How’s Life at Home? New Evidence on Marriage and the Set Point for Happiness

The design compares people who married against people who didn't, among those all unmarried ten years earlier, adjusting for their life satisfaction back then. That is better than a raw comparison but leaky. The early score carries only about a third of its weight forward, so most of any stable disposition remains unmeasured. The seven intervening years of jobs, health, and partner formation are uncontrolled. The analysis rests on the roughly 18% of panel members who stayed in the survey across the full decade, with no discussion of who dropped out. The married group in this model has been married at most seven years, yet the estimate is used as evidence about long-term effects. The persistence evidence is a one-time age snapshot that cannot observe anyone's trajectory, and the compositional shift in the unmarried pool (never-married at 25, divorced and widowed at 55) is acknowledged in a footnote and dismissed without a test. The one within-person long-term estimate comes from a subsample dropping 86% of people, with an unreported identifying group and an uncertainty range from a barely detectable effect to a large one. The friendship test compares kinds of marriages measured in the same interview as the outcome. Good practice: the authors check that happier people do marry more (they do), cluster standard errors, and report the stricter specification even while dismissing it.

Coefficients and standard errors are reported throughout, errors are clustered by individual where it matters, and the main effect size (about a third of a standard deviation) is plausible and consistent with prior work. But the central causal estimate depends on a single lagged control whose adequacy is never tested; when a more complete control lowered the estimate by 19%, the authors kept the higher figure as "probably more appropriate" with no discriminating test. The argument that fixed-effects models understate the marriage effect confuses imprecision with bias: rarely changing variables produce wide standard errors, not estimates pushed toward zero, and the paper's own full-sample long-term estimate is near zero with a tight standard error — the opposite of what the "too little variation" story predicts. The restricted-sample estimate (0.35) changes who is in the sample, not just the baseline, and its interval (roughly 0.07 to 0.62) does not support "large." The friendship mechanism is tested as a moderator, not a mediator; weighting by the 47% of married people who name their spouse as best friend, friendship accounts for under 30% of the average marriage effect. For cohabiters with the lag control, the friendship bonus is indistinguishable from zero, contrary to the summary. The world-poll regressions pool many countries per region with no country fixed effects and no reported clustering; the marriage-by-age terms add at most 0.004 to explained variance.

The paper's core claims — causation, persistence, mechanism, refutation of set points — are all about what happens inside one person over time. The evidence is almost entirely differences between people at one moment. The persistence claim is the clearest case: a mid-life peak in the married–unmarried gap is exactly what changing composition of the unmarried pool would produce with no lasting effect, and the ever-married versus never-married check that would partly address this is reported only as "smaller" with no numbers. The best-friend item is a report on marital quality given in the same interview as the satisfaction report; it is close to another measure of the outcome, and a satisfaction score from ten years earlier cannot control for marital quality that developed after the wedding. The set-point conclusion leans on differences in average life evaluation between countries, which say nothing about whether individuals revert after life events; the paper's own footnote notes that Cummins et al. 2014 reconcile set points with lasting circumstance effects. The descriptive supporting claims — the selection effect, the coefficient drop, cohabiters at three-quarters of the married gap, the regional sign pattern — are sound as descriptions of these datasets.

Every discretionary choice resolved toward a larger or more durable marriage effect: keeping the higher of two selection-adjusted estimates, restricting the fixed-effects sample and attributing the larger result to a cleaner baseline rather than a different population, and dismissing footnoted alternatives as "unlikely" without testing them. The abstract says causation holds "even after full allowance is made for selection effects"; one lagged score is a partial allowance, and the paper's own selection estimate confirms selection is present. "Flow largely through friendship" describes a moderation result as a mediation finding, and the numbers show the majority of the marriage effect accrues to people whose spouse is not their best friend. The same story — stable happier people marry more, unhappy marriages dissolve, the mid-life unmarried pool fills with the divorced and widowed — would produce every figure in the paper with no lasting effect of marriage, and the paper does not run the analysis that would separate the two. To its credit, the paper states a falsifiable prediction (adaptation implies the gap should peak at typical marrying ages), attempts an ever/never comparison, and reports that marriage is associated with lower life evaluations in Latin America and Sub-Saharan Africa.

This is an unrefereed NBER working paper with no pre-registration, no code, no shared data (two of three datasets are restricted-access), no conflict-of-interest statement, no sample-size reasoning, and no limitations section. The two discretionary specification choices that produce the headline "no adaptation" finding — dropping the stricter selection control and restricting the fixed-effects sample — are neither pre-committed nor tested. The identifying sample for the long-term estimate is not reported, nor are counts per age bracket in the charts, nor which waves carry the closest-friend question (the friendship regression grows from about 15,000 respondents in one wave to 50,000 observations). Caveats appear as scattered footnotes, each followed by a dismissal. Funding is disclosed. Question wording and scales are stated, which is adequate for secondary survey analysis.

The paper names and fairly describes the studies it argues against (Lucas et al. 2003, Clark and Georgellis 2013) and quotes Stutzer and Frey 2006 accurately. Two methodological arguments, however, lean on cited work that cuts the other way. Frijters and Beatton 2012 is cited only for a background correlation, while its main finding — that the age U-shape largely disappears when the same individuals are followed over time — undercuts the paper's argument that pre-marital baselines are the wrong comparison. Boyce 2010's three-fold larger estimate is presented as corroboration without noting that its method was the subject of a published symposium (Greene 2011; Breusch et al. 2011) concluding it produces misleading results. Beck 2001 and Plümper and Troeger 2007 describe loss of precision, not bias toward zero, so they support a claim the paper does not need. Twelve references are listed but never cited, including Stack and Eshleman 1998 (a 17-nation marriage study directly relevant to the international section) and Suh et al. 1996 (adaptation evidence never confronted). The Lim and Putnam 2010 entry gives the wrong journal, and the "only one previous paper" on friendship and marriage is never named.

No financial or industry conflict. Funding is from a non-profit academic institute; Gallup and the UK statistics office supplied data. The pressure is intellectual: the senior author co-edits the World Happiness Report and has spent years arguing that well-being is driven by social context rather than fixed by personality, and the paper's two headline conclusions are what that program predicts. Four citations are to his own work. This is the ordinary situation of a researcher testing his own theory, but each discretionary choice in the analysis resolved in the theory's favor without a test a reader could check. Against that, the paper reports results a pro-marriage advocate would omit: negative associations in two world regions, cohabiters getting most of the benefit, and the cohabiter friendship bonus disappearing under the selection control.