The effect of stratified sampling of items on the estimation of test score distribution parameters by multiple matrix sampling was studied. Item difficulty and/or interitem correlations were the bases of stratification. Various item iniverses were created by computer simulation and sampled according to several plans. The results indicate that stratification of items does not consistently improve the stability of parameter estimation. The results also show that the variance estimate used in many studies is biased for some item universes when difficulty stratification is used. A variance estimate developed in the current study removes this bias. (Author)