Skip to contents

Run `missRanger` at a Single `mtry` and Report Per-Column OOB Error

Source: R/missranger_mtry.R
missranger_oob_by_mtry.Rd

Fits [missRanger::missRanger()] once at a given `mtry`, returning both the completed data and a tidy, per-column out-of-bag error table. This is the unit of work swept over by [missranger_sweep_mtry()], and the `ranger`-based counterpart to [missforest_oob_by_mtry()].

Usage

missranger_oob_by_mtry(
  data,
  mtry,
  num.trees = 500,
  pmm.k = 0,
  maxiter = 10,
  seed = NULL,
  num.threads = 1,
  keep_forests = FALSE
)

Arguments

data

A data frame or tibble containing the columns to impute. Any identifier or time columns should already be removed; see [missranger_sweep_mtry()]'s `exclude` argument. Columns that are entirely `NA` are refused – see Details.

mtry

Number of variables randomly sampled as candidates at each split. A single positive integer.

num.trees

Number of trees per forest. Passed to [missRanger::missRanger()]. Default `500`, that function's own default.

pmm.k

Number of predictive-mean-matching donors. `0` (the default) disables PMM and imputes with the forest prediction directly.

maxiter

Maximum number of imputation iterations. Default `10`.

seed

Integer seed passed to [missRanger::missRanger()], or `NULL` (default) to leave it unseeded. See Reproducibility.

num.threads

Threads used by `ranger` *within* one fit. Defaults to `1`; see the Threading section, which explains why the `ranger` default is the wrong one here.

keep_forests

Whether to retain the fitted forests on the returned object. Default `FALSE`. See Details.

Value

A list of two elements, plus a third when `keep_forests = TRUE`:

`ximp`

A tibble of the completed (imputed) data.

`oob_error`

A tibble with columns `column`, `error_type`, `error`, and `mtry` – one row per *imputed* column. See Details.

`fit`

The `missRanger` object, when `keep_forests = TRUE`.

Details

**Error rows cover imputed columns only.** `missForest` run variablewise reports an OOB error for every column; `missRanger` fits forests only for columns that actually have missing values, so complete columns get no row. [missranger_sweep_mtry()] carries those columns through unchanged.

**Errors are not comparable across engines.** `missForest` reports raw MSE and PFC. `missRanger` reports `ranger`'s scaled OOB prediction error – `1 - R^2` for numeric targets and classification error for categorical ones – where roughly `1` means no better than predicting the mean. Compare `mtry` values within one engine; do not compare an `error` column from one against the other.

**All-`NA` columns are refused.** `missRanger` does not drop them the way `missForest` does; it silently returns them still entirely `NA`, producing a frame that looks imputed but is not. Such columns arise easily – a mapping-driven reader that fills absent source columns with a bare `NA` yields full-length logical columns that pass name-based schema checks.

**`keep_forests` is the seam for applying to new data.** `missForest` discards its forests, which is why it cannot be applied to a later batch. `missRanger` can retain them, making `predict()` on new data possible. This function only exposes the option; a fit/apply pair built on it is separate work.

Threading

`ranger` defaults to using every available core. Inside a parallel `mtry` sweep that nests one thread pool inside another and oversubscribes the machine badly – workers times cores threads competing for the same cores. `num.threads` therefore defaults to `1` here, leaving parallelism to the sweep. Raise it only when running a single fit on an otherwise idle box.

Reproducibility

Unlike `missForest`, `missRanger` takes its own `seed` and uses it to seed the fit directly, so a seeded fit is fully determined by `seed` regardless of ambient RNG state, worker count, or `future` plan. That makes seeded `missranger_*` sweeps reproducible by construction rather than by careful stream management. The architecture caveat still applies: identical seeds do not guarantee bit-identical results across x86_64 and arm64, because split selection compares floating-point impurity sums.

A seeded fit leaves the caller's random number stream as it found it: `missRanger` re-seeds the global generator, so the generator state is restored when the function exits, and later draws in the caller (a bootstrap loop, say) carry on as if the fit had not run. An unseeded fit draws from the caller's stream as usual.

See also

[missranger_sweep_mtry()], [missforest_oob_by_mtry()]

Examples

# \donttest{
if (requireNamespace("missRanger", quietly = TRUE)) {
  df <- data.frame(
    a = c(1, 2, NA, 4, 5, 6, 7, 8),
    b = c(2, NA, 6, 8, 10, 12, 14, 16),
    c = c(5, 4, 3, NA, 1, 2, 3, 4)
  )
  res <- missranger_oob_by_mtry(df, mtry = 1, num.trees = 50, seed = 1)
  res$oob_error
}
#> # A tibble: 3 × 4
#>   column error_type error  mtry
#>   <chr>  <chr>      <dbl> <int>
#> 1 a      1-R2       1         1
#> 2 b      1-R2       0.446     1
#> 3 c      1-R2       1.10      1
# }