Estimates of treatment effects using observational data can be biased due to confounding, model misspecification, and other reasons. A placebo test offers a complementary diagnostic for evaluating these threats to inference by checking for a relationship that should be found in the data if the main estimates were biased, but should be absent otherwise. Drawing on a comprehensive survey of recent empirical work in political science, this paper defines placebo tests, introduces a typology of tests, and analyzes what makes them informative (both in ideal and non-ideal circumstances). We discuss examples of placebo tests that effectively address different types of bias; we also point out tests and types of tests that we argue are largely uninformative, and we highlight the problem of null hacking in the design of placebo tests.