A Bootstrap Approach to Time Invariance in Panel Data

Researchers in political science are frequently interested in hypothesis testing in panel data environments with variables that never or very rarely change over time. This paper shows that common corrections for serial correlation in linear and generalized linear models do not perform well in these circumstances, leading to false positives as much as 50% of the time for rarely changing variables and having widely varying performance depending on the severity of temporal dependence. I propose a bootstrap approach that resamples individual panels and show via Monte Carlo simulation that it achieves correct coverage for rarely varying and time invariant variables. I conclude with a reanalysis of three recent papers in top journals