Deaths of Despair Are Two Different Crises

In 2015, Anne Case and Angus Deaton gave us a framework for understanding this kind of suffering. White Americans without college degrees were dying younger, driven by a simultaneous rise in suicide, drug overdose, and alcoholic liver disease. They called these "deaths of despair" — a unified syndrome caused by economic dislocation, status loss, and the erosion of the social structures that gave working-class life its meaning.

The framing was clarifying. If three death types shared a common cause, they should respond to common interventions. Economic renewal, social investment, restored dignity of work. It shaped a decade of policy thinking. It shaped funding decisions, program design, and the political vocabulary of rural suffering.

The framing also made a testable prediction. If suicide, overdose, and alcoholic liver disease share a root cause, they should cluster together geographically. Counties with high suicide rates should also have high overdose rates. The unhappiest places should look the same across all three axes.

I ran that test using county-level mortality data from 2018 to 2024. The prediction fails.

The result: they don't co-cluster

The standard tool for measuring geographic co-clustering between two variables is bivariate Moran's I — it asks whether high values of X tend to be spatially proximate to high values of Y. A score near 1 means tight co-clustering; near 0 means spatial independence; negative means they anti-cluster. (The full analysis and replication code are on GitHub.)

If the unified mechanism is real — if economic dislocation causes people to drink, overdose, and die by suicide — then the most economically dislocated counties should be elevated on all three axes. Testing geographic co-clustering is testing that prediction directly.

Suicide × drug overdose: BV I = 0.016, p = 0.15. Not statistically significant. Not even close.

To calibrate what this means: BV I(suicide × alcoholic liver disease) = 0.341. The observed suicide–overdose score is 4.7% of that benchmark — not a weak version of the same phenomenon, but effectively a different category.

The p-value framing understates how far this is from what the unified thesis predicts. A Bayesian frame is more informative: rather than testing against a null of zero co-clustering, we can test against what we actually observed when the deaths-of-despair thesis was being developed. Under the conservative 1999–2005 prior (BV I = 0.177, the pre-epidemic baseline when OxyContin was beginning to spread), the current data is 236× more consistent with geographic independence than with that level of co-clustering. Under the 2006–2012 peak prior (BV I = 0.277, when prescription opioids created genuine geographic overlap), it is 520,000×. A Bayes factor above 100 is conventionally described as decisive evidence. These are not close calls.

The small number of counties elevated on both axes (pink) are mostly border counties at the geographic edges of each regime — not evidence of a shared geographic core.

The two crises occupy almost entirely different parts of the country. That's not a statistical nuance — it's visible at a glance. Purple Mountain West; teal Appalachia and New England. The map doesn't look like a unified syndrome. It looks like two different problems that got a shared name.

Alcoholic liver disease breaks the taxonomy

If suicide and overdose don't co-cluster, where does alcoholic liver disease fall? The Case and Deaton framing groups all three together. You'd expect liver disease to sit somewhere in the middle — correlated with both.

It doesn't. BV I(suicide × alcoholic liver disease) = 0.341, p = 0.001. BV I(overdose × alcoholic liver disease) = −0.003, p = 0.435.Alcoholic liver disease is coded as K70 in ICD-10. The overdose measure uses codes X40–X44 (accidental poisoning). The analysis was also run with the combined measure (X40–X44 + Y10–Y14 undetermined intent). The independence finding is slightly stronger with X40–X44 alone. Liver disease goes with suicide, not overdose, across all four time periods tested.The four periods are 1999–2005, 2006–2012, 2013–2019, and 2018–2024. The temporal detail is in the smoking gun section below.

That result reorganizes the taxonomy. The "three deaths of despair" split 2-1, not 1-2. What Case and Deaton called a unified syndrome is more accurately described as two distinct regimes:

Regime A — Chronic self-destruction: suicide + alcoholic liver disease. Mountain West, Great Plains, tribal-land counties. Driven by elevation, gun access, isolation.

Regime B — Supply-chain poisoning: drug overdose. Appalachia, Rust Belt, pockets of New England. Driven by pharmaceutical targeting geography and the subsequent fentanyl supply chain.

The map also shows Regime C (elevated on all three axes) and Regime D (low on all three axes); neither clusters geographically in a way that changes the A/B argument, and both are small enough to be consistent with noise.

Portrait of Regime A

Regime A counties are older (median age 42.4 vs. 40.6), whiter (84.7% vs. 78.4%), and more likely to be high-altitude. Three variables — elevation, gun access, and racial composition — each carry independent signal. None of these findings are artifactual: the elevation and firearm relationships hold on partial correlations controlling for age composition. (These are crude rates, not age-adjusted — Regime A counties skew older.)

Elevation is the headline predictor (altitude suppresses serotonin synthesis — a biological mechanism, not a metaphor for mountain isolation). Spearman ρ between elevation and suicide rate = 0.384; between elevation and alcoholic liver disease = 0.373. Elevation and overdose: ρ = −0.037, not significant. The relationship is strong up to roughly 2,000 meters and flattens above it — resort towns and ski counties dilute the signal at extreme altitudes. This kink is statistically significant for both suicide (F = 36, p < 0.001) and alcoholic liver disease (F = 24.5, p < 0.001), and absent entirely for overdose (F = 0.00, p = 0.95).

Gun access amplifies lethality without explaining the geography directly. The firearm suicide fractionThe firearm suicide fraction — suicides by firearm as a share of all suicides — is a validated county-level proxy for gun ownership (Cook & Ludwig 2000). correlates with suicide rate at ρ = 0.291 (p < 0.001) and with alcoholic liver disease at ρ = 0.138 (p < 0.001). Its correlation with overdose: ρ = 0.006, not significant. Regime A counties have a median firearm fraction of 0.656 versus Regime B's 0.605 (p < 0.001). More firearms doesn't create more overdoses; it makes existing self-destructive impulses more lethal.

One sub-population deserves explicit attention before going further. Regime A contains two distinct groups: white rural Mountain West counties where altitude, isolation, and gun access are the primary mechanisms, and tribal-land counties where structural dispossession, healthcare deserts, and historical disruption are more plausible drivers. The same regime classification covers both. The same interventions will not. Interventions in tribal-land counties need to engage tribal health systems and address historical trauma rather than treating this as a variant of the altitude-isolation-gun-access mechanism. American Indian and Alaska Native (AIAN) counties in Regime A number 48 in this dataset — directionally consistent with the regime classification, but underpowered for separate analysis.

A spatial regression with elevation, firearm fraction, % AIAN, education, and income explains 42% of suicide variance (R² = 0.42). Residual spatial autocorrelation remains at Moran's I = 0.24 on the residuals — unmeasured isolation effects and social network structure still carry information the observable predictors miss.

Portrait of Regime B

Regime B counties have slightly higher median incomes ($64,534 vs. $60,867 for Regime A), are slightly less white (78.4%), and less rural. They share one striking feature: their current overdose geography was largely set 15 years ago.

The 2006–2012 overdose rate — the OxyContin era — predicts the 2018–2024 overdose rate with a z-statistic of 21.2 (p < 0.001) in the spatial regression. That single variable is more predictive than any demographic characteristic. The fentanyl supply chain didn't create new geography; it concentrated into the territories where prescription opioids had already established distribution infrastructure and user populations. A county that was heavily targeted by pharmaceutical distributors in 2009 is, with striking regularity, still a high-overdose county today.

The spatial regression for overdose reaches R² = 0.41 — comparable to the suicide model — but residual spatial autocorrelation is much higher: Moran's I = 0.54 on the residuals, versus 0.24 for the suicide model. This gap matters. Even after accounting for every observable county characteristic, overdose geography retains strong clustering that demographic or economic variables can't explain. The supply-chain structure is present in the data even when you subtract everything else. Supply-chain network topology — which counties are hubs, which are downstream — doesn't reduce to observable demographic or economic variables. The demographic profile of a county tells you far less about its overdose rate than its position in a distribution network does.

The network is the variable.

The temporal smoking gun

Case and Deaton weren't working from bad data. They published in 2015 using data through approximately 2013, and the co-clustering they implied was real — for a window.

BV I(suicide × alcoholic liver disease) held stable throughout all four periods, ranging from 0.34 to 0.46. That relationship is structural. The suicide–overdose relationship was contingent on a supply-chain decision that was later reversed.

OxyContin was distributed broadly enough in the 2000s that high-overdose counties partially overlapped with the high-suicide Mountain West. As prescription crackdowns pushed users to heroin and then fentanyl, supply chains concentrated geographically — into Appalachia and New England, regions with existing distribution infrastructure — and the geographic overlap with the suicide belt collapsed. Case and Deaton published at the peak of artificial co-clustering, exactly as the split was beginning.

Policy implications

Before drawing the policy conclusion, let me grant Case and Deaton the strongest version of their mechanism. The geographic split challenges the unified policy thesis more confidently than it challenges the unified causal thesis. It's possible — the data here can't rule it out — that economic dislocation and social dissolution are still the common origin of both regimes, with local context determining how that pain expresses itself: guns and alcohol where they're available and culturally normalized, opioids where a supply chain ran. On this reading, Case and Deaton's mechanism is intact and geography is downstream of it.

What the data rules out is the idea that the two crises manifest in the same places. They don't. And if you're designing an intervention — a program, a policy, a budget line — the geographic reality is what matters. The argument here is not that despair has no economic dimension. It's that the map of despair has split, and the split has policy consequences regardless of whether the root cause is shared.

The most important finding for policy is the one that sounds most like a null result. Poverty rates across Regime A and Regime B counties: 13.9% vs. 14.3%, p = 0.76. Not significantly different.

This doesn't mean poverty is irrelevant to despair — it's not a claim about despair as a psychological experience. It means poverty does not differentiate the two regimes. If economic dislocation were the common cause driving both suicide geography and overdose geography, you'd expect the most economically distressed counties to be elevated on both axes. They're not. The counties with the highest suicide rates and the counties with the highest overdose rates have nearly identical poverty rates, and they're mostly different counties.

Poverty does not predict which regime a county falls into — which means economic investment cannot be targeted by regime. Both crises may well benefit from improved economic conditions. But the counties with the worst poverty aren't the counties with the highest suicide rates or the highest overdose rates; they're spread across both regimes roughly equally. A unified "deaths of despair" policy response — broader economic investment, workforce development, wage supports — has no geographic lever to pull.

What Regime A needs: mental health infrastructure in rural areas that currently have none; lethal means counseling and safe storage programs; recognition that altitude-linked mood effects are a real biological mechanism, not a metaphor; interventions that work at the scale of county and tribal land, not city. What Regime B needs: harm reduction infrastructure (naloxone access, fentanyl test strips, syringe services) concentrated in the specific counties where the supply chain runs; treatment capacity scaled to those same geographies; disruption of the distribution networks that the 2006–2012 pharmaceutical era established and fentanyl has since inherited.

These are different programs targeting different places. That county in rural Montana and that county in eastern Ohio don't share a diagnosis. Treating them as if they do is how neither of them gets better.