The Protective Effect of Marriage for Survival: A Review and Update

The data are the paper's strength: six pooled waves of a national household survey linked to government death, earnings, and disability records, with covariates updated every year and a sample roughly ten times the size of previous panel studies. For the core question — is there a mortality gap between married and unmarried adults after adjusting for observed characteristics — this is a strong design.

There is no randomization, no natural experiment, and no within-person comparison, so protection and selection cannot be separated. The paper acknowledges this once and then uses "protection" throughout. Health is measured only by receipt of federal disability benefits (2–8% of the under-65 sample) and not at all after 65, so "otherwise similar" means similar in demographics and income, not health. People living with a partner but not married are coded "unmarried," diluting the contrast in the group where cohabitation was most common. Death records are matched at 93–96% for ages 65 and over but only 70–86% under 65. Social Security learns of deaths largely through benefit claims, and a surviving spouse is the most common claimant, so deaths of never-married and divorced people are plausibly undercounted. The survey excludes nursing homes, so frail unmarried people who are institutionalized vanish from the data while frail married people kept at home by a spouse die in view.

The comparison between never-married and widowed people, which underpins the paper's methodological recommendation, is the most affected. The groups barely overlap in age (averages of 36–38 versus 72), so the comparison is carried by the model's assumption that mortality rises in a straight line with age across 35 years, not by comparing people of the same age. The missing deaths fall almost entirely on the never-married side, pushing that group's estimated mortality down toward the widowed. The design cannot fairly compare these groups, yet the paper reports that they don't differ.

The modeling framework is appropriate: discrete-time logistic hazards on person-years, with covariates measured at the start of each year. The paper advocates and runs the right between-group tests: pooling sexes with full gender interactions and running Wald contrasts, and testing unmarried subgroups directly against each other rather than only against the married reference. All coefficients and standard errors are reported. The uncontrolled odds ratios match national vital statistics, a useful external check.

The headline male premium (p = .003) comes from a model with no age × unmarried interaction. The paper's own preferred model shows that interaction is needed: the unmarried effect declines steeply with age for both sexes. In that model the gender contrast is p = .070; at 65+ it is p ≈ .5–.8. Since unmarried men average age 43 and unmarried women 52 (a third of unmarried female person-years are widows averaging 72), omitting the age interaction mechanically inflates the coefficient for the younger group. The abstract reports the misspecified model's result.

The null on never-married versus widowed is a non-rejection with no equivalence bound. At 65+, the standard errors are wide enough that the data cannot exclude a 25–50% difference in mortality odds between the groups. Of 36 between-status Wald tests, seven reject at .05 (divorced vs. widowed men p = .002; divorced vs. never-married at 65+ for both sexes; widowed vs. divorced men 65+ p = .008; two age-interaction contrasts). The tests are correlated, but seven rejections is more than chance alone would produce. The paper calls these "mixed evidence" while treating non-rejections as evidence of equality and prior studies' rejections as chance. No multiplicity correction is applied in either direction. Standard errors are unadjusted for household clustering (spouses are in the same sample), so the borderline p-values are less trustworthy than they appear. The "protection ceases at 84/89" figures are straight-line extrapolations with no confidence intervals, and the men's figure doesn't reproduce from the model the text cites. The recommended procedure — pool the unmarried into one group, then split only if a statistical test says they differ — is circular, because the test's properties depend on whether you've already pooled. And applied to this paper's own data, the test says to split.

The variable "unmarried" measures legal status, not the social arrangement the theory is about. Cohabitors are coded unmarried even though they have the hypothesized exposure (a co-resident partner who monitors health and shares income). The paper never mentions this.

The label "marriage protection effect" names an adjusted association. Nothing in the design distinguishes protection from selection, and the monotone shrinkage of the estimate as controls are added is at least as consistent with "more controls would remove more" as with "a real effect remains." The male premium may be partly or wholly an age-composition artifact, since the effect declines with age and unmarried men are a decade younger on average than unmarried women. The never-married versus widowed "no difference" compares groups in different decades of life via a straight-line age term, in a sample where the younger group's deaths are systematically undercounted. The paper's own between-group tests contradict its claim that prior reports of subgroup differences are chance artifacts: they find the divorced/separated differing from other unmarried groups in a pattern matching the studies the paper dismisses. The core descriptive finding — a married/unmarried gap in the following year — is what the design measures and is sound.

The paper sets up two competing explanations for the messy prior literature — real differences hidden by bad data, or no real differences manufactured by bad statistics — and concludes both are true. This framework is unfalsifiable: if prior studies found differences, the data were noisy; if they didn't, the statistics were wrong. Either way the paper's conclusion holds.

Its central methodological point is correct: finding one coefficient significant and another not is not a test of whether they differ. But the paper commits the mirror error, reading a failed test between never-married and widowed as "little evidence of mortality differences" with no statement of what size difference could have been detected. The same asymmetry applies to rejections: the paper's own seven become "mixed evidence," prior studies' become "chance." The authors could have tested their claim that annually updated data explain their clearer results by re-running the models on first-observed marital status only. They didn't. The tables are honest and complete. The summary layer — abstract, discussion, title — does not match them.

Funding (Social Security Administration contracts to RAND and the Urban Institute) is disclosed with contract numbers. The paper concedes the selection alternative, states the absence of health measures after 65, names the nursing-home exit problem with its correct direction, and reports the unadjusted standard errors and incomplete death linkage. The linked administrative data are restricted-access and no code is shared, which was normal for demography in 2011 but means results cannot be reproduced. The authors made several choices after seeing results they don't show: disability-duration variables dropped for non-significance, year × marital status interactions tested but unreported, and the key check that missing deaths don't depend on marital status described only as "results not shown." The paper never states what size difference it could detect, which matters because several conclusions are about the absence of differences.

Citations are accurate where checkable. Manzoli 2007 is cited correctly but framed one-sidedly. The paper uses it to say support for marriage protection is "not overwhelming" because only half the underlying studies individually reached significance, when the pooled estimate is precise and strong — the small-sample pattern the paper elsewhere says should not be read as absence of effect. The paper describes Rogers 1995 as the only prior study to formally test gender differences, but Kaplan and Kronick 2006, cited in the same section, appear to report such a test. The literature on bereavement timing (the widowhood penalty spikes in the months after a spouse's death and then fades), for example Martikainen and Valkonen 1996, is absent. A single "widowed" state averaging a sharp early spike with a long low tail can look like the never-married rate. The paper cites Lund 2002 but never engages cohabitation.

Government-funded work with no commercial ties. The funder has no evident stake in whether marriage protects survival. One author works for the Social Security Administration, which also owns the data. The paper's favorable judgment about the completeness of the agency's death records, which bears directly on the never-married versus widowed comparison, rests on comparisons that are not shown and cannot be checked without the agency's data access. That is ordinary academic self-interest slightly amplified by the data being the funder's.