Marriage and health: Selection, protection, and assortative mating

The panel design is well-suited to the descriptive question: how much of the raw health difference between married and unmarried people is explained by stable personal characteristics? The PSID tracks the same individuals for up to 30 years, and the convergence of results across two independent estimation methods (standard fixed effects and grouped fixed effects) is genuinely informative. For the causal question — does marriage itself protect health — the design is fundamentally limited. The methods remove only characteristics that are stable over a person's lifetime. Factors that change over time, such as job loss, retirement, bereavement, or shifts in social networks, remain in the estimates and could explain the residual difference at older ages. A person widowed at 55 simultaneously loses a spouse and experiences grief-related health decline; the method cannot separate these.

The measurement setup introduces a further problem. In the main dataset, the household head reports health status for both themselves and their spouse. Married people's health is assessed by someone else; unmarried people self-report. If partners rate each other's health differently than people rate their own, this creates a systematic measurement difference between groups that has nothing to do with actual health. The paper acknowledges this issue only for its spousal health correlation analysis. When the authors switch from self-reported health to more objective chronic conditions, the correlation drops by about 40% — consistent with substantial reporting artifact in the self-reported measure. This check was never applied to the main estimates.

The main estimation methods are appropriate and competently executed. The progression from ordinary regression to fixed effects to grouped fixed effects is well-structured, and the grouped fixed effects estimator — which allows stable personal characteristics to affect health differently at different ages — is a genuine methodological contribution.

Beyond the descriptive decomposition, the statistical work has fundamental problems. The system-GMM dynamic panel model, presented as addressing reverse causation, is reported without the standard diagnostic tests for instrument validity or specification correctness. Its identifying assumption — that health shocks affect marriage only with a lag, not in the same year — is asserted rather than tested, and a serious illness could plausibly affect both health and marital stability within the same year. The model also uses only data from 1985–1997, a different era than the full sample.

The mechanism analyses — preventive care, smoking, health expenditures, health insurance — all use simple cross-sectional comparisons with no adjustment for stable personal characteristics. The paper's own core analysis demonstrated that such comparisons overstate the marriage-health link by roughly half. The smoking claim, that getting married raises the probability of quitting by 23 percentage points, comes from a raw comparison with no adjustment for education, income, age, or any other characteristic. The health insurance analysis splits the sample by insurance status, but insurance is itself a direct benefit of marriage (through spousal coverage), making the comparison uninterpretable as evidence about mechanism. No power analysis is reported, which matters most for the claim that the marriage-health effect is similar for men and women — the study may not have been large enough to detect a real gender difference.

The paper's central interpretive framework — selection explains the young-age difference, protection explains the old-age difference — treats a statistical residual as evidence of a causal mechanism. The fixed effects the paper removes and labels "innate health" capture everything stable about a person: not just health endowment, but personality, conscientiousness, childhood environment, family wealth, and genetics beyond health. Calling this "innate health" and connecting it to evolutionary biology is a naming choice, not a finding. The statistical results would be identical if the fixed effects primarily captured conscientiousness or socioeconomic background.

The residual 5-percentage-point difference at older ages — the basis for the "protective effect" claim — absorbs an unknown mixture of marriage effects, time-varying confounding, reverse causation, and measurement asymmetry. The design cannot separate these. The causal protection claim has no evidence that survives the paper's own methodological standards, and the insurance-as-mechanism claim conditions on a variable that marriage itself causes. The other two core claims — that selection plays a role at younger ages, and that a residual difference exists at older ages — are better supported as descriptions, though the specific interpretation (selection on health rather than on correlated traits) is not established.

The descriptive reasoning is sound: the paper demonstrates clearly that accounting for stable individual differences eliminates the health difference at young ages and halves it at older ages, and this finding holds across estimation methods. The reasoning breaks down at the interpretive level. The paper treats the residual difference at older ages as evidence of a "protective effect," but the methods can only establish that the difference is not explained by stable personal characteristics — it could be driven entirely by changing circumstances that correlate with both health and marital status at older ages.

The most consequential reasoning flaw is the methodological inconsistency between the main analysis and the mechanism evidence. The paper demonstrates that naive comparisons overstate the marriage-health link by roughly half, then uses exactly those naive comparisons — and sometimes no controls at all — for every piece of evidence about how marriage might protect health. The paper never acknowledges this tension. The abstract and conclusion also progressively strengthen the hedges found in the body text: "there might be a protective effect" becomes "protective effects of marriage on health" and "the beneficial effect of marriage" in the conclusion.

Funding is disclosed and comes entirely from European public research bodies. The paper was published in a peer-reviewed economics journal. Both underlying datasets are publicly available, and variable definitions are documented in detail. No analysis code or replication dataset is provided, and for a paper with multiple complex estimation strategies (including a custom iterative algorithm), the absence of code is a meaningful barrier to replication. There is no dedicated limitations section. The limitations the paper does acknowledge — such as the possibility of reverse causation — are less important than those it leaves unaddressed, particularly the methodological inconsistency between its main analysis and its mechanism evidence, and the proxy-reporting asymmetry in the main estimates.

The methodological citations (Bonhomme and Manresa 2015 for the grouped fixed effects estimator, Arellano and Bover 1995 for the dynamic panel model) are appropriate and accurately described. Williams 2003, cited to support the claim that marriage benefits men and women equally in terms of physical health, actually studied psychological well-being — a different domain. Lillard and Panis 1996 is selectively represented: the paper mentions their finding of negative selection into marriage but omits their parallel finding of positive selection on unobserved characteristics, which would support the paper's own framework. The evolutionary biology citations (Buss 1994, Dawkins 1989) are decorative — they provide narrative flavor for the "innate health" interpretation but no empirical evidence. The paper also does not engage with literature on divorce as a cause of poor health, which matters because the "unmarried" group at older ages increasingly includes divorced and widowed individuals.

The three authors are affiliated with public academic institutions in Spain. Funding comes entirely from European public research bodies — a European Research Council grant, the Spanish Ministry of Economy, and the Catalan regional government. The data sources are publicly available U.S. government surveys. There are no identifiable conflicts of interest.