Abstract
We look at the relationship between crop revenues and mortality in the Midwest from 1980–2019. For identification, we combine an exposure design with a two-way (that is, county and year) fixed-effects estimator. On average, a decrease in soybean revenue is associated with an increase in mortality. A 10 percent decrease in soybean revenues is associated with a 0.1 percent increase in the age-adjusted all-cause death rate, or about 170 more deaths throughout the Midwest in 2024. Our findings are driven by individuals 65 and older, by women, and they appear mediated by cardiovascular disease and mental health-related issues.
I. Introduction
What is the relationship between economic shocks and mortality? We examine the supply-side (that is, production) channel by testing whether negative shocks to producer revenues for economically important commodities are associated with increased mortality by studying the relationship between producer revenues and age-adjusted all-cause mortality in the US Midwest, whose population of nearly 70 million represented more than one-fifth of the US population in 2024.1 Because the US Midwest is heavily dependent on agriculture, and because the effects of shocks to the agricultural sector are more easily identifiable than those of shocks to either the industrial or services sectors, we focus on agricultural commodities—specifically, corn and soybeans, the dominant agricultural commodities in the Midwest (United States Department of Agriculture 2024).
Agricultural producers are entrepreneurs whose livelihoods depend on their crop revenues, which are the product of two variables: (i) how much of each crop producers grow and sell and (ii) at what price they do so. In the US Midwest, agricultural producers tend to produce exclusively for the market rather than for subsistence, and while they have some control over how much they produce (and thus over how much they can sell), all of them are price-takers, with no control over commodity prices.
Though there are a number of financial instruments and programs agricultural producers can use to hedge against revenue fluctuations (for example, futures and options, crop insurance, trade adjustment assistance, market facilitation payments, and others; see Tack and Yu 2021), we take these instruments and programs as given in our analysis and focus here on the residual effects of agricultural commodity revenue shocks. That is, we focus on unanticipated commodity revenue shocks that are not covered by any such financial instruments or programs, if only because the data sources that do include information on the use and availability of those instruments or programs (for example, the USDA’s Agricultural Risk Management Survey) are not representative at the county level. These residual fluctuations are conceptually similar to basis risk in the insurance literature.
Unexpected negative shocks to commodity revenues cause unanticipated decreases in profits that, should they be severe enough, may lead to health issues, both mental (for example, depression) and physical (for example, heart attacks), which can ultimately lead to death. Because the effects of shocks to commodity revenues may also be felt by others who are not themselves agricultural producers but whose livelihoods also depend on how well the agricultural sector is doing, we look at the overall population instead of focusing more narrowly on agricultural producers.
Specifically, we study the relationship between commodity revenues and mortality at the county level for the period 1980–2019 in a sample of 644 rural counties (out of 1,055 total counties) across the 12 states of the US Midwest.2 As outcome variable, we look at the age-adjusted all-cause death rate.3 As treatment, we rely on an exposure design by interacting (i) state-level prices for corn and soybeans with (ii) how much of each commodity is grown within a county. By interacting a plausibly exogenous price measure with county-level production (that is, a measure of local exposure to commodity prices), our treatment variable thus captures revenues from each, corn and soybeans, within a given county in a given year.4,5
For identification, we further rely on a two-way (that is, county and time) fixed-effects estimator, and we also estimate specifications where the production of each commodity is lagged.6 This allows one to look at how changes in commodity revenues—that is, unanticipated shocks to commodity revenue—are associated with age-adjusted all-cause mortality. Interacting commodity prices with how much of each commodity is grown in a given county allows focusing on the effects of those changes in the prices of those commodities via the production channel.7
We find a robust negative and statistically significant relationship between commodity revenues and mortality. Specifically, for the average county–year in our data, we find that a 10 percent decrease in soybean revenue is associated with an increase in the age-adjusted death rate of about 0.1 percent, or about 0.25 additional death in the average county (8.54 age-adjusted deaths per 1,000 persons × 0.001 × 28,900 persons = 0.246). Given that there are 690 rural counties in the Midwest, a back-of-the-envelope calculation combining our estimate for the period 1980–2019 with 2024 population figures would mean that a 10 percent decrease in soybean revenues would have been associated with more than 170 more deaths throughout the Midwest in 2024. Depending on the specification, corn revenues are sometimes also negatively and significantly associated with mortality, but that finding is much less robust than that for soybean revenues.
For robustness, we also estimate specifications in which we instrument revenues from corn and soybean or their total with a measure of drought occurrence and severity. Droughts—severe ones in particular—depress revenues from agricultural commodities (that is, the instrument should be relevant) while presumably only affecting mortality through revenues from commodities (that is, the exclusion restriction should hold). Instrument relevance is borne out in the data, and looking at local average treatment effect (LATE) estimates instead of average treatment effect (ATE) estimates, we find that for county–year observations where commodity revenues declined in response to drought conditions (that is, the “compliers” in our two-stage least squares [2SLS] setup), a 10 percent decrease in corn or soybean revenues is associated with an increase in the age-adjusted death rate of about 0.5 percent, or about 1.2 additional deaths in the average county (8.54 age-adjusted deaths per 1,000 persons × 0.005 × 28,900 persons = 1.234). Given that there are 690 rural counties in the Midwest, another back-of-the-envelope calculation combining our estimate for the period 1980–2019 with 2024 population figures would mean that a 10 percent decrease in soybean revenues would have been associated with more than 850 more deaths throughout the Midwest in 2024. Our findings are driven by individuals aged 65 or older and by women.
Moreover, our findings are further supported by falsification tests in which we look at the subset of urban counties in the Midwest or at whether there is a relationship between unanticipated shocks to commodity revenues and certain conditions that have no a priori reason to respond to commodity revenue shocks (for example, parasitic diseases, congenital malformations). Finally, our results appear mediated by deaths from cardiovascular disease (CVD) and by deaths from mental illness in the overall sample.
While our findings appear consistent with the well-known “deaths of despair” narrative put forth by Case and Deaton (2020), looking at treatment effect heterogeneity paints a more nuanced picture. Indeed, negative shocks to soybean revenues are associated with more deaths from CVD and alcohol-related causes among women. But positive shocks to corn revenues are associated with more deaths from mental health- and drug-related causes among men. Taking these results at face value, it seems the deaths of despair narrative holds for women, whereas a more procyclical story holds for men. That said, the size of our estimation sample decreases considerably for some causes of death because of data reporting rules regarding mortality.
There is a well-known literature on the relationship between economic conditions and mortality. Ruhm (2000) finds that most sources of mortality are procyclical, with suicide being an exception. More recently, Ruhm (2015) finds that mortality went from being procyclical to being at best weakly related to economic conditions. He thus cautions researchers against using fewer than 15 years of data to study the association between economic conditions and mortality.8 Summarizing the literature on economic conditions and mortality, Ruhm (2016) concludes that national recessions translate into improved health conditions. More recently, Hollingsworth, Ruhm, and Simon (2017) find a positive relationship between increases in the unemployment rate at the county level and deaths from opioid use, and Schwandt (2018) find that wealth shocks, measured by the interaction of stock holdings and stock market changes, negatively affect health outcomes, both physical and mental. Similarly, Pierce and Schott (2020) look at the effect of trade shocks on mortality and find that areas more exposed to trade shocks are more likely to see a rise in mortality due to drug overdoses.9
Turning specifically to the agricultural sector, the literature on adverse shocks and mortality often focuses on farmer suicides and other self-inflicted deaths. Carleton (2017), for instance, establishes a causal link between high temperatures and farmer suicides in India, and Christian et al. (2019) establish a similar link between agricultural productivity changes and farmer suicides in Indonesia.10 More recently, Proctor and Hopkins (2024) look at the relationship between stress and alcohol use in a sample of US farmers and find that higher stress is associated with binge-drinking behavior, among other findings.
More broadly in terms of time, geography, or both, Brueckner and Schwandt (2015) find that rising incomes lead to growing populations via reductions in infant mortality. Using data on the Swedish agricultural revolution that took place between 1750 and 1860, Dribe, Olsson, and Svensson (2017) find that mortality responds to harvest fluctuations, especially to harvest failures, and to fluctuations in the prices of staple crops. Fishback (2017) concludes that the spending and lending policies adopted by Franklin D. Roosevelt’s administration in the 1930s as part of the New Deal decreasesd the incidence of various types of mortality. Finally, in an analysis that is perhaps the closest in spirit to ours, Singhal and Tarp (2025) look at the impacts of coffee price volatility on the well-being of coffee producers in Vietnam and find that an increase in the volatility of the international price of coffee is associated with an increase in the psychological distress of coffee producers, a finding that appears mediated by increased alcohol consumption, a greater cognitive load, worsened expectations about the future, and a reduction in social capital.
We contribute to the literature on economic shocks and mortality by focusing on the relationship between shocks to commodity revenues and mortality in the US by focusing on shocks that affect the supply side of the economy, that is, producers as well as others whose income is correlated with that of producers. To our knowledge, no other study has looked at the relationship between commodity revenue shocks and mortality via producer income in the economy at large. This matters because while the effects of commodity revenue shocks affect the producers of those commodities directly, they also affect those who depend on those commodities indirectly. In the context we study, while changes in revenues from corn and soybean certainly affect the producers of those commodities, they also affect the livelihoods of others who depend on the financial well-being of agricultural producers, such as farm laborers, agricultural implement dealers, extension agents, and so on.
We begin in Section II by presenting our empirical framework and discussing the details of our estimation and identification strategies. Section III presents the data we use in our analysis and discusses some summary statistics. In Section IV, we start by presenting some ancillary results before moving on to our core results and the results of a number of robustness checks. Section V summarizes and offers some concluding remarks.
II. Empirical Framework
We first discuss our estimation strategy and then our identification strategy. In doing so, we also discuss the various robustness checks and falsification tests we conduct, as well as the additional analyses we run to determine which cause of death appears to drive the relationship between commodity prices and mortality.
A. Estimation Strategy
We estimate the following equation using an unbalanced panel of 644 rural counties across 12 Midwestern states for the period 1980–2019 (N = 18, 500 county-year observations) by ordinary least squares:
1where y is the age-adjusted death rate in county i in year t, pct is the price (either state-level or global, depending on the specification) of commodity c (corn or soybean, depending on the specification) in year t, qict denotes how much of crop c is grown in county i in year t, δ is a vector of county fixed effects, τ is a vector of year fixed effects, and ϵ is an error term with mean zero. The vector xict of control variables includes the average commodity revenues in neighboring counties (to avoid bias from a potential violation of the stable unit treatment value assumption, wherein revenues in neighboring counties cause a change in mortality in a given county),11 as well as lagged county-level employment (to control for shocks to employment and for migration in and out of a county) and lagged county-level income. We use lagged values of those two variables because they are likely to be affected by the treatment, so including their contemporaneous values would bias our estimate of the treatment variable of interest.
We also estimate an additional specification of Equation 1, which is such that
2Equation 2 is nearly identical to Equation 1, except that in Equation 2, we construct our treatment variable using qic,t−1 instead of qict. We do so in an effort to exogenize our treatment variable by avoiding reverse causality issues due to the fact that at our level of aggregation (that is, annual), the quantity cultivated for each crop in a given county in a given year might respond to a change in the price of that same commodity in the same year at the state and global levels.12
In our preferred specification, our main source of identification of a change in commodity revenue facing county i in year t is the variation in the interaction of the price of that commodity over time (prices do not vary between counties in a given state in a given year) and how much of that commodity is cultivated between counties in a given year and within a county over time. In other words, identification here comes from unexpected shocks to revenues from corn or soybeans between counties and over time.13
We estimate several variants of Equation 1. In addition to the two-way fixed-effects (TWFE) specification in Equation 1, to properly account for the passage of time, we also estimate versions of Equation 1 with overall, state-specific, or county-specific trends linear or quadratic trends. Moreover, in an effort to further exogenize our treatment variable, we estimate specifications where we use global commodity prices instead of state-level prices. The latter, by virtue of not being dependent on local economic conditions, are less likely than the former to suffer from bias arising from reverse causality.
It is a well-known agronomic fact that when corn and soybean are grown in rotation,14 soybeans replenish the soil nitrogen depleted by corn. Thus, since corn and soybean are complements in production, we estimate separate versions of Equation 1 and Equation 2 for corn revenues and for soybean revenues. In preliminary work, we also estimated versions of Equations 1 and 2 where we combined corn and soybean revenues into a measure of total commodity revenues, to which our results are robust.
Finally, in an effort to further exogenize our treatment variable, we also estimate versions of Equation 1 where we instrument our treatment variable, that is, ln(pct × qict), with a measure of recent drought severity in county i. We discuss the logic behind this instrumental variable in Section II.B below, but, briefly, by depressing the yields of corn and soybean, the occurrence of drought and how severe it is will reduce corn and soybean revenues. Assuming that droughts only affect mortality through corn and soybean revenues, it is possible to estimate the local average treatment effect of commodity revenue shocks on mortality.
Given the observational nature of our data and our TWFE research design, we follow the recommendations in Abadie et al. (2023) and MacKinnon, Nielsen, and Webb (2023) and cluster standard errors at the county level throughout. For robustness, we also present population-weighted results for our core specifications.
Because we regress a logarithm on a logarithm, our coefficient of interest is an elasticity, which allows us to easily quantify the economic significance of our results. For all estimates of β obtained from Equation 1 and the various specifications just discussed, we test the null hypothesis that H0: β = 0 versus the alternative hypothesis that HA: β ≠ 0. Rejecting the null hypothesis in favor of the alternative hypothesis and finding that β̂ is negative and statistically significant would thus lend support to the hypothesis that decreases in commodity prices are associated with increases in mortality.
B. Identification Strategy
Our research design consists of: (i) an exposure design similar to that in Dube and Vargas (2013), who study the effect of commodity price shocks on conflict in Colombia, and (ii) a TWFE estimator.
An issue with the TWFE estimator is that when there is heterogeneity in the treatment effect of interest, the TWFE may not return an average of county-level estimated treatment effects unless some units are left untreated (Sun and Shapiro 2022). In our data, no county–year observation has a value of corn or soybean revenues equal to zero, so our TWFE results are presented under the caveat that they are only valid in the absence of treatment effect heterogeneity. In other words, we assume that treatment effects are homogeneous.
As discussed in the previous section, we also present the results of 2SLS specifications in which we rely on measures of drought severity (that is, dummies for whether a given county has experienced an exceptional drought, extreme drought, severe drought, or moderate drought in a given year) as instruments for corn and soybean revenues in a given county in a given year. Here, the relevance of the instrumental variables (IVs) stems from the fact that a drought directly affects yields negatively, which translates into decreased revenues from the affected commodities.
While the relevance of our IVs is testable, whether those same variables meet the exclusion restriction—that is, whether the occurrence or duration of drought affects mortality only through commodity prices—is not. We argue that once county-specific time-invariant heterogeneity and year-specific county-invariant heterogeneity are both accounted for, the only way drought affects mortality is through corn and soybean revenues. In other words, we assume that when both (i) county-specific time-invariant characteristics and (ii) year-specific county-invariant characteristics are held constant, drought conditions do not affect mortality through, say, respiratory illnesses.15 While this is a strong assumption, we view our IV results as complementary to those of our preferred core TWFE specifications.
III. Data and Summary Statistics
The data we use for our analysis cover the period 1980–2019 and come from various sources (Bellemare 2025). The US Centers for Disease Control and Prevention’s (CDC) National Center for Health Statistics (NCHS) provide mortality data at the county level, including the Compressed Mortality File (1980–1998) and the Multiple Cause of Death File (1999–2019).16 We use age-adjusted death rates for all ages and causes in our core analysis, but we look at specific causes of death (for example, deaths from cardiovascular disease and deaths related to mental health) when assessing the mechanisms whereby our core results might be caused. We also use death from certain infections and parasitic diseases and from congenital malformations, deformations, and chromosomal abnormalities for the purpose of implementing a placebo outcome test (Eggers, Tuñón, and Dafoe 2024). An earlier version of this paper also used all-cause death rates in addition to all-cause age-adjusted death rates, with nearly identical findings.
Data on corn and soybeans (that is, price and production) in the US Midwest are from the US Department of Agriculture’s National Agricultural Statistics Service. Annual commodity prices are only available at the state level, but annual commodity production is available at the county level. International primary commodity price data are from the International Monetary Fund’s (IMF) Primary Commodity Price System. For our primary analysis, we use the commodity values for corn and soybeans by multiplying production data with state-level prices. We use the GDP deflator from the World Bank to express nominal prices in real (that is, 2015 US dollar) prices.
For droughts, we use the monthly Palmer Drought Severity Index (PDSI) at the county level provided by the Cooperative Institute for Climate and Satellites–North Carolina. The PDSI is a standardized index that measures relative dryness, which is estimated via temperature and precipitation data. The PDSI ranges from −10 to +10, from dry to wet. We calculate the average monthly data to determine the annual drought. The drought classification is based on the US Drought Monitor.
To define rural versus urban counties, we use 1990, 2006, and 2013 data from the Urban–Rural Classification Scheme provided by the CDC’s NCHS. We define urban counties as counties classified as urban in any of 1990, 2006, or 2013. Every other county in our data is defined as rural. Data on county-level annual employment is from the Quarterly Census of Employment and Wages (QCEW) by the US Bureau of Labor Statistics (BLS). Data on county-level annual personal income is from the US Bureau of Economic Analysis (BEA).
Online Appendix Table A1 presents descriptive statistics for the 644 counties we retain for analysis for the period 1980–2019 for our outcome of interest (that is, age-adjusted death rates), for our variables of interest (that is, commodity revenues for corn and soybeans), for our IVs (that is, drought severity), and for our control variables. Online Appendix Table A2 presents descriptive statistics for the variables we use for our mediation analyses, as well as for our falsification tests (that is, tests in which we use a “fake” outcome to test whether our core results are spurious). Online Appendix Tables A3 and A4 present detailed descriptions of the variables we use in our analysis.
Before presenting regression results, we present some visual representations of our data. Figures 1 and 2 show the spatial distribution of within-county corn and soybean average revenues over time. Unsurprisingly, given that corn and soybeans are often complements in production, there appears to be a spatial correlation between corn and soybean revenues across counties.
County Corn Revenues, 1980–2019
County Soybean Revenues, 1980–2019
Figures 3 and 4 plot the relationship between mortality and commodity revenues, with the difference being that the variables in Figure 4 are population-weighted, but the variables in Figure 3 are not. In both figures, for both corn and soybeans, there seems to be an unconditional negative correlation between commodity revenues and age-adjusted death rates.
Mortality Rate and Commodity Revenue Based on State Prices and County Production, 1980–2019
Population-Weighted Mortality Rate and Commodity Revenue Based on State Prices and County Production, 1980–2019
IV. Results and Discussion
Turning to our empirical results, we first present and discuss our core empirical results and the results of robustness checks. Having established that our core results are robust, we then discuss the mechanisms whereby our results operate, as well as the results of falsification tests.
A. Ancillary Results
Before discussing our core results, we briefly discuss the results of ancillary regressions that help put our main results in context.
Because the price and quantity of a commodity are likely to be related to each other, it is natural to ask what is the relationship between the quantity of corn or soybeans cultivated in a given county in a given year and the price of corn or soybeans in the same year for the state in which that county is located. Online Appendix Table A5 shows the relationship between the price and quantity of corn (Column 1) and soybeans (Column 2) conditional on county and year fixed effects. Because we look at the logarithm of both prices and quantities, the estimated coefficients are elasticities, and they suggest that a 1 percent increase in the state-level price of a given commodity in a given year is associated on average with a 2.5 percent (2.9 percent) decrease in the quantity of corn (soybeans) produced in the average county in the same year. In other words, conditional on county and year fixed effects, there appears to be an inverse relationship between the price and quantity cultivated for each commodity.
It is also natural to ask about the relationship between revenues from corn and soybeans in a given state in a given year and either income or farm income in a given county in the same year. Online Appendix Table A6 shows the relationship between farm income and revenues from corn (Column 1) and soybeans (Column 2), as well as the relationship between income and revenues from corn (Column 3) and soybeans (Column 4), conditional on county and year fixed effects. Unsurprisingly, the association between farm income and the revenue from each commodity is positive and statistically significant, although the magnitude of that association is almost five times as high for corn (Column 1) as it is for soybeans (Column 2). Here, too, the estimated coefficients are elasticities. What is perhaps more surprising is that while corn revenues are not statistically significantly associated with income overall (Column 3), the relationship between soybean revenues and income overall (Column 4) is positive and statistically significant, and its magnitude exceeds that of the relationship between soybean revenues and farm income (Column 2). Given that our research design is not aimed at causally identifying the relationships presented in this table, the coefficients in Online Appendix Table A6 may be biased.
Similarly, Online Appendix Table A7 looks at the relationship between commodity revenues and employment in agriculture and employment overall. While the estimation samples are not the same between the two measures of employment—the number of observations for employment in agriculture is smaller than for employment overall because the Bureau of Labor Statistics suppresses data to avoid identifying specific employers—it is notable that employment in agriculture does not seem to be significantly associated with either corn and soybean revenues conditional on county and year fixed effects, but that total employment is negatively and significantly associated with corn revenues (Column 3) but positively and significantly associated with soybean revenues (Column 4). Again, given our research design, these findings should be treated with caution, but this latter finding—the different signs in Columns 3 and 4 of Online Appendix Table A7—may help explain some of our findings below.
The results in Online Appendix Tables A6 and A7 help elucidate whether our results are driven by income effects for producers or labor-market spillovers for the population in general. The results in Online Appendix Table A6 show that commodity revenues are positively associated with both farm and overall income in a given county in a given year. The results in Online Appendix Table A7 show that shocks to either corn or soybean revenues are not significantly associated with employment in agriculture, but that corn revenue shocks are negatively associated with overall employment, while soybean revenues are positively associated with overall employment. Thus, it seems that there are both income effects for producers and labor market spillovers at play, with the latter exhibiting both pro- and countercyclical tendencies.
B. Core Results
Turning to our core results, Table 1 shows the results of a regression of the logarithm of the age-adjusted death rate on the logarithm of corn (Columns 1–4) or soybean (Columns 5–8) revenues based on state-level prices conditional on county and year fixed effects, on lagged employment and income, as well as on the contemporaneous average of the treatment variable in neighboring counties. In Columns 1, 2, 5, and 6, we show the results of regressions that omit control variables (that is, lagged employment in a given county in a given year and lagged income in the same county in the same year). In Columns 2, 4, 6, and 8 we weight each county–year observation by the size of its population.
Two-Way Fixed-Effects Estimates for Correlation Between Rural Mortality and Commodity Revenue, 1980–2019
The results in Table 1 suggest that while there is no statistically significant relationship between corn revenues and mortality (Columns 1–4), there is a persistent negative relationship between soybean revenues and mortality (Columns 5–8) that is robust to omitting or including control variables and to weighting observations by population or not. Specifically, for a 10 percent decrease in soybean revenues, there is an associated increase in mortality of 0.1 percent in the average rural county in the US Midwest for the period 1980–2019. The same cannot be said for the relationship between corn revenues and mortality. Although the sign of the coefficient of interest is the same in Columns 1–4 as in Columns 5–8, the magnitude of the estimated coefficients is roughly one order of magnitude smaller for corn than it is for soybeans. We speculate below about the reasons why that is.
The results in Table 1 are robust to using global commodity prices instead of state-level prices (Online Appendix Table A8), to using linear or quadratic trends that are either for the overall sample or state- or county-specific (Online Appendix Tables A9 and A10, respectively, for corn and soybeans). Online Appendix Table A8 also shows commodity revenue based on state-level prices and lagged production, and our core results are also robust to doing so. Breaking down the results in Table 1 by decade by interacting the logarithm of revenues for soybean with indicator variables for the 1990s, the 2000s, and the 2010s (that is, leaving the 1980s as the reference category), the negative relationship between soybean revenues and mortality in Table 1 is present for every decade (Online Appendix Table A11, Column 2), as well as overall, and it seems that the negative relationship between commodity revenues and age-adjusted all-cause mortality increased monotonically in magnitude from the 1990s to the 2000s and from the 2000s to the 2010s for both corn and soybeans. Interestingly, there is a similar relationship in Online Appendix Table A11 between corn revenues and mortality (Column 1), and the estimated coefficient for each of the 1990s, 2000s, and 2010s is larger in magnitude for corn revenues than it is for soybean revenues, but this does not translate into overall significance in Table 1.
Table 2 shows the results of the specification in Table 1 broken down by age group (that is, ages 0–14, 15–64, and 65+). The sample size is smaller for the age 0–14 category, because the CDC suppresses low-count mortality data to prevent identification of individuals in that age category. While the association between corn or soybean revenues and mortality in the age 0–14 and age 15–64 categories is not statistically significant, the associations between both corn revenues and mortality as well as soybean revenues and mortality in the age 65+ category are both negative and statistically significant. The results in Table 2 thus suggest that our core results are driven by the age 65+ category, and while this translates into a similar result for soybean in the full sample, it does not for corn revenues.
Two-Way Fixed-Effects Estimates of Rural Mortality and Commodity Revenue by Age, 1980–2019
Table 3 shows the results of the specification in Table 1 broken down by gender. While the association between corn revenues and mortality (Columns 1 and 3) is not statistically significant for either gender, the association between soybean revenue and mortality is negative and statistically significant for women (Column 2). The results in Table 3 thus suggest that our core results are driven by women.
Two-Way Fixed-Effects Estimates of Rural Mortality and Commodity Revenue by Gender, 1980–2019
Table 4 shows results for our 2SLS specifications, where we instrument corn revenues with the occurrence of an extreme or exceptional drought (Column 1) or with a moderate, severe, extreme, or exceptional drought (Column 3). In both cases (Columns 2 and 4), this causes the association between corn revenues and mortality to become negative and statistically significant. Similarly, we instrument soybean revenues with the occurrence of an extreme or exceptional drought (Column 5) or with a moderate, severe, extreme, or exceptional drought (Column 7). Only in the former case (Column 6) does this cause the negative relationship between soybean revenue and mortality in Table 1 to increase in magnitude. While the results in previous tables were average treatment effects if one believes our research design, the results in Table 4 are local average treatment effects if one believes that our IVs are valid. That is, they represent the effect of a change in commodity revenue on mortality in those counties where commodity revenues changed in response to a drought. The F-statistics for tests of weak instruments exceed the usual threshold of about 10–13 for all four specifications.
Instrumental Variables Estimates for the Impact of Commodity Revenue on Rural Mortality, 1980–2019, Using Drought Level as Instrumental Variable
Table 5 conducts something akin to a mediation analysis by keeping the specification in Table 1 to look at whether our core results are driven by specific causes of death, that is, cardiovascular disease (Columns 1–4) and from mental health issues (Columns 5–8). The results in Table 5 suggest that the negative association between revenues from soybeans and mortality is driven by both cardiovascular disease (Columns 3 and 4) and mental health issues (Columns 7 and 8). It is noteworthy that while our core results suggest that there is no relationship between corn revenues and mortality, the results in Table 5 suggest a statistically significant negative relationship that is large in magnitude between corn revenues and deaths from mental health issues. Here, however, sample sizes are smaller than in our core analysis because the CDC suppresses low count mortality data to protect privacy.
Two-Way Fixed-Effects Estimates of Rural Mortality and Commodity Revenue by Cause, 1980–2019
While Table 5 looks at deaths from cardiovascular disease and mental health issues, Online Appendix Table A12 looks at deaths from self-harm (Columns 1 and 2), alcohol (Columns 3 and 4), and drugs (Columns 5 and 6). While those causes of death do not seem related to commodity revenues at first glance, Online Appendix Tables A13 and A14 break down results for specific causes of death by gender. Online Appendix Table A13 shows that our core result for soybean operates through cardiovascular disease among women (Column 3), but Online Appendix Table A14 also suggests that corn revenues may also be related to mortality among women through alcohol-related deaths (Column 3). Interestingly, both Online Appendix Tables A13 and A14 show that in some cases, an increase—not a decrease—in corn revenues is associated with an increase in mortality—in Online Appendix Table A13 through mental health-related deaths and in Online Appendix Table A14 through drug-related deaths. We discuss these results below, when putting our results in the context of the “deaths of despair” narrative (Case and Deaton 2020).
Finally, we conduct two falsification tests to support the notion that our results are not spurious and driven by unobserved confounders. In both Online Appendix Tables A15 and A16, the idea is to find a lack of statistical significance in an effort to ensure that our core results are not spurious. In Online Appendix Table A15, we look at the relationship between corn revenues (Columns 1 and 2) and soybean revenues (Columns 3 and 4) and all-cause age-adjusted mortality, both without (Columns 1 and 3) and with controls (Columns 2 and 4). The only significant coefficient here (Column 1) shows the opposite of what our core results show, as it suggests a positive relationship between corn revenues and mortality, but only in the absence of controls. In Online Appendix Table A16, we look at an outcome that should not a priori respond to changes in corn or soybean revenues, namely, deaths from certain infections and parasitic diseases and congenital malformations, deformations, and from chromosomal abnormalities. Here, we see no significance anywhere, which is what we would expect.
In terms of economic significance, our core results suggest that for the average county–year in our data, a 10 percent decrease in soybean revenue is associated with an increase in the age-adjusted death rate of about 0.1 percent, or about 0.25 additional death in the average county (8.54 age-adjusted deaths per 1,000 persons × 0.001 × 28,900 persons in the average rural county in 2024 = 0.246). Given that there were 690 counties classified as rural in the Midwest in 2024, this means that a 10 percent decrease in soybean revenues would have been associated with about 170 more deaths throughout the Midwest in 2024. Taking our 2SLS results for soybean at face value suggests that the local average treatment effect is such that for a 10 percent decrease in soybean revenues in 2024, there would have been an additional 850 deaths throughout the Midwest in 2024.
To summarize, our results show a robust negative relationship between soybean revenue and mortality. While our results also occasionally show a negative relationship between corn revenue and mortality, that relationship is not robust.
If it is indeed the case that there is a negative relationship between soybean revenue and mortality but not between corn revenue and mortality, this prompts the question: Why soybeans, but not corn? One possible answer emerges in Online Appendix Table A7, which shows that total employment in a county is positively associated with soybean revenues, but negatively associated with corn revenues. This suggests that negative shocks to soybean revenues may lead to increased mortality via job loss but that the opposite holds for corn revenues. Another possible answer is that soybean is a relatively new crop in the US Midwest—farmers began planting soybeans only when the wheat frontier began moving north in response to a changing climate—and producers get to choose whether and how much insurance coverage to buy for each crop. If they are less familiar with soybeans, their soybean insurance purchase decisions may be suboptimal relative to their corn insurance purchase decisions, which would lead to their being exposed more to fluctuations in soybean revenues than in corn revenues. Both of these explanations are speculative and thus deserving of further research.
Finally, there remains the question of whether mortality from fluctuations in commodity revenues is procyclical, as in Stevens et al. (2015) and Ruhm (2015), or countercyclical, as in Case and Deaton (2020). If there is no relationship between corn revenue and mortality and there is only a negative relationship between soybean revenue and mortality, then our overall results contribute to the countercyclical “deaths of despair” narrative put forth by Case and Deaton (2020). But if there is indeed a relationship between corn revenues and mortality, then the results in Online Appendix Tables A13 and A14 suggest that the likely effect of fluctuations in commodity revenues on mortality are gendered and both procyclical (via the positive relationship between corn revenues and deaths from drugs and mental health issues among men) and countercyclical (via the negative relationship between soybean revenues and deaths from CVD and alcohol among women), which paints a considerably more nuanced picture than the traditional “deaths of despair” narrative.
V. Summary and Concluding Remarks
We have looked at the relationship between revenues from agricultural commodities and mortality in the rural Midwest for the period 1980–2019. To do so, we have relied on a combination of: (i) an exposure design interacting state-level commodity prices with county-level quantities cultivated for each commodity and (ii) a two-way (that is, county and year) fixed-effects estimator to look at the relationship between corn or soybean revenues on the one hand and age-adjusted all-cause death rates on the other hand. Because the TWFE may be biased in the presence of treatment effect heterogeneity, we have also supplemented our core TWFE results with a number of robustness checks involving (i) global instead of state-level prices, (ii) lagged instead of contemporaneous production levels, (iii) full-sample, state-specific, or county-specific linear or quadratic trends, and (iv) measures of drought severity as IVs. This is consistent with the hypothesis that economic shocks can cause mortality and, more specifically, with the findings in Carleton (2017); Christian, Hensel, and Roth (2019); and Singhal and Tarp (2025).
We find a robust, statistically significant, and negative relationship between soybean revenues and mortality. That is, as soybean revenues decrease, mortality increases. While this relationship holds for all-cause age-adjusted mortality, we find that it is mediated by deaths from cardiovascular disease and mental health issues. Digging deeper, we also find that our finding is driven by individuals aged 65 or older and by women. Though we find a negative relationship between corn revenues and mortality, that relationship is not as robust as that for soybean revenues.
We speculate that this differential effect between corn and soybean revenues may be due to either: (i) a procyclical relationship between corn revenues and employment but a countercyclical relationship between soybean revenue and employment or (ii) the different insurance purchases and coverage choices made by farmers when it comes to soybeans, a crop that is considerably newer to the US Midwest than corn. If it is indeed true that soybean revenue shocks cause mortality but corn revenues do not, this warrants further research into the mechanisms behind this finding.
If both shocks to corn and soybean revenues cause mortality, then our findings become much more nuanced. Indeed, we find that while decreased soybean revenues seem to translate into increases in mortality for women via cardiovascular disease and alcohol use, increases in corn revenues seem to translate into increases in mortality for men via mental health and drug use. Thus, for the reader interested in knowing whether mortality from agricultural commodity revenue fluctuations is procyclical, as posited by Stevens et al. (2015) and Ruhm (2015), or countercyclical, as famously hypothesized by Case and Deaton (2020), the answer is that it depends on which source of mortality one chooses to look at—as well as on mortality for whom.
Our work is limited both in terms of internal validity and of external validity. On the internal validity front, while we do find a robust negative relationship between soybean prices and mortality, our work relies on observational data, and our results thus fall short of the gold standard of experimental evidence and should be treated as such. Moreover, an important recent literature has questioned the usefulness of TWFE designs in the presence of treatment effect heterogeneity. Against that, we present the results of a battery of robustness checks.17 On the external validity front, our results are only valid for corn and soybean in the US Midwest for the period 1980–2019, and it is unclear whether they apply to other crops and other contexts.
The limitations above notwithstanding, if one were to grant internal validity to our findings, those findings would have clear implications for policy. Given our finding that soybean price decreases seem to translate into increased mortality from cardiovascular diseases and from mental health-related causes, one obvious policy implication would be to invest in better prevention, early detection, and treatment of those conditions in rural areas. Given the longer-than-average distances between individuals and their healthcare providers in rural areas, this may mean encouraging more rural health infrastructure or the deployment of mobile health care providers in rural areas.
It might also be possible to invest in additional income-support programs for farmers in the Midwest. While nothing in our findings suggests that the adverse effects of commodity revenue shocks are limited to farmers, and, given data limitations, we cannot identify which groups of individuals are the ones more likely to suffer from increased mortality because of soybean revenue shocks, the presumed mechanism here is that depressed farm incomes lead to mortality, so if those incomes were less likely to suffer unexpected decreases, there would likely be less mortality as a result. Less obviously, given that our results take into account the myriad of financial instruments and programs agricultural producers already have access to, our results may also offer a cautionary tale against curtailing access to those financial instruments or reducing income-support programs for agricultural producers.
The limitations above also have implications for future research. While it seems unlikely that one can improve on the internal validity of our results in the absence of data at a level below that of the county or of an experimental design, it is certainly possible to improve in terms of external validity by conducting similar studies in other contexts, such as for shocks to revenues from other commodities or for shocks to the manufacturing and services sectors.
Acknowledgments
The authors thank three reviewers as well as the participants of the January 2025 NBER “Determinants of Mortality” conference and Ellen Meara for useful comments and suggestions. They also thank Mook Bangalore, Tamma Carleton, Kathryn Grace, Jason Fletcher, Alex Hollingsworth, Terry Hurley, Jason Kerwin, Matthew MacLachlan, Raahil Madhok, Conner Mullally, and seminar participants at Cornell, Florida, Georgia, OARES, Penn State, and Southern Methodist for comments that helped improve earlier versions of the manuscript. Bellemare is grateful to NIFA for funding this work through grant MIN-14-184. All errors are those of the authors. The data used in this article are available online (https://doi.org/10.7910/DVN/FNKSGI).
Footnotes
Contributions were originally presented at the Determinants of Mortality conference hosted by the National Bureau of Economic Research with financial support from the National Institute on Aging (P30AG012810).
↵1. In an earlier version of this paper, we used the same empirical setup but purported to look at the relationship between commodity price shocks and mortality. As it turns out, quantities cultivated seem to respond to prices (Online Appendix Table A5), so we cannot isolate the relationship between commodity price shocks and mortality—only the relationship between commodity revenue shocks and mortality.
↵2. The counties we retain for analysis are those for which we have data on both our treatment and outcome variables for the period 1980–2019. The 12 states in the Midwest are Illinois, Indiana, Iowa, Kansas, Michigan, Minnesota, Missouri, Nebraska, North Dakota, Ohio, South Dakota, and Wisconsin.
↵3. The age-adjusted death rate is the death rate adjusted ex ante for the age distribution in the population of interest.
↵4. While it is possible that commodity prices at the state level are not exogenous to county-level mortality, our results are robust to using global commodity prices to further exogenize our treatment variable.
↵5. To assess the robustness of our findings and avoid the bad control problem—each price being dependent on the other given complementarities between corn and soybean—we look separately at the relationship between mortality and corn revenues and the relationship between mortality and soybean revenues. In an earlier version of this paper, we also combined corn and soybean revenues into a single measure of commodity revenues. Results are robust to doing so.
↵6. Some colleagues suggested interacting each commodity’s price with the average or initial amount cultivated of that commodity in a county. Since state-level (global) prices do not vary across counties within a state (between states) in a given year, however, we cannot do as suggested as that would drastically reduce the amount of identifying variation.
↵7. While some data sets allow looking directly at farm revenue, using those data sets would be second-best. The US Census of Agriculture has detailed longitudinal information on farms, but it is conducted only every five years, which would seriously reduce the statistical power of our analysis. Similarly, the US Department of Agriculture’s Agricultural Risk Management Survey (ARMS) has detailed financial information on the farms it samples, but it is not a longitudinal data set. Moreover, neither the Census of Agriculture nor the ARMS are publicly available, which limits the transparency and replicability of any empirical work done using either source. Similarly, while the National Longitudinal Mortality Study has information on mortality and occupation from 1970–2010, the limitation of the public-use data is that those data can only identify residency at the state level and not at the county level.
↵8. In our analysis, we have more than twice as many years of data.
↵9. A related strand of literature is the literature on how changes in international trade policy and technology have affected labor markets in the US (Autor, Levy, and Murnane 2003; Autor, Dorn, and Hanson 2013, 2019; Autor et al. 2014, 2020).
↵10. While Carriere, Marshall, and Binkley (2019) also look at farmer suicides, their analysis does not have a proper research design, which makes it more descriptive as it cannot establish a causal relationship.
↵11. This allows ruling out between-county contemporaneous spillovers, or that changes to farm revenues in neighboring county j cause changes in the death rate in county i in year t, but it does not allow ruling within- or between-county spillovers over time.
↵12. Some have suggested interacting pct with the average or initial value of qict in each county, but the fact that we rely on state-level and global prices for our analysis means that doing so would leave too little identifying variation. If we interacted initial or average quantities cultivated in each county with state- or global-level prices, the only variation within each county would stem from changes in state-level prices over time or in global-level prices over time.
↵13. In a previous version of this paper, we also presented results for the estimators suggested or developed by Millimet and Bellemare (2025), finding our results to be robust to them. Given the many results in this paper and the fact that the estimators suggested by Millimet and Bellemare (2025) rely on differencing and we now make extensive use of lagged variables for identification, we omit those additional robustness checks from this version both for brevity and because it is not clear what differencing estimators identify when dealing with lagged variables.
↵14. Corn and soybean can also be grown together, but this is less common.
↵15. In an earlier version of this manuscript, we also ran the 2SLS specifications using death rates that omitted respiratory illnesses as our outcome of interest, and the results were robust.
↵16. The mortality data from CDC consists of county-level mortality and population files. Deaths of nonresident aliens and fetal deaths are not included.
↵17. An earlier version of this paper included many more robustness checks which are omitted for brevity.
- Received March 2025.
- Accepted October 2025.
This open access article is distributed under the terms of the CC-BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0) and is freely available online at: https://jhr.uwpress.org.
