What the divorce-prediction studies actually showed
A programme of laboratory research reported that a short recorded conversation could sort married couples into those who would later separate and those who would not. A re-analysis published in 2001 argued that the headline accuracy was an artefact of how the models were tested.
The design
From the 1980s onward, John Gottman and Robert Levenson recorded married couples discussing an area of continuing disagreement under laboratory conditions. Sessions were short, commonly around fifteen minutes. Trained coders classified facial expression, vocal tone, and speech content in fine-grained units using observational coding systems, while physiological measures such as heart rate and skin conductance were recorded in parallel. Couples were then followed for periods ranging from four to roughly fourteen years, and their marital status recorded at follow-up.
Two papers carry most of the weight of the later discussion: Gottman and Levenson's 1992 report in the Journal of Personality and Social Psychology, which followed a sample of married couples over several years, and a 1998 paper with James Coan, Sybil Carrère, and Catherine Swanson in the Journal of Marriage and the Family, based on newlywed couples observed shortly after marriage.
The four behaviors
The most widely repeated output of this work is a set of four conflict behaviors: criticism directed at a partner's character rather than at a specific action; contempt, meaning expressions that convey disdain or superiority; defensiveness, meaning the rejection of responsibility through counter-complaint; and withdrawal from the interaction, described in these papers as stonewalling. Contempt was reported as the strongest single discriminator between couples who remained married and those who did not.
These are descriptive categories applied by coders to recorded behavior. They were not derived from a theory of causation, and the papers did not test whether the behaviors produce dissolution or accompany it.
The accuracy figure
The published analyses reported classification accuracy above ninety per cent, and in some models higher still. That figure travelled a long way. It was compressed in press coverage and in subsequent popular writing into the claim that the future of a marriage could be read from a quarter of an hour of conversation.
The objection
In 2001, Richard Heyman and Amy Smith Slep published a paper in the Journal of Marriage and Family titled “The hazards of predicting divorce without crossvalidation.” Their argument was statistical rather than substantive.
The models had been fitted with the outcome already known. Variables were selected, weighted, and given cut-off points that separated the two groups within that particular set of couples. A model constructed that way will necessarily describe the sample it was constructed on; the question is whether it performs on couples it has never seen. Heyman and Slep applied the approach to independent data and reported that accuracy fell substantially once the model had to classify a sample it had not been fitted to.
Two further constraints belong in the same paragraph. The samples were small for a prediction problem — dozens to low hundreds of couples — and they were drawn largely from one region of the United States, skewing white, middle class, and heterosexual. Nothing in the design supports extension to populations unlike those. The journalist Laurie Abraham set out the same cross-validation argument for a general readership in Slate in 2010, which is where most non-specialists encountered it.
What survives
The narrower claim holds up better than the headline. Observable behavior during conflict discussion is associated with later separation, and contempt is the category with the most consistent association across studies. That is a statement about a correlation observed in groups of couples over time. It is not an actuarial instrument, and the papers themselves are more careful on this point than their reception was.
What the evidence does not establish
That the future of any individual couple can be read from a recorded conversation. That these behaviors cause separation, since the designs are observational and cannot exclude the reverse direction, in which a deteriorating relationship produces the behavior. That the coding systems generalise to couples unlike those sampled. That a figure obtained without cross-validation describes performance on new cases.
Sources
- Gottman, J. M., & Levenson, R. W. (1992). Marital processes predictive of later dissolution: Behavior, physiology, and health. Journal of Personality and Social Psychology, 63(2), 221–233.
- Gottman, J. M., Coan, J., Carrère, S., & Swanson, C. (1998). Predicting marital happiness and stability from newlywed interactions. Journal of Marriage and the Family, 60(1), 5–22.
- Heyman, R. E., & Slep, A. M. S. (2001). The hazards of predicting divorce without crossvalidation. Journal of Marriage and Family, 63(2), 473–479.
- Abraham, L. (2010). Can you really predict the success of a marriage in 15 minutes? Slate, 8 March 2010.