Heightened mortality after the death of a spouse: Marriage protection or marriage selection?

1. What this paper claims

After a spouse dies, the survivor's own death risk rises. The paper argues this is a causal harm for men — losing a wife shortens a husband's life — but for women it is mostly an illusion created by couples sharing the same unhealthy habits.

2. What evidence it offers

An observational study of about 72,000 married white Americans aged 50–70, surveyed around 1980 and tracked against death records for nine years. The identifying idea is a natural experiment: sort spouses' deaths into causes that track income, education and occupation (heart disease, lung cancer, cirrhosis) and causes that don't (leukemia, pancreatic cancer, breast cancer). If the survivor's excess mortality is just shared lifestyle, it should vanish after the "uninformative" deaths. If losing a spouse is itself harmful, it should show up equally after both.

For men the excess risk is about the same after both kinds of death (36% vs 29%). For women it is 4% after uninformative deaths and 22% after informative ones. The paper reads the first as causation and the second as selection.

Three things undercut this. First, "not predicted by income and education" is not the same as "unrelated to the couple's habits" — pneumonia, pancreatic cancer, accidents and breast cancer are all on the uninformative lists and all linked to smoking, drinking or obesity, which spouses share. Second, the women's uninformative estimate rests on about 65 deaths; its plausible range runs from a 19% lower risk to a 34% higher risk and includes the 22% seen after informative deaths. Third, the two estimates could not be told apart for men or for women, yet the paper draws opposite conclusions. The body calls the women's finding "suggestive but not definitive"; the abstract states it as a result.

3. Rubric summary

Study design (2)

The comparison group (still-married people) is right and the natural-experiment idea is sound in principle. Execution fails at the sorting step. The dataset has no smoking, drinking, weight or health variables, so "uninformative" could only be defined against income, education and occupation — not against the shared behaviors that drive the selection worry. More than half the "uninformative" deaths are behavior-linked. The two death groups also differ in how people die: uninformative causes are mostly slow cancers with months of caregiving, informative causes are mostly sudden heart attacks and strokes. The paper's own cited mechanisms (caregiver exhaustion, acute shock) would make grief itself differ across these groups, so the variable meant to isolate selection also changes the causal exposure. Marital status is observed once at baseline, so divorce and remarriage over nine years are invisible. The women's key estimate is built on too few deaths to distinguish "no effect" from "the same effect as men." Good practices — progressive controls, an independent preventability classification, leave-one-cause-out checks — do not repair these problems.

Statistical analysis (2)

Cox proportional hazards with a time-varying widowhood indicator and SES controls is standard. The women's uninformative-cause HR is 1.04 with a 95% CI of roughly 0.82–1.33; the informative-cause HR is 1.22 (CI 1.10–1.35). Power to detect the pure-selection alternative is about 30%. The headline men-versus-women contrast (1.36 vs 1.04) is never tested; computed from the reported SEs it gives p ≈ 0.08. Non-rejection of equality (p = 0.65 for men, p = 0.22 for women) is read as "equal" in one case and "different" in the other. The classification procedure fails its own rule: the pooled husbands' uninformative group rejects SES independence (joint p = 0.014), and the "negligible" marginal effects the paper cites are the same relative gradient as the informative group once base rates are considered. The minimum cell size was chosen as the smallest cell that happened to reject. The paper's power argument ("we found an effect in men with fewer widowers") confuses the number of bereaved with the number of deaths among the bereaved; men in this age band die at twice the rate, so the men's estimate rests on roughly twice as many events. The men's uninformative HR (1.36, CI 1.14–1.62) does exclude 1, so some excess mortality survives after SES-unrelated deaths — the one adequately supported finding.

Construct validity (2)

The design needs causes of death that carry no information about the survivor's unobserved health risk. What it measures is causes not predicted by the decedent's income, education and occupation. A footnote makes this substitution quietly. It matters because the selection story is about health behavior — spouses of smokers are three times as likely to smoke — and the uninformative lists contain smoking-, alcohol- and obesity-linked causes. The paper concedes contamination would push the men's two estimates toward each other, which is the pattern it then calls causal. A second unstated assumption is that losing a spouse harms the survivor equally regardless of how the spouse died; the uninformative and informative groups differ in death trajectory. The prediction under pure selection ("substantially smaller") is never quantified, so no result could falsify either story. Descriptive claims — the two estimates are close for men, indistinguishable for women — are sound. The causal claim for men is a stretch; the selection claim for women is disconnected from the evidence.

Reasoning quality (2)

The argument's structure is clever and the body hedges appropriately. Two inconclusive comparisons are read in opposite directions by sex. The paper acknowledges that behavior-linked causes in the men's uninformative group would bias toward equality, then treats equality as causal. The sorting rule is stated and then not followed for husbands' deaths. The time pattern is read backward. Women show the fading effect that prior literature associates with grief (40% excess risk in the first six months falling to 13% after two years); men show a flat 30% that never fades. A permanent elevation looks more like shared risk than a grief shock, yet the paper takes it as support for causation. The power argument is wrong on its face. The closing policy point about undervaluing benefits to widows applies to women, the group the paper says shows no causal effect.

Transparency (3)

Source data are public government files and funding is disclosed. No code or derived dataset is shared, so reproducing the spouse-matching and cause groupings requires re-deriving both. The per-cause test results that produced the uninformative lists are not printed, so no single classification decision can be checked. The limitations paragraph is honest about unobserved divorce and remarriage but silent on the husbands' group failing the stated rule and on behavior-linked causes in the lists. Several numbers differ between text and tables (widowhood rates, the men's equality-test result, a repeated hazard ratio, outcome means copied from one table to another describing groups one-third the size).

Citation integrity (3)

No conclusion rests on an external citation; the core results are the authors' own. Background citations check out where checkable. The reference list contains two wrong-paper entries — the Cox model citation points to an unrelated Nature article, and the only cited prior evidence for a sex difference in bereavement timing points to a study of cognitive assortative mating — plus a duplicated entry counted as two studies. Several risk-factor citations are off-topic (a cocaine review cited for alcohol; prostate bone-loss trials cited for what causes prostate cancer). More important is what's missing: Smith & Zick 1996 used the same by-cause-of-death strategy, Christakis & Allison 2006 used spousal illness as a shock with diagnosis-specific magnitudes, and the cited Schaefer 1995 and Lichtenstein 1998 papers attacked the selection problem with shared-environment and twin designs — none is engaged for what it found about selection versus causation.

Independence (5)

Foundation-funded, academically affiliated, no commercial or advocacy stake. The conclusions serve no interested party: a marriage-promotion advocate would want both results causal, a skeptic would want both to be selection. The interpretive weaknesses point toward ordinary publication incentives, not conflict of interest.

4. What's good

The identifying idea — using the spouse's cause of death to separate shared risk from loss itself — is a real attempt at causal inference with observational data. The men's result holds as far as it goes: widowers die at elevated rates even after wives' deaths from causes unrelated to income or education. The paper volunteers its main weaknesses, shows the raw effect halving as controls are added, replicates with an independent preventability index, and reports what happens when each cause is moved between groups.

5. What's wrong with it

Three problems recur across the design, the measurement, the statistics and the reasoning.

The sorting instrument measures the wrong thing. "Not predicted by income and education" is not "unrelated to the couple's habits," and the uninformative lists are full of behavior-linked causes. The paper says this contamination would push the men's estimates toward equality, then calls equality causal.

The women's result is a small-sample null read as a finding. About 65 deaths, a range of plausible values that includes the men's estimate and the women's informative-cause estimate, a swing from 0.87 to 1.16 depending on which single cause is moved. The body says "not definitive"; the abstract says it "indicates" selection.

The same inconclusive statistical outcome is read two ways. Neither pair of estimates could be told apart. For men that becomes "equal, therefore causal"; for women, "different, therefore selection." The men-versus-women difference itself is never tested and would not clear conventional thresholds if it were.

Secondary: the husbands' uninformative group fails the paper's own classification rule; the death groups differ in how people die, so grief itself may differ; and numbers are inconsistent across text and tables.

6. What would change our mind

• A dataset with baseline smoking, drinking, weight and health status for both spouses, so "uninformative" could be defined against shared behavior rather than income — the same men's result there would substantially raise confidence.
• A sample with several hundred deaths among widows of uninformative-cause husbands; if the estimate stays near 1.0 with a tight range, the women's selection story gains real support.
• A direct test of the sex difference that clears conventional thresholds.
• Separating sudden from prolonged spousal deaths within each cause group; if the bereavement effect tracks trajectory rather than informativeness, the identification collapses.
• Engagement with the twin and shared-environment studies (Lichtenstein 1998, Schaefer 1995) and the later by-cause work (Elwert & Christakis 2008); agreement with those designs would raise confidence, disagreement would lower it.
• Marital-history data that resolve divorce and remarriage; if remarriage dilution is large for men, the true men's effect is larger than estimated.