Statistics Against Fate

The average preference among women for slightly older partners doesn’t prescribe a choice to any particular woman; a raised divorce rate under a certain style of behaviour doesn’t assign divorce to every couple. But individuality doesn’t abolish probability either. Group data set a starting estimate, which should then be updated with the history of the specific person.

In Bayesian language, which usually sounds more frightening than the idea itself, general statistics are the prior expectation. If only a few facts about a person are known, base rates are useful: they protect against stories that are too easy to remember and against confidence built on one familiar example. Then your own observations arrive. A partner keeps promises for years, can admit a mistake, doesn’t use dependence as leverage and stays reliable in a crisis — that data should weigh more than a general study about his age, sex or profession. And conversely, a flattering self-description and membership of a “good” group don’t override repeated specific behaviour. Statistics aren’t there to stop you seeing the person, but so that observation doesn’t start from nothing and the general forecast gets replaced by the particular one in time. Summary analyses of longitudinal research agree with this only in part: individual differences in relationship satisfaction persist noticeably over time, and the strongest indicators of current relationship quality are often found inside the couple — in the partner’s perceived commitment to the relationship, in appreciation, conflict and sexual satisfaction. Changes in quality over time are predicted much less well by the same models.

An average is most useful when personal information is almost entirely absent. As personal information accumulates, its weight is reduced by actual conduct, by repetition and by context. An exception doesn’t abolish the distribution, and a distribution doesn’t replace the person you can observe. The error comes when a single happy example is declared a universal rule, or a group probability is turned into a verdict on someone whose history can already be studied directly.

A descriptive frequency doesn’t create a norm. The average number of sexual encounters doesn’t oblige a couple to raise their figure, a statistical link between marriage and wellbeing doesn’t turn every marriage into therapy, and a common distribution of housework doesn’t make it fair. A normative decision needs a goal and a principle of distribution. A variable can raise the average outcome and worsen the position of a particular participant: specialisation can increase both total income and dependence; a high cost of divorce can protect specific investments and hold people inside violence; control can lower one risk and produce another.

Couple relationships complicate the forecast further, because the object reacts to the forecast itself. Someone told that their relationship is doomed may stop investing and help the prediction come true; a couple that sees a risk may change the rules and make the outcome less likely. Unlike a planet’s orbit, a union contains participants who read the model, take offence at it, use it as a weapon or rebuild their behaviour around it. Here the formula enters the environment it describes. “Whoever loves less holds the power” can create the coldness it then takes as proof of its own correctness; “real love takes no effort” turns ordinary coordination work into a sign of a wrong choice. A good model has to account for the prediction and for the consequences of people starting to live as though it were a law.