Use of Logistic Regression to Combine Two Causality Criteria for Signal Detection in Vaccine Spontaneous Report Data

PURPOSE We evaluated the use of logistic regression to model the probabilities of spontaneously reported vaccine-event pairs being adverse reactions following immunization (ARFI), using disproportionality and unexpectedness of time-to-onset (TTO) distributions as predictive variables and the presence of events in the global product information as a dependent variable. METHODS We used spontaneous reports of adverse events from eight vaccines and their labels as proxies for ARFIs. Three logistic regressions were built to predict ARFIs based on different combinations of the proportional reporting ratio (PRR; disproportionality measure) and two Kolmogorov-Smirnov (KS) tests ('between vaccines' and the 'between events') of TTO distribution: model 1, using the PRR estimate and its 95% lower confidence interval (CI) limit; model 2, using the p values of the two KS tests; and model 3, using the PRR (point estimate and lower CI limit) and both KS tests. The performance of the regressions (model fit statistics, calibration, and discrimination) was measured on 100 bootstrap samples. RESULTS Model 3, using two quantified causality criteria, provided the best performance for all measures. The p value of the 'between vaccines' KS test was the most significant predictive factor. Model 1 had the worst performance. CONCLUSIONS Logistic regression allows estimation of the probability of a vaccine-event pair being an ARFI using two causality criteria at the population level assessed in spontaneous report data: the strength of association (disproportionality measure) and temporality (TTO distribution tests). Logistic regression combines and weights these causality criteria based on their respective ability to predict known safety issues.