MC1 USE OF A MODIFIED RECURSIVE PARTITIONING AND AMALGAMATION (RECPAM) TECHNIQUE IN OUTCOMES RESEARCH Katz LM, Doyle JJ, Bergemann R, Siegartel LR, Stern LS, Chalfin D, Danis M, Rapoport J, Levy M Analytica International, New York, NY, USA; Analytica International, Loerrach, Germany; Montefiore Medical Center, Port Washington, NY, USA; National Institutes of Health, Bethesda, MD, USA; Mount Holyoke College, South Hadley, MA, USA; Brown University, Providence, RI, USA OBJECTIVES: The recursive partition and amalgamation (RECPAM) technique combines statistical discrete event modeling with decision tree analysis. This method has applications to outcomes research since it integrates advantages of logistic regression, survival analysis, and decision tree-building. Prediction of outcomes can be challenging when assessing complex interrelationships between multiple factors on several group levels (e.g., patients and institutions); benefits of such an integrated approach are explored. METHODS: A database of 111,907 ICU patients was randomly partitioned: one sample for model creation, the other for validation. Two generalized linear models predicting the primary outcome [e.g., length of stay (LOS)] used patientand institution-level variables as predictors, respectively. Type-3 sum of squares determined the 10 most significant variables from each. These 20 variables were then predictors in an overall model. To determine pathways recursively, the variable most associated with the outcome (Fstatistic) was used as a splitting point to segregate the sample into groups at the next tree level. Once a variable was used, it was removed from future iterations along the specified path. Nodal outcomes (e.g., survival) were compared within tree levels (chi-square) and those without significant differences were amalgamated. Successive splitting and amalgamating were performed until the sample in each node was reduced to 250, the sample represented was 10,000 combined, or >90% of predictors were represented. RESULTS: The RECPAM-based methodology applied to an ICU database showed that surgery and procedures (any vs. none), and the Simplified Acute Physiology Score (SAPS II) were the most highly associated with survival and LOS. CONCLUSIONS: RECPAM can be used to examine complex interactions, identifying key predictors and streamlining data. This technique allowed for development of a model to determine how combinations of multiple group-level variables affect patient outcomes. MC2 IMPACT OF THE PROPENSITY SCORE ESTIMATION METHOD WHEN MATCHING PATIENTS TO REDUCE RECRUITMENT BIAS IN OBSERVATIONAL STUDIES Riou França L, Payet S, Le Lay K, Launois R REES France, Paris, France OBJECTIVES: To reduce recruitment bias in an observational study by the propensity score (PS) matching method. METHODS: PREMISS is a prospective, multicentric pre-post study aiming at the evaluation of drotrecogin alfa (DA) in the treatment of severe sepsis with multiple organ failure. In observational studies, there is a need to control for recruitment bias. We decided to improve patient comparability by performing a PS optimal matching. PS was estimated using logistic regression after performing multiple imputations to handle missing patient characteristics. Each control was matched to a DA patient so that the sum of the PS distances was minimal. Three PS models were estimated and their performance compared: the first included all patient characteristics, the second only those being unbalanced, the third completed the second with interaction terms. RESULTS: A total of 1096 patients were retained in the whole sample, with strong evidence of recruitment bias. Forty-six initial characteristics were measured. A total of 22.8% of the patients had at least one missing characteristic. Once matched, respectively 76.6%, 79.4%, and 68.2% of patients were retained in the sample for PS models n° 1, 2, and 3. The first PS model performed better in reducing recruitment bias, only two indicators are slightly unbalanced (versus 4 and 5 indicators in the second and third models). In the resulting matched sample, adjusting for organ dysfunction (i.e. the LODS quartiles), survival estimates were similar, at the cost of multiple adjustments, to those obtained in the whole sample. CONCLUSION: Despite the growing popularity of the PS approach, the best method to estimate it remains unclear. In this particular study, simply including all patient characteristics, with no interaction terms, yielded the best results. More complicated processes, implying variable selection and interaction terms, were counterproductive.