Consider better selection of potential basis functions for GLMM mode

[jr-leary7]

• currently, basis functions for GLMM mode are selected via the following algorithm:

  1. identify a maximum of $M = 5$ basis functions per individual subject
  2. refit all subject-level basis functions on the entire dataset
  3. use a negative-binomial LASSO model to remove collinear / uninformative basis functions by fitting $n_\lambda = 50$ possible models across values of the penalty parameter $\lambda$, choosing the "best" model as the one that minimizes AIC
  4. using the set of retained basis functions from the "best" model, fit a negative-binomial GLMM with random intercepts and slopes for each basis function

• this might not be ideal, as I've seen cases where the number of retained basis functions is "too high" leading to estimation issues and weird, jagged-looking fitted values when visualized

• consider maybe choosing the "best" model as the one that is most sparse ? or perhaps do a combined ranking of sparsity and AIC and choose the model with the best mean rank ?

[jr-leary7]

in commit 30e5d92 i switched the "best" model selection from AIC to BIC as BIC imposes a harsher penalty on the number of covariates