In order to detect adverse drug reactions, large observational drug safety studies are necessary as randomized clinical trials rarely have enough power. However, in order to obtain reliable results the issue of confounding, especially confounding by indication, should be addressed. We proposed a multiple propensity score, which is an extension of the propensity score, to reduce the bias in a dose-response analysis in a drug safety study. The multiple propensity score has similar properties to the propensity score in Rosenbaum and Rubin.(1) Using the propensity score for bias reduction of the risk ratio was considered. We used the multiple propensity score in a study of the dose-response relationship between diclofenac prescriptions and hospitalization for gastrointestinal bleeding and perforation using a record linkage database. We found that the unadjusted risk ratios were biased downwards due to confounders and that this bias was reduced by using stratification based on the multiple propensity score.