Use base::rbind over plyr::rbind.fill

R/helper-margins.r

@@ -75,9 +75,15 @@ add_margins <- function(df, vars, margins = TRUE) {
   # Loop through all combinations of margin variables, setting
   # those variables to (all)
   margin_dfs <- llply(margin_vars, function(vars) {
-    df[vars] <- rep(list(factor("(all)")), length(vars))
+    for (var in vars) {
+      # Need '[]' to coerce '(all)' into the existing structure of the LHS,
+      #   as opposed to overwriting with character. In particular, that inserts
+      #   '(all)' as an element of the existing factor, e.g. respecting the
+      #   is.ordered()-ness of that factor as well.
+      df[[var]][] <- "(all)"
+    }
     df
   })
 
-  rbind.fill(margin_dfs)
+  bind_rows(margin_dfs)
 }

R/melt.r

@@ -60,7 +60,7 @@ melt.default <- function(data, ..., na.rm = FALSE, value.name = "value") {
 #' melt(list(list(1:3), 1, list(as.list(3:4), as.list(1:2))))
 melt.list <- function(data, ..., level = 1) {
   parts <- lapply(data, melt, level = level + 1, ...)
-  result <- rbind.fill(parts)
+  result <- bind_rows(parts)
 
   # Add labels
   names <- names(data) %||% seq_along(data)

R/utils.r

@@ -52,3 +52,21 @@ normalize_melt_arguments <- function(data, measure.ind, factorsAsStrings) {
 is.string <- function(x) {
   is.character(x) && length(x) == 1
 }
+
+# base-only drop-in for rbind.fill()
+bind_rows <- function(dfs) {
+  df_sizes <- lengths(dfs)
+  if (length(unique(df_sizes)) > 1L) {
+    # NB: rbind() _does_ use name matching for columns, so we don't need to
+    #   reorder all the ourselves.
+    out_names <- character()
+    for (df in dfs) {
+      out_names <- c(out_names, setdiff(names(df), out_names))
+    }
+    # for (df in dfs) doesn't work; 'df' gets copy-on-written.
+    for (ii in seq_along(dfs)) {
+      dfs[[ii]][setdiff(out_names, names(dfs[[ii]]))] <- NA
+    }
+  }
+  do.call(rbind, dfs)
+}