Image

Health Economics Seminar with Amanda Kowalski

Talk Title: Counting Defiers in Health Care: A Design-Based Model of an Experiment Can Reveal Evidence Against Monotonicity (joint with Neil Christy)

Amanda Kowalski, the Gail Wilensky Professor of Applied Economics and Public Policy at the University of Michigan Department of Economics, is a health economist who specializes in bringing together experiments, models grounded in context-specific knowledge, and econometric techniques to answer questions that inform current debates in health policy.

Professor Kowalski’s recent research analyzes experiments and clinical trials with the goal of designing policies to target insurance expansions and medical treatments to individuals who will benefit from them most. Her previous research has explored impacts of health insurance through Medicaid expansions, the Affordable Care Act, the Massachusetts health reform of 2006, and employer-sponsored plans.  She has also examined impacts of health spending on at-risk newborns.

Paper info: We show that a design-based model of an experiment with a binary intervention and outcome can reveal empirical evidence against a “monotonicity” assumption that the intervention affects all subjects in weakly the same direction. The canonical sampling-based model cannot reveal evidence against monotonicity, but we show that design-based models and other sampling-based models can. We use statistical decision theory to propose a maximum likelihood decision rule that does not assume monotonicity and establish conditions for its optimality.  Under these conditions, the performance of our rule relative to a design-based rule that assumes monotonicity increases with the sample size across all possible experiments with 4 to 40 subjects. In a given experiment, we quantify evidence against monotonicity with a likelihood ratio.  With the aid of figures that we develop to visualize potential outcomes, we illustrate evidence against monotonicity in a real experiment that examines the impact of a health care intervention. Even though the experiment shows a large and statistically significant average effect in one direction, our rule reveals positive counts of compilers who respond in that direction and defiers who respond in the other.