Attrition occurs when study participants who were assigned to the treatment and control conditions do not provide outcome data and thus do not contribute to the estimation of the treatment effects. It is very common in experimental studies in education as illustrated, for instance, in a meta-analysis studying "the effects of attrition on baseline comparability in randomized experiments in education" (Valentine & McHugh, 2007) that found that 119 of 367 randomized education experiments reported student-level attrition. Shadish et al (1998) called attrition the Achilles' heel of randomized experiments. Attrition reduces statistical power by decreasing sample size. It compromises external validity when those who do not contribute data are unrepresentative of the original sample and thus degrade the representativeness of those who remain in the sample (Orr, 1999). The main purpose of this study is to elaborate a model of the relationships between attrition and effect estimates and to use that model to guide Monte Carlo simulations that examine the sources and magnitude of attrition bias under various assumptions for randomized experiments and cluster randomized experiments. This study modeled the sources of attrition bias under various assumptions for completely randomized controlled trials (RCT) and (to be provided by the time of the SREE meeting) cluster randomized controlled trials (CRT). The overall bias is associated with both the overall attrition rate and the differential attrition rate, and both the overall and differential correlations between y and z for the treatment and control groups. In addition, these results show that bias can be reduced by including baseline covariates in the impact estimate model if those covariates are correlated with both the latent propensity to respond and the outcome variable. (Contains 3 tables.)