lddmm

This major release introduces an improved sampler.

Model and defaults

• Every participant has one random effect curve for correct responses and one for incorrect responses (random_effects = "correct-incorrect", the new default). A single curve for both is available with random_effects = "shared".
• The maximum number of clusters at each block defaults to min(K^2, 2K) (8 for 4 categories) and can be set with z_max.
• Spline coefficients that start a curve have the normal prior of article eq. (9), with mean and variance set from the data (hypers$mu_beta0_*, hypers$sigma2_beta0_*).
• The smoothness of the fixed effects has a half-Cauchy prior on its standard deviation (scale hypers$s_sigma_*). The upper limits on the variance parameters are documented hyperparameters (hypers$sigma2_1_max, hypers$sigma2_us_max, hypers$sigma2_ua_max).

Improved sampler

• A single implementation serves all boundary models, and the cluster-label and random-effect updates run in C++. Fits are about 12-21 times faster on the tone learning data.
• The cluster labels are updated with an exact locally informed Hamming ball sampler, and the cluster-specific spline coefficients with proposals built from the prior of article eq. (9).
• New joint moves improve mixing: a translation between fixed and random effects, a rescaling of the random effects together with their variances, and shifts of the offsets together with the drift and boundary levels.
• Coefficients of unused clusters are drawn close to those of the clusters in use (hypers$pseudo_sd), so that curves move between clusters more easily.
• Random effects start from each participant's moment estimates, and proposal scales are tuned during burn-in only, in batches of 50 iterations.