Marital status and mortality: The role of health

The joint model is a genuine advance over comparing married and unmarried men, and using full marriage histories rather than current status is a strength. But the design substitutes modelling assumptions the data cannot check for experimental control. Health is unobserved for the great majority of marriage decisions modeled. The four childhood variables that identify the health-to-marriage effect must have no direct effect on marriage timing, yet childhood poverty and low parental education are established predictors of early marriage in their own right. Results move substantially depending on which equations are included: the effect of health on divorce flips from near zero to clearly negative when the marriage equation is added. Never-married men are included only if they head their own household, so the group is pre-selected toward the independent, solitary men the paper later speculates die from isolation. The survival-to-1984 correction the appendix calls necessary is neutralized by fixing the relevant parameter at zero. The design supports the descriptive associations but not the causal conclusions.

The machinery — continuous-time hazard models, an ordered probit for the five-category health scale, full-information maximum likelihood with correlated random effects — is competent and appropriate. The problem is what is inferred from it. The conclusions about bias direction and "zero net benefit" come from comparing coefficients across two specifications on the same data; of six marital-status health comparisons, five differ by less than one standard error and none is formally tested. The never-married health deficit moved from about −0.15 to about −0.10 while its uncertainty doubled, and the paper reads this as proof that first marriage brings no benefit. The divorced mortality mediation claim rests on 20 deaths, with the adjusted estimate's confidence interval comfortably including the unadjusted value. Instrument strength and overidentification tests are not reported. The robustness check for the exclusion restrictions is described as "affected only mildly" without a single number. The correlation between marriage and dissolution heterogeneity is 0.96, close enough to the boundary that the two components are barely distinguishable. Transitory health shocks — 36% of health variance by the paper's own estimate — are excluded from the endogeneity correction entirely.

The construct at issue is "a man's perceived health at the time he decides to marry." This is observed for zero marriage events: even inside the 1984–90 window the model uses a long-run health index rather than the actual report, and for most marriages the index is childhood background projected forward. The paper's own exploratory two-step analysis shows the predicted-from-background part of health and the person-specific residual part enter the marriage equation with opposite signs; the final model forces them to share one coefficient, so the "adverse selection" result is carried by exactly the component most exposed to the alternative explanation. The mortality mediation claims are stretched rather than disconnected: coefficient shrinkage is read as "divorce worsens health," but the paper's own dissolution results show poor health raises separation risk, and the 370 separations inside the health window — where before-and-after trajectories could have been examined — were not used that way. Self-rated health is treated as equally valid across groups, yet never-married men rate their health about the same as married men while dying at 2.6 times the rate, which is what miscalibration would look like. The descriptive associations hold; the headline construct does not.

The three-way framing — protection, positive selection, adverse selection — is a real contribution, and the observation that a protective effect of marriage creates an incentive for the unhealthy to marry is correct and worth making. But the conclusion that men "base their marriage decisions in part on the potential health benefits" attributes motivation that nothing in the data can speak to, and the authors acknowledge the model cannot separate a man's choice from his attractiveness to partners. Failures to detect an effect are repeatedly stated as demonstrations that the effect is zero. The two-step results that contradict the joint model's structural constraint are presented in a footnote as confirmation of it. The summary states that unhealthy men "remain married longer" while the results section says healthier men have lower divorce risk; the results section matches the estimates. The abstract carries none of the caveats that appear in the body. Specification sensitivity is interpreted as the fuller model correcting the simpler one rather than as fragility.

The data source is public, funding is disclosed, the survey question is quoted verbatim, and the model is described mathematically in an appendix. Coefficients and standard errors are reported where results are printed. Against that: no code, no software named, the full coefficient tables are "available on request," and the robustness checks, the joint test of mortality heterogeneity loadings, and the duration-dependence test are described in prose with no statistics. There is no limitations section; the caveats that appear concern measurement (self-rated health is one measure among many) rather than whether the central estimates can bear the weight placed on them. The paper never reports how many never-married men were excluded for not heading a household. For 1996 this is typical, but the novel result cannot be reproduced from what is printed.

Citations are used honestly and the literature review is unusually balanced: the paper cites the cross-national studies supporting positive selection and then Goldman 1993, which undercuts them; it reports Fu and Goldman 1994's finding that current health barely predicts marriage rather than burying it. None of the core findings depend on a cited source. Weaknesses are thin sourcing — an unpublished working paper, a mimeograph, "various medical studies" with no citation for one of three proposed protective mechanisms — and every methodological citation is the authors' own prior work. The validity of self-rated health as a mortality predictor rests on Idler and Kasl 1991 alone when other sources existed.

Funded by two National Institute on Aging grants; both authors at RAND, a nonprofit. No commercial interest turns on whether married men live longer, and the conclusions (no first-marriage benefit, unhealthy men marry sooner) are not what a marriage-promotion advocate would want. The only structural stake is methodological: the joint-estimation approach is the authors' own research program, later sold as commercial software, and the paper states it "reveals more than" the standard two-stage alternative. The design includes nothing that could show the method performing badly, and the specification sensitivity is interpreted in the method's favor.