Causal inference is central to educational research, where in data analysis the aim is to learn the causal effects of educational treatments on academic achievement, to evaluate educational policies and practice. Compared to a correlational analysis, a causal analysis enables policymakers to make more meaningful statements about the efficacy of educational treatments. The fundamental problem of causal inference is that, at a given time, each subject can be exposed to only one of the treatments (Holland, 1986). Causal inference becomes inaccurate whenever data violate certain assumptions that are often made in practice, including: (1) the usual assumption of no outliers in the potential outcomes, (2) the typical assumptions that the treatment assignments have no outliers, no hidden bias (e.g., Rosenbaum, 2002), no confounding, and satisfy the Stable Unit Treatment Value Assumption (SUTVA; Cox, 1958); (3) the usual assumption that the missing data values are either missing-at-random (MAR) or missing-completely-at-random (MCAR) (Little & Rubin, 2002; Ibrahim, Chen, Lipsitz, & Herring, 2005), and (4) the usual assumption that parameter estimation requires no penalty for the absolute size of regression coefficients. To address the four open issues of causal modeling, the authors introduce a Bayesian semiparametric causal model, which provides a semiparametric approach to the full Rubin (1978) Causal Model. The paper presents their semiparametric causal model in full detail. The authors then illustrate this model through the analysis of data from the Progress In International Reading Literacy Study (PIRLS), to infer the causal effects of a writing instructional treatment on the reading performance of low-income students. This analysis is performed in a typical context of an observational study where SUTVA is potentially violated by the interference of subjects within each classroom, with many covariates describing the student, teacher, classroom, and school, where hidden bias and confounding can be present, and where there are missing covariate, treatment assignment, and potential outcome data, that can either be randomly (MCAR or MAR) or nonignorably missing. (Contains 3 tables and 3 figures.