Largest `mtry` `missRanger` Will Accept For a Data Set
Source:R/missranger_mtry.R
missranger_max_mtry.Rd`missRanger` admits a **smaller** `mtry` than `missForest` on the same data, and the difference is easy to trip over: exceeding it surfaces as `ranger`'s catch-all `"User interrupt or internal error."`, which names neither `mtry` nor the real bound.
Details
`missForest` predicts every column from every other one, so its bound is simply `ncol(data) - 1`. `missRanger` builds its predictor pool up over the first iteration instead. It starts with `completed`, the set of **fully-observed** columns, imputes the targets in increasing order of missingness, and adds each one to `completed` as it goes. The first target therefore sees only the complete columns, and `mtry` must fit *that* pool:
* With `k` usable complete columns, the bound is `k`. * With none, the first target falls back to univariate imputation and the second sees a single predictor, so the bound is `1` – no matter how wide the data is.
Constant columns are excluded from the count, matching `missRanger`'s own handling of them as features.
The practical consequence: a sweep is only informative when some columns are complete. That is the usual shape of the intended use – imputing a domain's incomplete outputs against its fully-observed inputs – but a frame in which *every* column has missing values admits `mtry = 1` alone, and there is nothing to sweep.
Examples
# Two complete columns -> mtry may be 1 or 2.
missranger_max_mtry(data.frame(a = c(1, NA, 3), b = c(1, 2, 3), c = c(4, 5, 6)))
#> [1] 2
# Every column has a gap -> only mtry = 1 is admissible.
missranger_max_mtry(data.frame(a = c(1, NA, 3), b = c(NA, 2, 3), c = c(4, 5, NA)))
#> [1] 1