OBJECTIVES: To develop and test a three-way propensity score matching algorithm to provide pseudo-randomization of subjects into three groups to allow for comparable groups in a retrospective study. METHODS: Logistic regression using the generalized multinomial logit linking function was used to calculate estimates of the propensity score: the probability of having received three putatively interchangeable drugs from demographic (Race, Gender, Age) and comorbidities (Charlson Comorbidities Index) in a large, retrospective database. The most costly drug was used as the reference group, and the probability of each treatment group having received the reference drug was retained as the propensity score. In the initial analysis 23,912, 4,789, and 4,318 individuals were available in the three treatment regimens. Random subsets of ¼ and 1/10 the original sample were constructed for the purpose assessing multi-group propensity score matching (PSM) effectiveness in constructing comparable groups via pseudo-randomization with varying starting sample sizes. PSM was conducted using calipers ranging from 8 digits to one digit of propensity score. Assessment of among-group differences before and after PSM were conducted using Chi-square tests for categorical variables and GLM analysis, with difference scores and their confidence intervals for continuous variables. RESULTS: For all sample sizes, prior to propensity score matching, significant differences existed among the three treatment regimens for all variables: gender, race, age and comorbidities. Following PSM there were 3381 matched triplets in the full sample. There were no significant differences among groups for gender, age or comorbidities; there were significant but tiny differences that remained for racial representation. In the smaller samples, 966 and 416 matched triplets were retained. There were no significant differences on any variable. CONCLUSIONS: Logistic regression using a generalized multinomial logit link appears to provide a good propensity score from which pseudo-randomization into three groups can be performed in a retrospective sample.