fwb

New Features

• Added a new confidence interval type for confint(), fwb.ci(), and summary(): "cheap", for the "cheap" confidence interval described by Lam (2022). This interval maintains nominal coverage even with few bootstrap replications (as low as one!).

Breaking changes

• **Bootstrap weights now depend only on the seed.** Previously, when simple = TRUE, the weights were drawn inside each bootstrap replication and so depended on how the work was distributed: results changed with cl, with the number of workers, and with verbose, and reproducing a parallel run required set.seed(###, "L'Ecuyer-CMRG") (or, with a cluster object, parallel::clusterSetRNGStream() instead of set.seed()). fwb() and vcovFWB() now reserve one random number stream per replicate before any work is dispatched, and each replication draws its weights from its own stream. A plain call to set.seed() is all that is needed, whatever cl, verbose, or RNGkind() are set to, and parallel::clusterSetRNGStream() no longer has any effect. See vignette("fwb-rep").
• **This changes the bootstrap estimates when simple = TRUE** (the default for every wtype except "multinom"). Analyses that need to reproduce results from *fwb* 0.6.0 or earlier should install that version (e.g., using pak::pak("fwb@0.6.0")). Results with simple = FALSE are unchanged, so wtype = "multinom" continues to match boot::boot(., stype = "f") exactly. Note also that simple = TRUE and simple = FALSE no longer give the same estimates as each other, even for the weight types where they previously did.
• **vcovFWB() no longer reproduces sandwich::vcovBS()** when the same seed is set, and its results from earlier versions cannot be reproduced. Each replicate now draws its weights from its own random number stream, as in fwb() with simple = TRUE, so vcovFWB() gives the same results as vcov() applied to the output of fwb(., simple = TRUE) for every wtype, without holding the weights for all replicates in memory at once.
• <fwb> objects no longer store the cl argument, since nothing about the backend is needed to recover the weights any more. Previously, a cluster object kept in the result made the object unusable after parallel::stopCluster() or after being saved and reloaded.
• vignette("fwb-rep") has been rewritten around the new behavior. It no longer describes the several cases that used to require different treatment, and it explains how to reproduce results from earlier versions.
• **The multinomial bootstrap is now exhaustive when it can be.** There are only $n^n$ distinct multinomial bootstrap samples of $n$ units, so when R is at least that many, fwb() and vcovFWB() now use each of them exactly once instead of sampling, and set R to that number (reporting this in a message). The estimates are then the exact bootstrap distribution rather than a sample from it, and they no longer depend on the seed. This applies with strata (where the count is $\prod_s n_s^{n_s}$) and with cluster (where it is computed over clusters). It is the one respect in which wtype = "multinom" no longer matches boot::boot(., stype = "f"), which always samples; it reaches only very small samples, as $n = 6$ already has 46656 distinct samples.

Other changes

• Added a CITATION file, so citation("fwb") now returns the *Journal of Open Source Software* article describing *fwb* (doi:10.21105/joss.08880).
• Simultaneous inference with ci.type = "perc" is much faster and now exact. Previously the confidence level was found by numerical search, which could return a band whose coverage fell short of the one requested; it is now computed directly.