Çoklu Tedavi Etkinliğinin Karşılaştırılmasında Propensıty Skor Ağırlıklandırması

Abstract

ABSTRACT Demir, O. Propensity Score Weighting For The Comparison of Multiple Treatment Effects. Hacettepe University, Institute of Sciences, Ph.D. Thesis, Ankara, 2015. In non-randomized quasi-experimental or observational studies, unbalance may occur in treatment arm assignments due to individual differences. This unbalance leads to bias in determining treatment efficiency. This study suggests using propensity scores to eliminate unbalance. This unbalance’s elimination is ensured by weigthing on propensity scores which are the assignment probabilities of individuals in treatment arms. When there are more then two treatment arms, generalized boosted model and multinomial logistic regression are used to estimate the propensity scores. Performances of these methods on unbalance are compared via a Monte Carlo simulation. Also, treatment effects on outcome are investigated in these methods using causal effects. In the first part of simulation study, which is conducted for the comparison of unbalance according to methods, generalized boosted model is more successful than multinomial logistic regession. In the second part of the simulation study where the treatment effects on outcome are compared,, it is found that the treatment effects of two methods are different from each other. In the application part of the study, significant effect of age, MCV and FOLAT variables on vitamin B12 obtained from unweighted model which is formed with 8 covariates and living place is taken as treatment variable have changed after weighting. As a results; in situation that random assignment can not be done and the treatment effect on the dependent variable is investigated, under various scenarious, information about results from methods are provided. Key Words: Generalized boosted model, multinomial logistic regression, propensity score, propensity score weighting, average treatment effect (ATE), average treatment effect on the treated (ATT).