There is a longstanding debate on whether the analysis of covariance (ANCOVA) or the change score approach is more appropriate when analyzing non-experimental longitudinal data. In this article, we use a structural modeling perspective to clarify that the ANCOVA approach is based on the assumption that all relevant covariates are measured (i.e., covariates are sufficient to remove confounding) and provides biased estimates of the treatment effect in the presence of unmeasured confounding variables. By contrast, the change score approach offers the option of controlling for unobserved confounders but relies on strong assumptions about the effects of these unobserved confounders and does not allow for dynamic causal relationships (i.e., pretest affects treatment variable). Furthermore, we discuss four issues that received less attention in previous research: (1) the inclusion of additional covariates in the ANCOVA and change score approaches, (2) the correction for measurement error in the predictor and treatment variables, (3) multilevel designs when the treatment is a cluster-level variable, and (4) the estimation of treatment effects in the presence of heterogeneity and nonlinearity. Implications for the analysis of non-experimental longitudinal data in educational and psychological research are discussed.