Usage
set_all_stremr_options(fit.package = c("speedglm", "glm", "h2o"), fit.algorithm = c("glm", "gbm", "randomForest", "deeplearning", "SuperLearner"), bin.method = c("equal.mass", "equal.len", "dhist"), nbins = 10, maxncats = 20, maxNperBin = 500, lower_bound_zero_Q = TRUE, skip_update_zero_Q = TRUE)
Arguments
fit.package
Specify the default package for performing model fitting: c("speedglm", "glm", "h2o")
fit.algorithm
Specify the default fitting algorithm: c("glm", "gbm", "randomForest", "deeplearning", "SuperLearner")
bin.method
The method for choosing bins when discretizing and fitting the conditional continuous summary
exposure variable sA. The default method is "equal.len", which partitions the range of sA
into equal length nbins intervals. Method "equal.mass" results in a data-adaptive selection of the bins
based on equal mass (equal number of observations), i.e., each bin is defined so that it contains an approximately
the same number of observations across all bins. The maximum number of observations in each bin is controlled
by parameter maxNperBin. Method "dhist" uses a mix of the above two approaches,
see Denby and Mallows "Variations on the Histogram" (2009) for more detail.
nbins
Set the default number of bins when discretizing a continous outcome variable under setting
bin.method = "equal.len".
If left as NA the total number of equal intervals (bins) is determined by the nearest integer of
nobs/maxNperBin, where nobs is the total number of observations in the input data.
maxncats
Max number of unique categories a categorical variable sA[j] can have.
If sA[j] has more it is automatically considered continuous.
maxNperBin
Max number of observations per 1 bin for a continuous outcome (applies directly when
bin.method="equal.mass" and indirectly when bin.method="equal.len", but nbins = NA).
lower_bound_zero_Q
Set to TRUE to bound the observation-specific Qs during the TMLE update step away from zero (with minimum value set at 10^-4).
Can help numerically stabilize the TMLE intercept estimates in some small-sample cases. Has no effect when TMLE = FALSE.
skip_update_zero_Q
Set to FALSE to perform TMLE update with glm even when all of the Q's are zero.
When set to TRUE the TMLE update step is skipped if the predicted Q's are either all 0 or near 0, with TMLE intercept being set to 0.