Settings where one treatment is evaluated for many different outcomes are ubiquitous in applied economics, especially in field experiments and health economics. Typically, these outcomes are also correlated with each other, which makes selecting those outcomes for which the treatment is most relevant a difficult task. We propose a ‘reverse’ model selection procedure using the sparse-group lasso that evaluates all outcome equations simultaneously and regularises the treatment effect in some (or all) equations to zero. We account for three different kinds of correlation: correlation in the response to covariates, correlation of the value of groups of covariates and residual correlation in the error vector. We also show that the distribution of false positives is different depending on the correlation structure of the outcome vector. We show that our selection procedure is less conservative than traditional step-down methods in settings with many small effects within groups. We apply our estimator to field experiment data in a development context.
Professor Lena Janys is Assistant Professor for Econometrics the University of Bonn. Research interests are both applied and theoretical micro econometrics with applications in health economics.