Using probability of drug use as independent variable in a register-based pharmacoepidemiological cause-effect study-An application of the reverse waiting time distribution.

BACKGROUND: In register-based pharmacoepidemiological studies, each day of follow-up is usually categorized either as exposed or unexposed. However, there is an underlying continuous probability of exposure, and by insisting on a dichotomy, researchers unwillingly force a nondifferential misclassification into their analyses. We have recently developed a model whereby probability of exposure can be modeled, and we tested this on an empirical case of nonsteroidal anti-inflammatory drug (NSAID)-induced upper gastrointestinal bleeding (UGIB).

METHODS: We used a case-controls data set, consisting of 3568 cases of severe UGIB and 35 552 matched controls. Exposure to NSAID was based on 3 different conventional dichotomous measures. In addition, we tested 3 probabilistic exposure measures, a simple univariate backward-recurrence model, a "full" multivariable model, and a "reduced" multivariable model. Odds ratios (ORs) and 95% confidence intervals for the association between NSAID use and UGIB were calculated by conditional logistic regression, while adjusting for preselected confounders.

RESULTS: Compared to the conventional dichotomous exposure measures, the probabilistic exposure measures generated adjusted ORs in the upper range (4.37-4.75) while at the same time having the most narrow confidence intervals (ratio between upper and lower confidence limit, 1.46-1.50). Some ORs generated by conventional measures were higher than the probabilistic ORs, but only when the assumed period of intake was unrealistically short.

CONCLUSION: The pattern of high ORs and narrow confidence intervals in probabilistic exposure measures is compatible with less nondifferential misclassification of exposure than in a dichotomous exposure model. Probabilistic exposure measures appear to be an attractive alternative to conventional exposure measures.