Bootstrapping the Hausman Test in Panel Data Models

The Hausman test statistic in panel data models is asymptotically pivotal under the null hypothesis. It could therefore be refined using the bootstrap resampling technique. Edgeworth expansion shows that coverage of a bootstrap version of the Hausman test is second-order correct. The asymptotic vis-à-vis the bootstrap version of Hausman test are compared by Monte Carlo simulations. Results show that the bootstrap version has around 20% lower error in coverage at the null hypothesis. If size-corrected, it also outperforms the power of the asymptotic Hausman test by almost 10% if fixed effects are ”weak”. Results are robust on parameters of data generating process. 1e-mail: [email protected], tel.: +386 1 2518704, fax: +386 1 4256870 2e-mail: [email protected], tel.: +386 1 2521688, fax: +386 1 4256870 1