| 1 |
set.seed(20210110) |
| 2 |
|
| 3 |
skip_if_not_installed("carData") |
| 4 |
|
| 5 |
|
| 6 |
# Duncan example |
| 7 |
data(Duncan, package = "carData") |
| 8 |
Duncan$cluster <- sample(LETTERS[1:8], size = nrow(Duncan), replace = TRUE) |
| 9 |
Duncan_int <- lm(prestige ~ type * (income + education), data=Duncan) |
| 10 |
coefs_int <- coef(Duncan_int) |
| 11 |
coef_names_int <- names(coefs_int) |
| 12 |
Duncan_int_CR2 <- vcovCR(Duncan_int, type = "CR2", cluster = Duncan$cluster) |
| 13 |
|
| 14 |
Duncan_sep <- lm(prestige ~ 0 + type + type:(income + education), data=Duncan) |
| 15 |
coefs_sep <- coef(Duncan_sep) |
| 16 |
coef_names_sep <- names(coefs_sep) |
| 17 |
Duncan_sep_CR2 <- vcovCR(Duncan_sep, type = "CR2", cluster = Duncan$cluster) |
| 18 |
|
| 19 |
|
| 20 |
# STAR example |
| 21 |
skip_if_not_installed("AER") |
| 22 |
data(STAR, package = "AER") |
| 23 |
levels(STAR$stark)[3] <- "aide" |
| 24 |
levels(STAR$schoolk)[1] <- "urban" |
| 25 |
STAR <- subset(STAR, |
| 26 |
!is.na(schoolidk), |
| 27 |
select = c(schoolidk, schoolk, stark, gender, ethnicity, math1, lunchk)) |
| 28 |
|
| 29 |
lm_urbanicity <- lm(math1 ~ schoolk * stark + gender + ethnicity + lunchk, |
| 30 |
data = STAR) |
| 31 |
|
| 32 |
|
| 33 |
CRs <- paste0("CR", 0:4) |
| 34 |
|
| 35 |
test_that("vcov arguments work", { |
| 36 |
VCR <- lapply(CRs, function(t) vcovCR(Duncan_sep, type = t, cluster = Duncan$cluster)) |
| 37 |
CI_A <- lapply(VCR, function(v) linear_contrast(Duncan_sep, |
| 38 |
vcov = v, cluster = Duncan$cluster, |
| 39 |
contrast = constrain_pairwise(1:3), |
| 40 |
level = .98)) |
| 41 |
CI_B <- lapply(CRs, function(t) linear_contrast(Duncan_sep, |
| 42 |
vcov = t, cluster = Duncan$cluster, |
| 43 |
contrast = constrain_pairwise(1:3), |
| 44 |
level = .98)) |
| 45 |
expect_equal(CI_A, CI_B, check.attributes = FALSE) |
| 46 |
|
| 47 |
}) |
| 48 |
|
| 49 |
test_that("constrain_() functions work.", { |
| 50 |
|
| 51 |
# Not worrying about CR4 here |
| 52 |
|
| 53 |
CI_A <- lapply(CRs[1:4], function(t) |
| 54 |
linear_contrast(Duncan_sep, vcov = t, cluster = Duncan$cluster, |
| 55 |
contrast = constrain_pairwise(1:3)) |
| 56 |
) |
| 57 |
|
| 58 |
CI_B <- lapply(CRs[1:4], function(t) |
| 59 |
linear_contrast(Duncan_int, vcov = t, cluster = Duncan$cluster, |
| 60 |
contrast = constrain_pairwise(2:3, with_zero = TRUE)) |
| 61 |
) |
| 62 |
|
| 63 |
CI_A <- lapply(CI_A, subset, select = -Coef) |
| 64 |
CI_B <- lapply(CI_B, subset, select = -Coef) |
| 65 |
|
| 66 |
expect_equal(CI_A, CI_B, check.attributes = FALSE) |
| 67 |
}) |
| 68 |
|
| 69 |
test_that("linear_contrast() works with lists.", { |
| 70 |
|
| 71 |
|
| 72 |
CIs <- linear_contrast(lm_urbanicity, |
| 73 |
vcov = "CR2", cluster = STAR$schoolidk, |
| 74 |
contrast = list( |
| 75 |
A = constrain_zero(2:3), |
| 76 |
B = constrain_pairwise(2:3, with_zero = TRUE), |
| 77 |
C = constrain_equal("ethnicity", reg_ex = TRUE), |
| 78 |
D = constrain_pairwise("ethnicity", reg_ex = TRUE) |
| 79 |
)) |
| 80 |
|
| 81 |
CI_A <- as.data.frame(subset(CIs, grepl("A\\.", Coef), select = -Coef)) |
| 82 |
CI_B <- as.data.frame(subset(CIs, grepl("B\\.[a-z]+$", Coef), select = -Coef)) |
| 83 |
expect_equal(CI_A, CI_B, check.attributes = FALSE) |
| 84 |
|
| 85 |
CI_C <- as.data.frame(subset(CIs, grepl("C\\.", Coef), select = -Coef)) |
| 86 |
CI_D <- as.data.frame(subset(CIs, grepl("D\\..+ethnicityafam$", Coef), select = -Coef)) |
| 87 |
expect_equal(CI_C, CI_D, check.attributes = FALSE) |
| 88 |
|
| 89 |
}) |
| 90 |
|
| 91 |
test_that("printing works", { |
| 92 |
CIs <- linear_contrast(Duncan_sep, vcov = "CR0", cluster = Duncan$cluster, |
| 93 |
constrain_pairwise(":education", reg_ex = TRUE)) |
| 94 |
expect_output(print(CIs)) |
| 95 |
|
| 96 |
CIs <- linear_contrast(Duncan_sep, vcov = "CR0", cluster = Duncan$cluster, |
| 97 |
constrain_pairwise(":education", reg_ex = TRUE), |
| 98 |
p_values = TRUE) |
| 99 |
expect_output(x <- print(CIs)) |
| 100 |
expect_true(all(c("p-value","Sig.") %in% names(x))) |
| 101 |
}) |
| 102 |
|
| 103 |
test_that("level checks work", { |
| 104 |
expect_error(linear_contrast(Duncan_sep, vcov = "CR0", cluster = Duncan$cluster, |
| 105 |
constrain_pairwise(":education", reg_ex = TRUE), |
| 106 |
level = -0.01)) |
| 107 |
expect_error(linear_contrast(Duncan_sep, vcov = "CR0", cluster = Duncan$cluster, |
| 108 |
constrain_pairwise(":education", reg_ex = TRUE), |
| 109 |
level = 95)) |
| 110 |
expect_output(print( |
| 111 |
linear_contrast(Duncan_sep, vcov = "CR0", cluster = Duncan$cluster, |
| 112 |
constrain_pairwise(":education", reg_ex = TRUE), |
| 113 |
level = runif(1)) |
| 114 |
)) |
| 115 |
}) |
| 116 |
|
| 117 |
test_that("CI boundaries are ordered", { |
| 118 |
lev <- runif(1) |
| 119 |
CI_z <- linear_contrast(Duncan_sep, vcov = "CR0", cluster = Duncan$cluster, |
| 120 |
constrain_pairwise(":education", reg_ex = TRUE), |
| 121 |
test = "z", level = lev) |
| 122 |
CI_t <- linear_contrast(Duncan_sep, vcov = "CR0", cluster = Duncan$cluster, |
| 123 |
constrain_pairwise(":education", reg_ex = TRUE), |
| 124 |
test = "naive-t", level = lev) |
| 125 |
CI_Satt <- linear_contrast(Duncan_sep, vcov = "CR0", cluster = Duncan$cluster, |
| 126 |
constrain_pairwise(":education", reg_ex = TRUE), |
| 127 |
test = "Satterthwaite", level = lev) |
| 128 |
expect_true(all(CI_t$CI_L < CI_z$CI_L)) |
| 129 |
expect_true(all(CI_t$CI_U > CI_z$CI_U)) |
| 130 |
expect_true(all(CI_Satt$CI_L < CI_z$CI_L)) |
| 131 |
expect_true(all(CI_Satt$CI_U > CI_z$CI_U)) |
| 132 |
}) |
| 133 |
|
| 134 |
test_that("linear_contrast() is consistent with Wald_test()", { |
| 135 |
|
| 136 |
skip_on_cran() |
| 137 |
|
| 138 |
lev <- runif(1) |
| 139 |
CIs <- lapply(CRs, function(v) |
| 140 |
linear_contrast(lm_urbanicity, vcov = v, cluster = STAR$schoolidk, |
| 141 |
contrasts = list( |
| 142 |
A = constrain_zero("ethnicity", reg_ex = TRUE), |
| 143 |
B = constrain_equal("ethnicity", reg_ex = TRUE) |
| 144 |
), |
| 145 |
test = "Satterthwaite", level = lev, p_values = TRUE)) |
| 146 |
Wald_tests <- lapply(CRs, function(v) |
| 147 |
Wald_test(lm_urbanicity, vcov = v, cluster = STAR$schoolidk, |
| 148 |
constraints = constrain_pairwise("ethnicity", reg_ex = TRUE, with_zero = TRUE), |
| 149 |
test = "HTZ", tidy = TRUE)) |
| 150 |
|
| 151 |
CI_pvals <- lapply(CIs, function(x) x$p_val) |
| 152 |
Wald_pvals <- lapply(Wald_tests, function(x) x$p_val[1:7]) |
| 153 |
expect_equal(CI_pvals, Wald_pvals, tolerance = 1e-6) |
| 154 |
|
| 155 |
}) |
| 156 |
|
| 157 |
test_that("linear_contrast() has informative error messages.", { |
| 158 |
expect_error( |
| 159 |
linear_contrast(Duncan_sep, vcov = "CR0", cluster = Duncan$cluster, |
| 160 |
constrain_pairwise(":education", reg_ex = TRUE), |
| 161 |
test = "all") |
| 162 |
) |
| 163 |
|
| 164 |
expect_error( |
| 165 |
linear_contrast(Duncan_sep, vcov = "CR0", cluster = Duncan$cluster, |
| 166 |
constrain_pairwise(":education", reg_ex = TRUE), |
| 167 |
test = "saddlepoint") |
| 168 |
) |
| 169 |
|
| 170 |
}) |
| 171 |
|
| 172 |
test_that("linear_contrast() works with scalar (length-1) contrasts.", { |
| 173 |
|
| 174 |
lm_int <- lm(math1 ~ 1, data = STAR) |
| 175 |
x0 <- linear_contrast(lm_int, vcov = "CR0", cluster = STAR$schoolidk, contrasts = 1) |
| 176 |
expect_s3_class(x0, "conf_int_clubSandwich") |
| 177 |
|
| 178 |
x2 <- linear_contrast(lm_int, vcov = "CR2", cluster = STAR$schoolidk, contrasts = 1) |
| 179 |
expect_s3_class(x2, "conf_int_clubSandwich") |
| 180 |
|
| 181 |
y0 <- linear_contrast(lm_int, vcov = "CR0", cluster = STAR$schoolidk, contrasts = matrix(1)) |
| 182 |
y0$Coef <- row.names(y0) <- "Contrast" |
| 183 |
expect_equal(x0, y0) |
| 184 |
|
| 185 |
y2 <- linear_contrast(lm_int, vcov = "CR2", cluster = STAR$schoolidk, contrasts = 1) |
| 186 |
y2$Coef <- row.names(y2) <- "Contrast" |
| 187 |
expect_equal(x2, y2) |
| 188 |
|
| 189 |
STAR$wt <- 1L |
| 190 |
wls_int <- lm(math1 ~ 1, weights = wt, data = STAR) |
| 191 |
z0 <- linear_contrast(wls_int, vcov = "CR0", cluster = STAR$schoolidk, contrasts = 1) |
| 192 |
expect_s3_class(z0, "conf_int_clubSandwich") |
| 193 |
|
| 194 |
z2 <- linear_contrast(wls_int, vcov = "CR2", cluster = STAR$schoolidk, contrasts = 1) |
| 195 |
expect_s3_class(z2, "conf_int_clubSandwich") |
| 196 |
|
| 197 |
}) |
| 198 |
|
| 199 |
test_that("linear_contrast() works with scalar (length-1) contrasts in metafor.", { |
| 200 |
|
| 201 |
skip_if_not_installed("metafor") |
| 202 |
suppressPackageStartupMessages(library(metafor)) |
| 203 |
|
| 204 |
dat <- dat.bangertdrowns2004 |
| 205 |
res_uni <- rma(yi, vi, weights = 1, data=dat) |
| 206 |
|
| 207 |
x0 <- linear_contrast(res_uni, vcov = "CR0", cluster = dat$id, contrasts = 1) |
| 208 |
expect_s3_class(x0, "conf_int_clubSandwich") |
| 209 |
y0 <- linear_contrast(res_uni, vcov = "CR0", cluster = dat$id, contrasts = matrix(1)) |
| 210 |
y0$Coef <- row.names(y0) <- "Contrast" |
| 211 |
expect_equal(x0, y0) |
| 212 |
|
| 213 |
x2 <- linear_contrast(res_uni, vcov = "CR2", cluster = dat$id, contrasts = 1) |
| 214 |
expect_s3_class(x2, "conf_int_clubSandwich") |
| 215 |
y2 <- linear_contrast(res_uni, vcov = "CR2", cluster = dat$id, contrasts = 1) |
| 216 |
y2$Coef <- row.names(y2) <- "Contrast" |
| 217 |
expect_equal(x2, y2) |
| 218 |
|
| 219 |
|
| 220 |
}) |
| 221 |
|
| 222 |
test_that("linear_contrast() works with multiple contrasts in metafor.", { |
| 223 |
|
| 224 |
skip_if_not_installed("metafor") |
| 225 |
suppressPackageStartupMessages(library(metafor)) |
| 226 |
|
| 227 |
dat <- dat.bangertdrowns2004 |
| 228 |
dat$grade_f <- factor(dat$grade) |
| 229 |
|
| 230 |
res_uni <- rma.mv( |
| 231 |
yi = yi, |
| 232 |
mods = ~ 0 + grade_f, |
| 233 |
V = vi, |
| 234 |
W = 1, |
| 235 |
data=dat, |
| 236 |
sparse = TRUE |
| 237 |
) |
| 238 |
|
| 239 |
contrast_mat <- constrain_pairwise("grade_f", coefs = coef(res_uni), reg_ex = TRUE) |
| 240 |
|
| 241 |
pairwise_contrasts <- linear_contrast( |
| 242 |
res_uni, |
| 243 |
vcov = "CR2", cluster = dat$id, |
| 244 |
contrasts = contrast_mat, |
| 245 |
p_values = TRUE |
| 246 |
) |
| 247 |
|
| 248 |
expect_s3_class(pairwise_contrasts, "conf_int_clubSandwich") |
| 249 |
|
| 250 |
pairwise_tests <- Wald_test( |
| 251 |
res_uni, |
| 252 |
vcov = "CR2", cluster = dat$id, |
| 253 |
constraints = contrast_mat, |
| 254 |
tidy = TRUE |
| 255 |
) |
| 256 |
|
| 257 |
expect_equal(pairwise_contrasts$df, pairwise_tests$df_denom) |
| 258 |
expect_equal(pairwise_contrasts$p_val, pairwise_tests$p_val) |
| 259 |
|
| 260 |
}) |