1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
|
get_comparisons_data <- function(
mfx,
variables,
cross,
...) {
newdata <- mfx@newdata
model <- mfx@model
modeldata <- mfx@modeldata
lo <- hi <- ter <- lab <- original <- rowid <- list()
# sometimes needed for extensions when get_data doesn't work
if (is.null(modeldata) || nrow(modeldata) == 0) {
modeldata <- newdata
}
# safety need for extensions not supported by `insight`
if (any(c("factor", "character") %in% mfx@variable_class)) {
idx <- mfx@variable_class %in% c("factor", "character")
first_cross <- names(mfx@variable_class[idx])[1]
} else {
first_cross <- NULL
}
# must use `as.data.table()` because `setDT()` does not handle columns with
# more dimensions (e.g., "idx" in {mlogit})
newdata <- as.data.table(newdata)
for (v in variables) {
args <- list(
model = model,
newdata = newdata,
variable = v,
cross = cross,
first_cross = identical(v$name, first_cross),
modeldata = modeldata,
mfx = mfx
)
args <- append(args, list(...))
# logical and character before factor used to be important; but I don't think so anymore
if (check_variable_class(mfx, v$name, "logical")) {
fun <- get_comparisons_data_logical
} else if (check_variable_class(mfx, v$name, "character")) {
fun <- get_comparisons_data_character
} else if (check_variable_class(mfx, v$name, "categorical")) {
fun <- get_comparisons_data_factor
} else if (check_variable_class(mfx, v$name, "numeric")) {
fun <- get_comparisons_data_numeric
} else if (check_variable_class(mfx, v$name, "integer")) {
fun <- get_comparisons_data_numeric
} else {
msg <- sprintf(
"Class of the `%s` variable is class is not supported.",
v$name
)
stop(msg, call. = FALSE)
}
tmp <- do_call(fun, args)
lo[[v$name]] <- tmp$lo
if (isTRUE(cross)) {
lo[[v$name]][[paste0("null_contrast_", v$name)]] <- tmp$contrast_null
}
hi[[v$name]] <- tmp$hi
ter[[v$name]] <- tmp$ter
lab[[v$name]] <- tmp$lab
original[[v$name]] <- tmp$original
rowid[[v$name]] <- tmp$rowid
}
clean <- function(x) {
for (col in colnames(x)) {
# tobit1 introduces AsIs columns
if (inherits(x[[col]], "AsIs")) {
x[[col]] <- as.numeric(x[[col]])
}
# plm creates c("pseries", "numeric"), but when get_comparisons_data
# assigns +1 numeric, we lose the inheritance
if (inherits(x[[col]], "pseries")) {
x[[col]] <- as.numeric(x[[col]])
}
# strip labelled data which break rbindlist()
cl <- class(x[[col]])
if (length(cl) == 2 && cl[1] == "labelled") {
class(x[[col]]) <- class(x[[col]])[2]
}
}
return(x)
}
lo <- lapply(lo, clean)
hi <- lapply(hi, clean)
original <- lapply(original, clean)
# single contrast
if (!isTRUE(cross)) {
lo <- rbindlist(lo, fill = TRUE, ignore.attr = TRUE)
hi <- rbindlist(hi, fill = TRUE, ignore.attr = TRUE)
original <- rbindlist(original, fill = TRUE, ignore.attr = TRUE)
# long names to avoid user-supplied colname conflict
marginaleffects_ter <- unlist(ter, use.names = FALSE)
marginaleffects_lab <- unlist(lab, use.names = FALSE)
lo[, "term" := marginaleffects_ter]
hi[, "term" := marginaleffects_ter]
original[, "term" := marginaleffects_ter]
lo[, "contrast" := marginaleffects_lab]
hi[, "contrast" := marginaleffects_lab]
original[, "contrast" := marginaleffects_lab]
# cross contrast
} else {
# drop variables for which we have contrasts
for (i in seq_along(lo)) {
if (i == 1) {
# keep rowid and original data only in one of the datasets
idx_lo <- setdiff(names(variables), names(lo)[i])
idx_hi <- setdiff(names(variables), names(hi)[i])
idx_or <- setdiff(names(variables), names(hi)[i])
} else {
# exclude rowid and variables excluded from `variables`, for
# which we do not compute cross-contrasts
contrast_null <- grep(
"rowid|^null_contrast_",
colnames(lo[[i]]),
value = TRUE
)
idx_lo <- c(
setdiff(names(lo[[i]]), c(contrast_null, names(variables))),
setdiff(names(variables), names(lo)[[i]])
)
idx_hi <- c(
setdiff(names(hi[[i]]), c(contrast_null, names(variables))),
setdiff(names(variables), names(hi)[[i]])
)
idx_or <- c(
setdiff(names(original[[i]]), c(contrast_null, names(variables))),
setdiff(names(variables), names(original)[[i]])
)
}
lo[[i]] <- data.table(lo[[i]])[, !..idx_lo]
hi[[i]] <- data.table(hi[[i]])[, !..idx_hi]
original[[i]] <- data.table(original[[i]])[, !..idx_or]
lo[[i]][[paste0("contrast_", names(lo)[i])]] <- lab[[i]]
hi[[i]][[paste0("contrast_", names(hi)[i])]] <- lab[[i]]
original[[i]][[paste0("contrast_", names(original)[i])]] <- lab[[i]]
}
fun <- function(x, y) merge(x, y, all = TRUE, allow.cartesian = TRUE, sort = FALSE)
lo <- Reduce("fun", lo)
hi <- Reduce("fun", hi)
original <- Reduce("fun", original)
# faster to rbind, but creates massive datasets. need cartesian join on rowid
# lo <- cjdt(lo)
# hi <- cjdt(hi)
# if there are fewer null_contrast_* columns, then there is at least
# one always non-null variable type, so we keep everything
idx <- grepl("^null_contrast_", colnames(lo))
idx_df <- lo[, ..idx]
lo <- lo[, !..idx]
if (sum(idx) == length(variables)) {
idx <- rowSums(idx_df) < ncol(idx_df)
lo <- lo[idx]
hi <- hi[idx]
original <- original[idx]
}
}
# get_predict() is much faster if we only build the model matrix once
lo <- add_model_matrix_attribute(mfx, lo)
hi <- add_model_matrix_attribute(mfx, hi)
original <- add_model_matrix_attribute(mfx, original)
out <- list(lo = lo, hi = hi, original = original)
return(out)
}
|
