Fig. 3.

Effects of design sparsity and having more or less balanced designs for first and second level model parameter estimation accuracy. All plots show log root mean squared errors (RMSQE) comparing ground truth vs. Bayesian model parameter estimates of intercept and slope for first (individual) and second (group) level. We manipulated the noise variance to follow σ² = 0.01/(1 + 25 × (p − 1)²), with p = 1, …, 10 indicating the noise level. Red vs. blue lines indicate errors for large vs. small individual differences as a function of the first level noise parameter. Stronger noise mainly increases first level model errors. The log RMSQE is depicted for different designs with independent variation of loss of balance and sparsity. These results were obtained from averaging over 200 independent random realizations of the ensembles.