| 1 |
context("lmerMod objects") |
| 2 |
set.seed(20191217) |
| 3 |
|
| 4 |
skip_if_not_installed("lme4") |
| 5 |
skip_if_not_installed("nlme") |
| 6 |
skip_if_not_installed("mlmRev") |
| 7 |
|
| 8 |
suppressMessages(library(lme4, quietly=TRUE)) |
| 9 |
library(nlme, quietly=TRUE, warn.conflicts=FALSE) |
| 10 |
|
| 11 |
obj_A1 <- lmer(Reaction ~ Days + (Days | Subject), sleepstudy) |
| 12 |
obj_A2 <- lmer(Reaction ~ Days + (Days || Subject), sleepstudy) |
| 13 |
|
| 14 |
data(Orthodont, package="nlme") |
| 15 |
|
| 16 |
obj_B1 <- lmer(distance ~ age + (1 | Subject), data=Orthodont) |
| 17 |
obj_B2 <- lmer(distance ~ age + (age || Subject), data=Orthodont) |
| 18 |
|
| 19 |
data(egsingle, package = "mlmRev") |
| 20 |
egsingle <- within(egsingle, { |
| 21 |
size <- (size - mean(size)) / sd(size) |
| 22 |
}) |
| 23 |
|
| 24 |
obj_C1 <- lmer(math ~ year * size + female + black + hispanic + (1 | schoolid) + (1 | childid), |
| 25 |
data = egsingle) |
| 26 |
obj_C2 <- lmer(math ~ year * size + female + black + hispanic + (year | schoolid) + (1 | childid), |
| 27 |
data = egsingle, control = lmerControl(check.conv.grad = .makeCC("ignore", tol = 2e-3, relTol = NULL))) |
| 28 |
|
| 29 |
test_that("bread works", { |
| 30 |
|
| 31 |
expect_true(check_bread(obj_A1, cluster = sleepstudy$Subject, y = sleepstudy$Reaction)) |
| 32 |
expect_true(check_bread(obj_A2, cluster = sleepstudy$Subject, y = sleepstudy$Reaction)) |
| 33 |
expect_true(check_bread(obj_B1, cluster = Orthodont$Subject, y = Orthodont$distance)) |
| 34 |
expect_true(check_bread(obj_B2, cluster = Orthodont$Subject, y = Orthodont$distance)) |
| 35 |
expect_true(check_bread(obj_C1, cluster = egsingle$schoolid, y = egsingle$math)) |
| 36 |
expect_true(check_bread(obj_C2, cluster = egsingle$schoolid, y = egsingle$math)) |
| 37 |
|
| 38 |
expect_equal(as.matrix(vcov(obj_A1)), bread(obj_A1) * getME(obj_A1, "sigma")^2 / v_scale(obj_A1)) |
| 39 |
expect_equal(as.matrix(vcov(obj_A2)), bread(obj_A2) * getME(obj_A2, "sigma")^2 / v_scale(obj_A2)) |
| 40 |
expect_equal(as.matrix(vcov(obj_B1)), bread(obj_B1) * getME(obj_B1, "sigma")^2 / v_scale(obj_B1)) |
| 41 |
expect_equal(as.matrix(vcov(obj_B2)), bread(obj_B2) * getME(obj_B2, "sigma")^2 / v_scale(obj_B2)) |
| 42 |
expect_equal(as.matrix(vcov(obj_C1)), bread(obj_C1) * getME(obj_C1, "sigma")^2 / v_scale(obj_C1)) |
| 43 |
expect_equal(as.matrix(vcov(obj_C2)), bread(obj_C2) * getME(obj_C2, "sigma")^2 / v_scale(obj_C2)) |
| 44 |
|
| 45 |
}) |
| 46 |
|
| 47 |
test_that("vcovCR options work for CR2", { |
| 48 |
|
| 49 |
CR2_A <- vcovCR(obj_A1, type = "CR2") |
| 50 |
expect_equal(vcovCR(obj_A1, cluster = sleepstudy$Subject, type = "CR2"), CR2_A) |
| 51 |
expect_equal(vcovCR(obj_A1, type = "CR2", inverse_var = TRUE), CR2_A) |
| 52 |
expect_false(identical(vcovCR(obj_A1, type = "CR2", inverse_var = FALSE), CR2_A)) |
| 53 |
target <- targetVariance(obj_A1) |
| 54 |
expect_equal(vcovCR(obj_A1, type = "CR2", target = target, inverse_var = TRUE), CR2_A, check.attributes = FALSE) |
| 55 |
expect_equal(vcovCR(obj_A1, type = "CR2", target = target, inverse_var = FALSE), CR2_A, check.attributes = FALSE) |
| 56 |
|
| 57 |
CR2_B <- vcovCR(obj_B1, type = "CR2") |
| 58 |
expect_equal(vcovCR(obj_B1, cluster = Orthodont$Subject, type = "CR2"), CR2_B) |
| 59 |
expect_equal(vcovCR(obj_B1, type = "CR2", inverse_var = TRUE), CR2_B) |
| 60 |
expect_false(identical(vcovCR(obj_B1, type = "CR2", inverse_var = FALSE), CR2_B)) |
| 61 |
target <- targetVariance(obj_B1) |
| 62 |
expect_equal(vcovCR(obj_B1, type = "CR2", target = target, inverse_var = TRUE), CR2_B, check.attributes = FALSE) |
| 63 |
expect_equal(vcovCR(obj_B1, type = "CR2", target = target, inverse_var = FALSE), CR2_B, check.attributes = FALSE) |
| 64 |
|
| 65 |
CR2_C <- vcovCR(obj_C1, type = "CR2") |
| 66 |
expect_equal(vcovCR(obj_C1, cluster = egsingle$schoolid, type = "CR2"), CR2_C) |
| 67 |
expect_equal(vcovCR(obj_C1, type = "CR2", inverse_var = TRUE), CR2_C) |
| 68 |
expect_false(identical(vcovCR(obj_C1, type = "CR2", inverse_var = FALSE), CR2_C)) |
| 69 |
target <- targetVariance(obj_C1) |
| 70 |
expect_equal(vcovCR(obj_C1, type = "CR2", target = target, inverse_var = TRUE), CR2_C, check.attributes = FALSE) |
| 71 |
expect_equal(vcovCR(obj_C1, type = "CR2", target = target, inverse_var = FALSE), CR2_C, check.attributes = FALSE) |
| 72 |
|
| 73 |
}) |
| 74 |
|
| 75 |
test_that("vcovCR options work for CR4", { |
| 76 |
|
| 77 |
CR4_A <- vcovCR(obj_A1, type = "CR4") |
| 78 |
expect_equal(vcovCR(obj_A1, cluster = sleepstudy$Subject, type = "CR4"), CR4_A) |
| 79 |
expect_equal(vcovCR(obj_A1, type = "CR4", inverse_var = TRUE), CR4_A) |
| 80 |
expect_false(identical(vcovCR(obj_A1, type = "CR4", inverse_var = FALSE), CR4_A)) |
| 81 |
target <- targetVariance(obj_A1) |
| 82 |
expect_equal(vcovCR(obj_A1, type = "CR4", target = target, inverse_var = TRUE), CR4_A, check.attributes = FALSE) |
| 83 |
expect_equal(vcovCR(obj_A1, type = "CR4", target = target, inverse_var = FALSE), CR4_A, check.attributes = FALSE) |
| 84 |
|
| 85 |
CR4_B <- vcovCR(obj_B1, type = "CR4") |
| 86 |
expect_equal(vcovCR(obj_B1, cluster = Orthodont$Subject, type = "CR4"), CR4_B) |
| 87 |
expect_equal(vcovCR(obj_B1, type = "CR4", inverse_var = TRUE), CR4_B) |
| 88 |
expect_false(identical(vcovCR(obj_B1, type = "CR4", inverse_var = FALSE), CR4_B)) |
| 89 |
target <- targetVariance(obj_B1) |
| 90 |
expect_equal(vcovCR(obj_B1, type = "CR4", target = target, inverse_var = TRUE), CR4_B, check.attributes = FALSE) |
| 91 |
expect_equal(vcovCR(obj_B1, type = "CR4", target = target, inverse_var = FALSE), CR4_B, check.attributes = FALSE) |
| 92 |
|
| 93 |
CR4_C <- vcovCR(obj_C1, type = "CR4") |
| 94 |
expect_equal(vcovCR(obj_C1, cluster = egsingle$schoolid, type = "CR4"), CR4_C) |
| 95 |
expect_equal(vcovCR(obj_C1, type = "CR4", inverse_var = TRUE), CR4_C) |
| 96 |
expect_false(identical(vcovCR(obj_C1, type = "CR4", inverse_var = FALSE), CR4_C)) |
| 97 |
target <- targetVariance(obj_C1) |
| 98 |
expect_equal(vcovCR(obj_C1, type = "CR4", target = target, inverse_var = TRUE), CR4_C, check.attributes = FALSE) |
| 99 |
expect_equal(vcovCR(obj_C1, type = "CR4", target = target, inverse_var = FALSE), CR4_C, check.attributes = FALSE) |
| 100 |
|
| 101 |
}) |
| 102 |
|
| 103 |
|
| 104 |
test_that("CR2 and CR4 are target-unbiased", { |
| 105 |
expect_true(check_CR(obj_A1, vcov = "CR2")) |
| 106 |
expect_true(check_CR(obj_A2, vcov = "CR2")) |
| 107 |
expect_true(check_CR(obj_B1, vcov = "CR2")) |
| 108 |
expect_true(check_CR(obj_B2, vcov = "CR2")) |
| 109 |
expect_true(check_CR(obj_C1, vcov = "CR2")) |
| 110 |
expect_true(check_CR(obj_C2, vcov = "CR2")) |
| 111 |
expect_true(check_CR(obj_A1, vcov = "CR4")) |
| 112 |
expect_true(check_CR(obj_A2, vcov = "CR4")) |
| 113 |
expect_true(check_CR(obj_B1, vcov = "CR4")) |
| 114 |
expect_true(check_CR(obj_B2, vcov = "CR4")) |
| 115 |
expect_true(check_CR(obj_C1, vcov = "CR4")) |
| 116 |
expect_true(check_CR(obj_C2, vcov = "CR4")) |
| 117 |
}) |
| 118 |
|
| 119 |
|
| 120 |
CR_types <- paste0("CR",0:4) |
| 121 |
|
| 122 |
test_that("Order doesn't matter.", { |
| 123 |
|
| 124 |
skip_on_cran() |
| 125 |
|
| 126 |
# Model A1 |
| 127 |
check_sort_order(obj = obj_A1, dat = sleepstudy) |
| 128 |
|
| 129 |
# Model C1 |
| 130 |
check_sort_order(obj = obj_C1, dat = egsingle) |
| 131 |
}) |
| 132 |
|
| 133 |
|
| 134 |
test_that("clubSandwich works with dropped observations", { |
| 135 |
|
| 136 |
dat_miss <- sleepstudy |
| 137 |
dat_miss$Reaction[sample.int(nrow(sleepstudy), size = round(nrow(sleepstudy) / 10))] <- NA |
| 138 |
obj_dropped <- update(obj_A1, data = dat_miss, na.action = na.omit) |
| 139 |
obj_complete <- update(obj_A1, data = dat_miss, subset = !is.na(Reaction)) |
| 140 |
|
| 141 |
CR_drop <- lapply(CR_types, function(x) vcovCR(obj_dropped, type = x)) |
| 142 |
CR_complete <- lapply(CR_types, function(x) vcovCR(obj_complete, type = x)) |
| 143 |
expect_equal(CR_drop, CR_complete) |
| 144 |
|
| 145 |
test_drop <- lapply(CR_types, function(x) coef_test(obj_dropped, vcov = x, test = "All", p_values = FALSE)) |
| 146 |
test_complete <- lapply(CR_types, function(x) coef_test(obj_complete, vcov = x, test = "All", p_values = FALSE)) |
| 147 |
expect_equal(test_drop, test_complete) |
| 148 |
}) |
| 149 |
|
| 150 |
|
| 151 |
|
| 152 |
test_that("lmer agrees with lme", { |
| 153 |
|
| 154 |
data(BodyWeight, package="nlme") |
| 155 |
|
| 156 |
lmer_fit <- lmer(weight ~ Time * Diet + (1 | Rat), data=BodyWeight) |
| 157 |
lme_fit <- lme(weight ~ Time * Diet, data=BodyWeight, ~ 1 | Rat) |
| 158 |
|
| 159 |
expect_equal(coef_CS(lmer_fit), coef_CS(lme_fit)) |
| 160 |
expect_equal(nobs(lmer_fit), nobs(lme_fit)) |
| 161 |
expect_equal(model_matrix(lmer_fit), model_matrix(lme_fit), check.attributes = FALSE) |
| 162 |
expect_equal(residuals_CS(lmer_fit), residuals_CS(lme_fit), check.attributes = FALSE) |
| 163 |
expect_equal(v_scale(lmer_fit), v_scale(lme_fit)) |
| 164 |
|
| 165 |
p <- length(coef_CS(lmer_fit)) |
| 166 |
expect_equal(bread(lmer_fit) / bread(lme_fit), matrix(1, p, p), check.attributes = FALSE, tol = 10^-6) |
| 167 |
expect_equal(targetVariance(lmer_fit), targetVariance(lme_fit), check.attributes = FALSE, tol = 10^-6) |
| 168 |
expect_equal(weightMatrix(lmer_fit), weightMatrix(lme_fit), check.attributes = FALSE, tol = 10^-6) |
| 169 |
|
| 170 |
CR_lmer <- lapply(CR_types, function(x) vcovCR(lmer_fit, type = x)) |
| 171 |
CR_lme <- lapply(CR_types, function(x) vcovCR(lme_fit, type = x)) |
| 172 |
expect_equivalent(CR_lmer, CR_lme, tolerance = 10^-6) |
| 173 |
|
| 174 |
test_lmer <- lapply(CR_types, function(x) coef_test(lmer_fit, vcov = x, test = "All", p_values = FALSE)) |
| 175 |
test_lme <- lapply(CR_types, function(x) coef_test(lme_fit, vcov = x, test = "All", p_values = FALSE)) |
| 176 |
compare_ttests(test_lmer, test_lme, tol = 10^-10) |
| 177 |
|
| 178 |
constraints <- c(combn(length(coef_CS(lmer_fit)), 2, simplify = FALSE), |
| 179 |
combn(length(coef_CS(lmer_fit)), 3, simplify = FALSE)) |
| 180 |
Wald_lmer <- Wald_test(lmer_fit, constraints = constrain_zero(constraints), vcov = "CR2", test = "All") |
| 181 |
Wald_lme <- Wald_test(lme_fit, constraints = constrain_zero(constraints), vcov = "CR2", test = "All") |
| 182 |
compare_Waldtests(Wald_lmer, Wald_lme) |
| 183 |
|
| 184 |
}) |
| 185 |
|
| 186 |
|
| 187 |
test_that("Emply levels are dropped in model_matrix", { |
| 188 |
|
| 189 |
data(AchievementAwardsRCT) |
| 190 |
|
| 191 |
AA_RCT_females <- subset(AchievementAwardsRCT, sex=="Girl" & year != "1999") |
| 192 |
|
| 193 |
AA_RCT_females <- within(AA_RCT_females, { |
| 194 |
sibs_4 <- siblings >= 4 |
| 195 |
treated2001 <- treated * (year=="2001") |
| 196 |
}) |
| 197 |
|
| 198 |
lmer_fit <- lmer(Bagrut_status ~ year * school_type + |
| 199 |
father_ed + mother_ed + immigrant + sibs_4 + |
| 200 |
qrtl + treated2001:half + (1 | school_id), |
| 201 |
data = AA_RCT_females) |
| 202 |
|
| 203 |
betas <- fixef(lmer_fit) |
| 204 |
X <- model_matrix(lmer_fit) |
| 205 |
expect_identical(names(betas), colnames(X)) |
| 206 |
|
| 207 |
}) |
| 208 |
|
| 209 |
|
| 210 |
|
| 211 |
test_that("Possible to cluster at higher level than random effects", { |
| 212 |
|
| 213 |
n_districts <- 10 |
| 214 |
n_schools_per <- rnbinom(n_districts, size = 4, prob = 0.3) |
| 215 |
n_schools <- sum(n_schools_per) |
| 216 |
n_students_per <- 10 |
| 217 |
n_students <- n_schools * n_students_per |
| 218 |
|
| 219 |
# identifiers for each level |
| 220 |
district_id <- factor(rep(1:n_districts, n_schools_per * n_students_per)) |
| 221 |
school_id <- factor(rep(1:sum(n_schools_per), each = n_students_per)) |
| 222 |
student_id <- 1:n_students |
| 223 |
|
| 224 |
# simulated outcome |
| 225 |
Y <- rnorm(n_districts)[district_id] + rnorm(n_schools)[school_id] + rnorm(n_students) |
| 226 |
X <- rnorm(n_students) |
| 227 |
dat <- data.frame(district_id, school_id, student_id, Y, X) |
| 228 |
dat_scramble <- dat[sample(nrow(dat)),] |
| 229 |
|
| 230 |
# fit two-level model |
| 231 |
lme_2level <- lmer(Y ~ X + (1 | school_id), data = dat) |
| 232 |
|
| 233 |
# cluster at level 3 |
| 234 |
V <- vcovCR(lme_2level, type = "CR2", cluster = dat$district_id) |
| 235 |
expect_is(V, "vcovCR") |
| 236 |
expect_error(vcovCR(lme_2level, type = "CR2", cluster = dat_scramble$district_id)) |
| 237 |
|
| 238 |
# check that result does not depend on sort-order |
| 239 |
V_scramble <- vcovCR(lmer(Y ~ X + (1 | school_id), data = dat_scramble), |
| 240 |
type = "CR2", cluster = dat_scramble$district_id) |
| 241 |
expect_equal(as.matrix(V), as.matrix(V_scramble)) |
| 242 |
}) |