We construct a similar measure to summarize the destinations of the out-migrants from each state. To construct this measure, we weigh the life expectancy of male stayers from all destinations of $s$ by the share of out-migrants from $s$ who relocate to each state. As in panel a, panel b of Figure 7 shows that there is a positive relationship between the life expectancy of stayers in a given state and the destinations of its out-migrants. Out-migrants who move from “healthy” locations relocate (on average) to destinations where the life expectancy of stayers is also high.²⁰

The previous analysis abstracts from the possibility that different locations might attract in-migrants based on their health, even after controlling for the state of origin of its in-migrants. To quantify how migration, both flows in levels and flows based on specific purposeful sorting regularities, affects differences between state-based life expectancy measures, we perform a series of counterfactual exercises where we modify the extent to which the final destinations of out-migrants are tied to their original health and state of origin.

A full description of the analysis and results is contained in online appendix D. Briefly, we decouple purposeful selection of migration destinations and place effects by taking the set of migrants in the data and “shuffling” their destinations using three different assignment procedures independently of their health status at the time of migration: by observed “popularity” of the destinations for all migrants, by equalizing in-migration rates, or by equalizing net migration rates across states. In addition to assignment of migrants to destinations, the counterfactual exercise also allows “place effects” of destinations to be present or absent.²¹ We focus on the subset of 13 states in which the difference in male life expectancy measures at age 50 is statistically significant at conventional levels.

In the counterfactuals that use popularity and no place effects, we find substantial reductions in the difference between counterfactual state of birth and state of residence life expectancies compared to the empirical data in the subset of selected states. Allowing for constant place effects moves our counterfactuals farther away from the empirical data. We calculate that the cross-state standard deviation in life expectancy measures is reduced by 36% to 49% in our counterfactual reshuffling exercise, suggesting that the nonrandom sorting of out-migrants to destinations is an important component of the differences in state of birth and state of residence life expectancy estimates in our subset of 13 states. Importantly, many states in the South appear to be attracting in-migrants who have lower mortality risk than what would be expected in a counterfactual scenario where the destination choices of out-migrants do not depend on state of origin and mortality risk. The opposite pattern appears to hold for many states in the Midwest. One explanation of these patterns is that migrants likely seek out economic opportunities, which during our data window are generally better in the South than in the Midwest (Kennan and Walker 2011).

When we compare the results across settings of our counterfactual exercises, we find additional evidence that once we equalize the unobserved mortality risk and state of origin of in-migrants across destinations, equalizing the in-migration or net migration rates across states still contributes to reducing the difference in life expectancy measures across states. However, the contribution of the equalization of migration rates is quantitatively smaller than the equalization of unobserved mortality risk and state of origin of in-migrants. Together, our findings suggest that (1) maximizing life expectancy does not appear to be a primary driver of migrants' destination choices, (2) there is substantial variation in the composition of in-migrants across states, and (3) “place matters” for migrant life expectancy.

Conclusion

While place/state of residence is nearly universally used in research measuring geographic inequalities in mortality, we outline a set of conception issues in interpreting these measures that could limit their usefulness in both understanding the sources of mortality disparities and proposing solutions. An alternative aggregation is place/state of birth, which is not subject to issues of migration and uncertain timing of exposure effects, though both aggregations face limitations regarding the duration of exposure to the “place.” While more than 30% of individuals in recent cohorts die in a state they were not born in, it is unclear whether considering an alternative “place” of aggregation is consequential for estimated geographic mortality inequalities. We used the novel MDAC data set and estimated consequential differences by method of aggregation; the mean absolute deviation of the difference in life expectancy at age 50 measured by state of birth versus state of residence is 0.58 (0.50) years for men and 0.40 (0.29) years for women.

Further, we show that these differences in life expectancy measures are clustered geographically. The difference in life expectancy by state of residence and by state of birth is positive and statistically significant in states in the East South Central and South Atlantic divisions. In contrast, the difference in life expectancy measures is negative and statistically significant in states in the East North Central and Middle Atlantic divisions. States in the South have been recognized by a broad literature as the states with the lowest levels of life expectancy. Our findings imply that regional disparities in mortality outcomes are even more unequal when life expectancies are constructed by aggregating individuals on the basis of their state of birth than by the usual life expectancies by state of residence. Life course migration patterns serve to reduce these disparities, making them less visible in nearly all research in this literature.

While the culprit of the differences is life course migration, the specific patterns are complex and reflect flows (levels) of in-migration and out-migration in each state, as well as the differential health of individuals (characteristics) who stay, move in, and move out. Flows and characteristics of migrants also may interact with “place effects” of origins and destinations.

Overall, we do not find any evidence that out-migration and in-migration flows (in levels) are correlated with the life expectancy of stayers. Thus, the states where stayers have the highest levels of life expectancy do not appear to be receiving a higher or lower inflow of immigrants. The same result holds for out-migration flows. This result provides suggestive evidence that health factors might not be a strong pull or push factor in determining which locations send or receive more individuals, at least at the state level.

To provide some descriptive evidence about the relative importance of state-specific experiences of migration flows, we compute the life expectancy of the subpopulations of stayers, in-migrants, and out-migrants for each state. By comparing these life expectancies with that of stayers, we find evidence that the mortality advantage of in-migrants across states plays a more important role than the mortality advantage/disadvantage of out-migrants in determining differences between life expectancy by state of residence and by state of birth. For five states in the South, we document that the male life expectancy of in-migrants is more than two years higher than that of stayers.

Thus, rather than the raw magnitude of migration flows leading to differences in our two life expectancy measures, our results point to more subtle state-to-state migration processes that together lead to the differences in life expectancy. Destination decisions are not independent of geography and state of origin. “Healthy” states attract migrants from other “healthy” states and lose migrants to other “heathy” states, and the net effects of these processes differ by state but also cluster geographically. For example, in-migrants in the South live longer than stayers by a substantially wider margin than in states where stayers have higher life expectancies. Our counterfactual examinations suggest that our main results are driven by a combination of “place effects” and that different places receive different compositions of migrants in terms of state of origin and unobserved mortality risk.

These explanations of the differences between life expectancy calculated at the state of birth versus state of residence level do not prescribe whether and when to focus attention on one or the other measurement. We note that the interpretation of life expectancy by state of residence is complicated—this is well understood, but our study shows that the decision to focus on one versus the other is consequential for the measurement of geographic disparities in mortality. While it is the typical focus in the literature, measures of mortality disparities based on state of residence represent a combination of early-life health processes of nonmigrants, migration inflows and outflows, causal place effects, and the interactions of these processes and therefore blur the interpretation of life expectancies by state of residence. As we have described, these processes create difficulties in decomposition exercises that attempt to explain state-of-residence-based measures of life expectancies using (contemporaneously measured) state conditions and policies (e.g., Montez and Farina 2021), since there is selective migration. Contemporaneously measured conditions also fail to capture life course exposures to “place” conditions that may accumulate over time, lead to migration, or be experienced in critical early time periods, such as in utero. The measure of life expectancy by state of birth is more easily interpretable, is consistent with the theoretical assumptions of closed populations used in the construction of life tables, and has a higher correlation with the life expectancy of stayers, which can be a reasonable proxy for the causal place effects of locations under certain conditions.

Finally, many papers in the literature have focused on measuring how life expectancy by place of residence has evolved over time across different locations. For example, Dwyer-Lindgren et al. (2016) documented that many counties in Florida saw important gains in life expectancy between 1980 and 2014. We point out that the interpretation of these changes is even more challenging than the interpretation of life expectancy by state of residence at a given time. Our framework can be extended to incorporate multiple periods of time to show that an observed improvement in the life expectancy by state of residence in a given location can be driven by intertemporal changes in the causal place effects of locations, in the in-migration and out-migration rates in the location, and in the relative selection of in-migrants and out-migrants (Dowd and Hamoudi (2014) coined “lagged selection bias” to refer to changes in the composition of migrant populations). We argue that substantially more research is required in this area to disentangle intertemporal changes in “place effects” from changes in migration patterns. This task is crucial in order to evaluate which public policies or government programs are contributing to the improvement of the mortality outcomes of residents (Miller et al. 2019; Montez et al. 2020).

Acknowledgments

The authors thank Hans Schwarz for exceptional research assistance. They would also like to thank Bruce Weinberg and participants at the 2019 Research Data Center Annual Conference, the 31st Annual Colloquium on Aging of the Institute on Aging of the University of Wisconsin–Madison, the 2019 Southeastern Demographic Association Annual Meeting, the La Follette School Seminar Series, and the University of Wisconsin–Madison Health/Aging/Place working group. Any opinions and conclusions expressed herein are those of the authors and do not reflect the views of the U.S. Census Bureau. All results have been reviewed to ensure that no confidential information is disclosed. The statistical summaries reported in this document have been cleared by the Census Bureau’s Disclosure Review Board release authorization numbers CBDRB-FY19-304, CBDRB-FY20-CES004-090, CBDRB-FY20-092, and CBDRB-FY21-CES004-021. The authors would like to acknowledge financial support from National Institute on Aging (NIA) grant R01AG060109 and the Center for Demography of Health and Aging (CDHA) at the University of Wisconsin–Madison under NIA core grant P30 AG17266.

Notes

References