Preprint characterises order-bias in correlation estimates for assortative mating studies

A preprint posted to bioRxiv examines a statistical problem specific to studies of assortative mating — the tendency of individuals to pair with partners who resemble them on a given trait. Such studies commonly quantify parent–parent similarity using Pearson correlation, but the assignment of which parent is labelled 'first' or 'second' within a pair is often arbitrary.

The authors demonstrate that this arbitrariness introduces order bias into standard Pearson correlation estimates when the paired variables are exchangeable — that is, when the joint distribution is symmetric with respect to reordering. If one variable has a systematically lower expectation than the other (for example, paternal versus maternal values for a trait that differs on average between sexes), the bias can be substantial and directional rather than random noise.

The preprint characterises the mathematical form of this bias and proposes corrected correlation estimators that are robust to within-pair reordering. The work has implications for the interpretation of published assortative mating estimates in humans and other species, and for the design of future studies examining social homogamy, genetic assortative mating, and partner similarity in complex traits such as height, educational attainment, or polygenic risk scores.

This work is primarily relevant to statistical geneticists, quantitative geneticists, and researchers working on the genetic architecture of complex traits where partner correlation is a meaningful parameter. It has not yet been peer-reviewed.