Meta-analysis of marital dissolution and mortality: Reevaluating the intersection of gender and age

The pooling itself is competent: four databases, nine rounds of citation chasing, 35 translated papers, two independent coders, and a random recheck that found no major errors. Results are split by how well each source study adjusted for other factors, and the headline barely moves when questionable estimates are excluded. That is enough to support the 30% figure as a description of the published literature.

Age, gender, recency, and culture are all inferred from study-level averages rather than from people. The recency test is the clearest failure: participants are classified as divorced at the start of each study with no record of when the divorce happened, so a study following people for two years and one following them for twenty both start with a mix of recent and decades-old divorces. The culture test rests on about 10% of estimates from roughly seven independent data sources, most of the South Asian data coming from two papers on one Bangladeshi surveillance site, and nothing from the Arab world the hypothesis names. The gender comparison pools men from occupational and heart-trial cohorts against women from cancer and nursing cohorts. Young samples are almost all unadjusted census data while old samples are adjusted cohort studies, so differences in adjustment are free to masquerade as an age effect. Several source publications reanalyze the same 1950 US census, and the paper does not say whether that overlap was corrected.

Random-effects models by maximum likelihood, inverse-variance weights, Q and I² for heterogeneity, and Egger and Peters tests for small-study effects are appropriate choices in outline, and confidence intervals are reported throughout. The pooled adjusted HR of 1.30 is plausible against prior work and stable under the paper's sensitivity checks.

The heterogeneity statistics are internally impossible: overall I² is 89.8%, yet the Q-test p-values for the adjusted, age-adjusted, and unadjusted strata are 0.9999, 0.92, and 0.29. The unadjusted stratum contains census-based estimates with tens of millions of person-years and HRs from 1.05 to 3.02; a correctly specified Q would be astronomically significant. A p of 0.9999 across 178 estimates means observed spread is far below sampling expectation, which happens only when input standard errors are inflated — most likely by the appendix variance-adjustment procedure or the regression-imputed standard errors (40.5% of estimates). Those same standard errors drive the inverse-variance weights, so the weighting of every pooled estimate is suspect. The paper reads these p-values as evidence that heterogeneity was "adequately accounted for," which is not what a Q test tests, while reporting a highly significant residual variance component in every meta-regression.

The parsimonious model was selected by backward elimination from roughly 40 predictors on 625 non-independent observations (eight publications supply 30–40 estimates each; five reanalyze the same census), so its p-values and intervals are not valid at face value. Elimination dropped every adjustment-level indicator except general health, letting adjustment level load onto age. The "6% per decade" age effect is the slope for all-female samples in a model with a sex-by-age interaction; the implied slope for all-male samples is 0.94 × 0.88 ≈ 0.83, a 17% decline per decade. The gender difference among adjusted estimates (1.37 vs 1.22) has overlapping intervals and no direct test; recomputing gives z ≈ 2.0 under an independence assumption that does not hold. Table 2 and Table 3 give incompatible age distributions (6.1% vs 39.7% of estimates at mean age 80+). The gender × follow-up p-value is 0.0790 in the text and 0.790 in the table. Roughly 150 subgroup analyses and 40 predictors were run with no multiplicity correction.

The headline association measures what it says: the mortality difference between people who were divorced and people who were married at the start of a study. Two caveats even there — "divorced at baseline" lumps last year's divorce with one from thirty years ago, and remarriage during follow-up is never tracked.

The moderator claims ask study-level averages to answer individual-level questions. No individual's age is observed; the age-by-gender curve is a pattern across studies with different average ages, fit largely on unadjusted census data at the young end (one adjusted estimate below age 40). The recency hypothesis is about time since divorce; the variable is how long a cohort was observed after a snapshot of marital status. A null on the second says nothing about the first. Stigma is measured in no study; the proxies are region and national divorce rate, entered in the same model so the divorce rate can only pick up variation within regions. Within the "traditional" regions the adjusted estimates point in opposite directions (East Asia well above Scandinavia, Bangladesh well below), and the paper averages this into "no difference."

The paper opens with a page on why the divorce–mortality link might be spurious — sick, depressed, and unemployed people divorce more — and never returns to it. Every hypothesis speaks of "harmful effects of marital dissolution," which the design cannot establish. The one piece of evidence bearing on selection runs against the causal reading: studies adjusting for baseline health find effects about 11% smaller, and the paper reports this without discussing it.

The recency and culture null results are read as refutations when the variables tested do not measure the things the hypotheses are about. The paper says the habituation explanation "is refuted" by a non-significant result on a variable that does not measure time since divorce. Regional averages are called "approximately equal" while spanning a range wider than the entire headline effect, and the two regions that did differ are dismissed with "should be treated with caution" and no reason. The 6% age slope is presented as general when it applies to all-female samples. The paper concedes that ratio measures compress toward 1 when death rates are high in both groups, which could manufacture the age decline, then sets this aside with "it is still likely" the effect is real, supported by no calculation. Social support is offered as the explanation for the sex-by-age pattern with the caveat, made once, that the data cannot test it.

Funding is disclosed (a single NIH grant, no funder role). The paper passed peer review and the appendix gives the full Medline search string, the coding variables, and descriptions of the conversion and imputation procedures. Two coders worked independently and a recheck found no major errors, though no agreement statistic is given.

None of the 625 coded estimates, their standard errors, or their moderator codes are shared; the study table gives only per-publication averages. No analysis code. No registration, and no way to confirm the five hypotheses were fixed before analysis. No conflict-of-interest statement. The limitations section covers publication bias, missing moderators, and thin developing-world data but omits the ecological-inference problem, the mismatch between follow-up duration and time since divorce, and the multiple reuse of the 1950 census. The paper contradicts itself: the two tables disagree on age composition by hundreds of estimates, imputation counts differ between text and table, a key test result differs by a factor of ten between text and table, and the summary bullets state the opposite of the results on recency. Without the underlying data, none of these can be resolved.

Where sources could be checked against full text, the paper does not misstate them, and the comparison to Manzoli 2007 is accurately reported in its essentials. Lillard and Waite 1995 is cited for economic factors explaining the effect without noting that this held for women, not men — a qualifier that matters for the paper's own argument. Williams and Umberson 2004, a study of self-rated health rather than death, is used to generate predictions in both directions. Manzoli's gender test was on the pooled unmarried category, not divorced people specifically, so the "surprising null" the paper sets out to overturn overstates what Manzoli tested. Sbarra 2011, a meta-analysis of divorce and mortality published the year before, is not cited. The paper does not tell readers that its central gender-by-age pattern appears in the team's companion analyses of widowhood and singlehood from the same database, which would bear on whether the pattern is real or a property of ratio measures at high death rates. Several references have wrong years or misspelled author names.

Publicly funded academic work with no commercial interest in the outcome. The authors have an ordinary stake in their own research program — a large database yielding a series of papers, and an unusual method of pooling every estimate a study reports — and the paper is framed as correcting a rival meta-analysis. Text carried over from the team's widowhood paper ("men who lost their spouse") shows an assembly line, not a slant. Two of the five stated hypotheses came out against the authors and were reported as such, publication bias in the authors' own data was reported rather than omitted, and the conservative adjusted estimate was chosen as the headline over the larger unadjusted one.