Estimating causal effects from observational data with a model for multiple bias

Abstract

Conventional analyses of observational data may be biased due to confounding, sampling and measurement, and may yield interval estimates that are much too narrow because they do not take into account uncertainty about unknown bias parameters, such as misclassification probabilities. We used a simple, multiple bias adjustment method to estimate the causal effect of social anxiety disorder (SAD) on subsequent depression. A Monte Carlo sensitivity analysis was applied to data from the Early Developmental Stages of Psychiatry (EDSP) study, and bias due to confounding, sampling and measurement was modelled. With conventional logistic regression analysis, the risk for depression was elevated in the presence of SAD only in the older cohort (age 17-24 years at baseline assessment); odds ratio (OR) = 3.06, 95% confidence interval (CI) 1.64-5.70, adjusted for sex and age. The bias-adjusted estimate was 2.01 with interval limits of 0.61 and 9.71. Thus, given the data and the bias model used, there was considerably more uncertainty about the real effect, but the probability that SAD increases the risk for subsequent depression (OR > 1) was 88.6% anyway. Multiple bias modelling, if properly used, reveals the necessity for a better understanding of bias, suggesting a need to conduct larger and more adequate validation studies on instruments that are used to diagnose mental disorders.