Using independent covariates in experimental designs: Quantifying the trade-off between power boost and Type I error inflation