| Popis: |
We propose to develop power calculation methods that will improve inference in empirical social science research. This will enhance transparency by building stronger links between study design, pre-analysis plans, and ex post analyses. Currently, power calculations are missing from most pre-analysis plans and papers. More problematic, many power calculations are based on existing formulas and canned packages. These methods are designed for independent and identically distributed errors, contrary to the modern practice of clustering standard errors to allow for arbitrary dependence among groups of observations. Moreover, even when researchers employ simulation-based power calculation methods accounting only for one-way correlated error structures, when using panel datasets, these assumptions are likely incorrect, leading to over-rejection of null hypotheses. To our knowledge, there exists no work in economics assessing the impact of these issues on experimental design and transparency. We propose deriving analytic results to fully characterize these problems, investigating the extent to which they matter in the existing literature using real data and Monte Carlo simulations, and creating open-source statistical packages to allow researchers to easily perform cluster-robust power calculations and ex post analyses. |
