Life expectancy and active life expectancy by marital status among older U.S. adults: Results from the U.S. Medicare Health Outcome Survey (HOS)

The two-wave panel with linked death records is large and well documented, and letting marital status change over the projection improves on earlier work that froze people in their baseline status until death (Martikainen 2014). That supports the descriptive finding, but not the paper's conclusions. The projection is stratified by age and sex only; race and chronic conditions were collected but used only to check for dropout bias. About 23% of the sample was alive at follow-up but did not respond. The paper does not say how these people enter the model or whether they differed from responders in marital status — the variable that matters, since unmarried older people enter nursing homes more often and would plausibly be among the non-responders. The selection-bias analysis conditions on disability measured long after marriage, which cannot separate who married from what marriage did. Convergence at older ages is exactly what survivor selection produces: the frailest unmarried people have died by 85, leaving survivors who look like the married. The study cannot distinguish this from protection weakening with age.

The multi-state life-table method is standard and correctly described, with bootstrap standard errors for every estimate in the two main tables. Absolute death rates come from 2012 national life tables, with the survey contributing only relative hazards between groups; the men's estimate at 65 matches the national figure closely. The problems are in what is not reported. The adjusted analysis behind the "61% selection" figure appears only as two numbers in text for a single age, with no standard errors, confidence intervals, table, or description of the adjustment method. The 61% and 64% are ratios of point estimates rounded to one decimal; depending on how you round, the same raw numbers give anywhere from about 58% to 65%, so the precise figure is an artifact of decimal choices. The claim that active-life differences "disappear completely" after adjustment is stated without supporting data--literally marked "not shown." Handling of the 23% with missing follow-up is unstated, marital transition counts are never given, and roughly 4,700 people have no marital status in the baseline table without explanation. There is no multiple-comparison correction across 64 age-by-sex-by-outcome tests, which matters for the 0.1–0.2-year differences at the oldest ages that the paper reads as meaningful.

Marital status is a single checkbox; cohabiting partners count as unmarried, and marriage duration, quality, and living arrangement are unmeasured. The paper argues marriage helps through social support, income, and household environment, none of which are observed. "Active" means reporting no difficulty with any of six basic tasks, so "active life expectancy" is years free of reported difficulty, not years in good health. These are standard measures and the descriptive results are fairly labeled. The mismatch is the selection claim. "Selection" means health before marriage determined who married; the paper measures disability at an average age of 75 and treats whatever the adjustment removes as selection. The same shrinkage in the gap after adjusting for disability is equally consistent with marriage keeping people healthier, with shared causes like poverty producing both widowhood and disability, and with a disabled spouse raising the risk of widowhood. The composition of "unmarried" makes the selection story worse: about half are widowed — people who were once married, so the claim that unhealthy people fail to marry does not apply to them. The selection mechanism fits cleanly only the never-married, 4% of the sample. The paper calls its estimates "unbiased," but the model only fixes one problem — it lets marital status change instead of freezing it at baseline. It does nothing about confounding from income, education, smoking, or any other unmeasured difference between married and unmarried people.

From the abstract onward, the paper calls a raw difference adjusted for nothing but age and sex "protection." The paper's interpretive framework is unfalsifiable: any large difference is protection, any shrinkage after adjustment is selection, and whatever remains is still protection. The same paragraph says marriage protects and that selection accounts for roughly two-thirds of what it calls protection, and the paper does not reconcile the two. The paper acknowledges the circularity for active life expectancy — adjusting for disability shrinks a measure partly built from disability, "no surprise" — but does not see the same circularity for total life expectancy, where disability at 75 is equally likely to be a consequence of decades of marriage or widowhood rather than a pre-existing trait. The claim that earlier fixed-status studies "likely overestimated" marriage's effect is never tested; no fixed-status model is run on the same data. Confounding by income, education, smoking, or living alone is never raised as a competing explanation, and the age convergence is read as weakening protection without considering survivor selection or differential dropout. The novelty claim — "first study to examine ALE by marital status" — is contradicted by the paper's own literature review two pages earlier (Chiu 2019; Kaprio 1996; Martikainen 2014).

Routine items are covered: the survey instrument is public, the data are obtainable through a CMS data-use agreement, the authors declare no conflicts, and the limitations section is candid about the healthier-than-average sample, self-report, missing marital history, and the single-transition assumption. Everything needed to check the distinctive claims is missing: no code, no description of the disability adjustment, no transition counts, no explanation of how the 23% with no follow-up were handled, and a key supporting result marked "not shown." There is no funding statement or pre-registration, and the ethics section says only that the data were pre-existing. The limitations section does not mention that the selection estimate treats disability at 65 or older as a pre-marriage trait, or that the residual difference is unadjusted for anything.

The core findings come from the authors' own data, so no citations for those. The framing citations are padded and sometimes wrong. Two Verbrugge 1979 references are swapped, attributing the selection-protection framework to a government report on acute medical care. A list of "large, longitudinal samples" showing lower mortality among the married includes a 1950s vital-statistics analysis, a methods paper on marriage rates, and a study of disability, not mortality. Martikainen 2014 is summarized as finding never-married people had the shortest life expectancy, but the figures the paper quotes show that holds for men only. Rendall 2011, a cited review, reports little difference between never-married, divorced, and widowed people and describes selection as well studied, with several studies finding protection survives modeling of selection; the paper says selection has been examined "less often" and presents its estimate as novel. Lillard and Panis 1996, the standard reference for health selection into marriage, is absent. None of this changes the numbers, but the paper's account of the literature should not be repeated without checking.

No funding source is declared and the authors state no competing interests; both are academics using a federal survey obtained as a limited data set. The only visible pressure is academic novelty, which shows in the "unbiased" and "first study" language rather than in the direction of findings. The headline conclusion — that most of the marriage advantage is selection — runs against what a pro-marriage constituency would want.