Covariates and causal effects: the problem of context

This paper is concerned with understanding how causal effects can be identified in past data and then used to predict the future in light of the problem of context, or the fact that treatment always influences the outcome variable in combination with covariates. Structuralist and experimentalist views of econometric methodology can be reconciled by adopting notation capable of distinguishing between effects independent of and dependent on context, or direct and net effects. By showing that identification of direct and net effects imposes distinct assumptions on selection into covariates (i.e., exclusion restrictions) and explicitly constructing predictions based on past effects, the paper is able to characterize the tradeoff researchers face. Relative to direct effects, net effects can be identified in the past from more general data-generating processes (DGPs), but they can predict the future of less general DGPs. Predicting the future with either type of effect requires knowledge of direct effects. To highlight implications for applied work, I discuss why Local Average Treatment Effects and Marginal Treatment Effects of educational attainment are net effects and are therefore difficult to interpret, even when identified with a perfectly randomized treatment.