Fisher’s randomizationmode of statistical inference, then and now.

How is statistical inference possible when n = ? How can we infer without a sample from a population? How should we choose methods for assessing causal claims when we have low information (like a small sample, a binary outcome, a multilevel design with few clusters, or a weak instrument)? R. Fisher answered these questions in showing that valid small sam-ple hypothesis tests are possible, inference does not require a population, and choices about assessing causal effects can arise from design. is paper reframes and extends Fisher’s method, showing that it is a practical alternative for political scientists. As an example, we show how to assess treatment effects using a field experiment of the effect of newspaper advertising on aggregate turnout with only eight obser-vations. In the end, we produce confidence intervals using linear models, but requiring none of the standard assumptions of linear models to guarantee valid statistical inferences