Fig. 1

Benchmark procedure schematic. The input data are continuous vectors with known class assignments. Missing values are introduced completely at random. GMMs were then fit to the incomplete data in several ways: 1. by using MGMM, which allows for missing values and arbitrary covariance structures; 2. by using MixAll, which allows for missing values but assumes a diagonal covariance structure; 3. by imputing the missing values, then fitting a standard GMM. The GMMs were evaluated based on the adjusted Rand index between the predicted and true cluster assignments. This procedure was repeated \(N=\) times