P1-53 Collider-stratification bias complicates estimation of the strength of risk factors of disease recurrence

Knowledge of factors influencing recurrence risk is essential in the prevention of disease recurrence. In this paper, we show that correct estimation of the strength of such factors is, however, troublesome. We performed a simulation study of the recurrence risk of a fictional pregnancy-related disorder, Y. We assumed that there were four component causes of Y, X1 representing the determinant under study, X2 and X3 representing unmeasured determinants, and X4 representing pregnancy as a necessary condition for developing Y. We stipulated that each woman would become pregnant twice. RR of disease during the first pregnancy for X1+ (presence) vs X1- (absence) was 19.0. Attributable risk (AR) was 0.18. RR of recurrent disease for X1+ vs X1- during the second pregnancy, calculated among women with previous disease, was apparently 1.0, and AR was apparently 0.00. However, we show that real RR and AR were considerably higher (19.0 and 0.95, respectively). Our simulation shows that selection of a study population on the basis of previous disease can lead to underestimation of the strength of recurrence risk factors. The bias involved is a form of collider-stratification bias. We urge for extra caution in the interpretation of studies of recurrence risk factors.