The manuscript focuses on effects in nonrandomized studies with two outcome measurement occasions and one explanatory variable, and in which groups already differ at the pretest. Such study designs are often encountered in educational and instructional research. Two prominent approaches to estimate effects are (1) covariance analytical approaches and (2) latent change-score models. In current practice, both approaches are applied interchangeably, without a clear rationale for when to use which approach. The aim of this contribution is to outline under which conditions the approaches produce unbiased estimates of the instruction effect. We present a theoretical data generating model in which we decompose the variances of the relevant variables, and examine under which data generating conditions the estimated instruction effect is unbiased. We show that, under specific assumptions, both methods can be used to answer the general question of whether instruction has an effect. Another implication from the results is that practitioners need to consider which underlying data generating assumptions the approaches make, since a violation of those assumptions will lead to biased effects. Based on our results, we give recommendations for preferable research designs.