Merge pull request #272 from stan-dev/factor_not_levels

Do not automatically create an rvar_factor when rvar() is passed a numeric array with levels

R/rvar-.R

@@ -806,22 +806,11 @@ cleanup_rvar_draws <- function(x) {
     rownames(x) <- as.character(seq_rows(x))
   }
 
-  # if x is factor-like (character or with "levels" attr), make it a factor
-  x <- as_factor_if_factor_like(x)
-
-  x
-}
-
-#' convert factor-like objects (character vectors or numerics with "levels"
-#' attributes) to factors. Leaves non-factor-like objects alone.
-#' @param x a vector or array (including factor-like arrays with "levels" attributes)
-#' @noRd
-as_factor_if_factor_like <- function(x) {
+  # if x is a character array, make it a factor
   if (is.character(x)) {
     x <- while_preserving_dims(factor, x)
-  } else if (!is.factor(x) && !is.null(attr(x, "levels"))) {
-    x <- while_preserving_dims(factor, x, labels = attr(x, "levels"))
   }
+
   x
 }
 

R/rvar-factor.R

@@ -51,19 +51,25 @@
 #' # Unlike base factors, rvar factors can be matrices or arrays:
 #' rvar_factor(x_array, dim = c(2, 2))
 #'
-#' # If the input to rvar() is an array with a `"levels"` attribute, it
-#' # will automatically be treated as an `rvar_factor()`:
+#' # If the input to rvar_factor() is an array with a `"levels"` attribute, it
+#' # will use those as the levels of the factor
 #' y_array <- t(array(rbinom(3000, 1, c(0.1, 0.5, 0.9)) + 1, dim = c(3, 1000)))
 #' rvar(y_array)
 #' # with levels
 #' attr(y_array, "levels") = c("a", "b")
-#' rvar(y_array)
+#' rvar_factor(y_array)
 #'
 #' @export
 rvar_factor <- function(
   x = factor(), dim = NULL, dimnames = NULL, nchains = NULL, with_chains = FALSE, ...
 ) {
 
+  # to ensure we pick up levels already attached to x (if there are any), we
+  # need to convert x to a factor here if it has levels
+  if (!is.factor(x) && !is.null(attr(x, "levels"))) {
+    x <- while_preserving_dims(factor, x, labels = attr(x, "levels"))
+  }
+
   out <- rvar(
     x, dim = dim, dimnames = dimnames, nchains = nchains, with_chains = with_chains
   )

tests/testthat/test-rvar-factor.R

@@ -11,15 +11,15 @@ test_that("0-length rvar factors are constructured correctly", {
   expect_equal(draws_of(rvar(ordered(NULL))), ref)
 })
 
-test_that("rvar() input with a levels attribute creates an rvar_factor", {
+test_that("rvar_factor() input with a levels attribute preserves levels", {
   x_array <- matrix(1:10, nrow = 2)
   attr(x_array, "levels") <- letters[1:10]
 
   ref_draws <- factor(letters[1:10])
   dim(ref_draws) <- c(2, 5)
   dimnames(ref_draws) <- list(1:2, NULL)
 
-  x <- rvar(x_array)
+  x <- rvar_factor(x_array)
   expect_equal(draws_of(x), ref_draws)
   expect_true(inherits(x, "rvar_factor"))
 })

vignettes/rvar.Rmd

@@ -101,17 +101,17 @@ to constructing and manipulating `rvar`s are described below.
 
 You can also use `rvar`s to represent discrete distributions, using the `rvar_factor()`
 and `rvar_ordered()` subtypes. If you attempt to create an `rvar` using character
-values, it will automatically be treated as an `rvar_factor`:
+values or a `factor`, it will automatically be treated as an `rvar_factor`:
 
 ```{r rvar_factor}
 x <- rvar(sample(c("a","b","c"), 4000, prob = c(0.7, 0.2, 0.1), replace = TRUE))
 x
 ```
 
-Objects with a `levels` attribute will also automatically be converted to an
-`rvar_factor()`. This (along with automatic conversion of character values) 
-means output from `rstanarm::posterior_predict()` and `brms::posterior_predict()` 
-will be converted into `rvar_factor`s if wrapped in `rvar()`.
+Numeric arrays with a `"levels"` attribute can also be passed to `rvar_factor()`.
+This (along with conversion of character values) means output from 
+`rstanarm::posterior_predict()` and `brms::posterior_predict()` on categorical
+models can be passed directly to `rvar_factor()`.
 
 The default display shows the mode (as returned by `modal_category()`) and
 normalized entropy (`entropy()`), which is Shannon entropy scaled by the maximum