Randomized Experiments and Randomization Inference. Workshop, Mannheim Centre for European Social Research, September 16, 2019.
Randomization inference is a design-based approach to hypothesis testing, which relies on minimal assumptions and enables the researcher to "analyse as you randomize". Randomization inference considers what would have happened under all possible random assignments (all possible ways of assigning N number of units to treatment and control). Against the backdrop of all possible random assignments, is the actual experimental result unusual, and how unusual is it? Randomization inference is flexible and allows for the test of different sharp hypotheses, using a variety of test-statistics to obtain p-values, which have an intuitive interpretation: the share of random assignments that produce a test statistic as large or larger than the statistic obtained from the realised experiment. Randomization-inference-based p-values can differ from p-values obtained from conventional tests if samples are small and/or if test-statistics are not normally distributed. During the workshop, building on the potential outcomes framework, I will introduce participants to the logic of randomization inference, and discuss applied examples both on the white board and using the ri2 package in R.
Florian Foos is an Assistant Professor in Political Behaviour in the Department of Government at the London School of Economics and Political Science (LSE). His research focuses on partisan election campaigns, including electoral mobilization, opinion change and political activism of politicians. His methodological expertise includes the design, conduct, and analysis of randomized field experiments as well as natural and quasi-experiments.