Abstract
Polygenic risk scores (PRS) have potential utility as biomarkers of psychiatric disorders. However, the potential utility of some PRS has been much clearer than others. The schizophrenia PRS has been consistently associated with diagnosis, symptom severity, and other correlates of schizophrenia. Yet despite the close genetic correlation (r(g) = 0.69) between bipolar disorder and schizophrenia, the bipolar (BP) PRS has been inconstantly associated with clinical course and outcomes. We hypothesize that common sample selection strategies induce collider bias between the SZ and BP PRS, in turn moderating the association between the BP PRS and clinical outcomes. Collider bias is illustrated in three effects. First, it is shown that clinical characteristics used in sample ascertainment (i.e. case status, treatment history) are a function of the SZ and BP PRSs. Second, selecting on these clinical characteristics biases the correlation between the BP and SZ PRSs. Third, selection on these clinical characteristics in turn moderates the association between the BP PRS and clinical outcomes. Effects were tested in three samples: UK Biobank (N = 337,420), a population-based sample; PsyCourse (N = 1,594), a case-control cohort of individuals with mood and psychotic disorders; and the Suffolk County Mental Health Cohort (N = 378), a first-admission psychosis cohort. In all three samples, SZ and BP PRSs were significantly correlated with case status or treatment history. In addition, correlations between SZ and BP PRS were biased in samples selected by case status or treatment history. Finally, conditioning analyses on case status moderated, and in some cases reversed, associations between the BP PRS and clinical outcomes. It is important to understand the impact of this and other forms of selection bias in evaluating PRS as biomarkers of psychiatric disorders, particularly when the intended application is populations enriched for high genetic risk.