Fernando A. Quintana (Pontificia Universidad Catolica de Chile)
Discovering Interactions Using Covariate Informed Random Partition Models
Combination chemotherapy treatment regimens created for patients diagnosedwith childhood acute lymphoblastic leukemia have had great success inimproving cure rates. Unfortunately, patients prescribed these types oftreatment regimens have displayed susceptibility to the onset ofosteonecrosis. Some have suggested that this is due to pharmacokineticinteraction between two agents in the treatment regimen (asparaginase anddexamethasone) and other physiological variables. Determining whichphysiological variables to consider when searching for interactions inscenarios like these, minus a priori guidance, has proved to be achallenging problem, particularly if interactions influence the responsedistribution in ways beyond shifts in expectation or dispersion only. We propose an exploratory technique that is able to discoverassociations between covariates and responses in a very general way. Theprocedure connects covariates to responses very flexibly through dependentrandom partition prior distributions, and then employs machine learningtechniques to highlight potential associations found in each cluster. Weapply the method to data produced from a study dedicated to learning whichphysiological predictors influence severity of osteonecrosis multiplicatively.