The first variance estimation method uses U-statistics to provide unbiased estimates of the variance components [4]. This method lacks a positivity constraint and can lead to negative estimates of variance components and total variance. The second variance estimation method uses the non-parametric maximum likelihood estimate (MLE) of the distribution of readers and cases, which is the empirical distribution of readers and cases [2]. Efron and Tibshirani also refer to this estimation method as the “ideal” bootstrap. The MLE estimate of variance components and total variance cannot go negative. The tradeoff for positive variance estimates is a positive bias.
