Jake Bowers is Asst Professor of Political Science and Statistics at the University of Illinois at Urbana-Champaign. His research focuses his methodological work on the use of research design as a basis for statistical and causal inference.
Did the London bombing of July 2005 cause an increase in the civic activity of UK residents in the months shortly afterward? And, did newspaper advertisements to increase turnout in a number of 2005 mayoral elections in the USA have the hoped for effects? Research design in these two cases enables us to assess causal effects using simple, transparent, and valid tests. Valid tests however, may not be powerful. One can enhance the power of tests using covariates that past work has shown to be predictive of outcomes -- after all, social scientists know which variables predict civic activity -- but conventional covariance adjustment via a linear model requires many decisions and assumptions. A complex linear model may diminish the virtues of the simple experiment and simple analysis. In our paper, we develop a principled way to choose linear modeling strategies to maximize the power of the simple tests allowed by experiments. Our approach uses modern machine learning techniques like the adaptive lasso, but we depart from existing practice in that we do not pursue measures of outcome prediction and instead focus on enabling the most powerful treatment effect assessment. We show that one can use linear models to encode what one knows about outcomes without knowing exactly how covariates may predict them. And, we demonstrate that one can choose the specification of such models without prematurely assessing treatment effects. We hope that this methodology enables analysts to produce both valid and powerful tests of causal effects.