WHAT IS ALREADY KNOWN ON THIS TOPICHOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

Research should examine the protective factors that allowed high-income (and high-inequality) states to experience reductions in mortality while lower-income states experienced increases. Policymakers should enact proven population health interventions, whether or not they reduce income inequality. Policies that reduce inequality may still be important for other reasons.

Introduction

Numerous studies over the past 30 years have shown a strong relationship between income inequality and population health,1 as well as other negative societal outcomes.2 3 Following the publication of Wilkinson’s finding that income inequality is negatively associated with life expectancy at the national level,4 two potential causal pathways were posited as explanations for the correlation: psychosocial, where perceptions of place in society influence outcomes, and neo-material, where inequality is a marker for access to both private resources and public infrastructure.5 The relationship between income inequality and mortality at the US state level was widely studied in the 1990s and early 2000s.6–13 Since then, the annual number of publications and citations referencing income inequality and health have steadily grown, but the correlation with mortality in US states has received less attention, as authors have investigated a broader set of health outcomes and geographies.1 14

There are several reasons to revisit this relationship: US states provide a useful level of geographic aggregation for the study of subnational patterns of income inequality and population health, as there is significant variability in social policies and inequality patterns across states, partly due to states’ policy autonomy from the federal government. Some of the strongest correlations between inequality and mortality were found among US states in the 1980 and 1990 census, even when controlling for poverty, median income, household size and/or the rate of smoking.6–9 15 One paper conducted repeated cross-sectional correlations between inequality and mortality over the 1950–2000 period and found that the correlation was only significant in the latter half of the period, peaking with the 1990 census.12 A subsequent wave of studies examined variables that could mediate or confound the relationship, notably education,10 public expenditure13 and racial composition.16 More recent work has tended to examine narrower outcomes with several studies examining pregnancy-related, neonatal, infant and COVID-19 mortality.17–20 Understanding whether correlations with broader measures of mortality have persisted, strengthened or abated over the 30 years since then seems particularly relevant, as recent trends have identified growing disparities in life expectancy and mortality at sub-national levels.21 22

Further, there is a dearth of time series analysis examining inequality and mortality, especially at the subnational level. One study considered 10-year and 20-year changes in inequality and mortality over the 1950–1990 period, finding the relationship to be positive but not statistically significant in most specifications.23 One study examined the 1960–1990 period using a fixed-effects model and found a negative but not significant relationship11; another considered only 2000–2010 and found a modest correlation between increases in inequality and decreases in life expectancy.24 The data available for US states now allows for stronger time series analysis.

This paper investigates the relationship between US state-level income inequality and mortality over a 30-year period, providing time-series analysis and using the most current data available. Specifically, we investigate whether the relationship between income inequality and mortality persisted from 1989 to 2019, whether changes in inequality are correlated with changes in mortality rates, and whether a relationship exists with a 10-year or 20-year lag (using median income and inequality measures from 1969 and 1979). We conduct a preliminary analysis of factors that could account for our observed results.

Data and methods

We replicate the analysis conducted in a study by Ross et al that found a strong correlation between income inequality and mortality using 1990 census data,9 expanding it to cover a 30-year period. We study all 50 US states from 1989 to 2019, with measures taken every 10 years, yielding 200 observations. All data used in this study is publicly available.

Dependent variables

We use all-cause mortality rates as the dependent variable. Ross et al 9 divided populations into six age-sex groups: all infants (under 1), all children and youth (1–24), working-age males and females (25–64) and elderly males and females (65+). They present working-age male mortality as their primary outcome, which is much higher than working-age female mortality and was also more highly correlated with inequality.9 13 We believe it is equally important to consider women’s outcomes, so we add working-age female mortality as a second primary outcome and we consider total mortality as a secondary outcome. We present all other age-sex groups specified by Ross et al 9 in online supplemental material and test if there are significant variations within the working-age population by examining narrower age ranges (25–34, 35–44, 45–54 and 55–64) in a sensitivity analysis.

We use publicly available data from WONDER, operated by the US Centers for Disease Control and Prevention. Mortality is measured per 100 000 population by year and is averaged over 3 years (eg, 1988–1990 for the 1989 year) to reduce the effect of outlier years, except in 2019, when we use a single year of data to avoid the impact of the COVID-19 pandemic on mortality in 2020.

Independent variables

Our focal independent variable is the Gini coefficient of household income, as calculated by the Census Bureau. We multiply this by 100 for ease of interpretation to give a range from 0 (each household in a state has the same amount of income) to 100 (one household has all of the state’s income). Ross et al 9 used median share (the share of income received by the bottom 50% of the income distribution), but we use Gini as it is more common and the Census Bureau provides a more precise estimate of Gini than we are able to derive for median share. However, we check for sensitivity to our measure of inequality by using an estimate of median share and testing alternative inequality measures: the squared coefficient of variation, Theil-Bernoulli index and polarisation index.25 We follow Ross et al in controlling for median household income at the state level.9

State-level Gini coefficients are taken from the 1990 and 2000 censuses (and the 1970 and 1980 censuses in lagged analyses), where income is reported for the prior year. Thereafter, income data are obtained from the American Community Survey (ACS). We prefer the 5-year ACS for its larger sample size and temporal smoothing, so we use the 2011 vintage, which covers 2007–2011. However, we use the 1-year ACS for 2019, following the precedent in our mortality data of avoiding the COVID-19 pandemic.

Statistical analysis

We address three different questions about the association between income inequality and working-age mortality using different methods. First, we perform separate cross-sectional regressions between state-level Gini coefficients and mortality at each of the time points to assess changes in the association between inequality and mortality over time. Second, we examine the relationship between mortality and the level of inequality 10 or 20 years earlier, since the effect of exposure to an adverse health environment may take time to manifest as mortality.22 Third, we estimate the correlation between within-state changes in income inequality and changes in mortality rates. We bound our estimates by using one model with state-fixed and year-fixed effects and another with a lagged dependent variable.26 We control for state-level median household income in all models and use lagged median income when we examine lagged Gini. Following Ross et al,9 regressions are weighted by state population, using initial population for fixed effects regressions where the weights must be constant over time; unweighted versions are presented as a sensitivity test in online supplemental tables. We used Stata V.17 for all analyses.

Patients and public involvement

Patients and the public were not involved in the design of this research.

Results

Figure 1 shows descriptive statistics for each of the main variables from 1989 to 2019. Mortality rates for working-age adults in the USA changed course dramatically. Working-age mortality rates fell almost everywhere from 1989 to 1999, but rose almost everywhere from 2009 to 2019, though at different rates across states. High mortality states were more likely to see increases in the 1999–2009 period as well. The 30-year change in mortality rates varied from −47% for working-age males in New York to +32% for working-age females in West Virginia. All-age mortality continued to decline in most states in the final decade, but at a slower rate than in the first decade, leaving all-age mortality rates lower in all states in 2019 than in 1989.

Figure 1

State 10th, 50th and 90th percentiles for key variables, unweighted (n=50).

Income inequality, as measured by Gini, rose in every state during the period. Inflation-adjusted median household income also increased in almost all states, with only Georgia (−0.2%), Connecticut (0.0%) and Michigan (1.0%) seeing increases of less than 5%. However, almost all of that gain came in the final decade, as household median incomes increased by an average of 16.7% from 2009 to 2019 (as described earlier, the 2009 sample draws from 2007 to 2011, capturing much of the impact of the Great Recession, while the 2019 sample uses single-year data, avoiding the pandemic-induced recession).

Figure 2 shows the distribution of income inequality and working-age female mortality for each state and year, along with the results of our repeated cross-sectional regressions (we present a similar graph also including working-age male mortality and all-age mortality in online supplemental figure A1). It can be seen that mortality was positively associated with inequality in 1989 but that relationship evaporated and possibly even reversed by 2019.

Figure 2

Working-age (25–64) female mortality and inequality, 1989–2019 (n=50).

Table 1 displays the results of all 12 repeated cross-sectional regressions, showing the relationship between mortality rates and the Gini coefficient at each point in time. In all regressions, median income is expressed in thousands of constant 2019 dollars and Gini is multiplied by 100; the interpretation of coefficients is thus the change in deaths per 100 000 people associated with a $1000 increase in median income or a 0.01 increase in the Gini coefficient. Thus, a state with a Gini 0.01 higher in 1989 is associated with 9.6 more deaths per 100 000 working-age females, while in 2019 it is associated with 6.7 fewer deaths.

Table 1

Cross-sectional regressions of mortality on Gini and median income (n=50)

In 1989, the Gini coefficient is associated with higher levels of mortality for all three groups, but this association attenuates and potentially reverses over time, reaching inverse correlation by 2019. Over the same period, the inverse association between median income and mortality becomes significant and increases in magnitude. We present univariate regressions of mortality on Gini and median income in online supplemental tables A1 and A2; the R2 values for Gini approached zero over the time period, but dramatically increased for median income.

Table 2 shows the results of regressions with 10-year and 20-year lagged independent variables, where the dataset has been supplemented with 1969 and 1979 values for median income and the Gini coefficient. While coefficient sizes change, sign and significance results follow the trends in table 1 closely. For brevity, we present only working-age females here and full results in online supplemental table A3.

Table 2

Regressions of female 25–64 mortality on lagged Gini and median income (n=50)

Table 3 shows the results of state-fixed and year-fixed effects regressions in panel A. As a central estimate, a hypothetical increase of 0.01 in the Gini coefficient is associated with 16.0 fewer deaths per 100 000 population for working-age females, 19.3 fewer for working-age males and 24.8 fewer for all people. We do not advocate a causal interpretation of these large correlations, which represent 5.5% (female working-age), 3.5% (male working-age) and 2.6% (all-age) of mean baseline mortality. However, they add to the evidence that the positive correlations between inequality and mortality found in the 1980s and 1990s have not been predictive of inequality-mortality correlations in the period since.

Table 3

Time series analysis of mortality on Gini coefficient and median income

The results of the lagged dependent variable model are presented in panel B. As expected, there are very high levels of autocorrelation. While the point estimates for Gini are attenuated in this specification, the direction and significance of results remain unchanged. Median income is negatively correlated with mortality and statistically significant for all three population groups.

We present several sensitivity tests in our online supplemental tables. Disaggregating the working-age population into 10-year age bands and considering other age groups, we find that results remain similar (online supplemental tables A4–A7). Each of the regressions from tables 1 and 3 were then repeated with our four alternative inequality measures used in place of Gini and the results were broadly similar in direction and significance (online supplemental tables A8–A12). Running the cross-sectional regressions without population weights yields smaller decreases in the coefficients of the Gini variable over time, with coefficients close to zero in 2019 for working-age women and total mortality and positive but non-significant for working-age men. The time-series estimates of the coefficients of the Gini variable are similar in sign and significance to the results in table 3 (online supplemental tables A13 and A14).

Finally, we consider an exploratory question, that is, one that was not planned prior to our analysis of the data: To what extent do starting conditions or changes over the course of the period predict mortality at the end of the period? As shown in our fixed effects model, states that had greater increases in their Gini over time saw greater reductions in mortality. States that initially had high levels of median household income also saw greater reductions in mortality. However, these states were largely the same: states with higher initial income tended to have greater increases in inequality over the 30-year period (see online supplemental figure A2). In contrast, income growth over the period was not significantly correlated with changes in mortality (or with initial income).

We run a simple regression with 2019 mortality as the dependent variable and 1989 mortality, 1989 Gini, 1989 median income, change in Gini and change in median income as the independent variables. Initial mortality was highly predictive of final mortality for all-age mortality (coefficient of 1.01) and working-age females (1.07), but only moderately predictive for working-age males (0.57). Higher initial Gini, initial median income and increases in Gini were all associated with lower mortality, while changes in median income were not significantly associated with changes in mortality. The full results of this regression are available in online supplemental table A15. Divergences remain: for example, Kentucky and Louisiana had similar starting incomes and changes in inequality, but Louisiana saw a decline in female working-age mortality while Kentucky saw a notable increase.

Discussion

The strong, positive correlation between income inequality and mortality disappears after 1989 and may even be reversed by 2019, using median income as the only control variable. Among the three population groups studied, working-age male mortality had the strongest positive correlation with inequality in 1989 and the weakest negative correlation in 2019. In this period, the correlation between absolute income (captured by median household income) becomes much stronger than the correlation between income inequality and mortality, consistent with the findings of Couillard et al.22 They find that major causes of death for the working-age population have become highly correlated with state-level income over a similar period.

This paper adds to the existing literature by showing that the association between inequality and mortality has evolved over the last 30 years when examined using repeated cross-sections and time-series analyses. Our results are robust to the use of narrower age ranges, different measures of inequality, and the removal of population weights.

An inherent limitation of studying geographic areas over time is that the composition of their populations is not constant: a person who dies in one state may not have resided there for most of their life. Because this is an observational study, we cannot control for all possible confounding variables,27 nor state with confidence whether an omitted variable would be a confounder or a mediator. Our analysis is also limited as it only considers all-cause mortality and does not investigate the relationship between inequality and cause-specific mortality. Our objectives were to determine if the positive relationship between inequality and mortality observed in a number of prior studies persisted over time and whether within-state changes in inequality were correlated with within-state changes in mortality. We based our model on one of those studies9 to ensure consistency in approach. Future work using integrative multivariate analysis embodied in an agent-based model could address some of the challenges of potential omitted variable bias.28

Pickett and Wilkinson hypothesised that income inequality is not only correlated with, but causes poor population health outcomes, including mortality.29 At the US state level, we find that the correlation between income inequality and mortality ceased to exist sometime between 1989 and 1999. We also find the presence of an inverse correlation between change in inequality and change in mortality, consistent with what Mellor and Milyo found in their examination of the prior 30-year period.11 These two findings provide evidence against the Wilkinson and Pickett hypothesis in US states within the time frame we studied.

Does inequality matter for mortality? It is possible that the correlations observed in US states between 1980 and 1990 were strictly spurious, but it does not explain the presence of correlations observed at the national level,30 31 nor correlations with cause-specific mortality or other health outcomes.32–34 Reflecting on the observation that inequality and mortality were correlated in USA and UK metro areas, but not in Canada, Australia or Sweden,35 Wolfson and Beall asserted that these correlations must be contingent on additional societal factors.28

There are alternative possible explanations for our observations. Mortality trends in the USA changed dramatically for the worse in this period. Wealthier states, which also had higher inequality, may have been better able to protect residents, through investments in public services or otherwise, which would appear to abate the mortality-inequality relationship. Couillard et al 22 suggest that income may be serving as a marker for a broad array of interrelated health behaviours and policies that have diverged between higher-income and lower-income states. Income inequality could be the wrong measure of inequality as it pertains to mortality. If social mobility (inequality of opportunity across generations) or wealth inequality are salient and inequality previously served as a marker for these factors, that would create the relationship we observed.1 It is also possible that there could be threshold effects, that is, a threshold of income inequality above which adverse consequences on population health begin to disappear31; additional evidence would be required to establish that as a credible explanation for our observed results.

Dunn et al 1 laid out a research agenda to study the relationship between inequality and mortality; in this paper, we addressed the question of correlation at the state level over the last 30 years. Future research should rule out heterogeneous effects by subpopulation (eg, education and race) and check whether trends are similar at the metropolitan level, expanding beyond the USA, if possible. The research agenda also proposed examining whether inequality of opportunity, also known as social mobility, holds greater associations with mortality than inequality of outcome. Some work has begun in this area with initial findings in the USA suggesting correlations with all-cause mortality and deaths of despair.36 37 Finally, additional research is needed to clarify why states with similar characteristics had different outcomes, as with the example of Kentucky and Louisiana. Prevalence of opioids may be a factor,38 39 although more recent work40 has suggested that deaths of despair only account for part of the changes in mortality among groups in the last three decades.

What do these results mean for policymakers? Should inequality be ignored? The sum of the evidence makes us sceptical of claims that income inequality is consistently associated with mortality, but also makes us cautious of disregarding it altogether. Policymakers should thus continue to focus on proven population health interventions, such as access to effective health insurance, income support and taxes on cigarettes, some of which may impact income inequality and some of which may not.

Conclusion

In this study, we examined the correlation between inequality and mortality in US states over the 1989–2019 period. We found that the strong positive correlation which existed at the start of the period had disappeared; by the end of the study period, the correlation was commonly negative, although not always statistically significant. Our time series analysis also showed that greater increases in inequality over the course of the period were correlated with decreases in mortality.

The absence of evidence for a positive correlation after 1989 and the presence of an inverse correlation between change in inequality and change in mortality represents a significant reversal from the findings of a number of other studies. It also raises questions about the conditions under which income inequality may be an important policy target for improving population health.

Data availability statement

Data are available upon reasonable request. All data used in this study are publicly available and also available from the corresponding author.

Ethics statementsPatient consent for publication

Not applicable.

Ethics approval

This study uses only aggregate data and so is exempt from ethics approval, as confirmed by the McMaster Research Ethics Board prior to commencing work on this study.

Acknowledgments

Saud Haseeb assisted with sensitivity analyses and reviewed the manuscript. We are grateful for his contributions.