Simulation study on double bootstrap confidence intervals in linear models: case of outliers

The marginal effects in linear models have been of considerable interest in social science. Inferences about marginal effects have relied largely on asymptotic methods which have an assumption that the limiting distribution of the estimator is normal. We introduce bootstrap approach as an alternative way to construct confidence intervals and to estimate the sampling distributions of estimators of marginal effect in linear model. We illustrate the performance of traditional method and bootstrap procedure in case of bad outliers. We make use of double bootstrap procedure for confidence interval estimation. Results indicate that double bootstrap confidence intervals outperform traditional OLS intervals in presence of severe outliers in small samples.