R Under development (unstable) (2026-10-28 r90783 ucrt) -- "Unsuffered Consequences"
Copyright (C) 2026 The R Foundation for Statistical Computing
Platform: x86_64-w64-mingw32/x64

R is free software and comes with ABSOLUTELY NO WARRANTY.
You are welcome to redistribute it under certain conditions.
Type 'license()' or 'licence()' for distribution details.

R is a collaborative project with many contributors.
Type 'contributors()' for more information and
'citation()' on how to cite R or R packages in publications.

Type 'demo()' for some demos, 'help()' for on-line help, or
'help.start()' for an HTML browser interface to help.
Type 'q()' to quit R.

> library(BATTS)
> # predict()/fitted()/density_ratio() agree with eval_balance_weight() exactly.
> set.seed(31)
> x <- matrix(c(rnorm(40), rnorm(40, 0.6)), ncol = 1); g <- rep(0:1, each = 40)
> nd <- x[1:5, , drop = FALSE]
> 
> set.seed(7); fb <- boots(x, g, num_trees_max = 8, K_CV = 3, quiet = TRUE)
Warning message:
boots() is deprecated; use batts_boost() instead (see ?`BATTS-deprecated`).
boots() still works and gives the same result. This warning is shown once per session. 
> set.seed(8); fs <- batts(x, g, num_trees = 4, size_burnin = 5, size_backfitting = 13,
+                          output_BART_ensembles = TRUE, quiet = TRUE)
Warning message:
batts() is deprecated; use batts_bart() instead (see ?`BATTS-deprecated`).
batts() still works and gives the same result. This warning is shown once per session. 
> 
> # 1. type = "weight" is bit identical to the legacy entry point.
> legacy_b <- as.numeric(eval_balance_weight(fb, nd)$balancing_weight_boosting)
Warning message:
eval_balance_weight() is deprecated; use predict() instead (see ?`BATTS-deprecated`).
eval_balance_weight() still works and gives the same result. This warning is shown once per session. 
> stopifnot(identical(predict(fb, nd, type = "weight"), legacy_b))
> 
> legacy_s <- eval_balance_weight(fs, nd, is_Bayes = TRUE)$balancing_weight_BART
> stopifnot(identical(predict(fs, nd, type = "weight", summary = "none"), legacy_s))
> 
> # 2. The default is the log-density ratio, not the weight.
> stopifnot(identical(predict(fb, nd), 2 * log(legacy_b)),
+           !identical(predict(fb, nd), legacy_b))
> 
> # 3. Transformations are exact.
> w <- predict(fb, nd, type = "weight")
> stopifnot(identical(predict(fb, nd, type = "ratio"), w^2),
+           identical(predict(fb, nd, type = "log_ratio"), 2 * log(w)),
+           identical(density_ratio(fb, nd), w^2))
> 
> # 4. Each draw is transformed before summarizing, not the other way round.
> d <- predict(fs, nd, type = "weight", summary = "none")
> stopifnot(identical(predict(fs, nd, type = "ratio"), rowMeans(d^2)),
+           identical(predict(fs, nd), rowMeans(2 * log(d))),
+           !isTRUE(all.equal(rowMeans(d^2), rowMeans(d)^2)))
> 
> # 5. The median summary is invariant to the monotone transform only when the
> #    number of draws is odd, because median() averages the two middle values.
> stopifnot(ncol(d) == 13L, ncol(d) %% 2L == 1L)
> stopifnot(isTRUE(all.equal(predict(fs, nd, type = "ratio", summary = "median"),
+                            apply(d, 1, median)^2, tolerance = 1e-12)))
> 
> # 6. fitted() matches prediction at the training points.
> stopifnot(identical(fitted(fb, type = "weight"),
+                     as.numeric(fb$balance_weight_boosting_data)),
+           isTRUE(all.equal(fitted(fb), predict(fb, x), tolerance = 1e-10)),
+           identical(fitted(fs, type = "weight", summary = "none"),
+                     fs$balance_weight_BART_data))
> 
> # 7. Omitting newdata uses the training observations.
> stopifnot(identical(predict(fb), fitted(fb)),
+           identical(predict(fs), fitted(fs)),
+           identical(density_ratio(fb), fitted(fb, type = "ratio")))
> 
> # 8. Intervals.
> iv <- predict(fs, nd, interval = TRUE, level = 0.8)
> stopifnot(identical(dim(iv), c(5L, 3L)),
+           identical(colnames(iv), c("fit", "lower", "upper")),
+           all(iv[, "lower"] <= iv[, "fit"]), all(iv[, "fit"] <= iv[, "upper"]),
+           all(is.finite(iv)))
> 
> # 9. Invalid arguments are rejected.
> bad <- function(expr) tryCatch({expr; NA_character_}, error = conditionMessage)
> stopifnot(grepl("should be one of", bad(predict(fb, nd, type = "nope"))),
+           grepl("level must be", bad(predict(fs, nd, level = 1)), fixed = TRUE),
+           grepl("interval must be", bad(predict(fs, nd, interval = NA)), fixed = TRUE),
+           grepl("outside the fitted domain", bad(predict(fb, x + 1e6)), fixed = TRUE),
+           grepl("outside of the sample space", bad(eval_balance_weight(fb, x + 1e6)), fixed = TRUE))
> 
> # 10. A boosting fit rejects posterior-only arguments instead of ignoring them.
> stopifnot(grepl("no posterior draws", bad(predict(fb, nd, summary = "none")), fixed = TRUE),
+           grepl("summary", bad(predict(fb, nd, summary = "none")), fixed = TRUE),
+           grepl("no posterior draws", bad(fitted(fb, interval = TRUE)), fixed = TRUE))
> 
> # 11. Prediction consumes no random numbers.
> rng <- .Random.seed
> invisible(predict(fb, nd)); invisible(predict(fs, nd))
> invisible(density_ratio(fb, nd)); invisible(density_ratio(fs, nd))
> invisible(fitted(fb)); invisible(fitted(fs))
> invisible(predict(fs, nd, interval = TRUE))
> stopifnot(identical(rng, .Random.seed))
> 
> cat('predict()/fitted()/density_ratio() checks passed.\n')
predict()/fitted()/density_ratio() checks passed.
> 
> proc.time()
   user  system elapsed 
   0.17    0.14    0.29