Skip to contents

Common Features and Arguments of RTMB Wrapper Functions

Source: R/wrapper_functions.R
rtmb_wrappers.Rd

The RTMB wrapper functions (`rtmb_lm`, `rtmb_glm`, `rtmb_glmer`, `rtmb_fa`, etc.) share a unified interface designed to make Bayesian and Frequentist inference accessible through familiar R formulas and standard model specifications.

Details

All wrapper functions in this package are built upon the same core engine. This ensures that regardless of the model type, the workflow for estimation, summary, and expansion remains consistent.

1. Unified Inference Methods: Every model object returned by a wrapper function provides the following methods:

  • $optimize(): Performs MAP estimation (comparable to MLE).

  • $sample(): Performs MCMC sampling (NUTS) for full Bayesian inference.

  • $variational(): Performs Variational Inference (ADVI) for fast posterior approximation.

2. Regularization and Variable Selection: For rtmb_lm, rtmb_glm, and rtmb_glmer, you can handle high-dimensional predictors (where the number of variables is large relative to the sample size) using the penalty argument:

  • "none": Standard flat or weakly informative priors.

  • "rhs": Regularized Horseshoe prior for continuous shrinkage.

  • "ssp": Spike-and-Slab prior for sparse variable selection.

Note: When using regularization, you must specify y_range = c(min, max) to let the model set appropriate global scales for the priors.

3. Weakly Informative Priors (y_range): By providing the theoretical range of your response variable via y_range, the wrappers automatically construct "Weakly Informative Priors". These priors are designed to be broad enough to cover any reasonable value but narrow enough to stabilize the estimation and prevent the sampler from wandering into non-sensical parameter space.

4. Fixed vs. Random Effects: For mixed-effect models (e.g., rtmb_glmer), random effects are marginalized using the Laplace approximation during $optimize(laplace = TRUE). When using $sample(), random effects are treated as unknown parameters and sampled alongside fixed effects.

5. Null Model Creation (null): You can specify a null argument (e.g., null = "x1 ~ normal(0, 0.1)") in the wrappers to simultaneously create a restricted version of your model. This is particularly useful for computing Bayes Factors or performing model comparisons.