Sample Fit Reliability

Abstract: Researchers frequently test and improve model fit by holding a sample constant and varying the model. We propose methods to test and improve sample fit by holding a model constant and varying the sample. Much as the bootstrap is a well-known method to re-sample data and estimate the uncertainty of the fit of parameters in a model, we develop Sample Fit Reliability (SFR) as a set of computational methods to re-sample data and estimate the reliability of the fit of observations in a sample. SFR uses Scoring to assess the reliability of each observation in a sample, Annealing to check the sensitivity of results to removing unreliable data, and Fitting to re-weight observations for more robust analysis. We provide simulation evidence to demonstrate the advantages of using SFR, and we replicate three empirical studies with treatment effects to illustrate how SFR reveals new insights about each study.