Marriage protection and marriage selection—Prospective evidence for reciprocal effects of marital status and health

The prospective design — following the same women over time rather than comparing different groups at a single moment — is better than the cross-sectional studies that preceded it. The sample comes from a well-established national survey with a high initial response rate.

Splitting the sample by employment status conditions on a variable caused by both health and marital status. Among unmarried non-employed women, a large and growing share were out of work specifically because of health problems. Married non-employed women were disproportionately healthy homemakers supported by a husband's income. This compositional difference alone can produce the appearance of a marriage benefit that doesn't exist in the broader population. The paper never acknowledges this problem.

Beyond this, five-year measurement gaps mean that a woman who divorced in year one and spent four years unmarried counts as "married" for the entire interval. The authors acknowledge omitting several plausible confounders (religiosity, obesity, substance use). Thirty-two percent of the sample was lost by 1988 with no analysis of whether attrition was related to both health and marital status. The income mediation test treats family income — which mechanically includes a husband's earnings — as a variable that might explain why marriage helps, when it is partly just another way of measuring marital status itself.

The regressions are standard and the paper reports effect sizes with confidence intervals throughout, which was good practice for 1996. The national probability sample is large enough for the main comparisons.

Two problems are severe. First, the employment stratification that structures the entire analysis induces collider bias, as described above — a concern the statistical approach cannot address and the paper does not raise. Second, the paper's core interpretive claim — that marriage protects health through income and social support — is presented as a conclusion despite social support never being measured and income mediation failing in one of two time periods.

Multiple findings fail to replicate within the same dataset. The marriage selection effect (healthier women more likely to stay married) appears in the first five-year interval but disappears in the second, with no explanation offered. The income mediation result appears in the second interval but not the first, which the authors attribute to a measurement problem they cannot verify. No corrections were applied for the large number of interaction tests conducted. Several key analyses are described only as "data not shown," including the comparison that supposedly rules out marital disruption stress.

The paper's causal claims consistently outrun what its data can support. "Marriage protection" requires that marriage itself caused the health difference, but the study cannot rule out that unmeasured disadvantages (poverty, discrimination, chronic illness) independently cause both poor health and non-marriage. The five-year gaps between measurements mean the actual "dose" of marriage each woman received is unknown.

Social support, one of the two proposed mechanisms, was never measured. Income mediation produced contradictory results across the two time periods. Family income is a problematic mediator because it mechanically includes a husband's earnings, making the test partly circular. Yet the abstract and conclusion present this mechanism as established.

The marriage selection finding — that unhealthy women were less likely to marry — appeared in only one of two time periods and relies on a health measure (self-reported symptoms like fatigue and dizziness) that may not be visible to potential or current partners. The correlation between self-reported health and subsequent marital status could reflect shared causes rather than health influencing partner decisions.

The paper frames its analysis around two named hypotheses — "marriage protection" and "marriage selection" — and the entire structure is organized to confirm both. But a third possibility — that unmeasured structural disadvantages cause poor health, non-employment, and non-marriage simultaneously — fits the data at least as well and is never tested. The paper's own descriptive data show that unmarried, non-employed women were disproportionately Black, poorly educated, and low-income, exactly the profile where structural disadvantage could drive all three outcomes.

Results that support the hypotheses are accepted at face value; results that don't are explained away. When income mediation fails in the first interval, the paper attributes this to a deficient income measure. When it works in the second interval, it becomes evidence for the mechanism. When the selection effect vanishes in the second interval, the paper calls it unexplained but does not downgrade the claim. The abstract drops the hedging present in the body text, stating that marriage had "beneficial effects on health" — causal language an observational study cannot support.

The most obvious alternative explanation — that stratifying by employment creates a spurious association — is never discussed.

The underlying data come from a publicly available government survey. The health measure and variable coding are described clearly. The paper acknowledges some limitations: unmeasured confounders, the five-year measurement gaps, and the small part-time worker subgroup.

However, the most important methodological problem — collider bias from employment stratification — is never raised. Several supporting claims rest on analyses described only as "data not shown," including the finding that rules out marital disruption stress and the null results for employed women. No analysis code is provided, no power analysis justifies the sample for its many subgroup comparisons, and there is no conflict of interest statement (though none was standard in 1996).

The citation practices are generally competent. The paper is honest about contradictory evidence: it acknowledges studies finding no health-selection effect on marriage and notes its own selection finding's failure to replicate. Background citations are accurately represented.

Two problems stand out. The paper cites Umberson (1992) as evidence that social ties improve health by reducing risky behavior — but Umberson's own results suggest this mechanism primarily benefits husbands (wives monitor their health behavior), with little effect in the other direction. Citing it as support for marriage protecting women's health overstates what the source found. The core finding's only external precedent is a self-citation (Waldron and Jacobs 1989); no independent replication is cited. The paper also omits two relevant literatures available by the mid-1990s: research on marriage quality (unhappy marriages may not be protective) and rising cohabitation (unmarried women classified as having no partner may have had one).

The authors were at the University of Pennsylvania with no industry connection to the subject matter. Funding came from a small university research endowment. Marriage and health is a basic social-science question with no commercial stakeholder. The study used a pre-existing federal dataset the authors could not have designed to favor a particular result.