The standard regression technique for modeling binary response variables in education research is logistic regression. The odds ratios from these models are used to quantify and communicate variable effects. These effects are sometimes pooled together as in a meta-analysis. We argue that this process is problematic as odds ratios calculated from different models and studies are not directly comparable. As an alternative, we recommend the linear probability model for computing risk differences and a Poisson working model for computing risk ratios. These effect sizes are comparable across models and studies. However, standard approaches for estimating these models have their problems, hence, we motivate and present modified estimation techniques for estimating these models that mitigate these problems.