![]() | Only 14 pages are availabe for public view |
Abstract When conducting an impact assessment, a pivotal step is to measure the Counterfactual. In non-experimental settings, a comparison group is formed to provide an estimate of the counterfactual. One of the major challenges encountered when selecting such group is the presence of selection bias. This study aims to compare between two impact assessment methods, namely Propensity Score Matching (PSM) and Instrumental Variable (IV) in estimating the Average Treatment Effect on the Treated (ATT) under two types of selection bias; omitted variable bias and hidden bias. In the absence of real data, Monte Carlo simulation is used to generate data under two scenarios whereby different conditions are manipulated. Under each scenario, a total of 500 replications are generated to compute the ATT. Three matching algorithms, namely nearest neighbor matching, nearest neighbor within caliper and kernel matching are explored to form pairs between intervention and comparison groups. Under each method, three models with different number of variables, each varying in their association with the outcome and intervention, are used. The ATT is reported on the odds and risk difference scales. The performance of these methods is examined using Relative Bias (RB) and Mean Squared Error (MSE) |