WIRED: Weekly Informal REsearch Discussions Spring 2018
Wednesday, April 11th
KMC 8-170
Lunch will be available at 12:15 pm. The seminar begins at 12:30 pm.
"Causal Classification: Outcome Prediction vs. Treatment Effects"
Foster Provost and Carlos Fernandez
Can outcome predictive models perform better than treatment effect models at targeting for treatment effects? Turns out they can, even if we were to use the best possible causal model to estimate the treatment effect and even if there is no confounding. We present theoretical and simulation results showing that outcome prediction models perform better than treatment effect models when the target variable is difficult to predict and the desired outcome is relatively rare. And guess what? An important targeting application near and dear to our hearts falls right in that sweet spot!