Marital status and mortality in the elderly: A systematic review and meta-analysis

The review is competently built: prospective cohorts only, adjusted estimates only, two extractors blind to authorship, and pooling that allows for real differences between populations. The problems are in what went into the pool. About one input in ten measures something other than marital status — living with a partner, spousal support among the married, or confidant-network size — yet the paper calls the exposures "comparatively homogeneous." In 23 of 53 comparisons marital status was recorded once at baseline and never updated, so people widowed during a nine-year follow-up stayed counted as married. This dilutes the female estimate more than the male one, because married women are widowed far more often. Two Finnish register studies covering the entire population aged 80 and over (Martelin 1998; Martikainen 2005) both appear in the pooled analyses despite a stated rule against using the same people twice, and their narrow intervals give them outsized weight. The counts of individually favourable studies (12 of 14 for divorce, 11 of 17 for never-married) are reachable only by including studies whose intervals touch the no-effect line. None of this threatens the direction of the finding, but all of it makes the pooled figure less exact than it looks.

The core method — DerSimonian–Laird random-effects pooling by inverse variance, with I² for heterogeneity — is standard and appropriate. Several implementation choices weaken it. Odds ratios and hazard ratios are pooled as interchangeable, which inflates protective estimates from logistic-regression studies where cumulative mortality is high, as it is over 8–20 years in people over 65. Within-study estimates for widowed, divorced, and never-married groups were combined by fixed-effect pooling as though independent, though they share a reference group. Add the probable double-counting of the Finnish registers and of the UK Longitudinal Study in the widowed analysis (Breeze 1999 appears as both a full cohort and a bereavement sub-sample), and every pooled interval is too narrow. No subgroup difference was formally tested — gender, region, quality, and follow-up length are compared by eyeballing overlapping intervals — and no meta-regression was attempted despite I² of 60–80% in nearly every analysis. The "worst possible scenario" imputation assigns RR = 1.00 with interval 0.99–1.01 to eight excluded non-significant comparisons. Under random effects a tiny within-study variance does not produce overwhelming weight, and a non-significant result at n ≈ 2,000 is compatible with RRs from roughly 0.7 to 1.4, so this is a sensitivity analysis under one arbitrary assumption, not a bound. The two most extreme protective estimates (0.34 and 0.26) come from studies where the other sex was reported only as "non-significant" and excluded. The 5–15% range splices the upper limit of one analysis with the lower limit of another and is stated as a single interval; the abstract gives 9–15%.

The paper claims to measure the mortality difference between married and non-married elderly people. What it pools is a mixture: roughly forty comparisons of recorded marital status, three of co-residence with a partner, one of spousal support among married couples, and one of confidant-network size. The comparison group bundles people widowed last year with people divorced decades ago and people who never married. Sub-analyses separate these, but the headline figure averages over them. The divorced and never-married comparisons are the cleanest, since those statuses rarely change after 65. The widowed comparison pools two exposures (widowed at baseline versus bereaved during follow-up), and the baseline-only design biases the gender comparison toward finding no difference. The "true reduction between 5% and 15%" is causal language applied to an association the same paragraph says has not been shown to be causal, and neither endpoint bounds the true value. The argument that a stronger effect in the seven fully adjusted studies confirms marriage is protective does not follow: if healthier people select into marriage, better adjustment should shrink the effect, and finding the opposite in a small subgroup is unexplained rather than confirmatory.

The descriptive finding is well grounded and the discussion is unusually candid: it names health selection, reverse causality, the unvalidated quality scale, the unexplained disagreement between studies, and the form of publication bias its checks cannot detect. The trouble is the move from "associated with" to "protective effect." The abstract and results use causal wording throughout; the concession that cause and effect "has yet to be demonstrated" sits near the end. The recommendation that marital status is "essential" for identifying at-risk elderly is followed by the statement that "even a 5% reduction" would justify it, so the recommendation would have followed from almost any nonzero result. The paper never shows marital status adds predictive value beyond age and health. The 16% excess for divorced people versus 11% for the widowed and never-married is explained through dissolution stress and marital satisfaction, though the three intervals overlap. The paper does not raise why its gender comparison is tilted toward equality, nor engage the demographic literature arguing that adjustment for measured health cannot rule out selection.

The search strategy, inclusion criteria, and quality scale are fully specified. The extracted data are not: the extraction table and list of excluded papers are "available from the authors," per-study quality scores are never given, the 14 comparisons not split by sex appear only in a figure, and two results are "data not shown." No pre-registration, funding statement, or conflict-of-interest declaration appears. That reflects 2006 norms, but it means there is no way to tell whether the "worst-case" re-analysis or the cutoffs for "small" and "high-quality" studies were planned in advance. The limitations section is long and mostly honest, though it does not flag the mismatch between the abstract's 9–15% and the conclusion's 5–15%, the non-marital exposures in the main pool, or the probable Finnish double-count.

Methods citations — DerSimonian and Laird 1986, the Cochrane Handbook, the Newcastle–Ottawa Scale, Egger 1998, Sterne 2001 — are appropriate and correctly applied. The citations that matter most are the included cohort studies, and the extraction from them cannot be verified here. The pooled estimates are dominated by a handful of huge register datasets, and two of those appear to be counted twice: the Finnish 80+ population (Martelin 1998 and the 80+ stratum of Martikainen 2005) and the UK Longitudinal Study (Breeze 1999, as both full cohort and bereavement sub-sample). Hu and Goldman 1990 is cited only for framing, though it reports evidence of selection that cuts against the paper's lean toward a protective effect. The causal suggestion rests on a single citation (Cornman 2003) while the demographic work showing that adjustment for baseline health cannot rule out selection (Goldman 1993; Lillard and Panis 1996) goes unmentioned. A few supporting citations are mismatched: a claim that unmarried people are more depressed is sourced to a study of grandparents raising grandchildren (Minkler 1997).

There is no identifiable funding or industry conflict. The paper carries no funding or conflict statement, which was common in this journal in 2006. All four authors hold academic or public-institute posts in Italy, and the topic has no commercial stakeholder. The one lean is disciplinary — public-health researchers have reason to prefer a finding that social circumstances shape survival — and the paper handles it by laying out the selection explanation up front and running checks that could weaken its own result. The policy recommendation reaches further than the data support, but that is ordinary academic pull toward relevance rather than outside interest.