Studying the Strength of Prediction Using Indirect Mixture Modeling: Nonlinear Latent Regression with Heteroskedastic Residuals