The Rising Relative Risk of Mortality for Singles: Meta-Analysis and Meta-Regression

The pooled design establishes that the published literature consistently finds higher mortality among never-married people, reported by level of adjustment and led by the more conservative high-quality figure. Two authors screened and extracted independently, a random re-coding of 25 papers found no errors, and non-English work was translated. Beyond the pooled number, the design cannot support the historical, regional, and quality comparisons asked of it. The literature search was built in 2005 for a broader question about social isolation, not marital status, and its terms are withheld. The authors pooled cross-sectional census comparisons with cohort follow-ups without reporting how many of each there are. The 641 estimates come from only 95 studies — about seven per study — with several cells dominated by one source. The correction for this overlap is described only in an appendix not available with the paper. Most seriously, older estimates are overwhelmingly unadjusted and newer ones overwhelmingly adjusted, so the "rising over time" finding is inseparable from a change in how studies were done. The people extracting the numbers also assigned the quality rating, with no statement on whether they saw the results first and no measure of rater agreement.

The random-effects pooling and meta-regression are appropriate in principle, and hazard ratios with confidence intervals are reported throughout. Heterogeneity Q-test P values for the main strata are 0.95–0.99 — 50 studies across a century and seven regions showing less spread than sampling error alone would produce — yet the meta-regression on the same data reports R² = 0.58 with roughly twenty significant moderators. These are incompatible. The most plausible explanation is that the variance adjustment for overlapping person-years and the regression-imputed standard errors (263 of 641 estimates) inflated within-study variances, collapsing the model toward fixed-effects weighting dominated by census-scale samples. Every reported confidence interval inherits this uncertainty. The regression also produces a sign flip: raw strata show age-adjusted estimates below unadjusted ones (1.37 vs 1.51), but the model reports them 9% higher — instability in exactly the dimensions the novel claims depend on. Publication age and study age were entered together and only the latter is significant, so the coefficient may measure publication lag rather than calendar time. The paper does not address this. The gender × study-age interaction that the paper calls its most interesting finding has a confidence interval touching the null (1.00–1.03) in the parsimonious model. Peters' test finds bias in one direction while the sample-size coefficient points the other way. The Rosenthal fail-safe N values (845,374 and similar) are artifacts of pooling half a billion people and say nothing about the size of any overestimate. Key subgroup cells are nearly empty: the age 30–39 adjusted figure is 4 estimates from 3 studies; the 2000s decade is 6 estimates from 1 study; the low-quality adjusted stratum is 2 estimates from 1 study. The only figure illustrating the trend cites a table that does not exist in the paper.

The pooled comparison measures what it claims: how the death rate of people coded never-married compares with people coded married across the published literature. Where the measurement is loose — cohabiters included in the married group in some studies, marital status fixed at study start so later marriers stay "single" — the errors understate the gap, which strengthens the headline. The problems begin when between-study differences are read as facts about populations. "Single" means very different things across the pool: a never-married 25-year-old in 1990s Sweden will mostly marry later and is barely selected; a never-married 45-year-old in 1980s Japan belongs to a rare, heavily health-selected group. The paper codes both identically, then attributes the larger East Asian gap to social sanctions — but in countries where very few people stay unmarried, those who do are a more unusual and likely more health-selected group, which alone would produce a larger gap without any cultural mechanism. "Rising over time" compares old studies with new ones, and old and new studies differ in design, adjustment, and data source all at once. "Lower-quality studies report bigger effects" rests on a quality rating that may have been influenced by the size of the effect being rated. The pattern holds only in one stratum where the low-quality cell is a single study; among unadjusted studies, high-rated studies report larger gaps than average-rated ones. The paper presents itself as the first to cover non-elderly singles, but 85% of adjusted estimates concern people 60 and older. The age gradient is the strongest secondary finding, because much of the age variation comes from within studies. But the paper names two competing explanations (ratio compression at high death rates, and the changing meaning of "never-married" across the life course) and tests neither.

The paper's title, abstract, and closing paragraph state the rising trend flatly; the results section says it "is difficult to discern" and the decade figures do not show it. The trend exists only as one coefficient in a model with about twenty predictors. A nearly identical variable — years since publication rather than since data collection — shows no trend in the same model, which the paper does not explain. The claim that men and women are "now approximately equal" infers equality from the failure to detect a difference at a time point supported by one study. The discussion is candid about latent selection (heavy drinking, personality); the conclusion carries the protection reading forward anyway, on the strength of one null comparison between studies that did and did not control for baseline health. A formal test confirms publication bias. The paper then dismisses it with a calculation that tests whether the effect exists rather than whether it is overstated, and a comparison of quality tiers that says nothing about unpublished studies. The East Asia result is given a cultural mechanism no variable in the model measures. The limitations section warns against regional conclusions from small samples but applies the warning only to other regions. The closing sentence says ratios are "particularly high in younger cohorts," conflating age with birth cohort. The paper does present its raw stratified tables in full — which is how a reader can see the decade figures contradict the title — and the limitations section is more candid than the conclusions.

There is no protocol or registration, and the analysis is one of at least three carved after the fact from a database built in 2005 for a different question. The extracted dataset is not available, search strings are "available upon request," and no analysis code is shared. The list of included studies, the conversion of odds ratios and death-rate ratios into hazard ratios, the regression used to impute missing standard errors for 263 estimates, the variance adjustment for overlapping samples, the quality-score factor analysis, the trend plots, and the funnel plot all live in a web appendix that was not part of the evaluated text. These are the steps where the analysis could go wrong, and none can be checked. No funding source is named. Conflicts of interest are declared as none. The limitations section raises four real weaknesses — publication bias, uneven adjustment, the difficulty of rating quality, and over-representation of rich countries — and then argues each one away. It does not mention the tension between near-zero heterogeneity and a high-explanatory-power regression, the absence of a trend in the raw decade figures, the figure citing a nonexistent table, or the unblinded quality rating.

The core numbers come from the 95 pooled studies, so they do not depend on any cited work. The reference list does framing and interpretation, and it is uneven. Manzoli 2007, the prior meta-analysis on the elderly, is described accurately. But its most relevant finding — that correcting for publication bias roughly halved the married advantage — is omitted, even though the present paper argues its own publication bias is minor. The selection-versus-protection section cites several off-topic sources for its claim that most studies favor protection: a study of religious involvement among African Americans, a paper on psychological well-being rather than death, an edited sociology volume also cited elsewhere for the opposite hypothesis, and a reference with the wrong year. The explanation for women's rising risk leans on three essays from a feminist anthology (Funiciello 2003, Norsigian 2003, Braun-Levine 2003) presented as evidence for declining public assistance and health benefits. These are opinion pieces, nearly all American and about single mothers, while the pooled data span Scandinavia, Europe, and East Asia. A trade book by a marriage-movement advocate (Waite and Gallagher 2000) supports the claim that married people take fewer risks. The methodological citations are standard and correctly used, and the paper does cite one conflicting finding (Liu and Umberson 2008).

No funder, industry, or advocacy affiliation is identifiable. All four authors are academic sociologists, conflicts are declared as none, and nobody profits from a finding about marital status and mortality. Several choices cut against the authors' own interest in a dramatic result: they built the database for a broader question, limiting opportunities to select favorable studies, and they led with the lower quality-restricted figure rather than the overall one. The one pressure is the ordinary academic need for a novel headline: this is one of at least three papers from the same database, each needing its own distinctive result. That pressure shows in a title that says "rising" over a flat raw trend, but that is a question of evidence, not conflict.