Fast Exact/Conservative Monte Carlo Confidence Intervals

Monte Carlo tests about parameters can be "inverted" to form confidence sets: the confidence set comprises all hypothesized values of the parameter that are not rejected at level α. When the tests are exact or conservative – as some families of such tests are – so are the confidence sets. Because the validity of confidence sets depends only on the significance level of the test of the true null, every null can be tested using the same Monte Carlo sample, substantially reducing the computational burden of constructing confidence sets: the computation count is O(n), where n is the number of data. The Monte Carlo sample can be arbitrarily small, although the highest nontrivial attainable confidence level generally increases as the number of Monte Carlo replicates increases. When the parameter is real-valued and the P-value is quasiconcave in that parameter, it is straightforward to find the endpoints of the confidence interval using bisection in a conservative way. For some test statistics, values for different simulations and parameter values have a simple relationship that make more savings possible. An open-source Python implementation of the approach for the one-sample and two-sample problems is available.