Estimating classification accuracy for latent variable mixture models