| 1 |
context("impute_covariance_matrix") |
| 2 |
set.seed(20190513) |
| 3 |
|
| 4 |
K <- 20 |
| 5 |
N <- sum(1:K) |
| 6 |
dat <- data.frame( |
| 7 |
study = rep(LETTERS[1:K], 1:K), |
| 8 |
yi = rnorm(N), |
| 9 |
vi = rchisq(N, df = 2), |
| 10 |
ti = sample(1:(10 * N), N), |
| 11 |
si = sample(letters[1:3], N, replace = TRUE) |
| 12 |
) |
| 13 |
|
| 14 |
dat$v_study <- unsplit(tapply(dat$vi, dat$study, mean), dat$study) |
| 15 |
|
| 16 |
dat_scramble <- dat[sample.int(nrow(dat)),] |
| 17 |
|
| 18 |
test_that("impute_covariance_matrix() is deprecated.", { |
| 19 |
expect_warning( |
| 20 |
impute_covariance_matrix(vi = dat$vi, cluster = dat$study, r = 0.6) |
| 21 |
) |
| 22 |
}) |
| 23 |
test_that("impute_covariance_matrix error messages and missing argument handling are correct.", { |
| 24 |
|
| 25 |
expect_error(impute_covariance_matrix(vi = dat$vi, cluster = dat$study)) |
| 26 |
expect_error(impute_covariance_matrix(vi = dat$vi, cluster = dat$study, ti = dat$ti)) |
| 27 |
expect_error(impute_covariance_matrix(vi = dat$vi, cluster = dat$study, ar1 = 0.8)) |
| 28 |
|
| 29 |
V1 <- impute_covariance_matrix(vi = dat$vi, cluster = dat$study, r = 0.6) |
| 30 |
V2 <- impute_covariance_matrix(vi = dat$vi, cluster = dat$study, r = 0.6, ti = dat$ti) |
| 31 |
V3 <- impute_covariance_matrix(vi = dat$vi, cluster = dat$study, r = 0.6, ti = dat$ti, ar1 = 0) |
| 32 |
V4 <- impute_covariance_matrix(vi = dat$vi, cluster = dat$study, ti = dat$ti, ar1 = 0.5) |
| 33 |
V5 <- impute_covariance_matrix(vi = dat$vi, cluster = dat$study, r = 0.0, ti = dat$ti, ar1 = 0.5) |
| 34 |
|
| 35 |
expect_equal(V1, V2) |
| 36 |
expect_equal(V1, V3) |
| 37 |
expect_equal(V4, V5) |
| 38 |
}) |
| 39 |
|
| 40 |
test_that("impute_covariance_matrix returns correct correlations.", { |
| 41 |
r <- 0.7 |
| 42 |
V_single_r <- impute_covariance_matrix(vi = dat$vi, cluster = dat$study, r = r) |
| 43 |
r_list <- rbeta(K, 2, 2) |
| 44 |
V_multiple_r <- impute_covariance_matrix(vi = dat$vi, cluster = dat$study, r = r_list) |
| 45 |
|
| 46 |
check_correlation <- function(M, r) if (nrow(M) > 1) max(abs(cov2cor(M)[lower.tri(M)] - r)) else 0 |
| 47 |
check_singles <- sapply(V_single_r, check_correlation, r = r) |
| 48 |
expect_true(all(check_singles < 10^-14)) |
| 49 |
check_multiples <- Map(check_correlation, M = V_multiple_r, r = r_list) |
| 50 |
expect_true(all(check_multiples < 10^-14)) |
| 51 |
|
| 52 |
dat_scramble <- dat[sample(nrow(dat)),] |
| 53 |
V_mat <- impute_covariance_matrix(vi = dat_scramble$vi, cluster = dat_scramble$study, r = r) |
| 54 |
expect_equal(dat_scramble$vi, diag(V_mat)) |
| 55 |
|
| 56 |
V_resorted <- V_mat[order(dat_scramble$study), order(dat_scramble$study)] |
| 57 |
dat_unscramble <- dat_scramble[order(dat_scramble$study),] |
| 58 |
V_unscramble <- impute_covariance_matrix(vi = dat_unscramble$vi, cluster = dat_unscramble$study, r = r) |
| 59 |
expect_equal(V_resorted, unblock(V_unscramble)) |
| 60 |
|
| 61 |
}) |
| 62 |
|
| 63 |
|
| 64 |
|
| 65 |
test_that("impute_covariance_matrix works with AR1 argument.", { |
| 66 |
|
| 67 |
ar1 <- 0.5 |
| 68 |
V_mat <- impute_covariance_matrix(vi = dat$vi, cluster = dat$study, ti = 1:nrow(dat), ar1 = ar1) |
| 69 |
r_mat <- lapply(V_mat, cov2cor) |
| 70 |
r_sums <- sapply(r_mat, sum) |
| 71 |
check <- sapply(2:K, function(k) k + 2 * sum((1:(k-1)) * ar1^((k-1):1))) |
| 72 |
expect_equal(r_sums[-1], check, check.attributes = FALSE) |
| 73 |
|
| 74 |
V_mat_big <- impute_covariance_matrix(vi = dat$vi, cluster = dat$study, |
| 75 |
ti = 1:nrow(dat), ar1 = ar1, return_list = FALSE) |
| 76 |
r_mat_big <- cov2cor(V_mat_big) |
| 77 |
expect_equal(sum(r_mat_big), sum(check) + 1) |
| 78 |
v_sums <- sapply(V_mat, sum) |
| 79 |
expect_equal(sum(v_sums), sum(V_mat_big)) |
| 80 |
|
| 81 |
}) |
| 82 |
|
| 83 |
test_that("impute_covariance_matrix works with subgroup argument.", { |
| 84 |
|
| 85 |
X <- model.matrix(~ si + 0, data = dat) |
| 86 |
X_list <- by(X, dat$study, as.matrix) |
| 87 |
|
| 88 |
check_diag <- function(v, x) { |
| 89 |
XVX <- t(x) %*% v %*% x |
| 90 |
all.equal(sum(XVX), sum(diag(XVX))) |
| 91 |
} |
| 92 |
|
| 93 |
check_all_diag <- function(v_list,x_list) { |
| 94 |
all(mapply(check_diag , v = v_list, x = x_list)) |
| 95 |
} |
| 96 |
|
| 97 |
V1 <- impute_covariance_matrix(vi = dat$vi, cluster = dat$study, r = 0.6, subgroup = dat$si) |
| 98 |
V2 <- impute_covariance_matrix(vi = dat$vi, cluster = dat$study, r = 0.6, ti = dat$ti, subgroup = dat$si) |
| 99 |
V3 <- impute_covariance_matrix(vi = dat$vi, cluster = dat$study, r = 0.6, ti = dat$ti, ar1 = 0, subgroup = dat$si) |
| 100 |
V4 <- impute_covariance_matrix(vi = dat$vi, cluster = dat$study, ti = dat$ti, ar1 = 0.5, subgroup = dat$si) |
| 101 |
V5 <- impute_covariance_matrix(vi = dat$vi, cluster = dat$study, r = 0.0, ti = dat$ti, ar1 = 0.5, subgroup = dat$si) |
| 102 |
|
| 103 |
expect_true(check_all_diag(V1, X_list)) |
| 104 |
expect_true(check_all_diag(V2, X_list)) |
| 105 |
expect_true(check_all_diag(V3, X_list)) |
| 106 |
expect_true(check_all_diag(V4, X_list)) |
| 107 |
expect_true(check_all_diag(V5, X_list)) |
| 108 |
|
| 109 |
X_scramble <- model.matrix(~ si + 0, data = dat_scramble) |
| 110 |
|
| 111 |
V1 <- impute_covariance_matrix(vi = dat_scramble$vi, cluster = dat_scramble$study, r = 0.6, subgroup = dat_scramble$si) |
| 112 |
V2 <- impute_covariance_matrix(vi = dat_scramble$vi, cluster = dat_scramble$study, r = 0.6, ti = dat_scramble$ti, subgroup = dat_scramble$si) |
| 113 |
V3 <- impute_covariance_matrix(vi = dat_scramble$vi, cluster = dat_scramble$study, r = 0.6, ti = dat_scramble$ti, ar1 = 0, subgroup = dat_scramble$si) |
| 114 |
V4 <- impute_covariance_matrix(vi = dat_scramble$vi, cluster = dat_scramble$study, ti = dat_scramble$ti, ar1 = 0.5, subgroup = dat_scramble$si) |
| 115 |
V5 <- impute_covariance_matrix(vi = dat_scramble$vi, cluster = dat_scramble$study, r = 0.0, ti = dat_scramble$ti, ar1 = 0.5, subgroup = dat_scramble$si) |
| 116 |
|
| 117 |
expect_true(check_diag(V1, X_scramble)) |
| 118 |
expect_true(check_diag(V2, X_scramble)) |
| 119 |
expect_true(check_diag(V3, X_scramble)) |
| 120 |
expect_true(check_diag(V4, X_scramble)) |
| 121 |
expect_true(check_diag(V5, X_scramble)) |
| 122 |
|
| 123 |
}) |
| 124 |
|
| 125 |
test_that("impute_covariance_matrix works with smooth argument.", { |
| 126 |
|
| 127 |
V1 <- impute_covariance_matrix(vi = dat$v_study, cluster = dat$study, r = 0.6) |
| 128 |
V2 <- impute_covariance_matrix(vi = dat$v_study, cluster = dat$study, r = 0.6, ti = dat$ti) |
| 129 |
V3 <- impute_covariance_matrix(vi = dat$v_study, cluster = dat$study, r = 0.6, ti = dat$ti, ar1 = 0) |
| 130 |
V4 <- impute_covariance_matrix(vi = dat$v_study, cluster = dat$study, ti = dat$ti, ar1 = 0.5) |
| 131 |
V5 <- impute_covariance_matrix(vi = dat$v_study, cluster = dat$study, r = 0.0, ti = dat$ti, ar1 = 0.5) |
| 132 |
|
| 133 |
V1s <- impute_covariance_matrix(vi = dat$vi, cluster = dat$study, r = 0.6, smooth_vi = TRUE) |
| 134 |
V2s <- impute_covariance_matrix(vi = dat$vi, cluster = dat$study, r = 0.6, ti = dat$ti, smooth_vi = TRUE) |
| 135 |
V3s <- impute_covariance_matrix(vi = dat$vi, cluster = dat$study, r = 0.6, ti = dat$ti, ar1 = 0, smooth_vi = TRUE) |
| 136 |
V4s <- impute_covariance_matrix(vi = dat$vi, cluster = dat$study, ti = dat$ti, ar1 = 0.5, smooth_vi = TRUE) |
| 137 |
V5s <- impute_covariance_matrix(vi = dat$vi, cluster = dat$study, r = 0.0, ti = dat$ti, ar1 = 0.5, smooth_vi = TRUE) |
| 138 |
|
| 139 |
expect_equal(V1, V1s) |
| 140 |
expect_equal(V2, V2s) |
| 141 |
expect_equal(V3, V3s) |
| 142 |
expect_equal(V4, V4s) |
| 143 |
expect_equal(V5, V5s) |
| 144 |
|
| 145 |
V1 <- impute_covariance_matrix(vi = dat_scramble$v_study, cluster = dat_scramble$study, r = 0.6) |
| 146 |
V2 <- impute_covariance_matrix(vi = dat_scramble$v_study, cluster = dat_scramble$study, r = 0.6, ti = dat_scramble$ti) |
| 147 |
V3 <- impute_covariance_matrix(vi = dat_scramble$v_study, cluster = dat_scramble$study, r = 0.6, ti = dat_scramble$ti, ar1 = 0) |
| 148 |
V4 <- impute_covariance_matrix(vi = dat_scramble$v_study, cluster = dat_scramble$study, ti = dat_scramble$ti, ar1 = 0.5) |
| 149 |
V5 <- impute_covariance_matrix(vi = dat_scramble$v_study, cluster = dat_scramble$study, r = 0.0, ti = dat_scramble$ti, ar1 = 0.5) |
| 150 |
V1s <- impute_covariance_matrix(vi = dat_scramble$vi, cluster = dat_scramble$study, r = 0.6, smooth_vi = TRUE) |
| 151 |
V2s <- impute_covariance_matrix(vi = dat_scramble$vi, cluster = dat_scramble$study, r = 0.6, ti = dat_scramble$ti, smooth_vi = TRUE) |
| 152 |
V3s <- impute_covariance_matrix(vi = dat_scramble$vi, cluster = dat_scramble$study, r = 0.6, ti = dat_scramble$ti, ar1 = 0, smooth_vi = TRUE) |
| 153 |
V4s <- impute_covariance_matrix(vi = dat_scramble$vi, cluster = dat_scramble$study, ti = dat_scramble$ti, ar1 = 0.5, smooth_vi = TRUE) |
| 154 |
V5s <- impute_covariance_matrix(vi = dat_scramble$vi, cluster = dat_scramble$study, r = 0.0, ti = dat_scramble$ti, ar1 = 0.5, smooth_vi = TRUE) |
| 155 |
|
| 156 |
expect_equal(V1, V1s) |
| 157 |
expect_equal(V2, V2s) |
| 158 |
expect_equal(V3, V3s) |
| 159 |
expect_equal(V4, V4s) |
| 160 |
expect_equal(V5, V5s) |
| 161 |
|
| 162 |
}) |
| 163 |
|
| 164 |
test_that("impute_covariance_matrix works with unobserved factors.", { |
| 165 |
K <- 10 |
| 166 |
N <- sum(1:K) |
| 167 |
dat <- data.frame(study = rep(LETTERS[1:K], 1:K), |
| 168 |
yi = rnorm(N), |
| 169 |
vi = rchisq(N, df = 2)) |
| 170 |
levels(dat$study) <- LETTERS[1:(K + 3)] |
| 171 |
|
| 172 |
r <- 0.7 |
| 173 |
V_single_r <- impute_covariance_matrix(vi = dat$vi, cluster = dat$study, r = r) |
| 174 |
r_list <- rbeta(K, 2, 2) |
| 175 |
V_multiple_r <- impute_covariance_matrix(vi = dat$vi, cluster = dat$study, r = r_list) |
| 176 |
|
| 177 |
check_correlation <- function(M, r) if (nrow(M) > 1) max(abs(cov2cor(M)[lower.tri(M)] - r)) else 0 |
| 178 |
check_singles <- sapply(V_single_r, check_correlation, r = r) |
| 179 |
expect_true(all(check_singles < 10^-14)) |
| 180 |
|
| 181 |
dat_scramble <- dat[sample(nrow(dat)),] |
| 182 |
V_mat <- impute_covariance_matrix(vi = dat_scramble$vi, cluster = dat_scramble$study, r = r) |
| 183 |
expect_equal(dat_scramble$vi, diag(V_mat)) |
| 184 |
|
| 185 |
V_resorted <- V_mat[order(dat_scramble$study), order(dat_scramble$study)] |
| 186 |
dat_unscramble <- dat_scramble[order(dat_scramble$study),] |
| 187 |
V_unscramble <- impute_covariance_matrix(vi = dat_unscramble$vi, cluster = dat_unscramble$study, r = r) |
| 188 |
expect_equal(V_resorted, unblock(V_unscramble)) |
| 189 |
}) |
| 190 |
|
| 191 |
test_that("impute_covariance_matrix works with missing variances.", { |
| 192 |
skip_if_not_installed("robumeta") |
| 193 |
|
| 194 |
data(corrdat, package = "robumeta") |
| 195 |
dat_miss <- corrdat |
| 196 |
dat_miss$var[sample.int(nrow(corrdat), size = round(nrow(corrdat) / 10))] <- NA |
| 197 |
V_missing <- impute_covariance_matrix(dat_miss$var, cluster = dat_miss$studyid, |
| 198 |
r = 0.8, return_list = FALSE) |
| 199 |
non_missing_rows <- !is.na(dat_miss$var) |
| 200 |
V_missing <- V_missing[non_missing_rows, non_missing_rows] |
| 201 |
|
| 202 |
dat_complete <- subset(dat_miss, !is.na(var)) |
| 203 |
V_complete <- impute_covariance_matrix(dat_complete$var, cluster = dat_complete$studyid, |
| 204 |
r = 0.8, return_list = FALSE) |
| 205 |
|
| 206 |
expect_equal(V_missing, V_complete) |
| 207 |
}) |
| 208 |
|
| 209 |
|
| 210 |
|
| 211 |
|
| 212 |
test_that("pattern_covariance_matrix works.", { |
| 213 |
skip_if_not_installed("robumeta") |
| 214 |
skip_if_not_installed("metafor") |
| 215 |
|
| 216 |
data(oswald2013, package = "robumeta") |
| 217 |
dat <- metafor::escalc(data = oswald2013, measure = "ZCOR", ri = R, ni = N) |
| 218 |
|
| 219 |
# make a patterned correlation matrix |
| 220 |
p_levels <- levels(dat$Crit.Cat) |
| 221 |
r_pattern <- 0.7^as.matrix(dist(1:length(p_levels))) |
| 222 |
diag(r_pattern) <- seq(0.75, 0.95, length.out = 6) |
| 223 |
rownames(r_pattern) <- colnames(r_pattern) <- p_levels |
| 224 |
|
| 225 |
# impute the covariance matrix using patterned correlations |
| 226 |
V_list <- pattern_covariance_matrix(vi = dat$vi, |
| 227 |
cluster = dat$Study, |
| 228 |
pattern_level = dat$Crit.Cat, |
| 229 |
r_pattern = r_pattern, |
| 230 |
smooth_vi = TRUE) |
| 231 |
|
| 232 |
r_pattern1 <- r_pattern2 <- r_pattern |
| 233 |
|
| 234 |
# Recreate a constant covariance matrix |
| 235 |
|
| 236 |
r_pattern1[1:6,1:6] <- 0.6 |
| 237 |
|
| 238 |
V_pattern_const <- pattern_covariance_matrix(vi = dat$vi, cluster = dat$Study, pattern_level = dat$Crit.Cat, |
| 239 |
r_pattern = r_pattern1, smooth_vi = FALSE) |
| 240 |
|
| 241 |
V_impute_const <- impute_covariance_matrix(vi = dat$vi, cluster = dat$Study, r = 0.6) |
| 242 |
expect_equal(V_pattern_const, V_impute_const, check.attributes = FALSE) |
| 243 |
|
| 244 |
# Patterns work with excluded categories |
| 245 |
exclude <- 3:5 |
| 246 |
r_pattern2[3:5,] <- 0.3 |
| 247 |
r_pattern2[,3:5] <- 0.3 |
| 248 |
|
| 249 |
V_pattern_exclude <- pattern_covariance_matrix(vi = dat$vi, cluster = dat$Study, pattern_level = dat$Crit.Cat, |
| 250 |
r_pattern = r_pattern[-(3:5),-(3:5)], r = 0.3, smooth_vi = FALSE) |
| 251 |
V_pattern_full <- pattern_covariance_matrix(vi = dat$vi, cluster = dat$Study, pattern_level = dat$Crit.Cat, |
| 252 |
r_pattern = r_pattern2, smooth_vi = FALSE) |
| 253 |
expect_equal(V_pattern_exclude, V_pattern_full) |
| 254 |
|
| 255 |
# Patterns work with extra categories |
| 256 |
levs <- c(p_levels, LETTERS[1:5]) |
| 257 |
r_pattern3 <- matrix(-4, nrow = length(levs), ncol = length(levs), dimnames = list(levs, levs)) |
| 258 |
r_pattern3[p_levels,p_levels] <- r_pattern[p_levels, p_levels] |
| 259 |
|
| 260 |
V_pattern_extra <- pattern_covariance_matrix(vi = dat$vi, cluster = dat$Study, pattern_level = dat$Crit.Cat, |
| 261 |
r_pattern = r_pattern3, r = 0.3, smooth_vi = TRUE) |
| 262 |
expect_equal(V_pattern_extra, V_list) |
| 263 |
|
| 264 |
V_pattern_extra_exclude <- pattern_covariance_matrix(vi = dat$vi, cluster = dat$Study, pattern_level = dat$Crit.Cat, |
| 265 |
r_pattern = r_pattern3[-(3:5),-(3:5)], r = 0.3, smooth_vi = FALSE) |
| 266 |
expect_equal(V_pattern_extra_exclude, V_pattern_full) |
| 267 |
|
| 268 |
VS_pattern_extra_exclude <- pattern_covariance_matrix(vi = dat$vi, cluster = dat$Study, pattern_level = dat$Crit.Cat, |
| 269 |
r_pattern = r_pattern3[-(3:5),-(3:5)], r = 0.3, |
| 270 |
subgroup = dat$IAT.Focus, |
| 271 |
smooth_vi = FALSE, return_list = FALSE) |
| 272 |
VS_pattern_full <- pattern_covariance_matrix(vi = dat$vi, cluster = dat$Study, pattern_level = dat$Crit.Cat, |
| 273 |
r_pattern = r_pattern2, |
| 274 |
subgroup = dat$IAT.Focus, |
| 275 |
smooth_vi = FALSE, return_list = FALSE) |
| 276 |
|
| 277 |
expect_equal(VS_pattern_extra_exclude, VS_pattern_full) |
| 278 |
|
| 279 |
|
| 280 |
expect_warning( |
| 281 |
pattern_covariance_matrix(vi = dat$vi, cluster = dat$Study, pattern_level = dat$Crit.Cat, |
| 282 |
r_pattern = r_pattern3[-(3:5),-(3:5)], r = 4, |
| 283 |
subgroup = dat$IAT.Focus, |
| 284 |
smooth_vi = FALSE, return_list = FALSE) |
| 285 |
) |
| 286 |
|
| 287 |
V_npd <- pattern_covariance_matrix(vi = dat$vi, cluster = dat$Study, pattern_level = dat$Crit.Cat, |
| 288 |
r_pattern = r_pattern3[-(3:5),-(3:5)], r = 4, |
| 289 |
subgroup = dat$IAT.Focus, |
| 290 |
smooth_vi = FALSE, return_list = FALSE, check = FALSE) |
| 291 |
expect_true(inherits(V_npd, "matrix")) |
| 292 |
|
| 293 |
# pattern_covariance_matrix works with missing entries |
| 294 |
dat_miss <- dat |
| 295 |
dat_miss$vi[sample.int(nrow(dat), size = round(nrow(dat) / 10))] <- NA |
| 296 |
|
| 297 |
V_missing <- pattern_covariance_matrix(dat_miss$vi, cluster = dat_miss$Study, |
| 298 |
pattern_level = dat_miss$Crit.Cat, |
| 299 |
r_pattern = r_pattern, return_list = FALSE) |
| 300 |
|
| 301 |
non_missing_rows <- !is.na(dat_miss$vi) |
| 302 |
V_missing <- V_missing[non_missing_rows, non_missing_rows] |
| 303 |
|
| 304 |
dat_complete <- subset(dat_miss, !is.na(vi)) |
| 305 |
V_complete <- pattern_covariance_matrix(dat_complete$vi, cluster = dat_complete$Study, |
| 306 |
pattern_level = dat_complete$Crit.Cat, |
| 307 |
r_pattern = r_pattern, return_list = FALSE) |
| 308 |
|
| 309 |
expect_equal(V_missing, V_complete) |
| 310 |
|
| 311 |
|
| 312 |
dat_miss$Crit.Cat[sample.int(nrow(dat), size = round(nrow(dat) / 10))] <- NA |
| 313 |
|
| 314 |
expect_error( |
| 315 |
pattern_covariance_matrix(dat_miss$vi, cluster = dat_miss$Study, |
| 316 |
pattern_level = dat_miss$Crit.Cat, |
| 317 |
r_pattern = r_pattern, return_list = TRUE, |
| 318 |
check_PD = FALSE) |
| 319 |
) |
| 320 |
}) |
| 321 |
|