Create setting for Broken Adaptive Ridge logistic regression

Creates model settings for Broken Adaptive Ridge logistic regression using Cyclops and the BrokenAdaptiveRidge prior. initialRidgeVariance = "auto" first fits a ridge model with Cyclops cross-validation and uses the selected ridge variance to initialize BAR. penalty = "auto" cross-validates over a BAR penalty grid and refits using the penalty with the highest mean out-of-fold AUC. Automatic penalty tuning fits one model per fold and grid value, in addition to the final model and cross-validation refits, so it can be substantially slower than using a fixed penalty.

Usage

setBrokenAdaptiveRidge(
  initialRidgeVariance = "auto",
  seed = NULL,
  includeCovariateIds = c(),
  noShrinkage = c("(Intercept)"),
  penalty = "auto",
  penaltyRatio = 0.1,
  penaltyGridSize = 10,
  threads = -1,
  forceIntercept = FALSE,
  upperLimit = 20,
  lowerLimit = 0.01,
  tolerance = 2e-06,
  maxIterations = 3000,
  threshold = 1e-06
)

Arguments

initialRidgeVariance

Numeric prior starting variance, or "auto" to estimate this using ridge cross-validation.

seed

An option to add a seed when training the model.

includeCovariateIds

A set of covariateIds to limit the analysis to.

noShrinkage

A set of covariates which are forced into the model. The default is the intercept. This takes precedence over forceIntercept: a covariate listed here is excluded from the prior even when forceIntercept = TRUE.

penalty

Numeric BAR penalty, "bic" to use log(n) / 2, or "auto" to cross-validate over a penalty grid.

penaltyRatio

Minimum penalty in the automatic grid as a ratio of the log(n) / 2 starting penalty.

penaltyGridSize

Number of penalties to evaluate when penalty = "auto".

threads

An option to set number of threads when training model.

forceIntercept

Logical: Include the intercept coefficient in the prior, unless it is listed in noShrinkage. To penalize the intercept, set this to TRUE and remove the intercept from noShrinkage.

upperLimit

Numeric: Upper prior variance limit for grid-search.

lowerLimit

Numeric: Lower prior variance limit for grid-search.

tolerance

Numeric convergence tolerance passed to both Cyclops and BAR. In Cyclops, this controls the maximum relative change in its convergence criterion; in BAR, it controls the maximum absolute coefficient change between outer iterations.

maxIterations

Integer maximum iteration count passed to both the Cyclops optimizer and the BAR outer loop.

threshold

Numeric BAR threshold.

Value

modelSettings object

References

Fridgeirsson EA, Williams R, Rijnbeek P, Suchard MA, Reps JM. Comparing penalization methods for linear models on large observational health data. Journal of the American Medical Informatics Association. 2024;31(7):1514-1521. doi:10.1093/jamia/ocae109

Li N, Peng X, Kawaguchi E, Suchard MA, Li G. A scalable surrogate L0 sparse regression method for generalized linear models with applications to large scale data. Journal of Statistical Planning and Inference. 2021;213:262-281. doi:10.1016/j.jspi.2020.12.001

Examples

modelBar <- setBrokenAdaptiveRidge(seed = 42)