Collider Bias in Partner Preference Research: A Methodological Extension of Eastwick et al. (2024)

Observational research on partner preference-matching commonly conditions on a sample of (most desired) partners. When examined through the lens of causal inference, this configuration constitutes conditioning on a collider variable. This thesis introduces and applies directed acyclic graphs (DAGs) as a framework for assessing corresponding collider bias within the large-scale cross-cultural dataset of Eastwick et al. (2024, N = 10,358, 43 countries). Building on this, a three-layer methodological framework is developed to distinguish three distinct sources of bias: confounding by measured background characteristics, intertrait confounding among the 35 attributes, and residual collider restriction bias arising from the sampling procedure itself. The trait-by-trait level metric of Eastwick et al. (2024) was replicated, followed by a hierarchical multivariate extension entering all 35 ideal preferences and partner attributes simultaneously across four nested models. Results show that demographic and individual-difference covariates contribute low aggregate variance (H1), that multivariate adjustment substantially attenuates both main effects and preference-matching interactions relative to the trait-by-trait approach (H2), and that idiosyncratic matching effects remain near zero even after full adjustment (H3). The discussion situates these findings within the broader context of the research design. Taken together, the methodological evaluation raises the question whether the original preference-matching hypothesis (H0) is adequately testable under the conditions of the present study design. The conclusion identifies the analysis of unselected peer targets as the most direct empirical path toward disentangling residual collider bias from genuine absence of idiosyncratic matching.