ABSTRACT
While most regressions attempt to estimate causal effects of some treatment, regressions can also be used to forecast or predict an outcome, determine the factors that predict an outcome (not necessarily due to causal influences), and adjust outcomes for certain factors (e.g., to gauge relative performance). This chapter discusses the strategies to use when using regressions for these other objectives, in consideration of the sources of bias from Chapter 6. The discussion demonstrates that strategies are quite different across the different regression objectives, in terms of what control variables to include in a model (or, model selection). The case is made that the optimal strategies are:
For forecasting, throw in (nearly) the kitchen sink of variables;
For “determining predictors of an outcome,” use one explanatory variable per model (or, a set of variables representing a categorization, such as race/ethnicity);
For adjusting outcomes, strategies could differ, but be careful about using demographic factors, as it would hold different demographic groups to different standards.
