Fixed a bug in print.fwbci() due to incorrect ordering of the intervals which led them to be printed with incorrect labels. These have been corrected and printing is a little prettier. Thanks to Katya Zelevinsky.
Documentation and vignette updates.
Added a new confint() method for fwb objects.
Added a new strata argument to fwb() to perform stratified bootstrapping within levels of a stratification variable.
Added a new drop0 argument to fwb() to drop all units with weights of 0 in each bootstrap iteration.
Added a new .coef argument to vcovFWB(). A function can be supplied to extract a vector of coefficients from the fitted model in each bootstrap iteration if the default (stats::coef()) doesn't return a numeric vector (e.g., for nnet::multinom() models). An error message is now thrown if .coef doesn't return a numeric vector.
Added support for using future backend for fwb() by supplying cl = "future". Thanks to Katya Zelevinsky for the suggestion.
Added a new vignette on reproducibility and parallelization, which can be accessed at vignette("fwb-rep").
For fwb(), simple has a new default that is TRUE in most cases and FALSE when wtype is "multinom". This should not affect results but will reduce memory use for large datasets by avoiding computing all bootstrap weights simultaneously. Note that when there is randomness in the statistic supplied to fwb(), the argument to simple affects whether BCa confidence intervals can be computed. See the reproducibility vignette mentioned above for details.
A warning is now thrown when using fwb() with simple = TRUE with non-NULL cl when the random number generator kind is not "L'Ecuyer-CMRG". Under these circumstances, results may not replicate and the BCa confidence interval will be inaccurate. See the reproducibility vignette mentioned above for details.
Fixed a bug where the names of quantities produced by fwb() when statistic returns an unnamed vector were incorrect.
When BCa confidence intervals are requested, an error is thrown if the number of bootstrap replications is smaller than the sample size.
Documentation updates.
fwb() and vcovFWB() now take an additional argument, wtype, which specifies how the weights are drawn. The default, "exp" is still to draw weights from an $\text{Exp}(1)$ distribution but other options, namely "multinom" for multinomial integer weights (which reproduce boot::boot() results exactly), "poisson" for Poisson integer weights, and "mammen" for second-order accurate Mammen weights as recommended by Lihua Lei here. (#4)
New functions set_fwb_wtype() and get_fwb_wtype() allow one to set global defaults for the wtype argument of fwb() and vcovFWB()`.
Fixed bugs related to the index argument of various functions, including bugs when the estimated quantity is not given a name.
Some error messages may be clearer.