Does Marriage Protect Health? A Birth Cohort Comparison*

The three-model comparison — cross-sectional regression, random effects, then fixed effects — is well-established, and the PSID is one of few datasets long enough to attempt cohort comparisons. The overall sample of 12,373 people and 85,000 observations is substantial.

Two problems undermine the central claims. First, the rater-identity problem: in the PSID during this period, the male partner rated both his own and his spouse's health. For women, the person evaluating their health changes at the moment of marriage — the transition the study is trying to measure. The paper's one-sentence defense (spouse-rated health predicts death about as well as self-rated health) does not address whether the two raters give the same scores for the same person, which is what matters for a before-and-after comparison.

Second, the cohort-decline claim rests on roughly 370 people (perhaps 185 women) in the youngest birth cohort who had been married ten or more years — women who married by their early twenties in a generation defined by delayed marriage. They are not representative of their cohort. The paper acknowledges this in passing but draws the generational conclusion anyway.

Additional gaps: cohabitation is never measured despite being central to the theoretical framework (a person living with a future spouse is coded as "never married"); income is omitted despite the PSID being built around income data; and the youngest cohort is observed only to a median age of 30, making it impossible to separate cohort effects from age-at-observation effects.

The regression specifications are standard and the three-model comparison is sound in principle. The broad finding that coefficients shrink from cross-sectional to within-person models is well-powered and interpretable.

The problems cluster around the specific claims. The interaction supporting the headline conclusion about generational erosion comes from roughly 300 observations contributed by roughly 185 women. No power analysis addresses whether this cell can support the conclusion drawn from it. Confidence intervals are never reported — only coefficient estimates and significance markers — so the reader cannot judge precision in these small cells. The net effect for the youngest cohort (+0.07) is described as indistinguishable from zero, but it is also indistinguishable from a small remaining benefit.

With three duration categories, two cohort comparisons, three model types, and two genders, the paper tests 36 interaction terms with no correction for multiple comparisons. The headline conclusion is drawn from the one combination that produces a notable result: fixed-effects models for women with ten-plus-year marriages in the youngest cohort.

The rater-switching artifact is equally relevant here. If the earliest cohort's coefficient for women was partly an artifact of rater change (because fewer women cohabited before marriage), and later cohorts have less rater switching (because more cohabited first), the apparent "erosion" across cohorts could be the gradual disappearance of a measurement artifact.

The null finding for men is credible. Men rate their own health throughout the study, so the within-person comparison is clean. The main caveat is that "never married" includes men cohabiting with a partner, so the null could mean "marriage adds nothing beyond cohabitation" rather than "marriage adds nothing."

For women, the design cannot answer the question it poses. The rater switches at marriage for most women. The paper's only positive finding (a small benefit of long marriages) and its cohort-decline finding cannot be separated from the interaction between rater-switching rates and cohort membership.

The claim that selection explains most of the cross-sectional association is plausible — the attenuation pattern points in the right direction — but overstated. The within-person model could also attenuate real effects by being too blunt to detect gradual health changes, and the paper treats this as a footnote rather than a competing interpretation.

The paper frames methodology and cohort change as having "converged to falsify" the hypothesis that marriage improves health, treating two pillars of very different strength as equally established. The methodology pillar (coefficients shrink under stricter models) is well-supported across the full sample. The cohort pillar (the benefit is disappearing in younger generations) rests entirely on roughly 185 early-marrying women observed to a median age of 30. You could remove the cohort analysis and the methodology story would hold; you could remove the methodology analysis and the cohort story would be too thin to stand.

The most consequential unexamined alternative is rater switching. The finding that men show no marriage effect while women show a small one is presented as requiring a theoretical explanation about gender differences. A simpler explanation: for men, the rater stays the same before and after marriage; for women, it changes. The apparent cohort erosion could likewise reflect the disappearance of this artifact as premarital cohabitation becomes more common. Neither alternative is discussed.

"Falsify" — a strong epistemological claim — is applied to a finding from one gender, one duration category, one cohort, and a small unrepresentative cell.

The PSID is publicly available, so any researcher can attempt reproduction. The paper is peer-reviewed and acknowledges several limitations honestly.

Against this: no analysis code is provided, no pre-registration is mentioned, no funding or conflicts are disclosed, and no sample size justification addresses the small cells driving the key findings. The limitations section omits the two most threatening issues — rater switching for women and the fact that "never married" includes people cohabiting with a future spouse.

The most frequently cited empirical paper — Liu and Umberson (2008) — is selectively represented. That study found the health gap between married and never-married adults narrowed over time for men only, while the health penalty of divorce and widowhood actually worsened over time, especially for women. The paper cites it broadly as evidence that "the protective effects of marriage have eroded," taking the never-married-men finding and omitting the rest.

Hughes and Waite (2009) is cited for the narrow point that shorter marriage tenure reduces benefits, while its stronger finding — that marital disruption damages health with effects lasting years — goes unmentioned, though it complicates the paper's conclusion. A working paper (Guner, Kulikova, and Llull 2014) is cited five times and treated as a core reference despite not having passed peer review.

Two major meta-analyses published before this paper — on marital quality and health, and on social relationships and mortality — are absent. Both would complicate the conclusion that marriage has no causal effect on health.

Single-author academic study with no disclosed funding, no industry ties, and no structural conflicts. The conclusions push against federally funded marriage-promotion programs, making it mildly contrarian relative to the policy establishment rather than aligned with any sponsor.