THE EFFECT OF DIFFERENT RELATIVE LOGISTIC REGRESSION GENERATED PROPENSITY SCORE DISTRIBUTIONS ON THE PERFORMANCE OF PROPENSITY SCORE METHODS

Much education research involves evaluating the causal effects of interventions. The propensity score (PS) method, which is often used to account for selection bias, has become a popular approach to facilitating causal inference in quasi-experimental designs. Because the success of the application of PS conditioning methods is dependent on the estimated propensity scores, the relative PS distribution between the treated and control groups could be an important yet not well-known factor. The primary goal of this dissertation was to explore, via a simulation study, the relations between the relative PS distributions and the performance of selected PS matching methods. The results indicated that PS weighting (without trimming) tends to be robust to a variety of data conditions and produces more accurate and trustworthy TE and SE estimates. The performance of the methods and conclusions were then illustrated through an empirical data analysis using data selected from the Early Childhood Longitudinal Study Kindergarten Class of 2010-11 study, assessing the effect of having home computers on first grade students’ math achievement