| 1 |
context("multi-variate multi-level lme objects") |
| 2 |
set.seed(20190513) |
| 3 |
|
| 4 |
skip_on_cran() |
| 5 |
|
| 6 |
dat <- read.table(file="https://raw.githubusercontent.com/wviechtb/multivariate_multilevel_models/master/data.dat", header=TRUE, sep="\t") |
| 7 |
dat$pa <- rowMeans(dat[, grepl("pa", names(dat))]) |
| 8 |
dat$na <- rowMeans(dat[, grepl("na", names(dat))]) |
| 9 |
|
| 10 |
# keep only variables that are needed |
| 11 |
dat <- dat[, c("id", "sex", "beep", "pa", "na")] |
| 12 |
|
| 13 |
# keep only the first 10 IDs |
| 14 |
dat <- subset(dat, id <= 10) |
| 15 |
|
| 16 |
# change into very long format |
| 17 |
dat <- reshape(dat, direction="long", varying=c("pa","na"), v.names="y", idvar="obs", timevar="outcome") |
| 18 |
dat$obs <- NULL |
| 19 |
dat <- dat[order(dat$id, dat$beep, dat$outcome),] |
| 20 |
rownames(dat) <- 1:nrow(dat) |
| 21 |
dat$outnum <- dat$outcome |
| 22 |
dat$outcome <- factor(dat$outcome) |
| 23 |
|
| 24 |
|
| 25 |
library(nlme, quietly=TRUE, warn.conflicts=FALSE) |
| 26 |
|
| 27 |
MVML_full <- lme(y ~ outcome - 1, |
| 28 |
random = ~ outcome - 1 | id, |
| 29 |
weights = varIdent(form = ~ 1 | outcome), |
| 30 |
correlation = corSymm(form = ~ outnum | id/beep), |
| 31 |
data = dat, na.action = na.omit) |
| 32 |
|
| 33 |
MVML_diag <- lme(y ~ outcome - 1, |
| 34 |
random = ~ outcome - 1 | id, |
| 35 |
weights = varIdent(form = ~ 1 | outcome), |
| 36 |
data = dat, na.action = na.omit) |
| 37 |
|
| 38 |
gls_full <- gls(y ~ outcome - 1, |
| 39 |
weights = varIdent(form = ~ 1 | outcome), |
| 40 |
correlation = corSymm(form = ~ outnum | id/beep), |
| 41 |
data = dat, na.action = na.omit) |
| 42 |
|
| 43 |
objects <- list(MVML_full = MVML_full, MVML_diag = MVML_diag, gls = gls_full) |
| 44 |
|
| 45 |
CR2_mats <- lapply(objects, vcovCR, type = "CR2") |
| 46 |
|
| 47 |
test_that("bread works", { |
| 48 |
bread_checks <- lapply(objects, check_bread, cluster = dat$id, y = dat$y) |
| 49 |
expect_true(all(unlist(bread_checks))) |
| 50 |
|
| 51 |
obj_vcovs <- lapply(objects, vcov) |
| 52 |
obj_bread <- lapply(objects, function(obj) obj$sigma^2 * sandwich::bread(obj) / v_scale(obj)) |
| 53 |
expect_equal(obj_vcovs, obj_bread) |
| 54 |
}) |
| 55 |
|
| 56 |
|
| 57 |
test_that("vcovCR options work for CR2", { |
| 58 |
|
| 59 |
expect_identical(vcovCR(MVML_full, cluster = dat$id, type = "CR2"), CR2_mats[["MVML_full"]]) |
| 60 |
expect_equal(vcovCR(MVML_full, type = "CR2", inverse_var = TRUE), CR2_mats[["MVML_full"]]) |
| 61 |
expect_false(identical(vcovCR(MVML_full, type = "CR2", inverse_var = FALSE), CR2_mats[["MVML_full"]])) |
| 62 |
target <- targetVariance(MVML_full) |
| 63 |
expect_equal(vcovCR(MVML_full, type = "CR2", target = target, inverse_var = TRUE), CR2_mats[["MVML_full"]]) |
| 64 |
attr(CR2_mats[["MVML_full"]], "inverse_var") <- FALSE |
| 65 |
expect_equal(vcovCR(MVML_full, type = "CR2", target = target, inverse_var = FALSE), CR2_mats[["MVML_full"]]) |
| 66 |
|
| 67 |
}) |
| 68 |
|
| 69 |
|
| 70 |
test_that("CR2 is target-unbiased", { |
| 71 |
|
| 72 |
CR2_checks <- mapply(check_CR, obj = objects, vcov = CR2_mats) |
| 73 |
expect_true(all(CR2_checks)) |
| 74 |
}) |
| 75 |
|
| 76 |
|
| 77 |
CR_types <- paste0("CR",0:3) |
| 78 |
|
| 79 |
test_that("Order doesn't matter.", { |
| 80 |
|
| 81 |
check_sort_order(MVML_full, dat, seed = 20200530) |
| 82 |
check_sort_order(MVML_diag, dat, seed = 20200530) |
| 83 |
|
| 84 |
}) |
| 85 |
|
| 86 |
|
| 87 |
test_that("clubSandwich works with dropped observations", { |
| 88 |
|
| 89 |
dat_miss <- dat |
| 90 |
dat_miss$y[sample.int(nrow(dat), size = round(nrow(dat) / 10))] <- NA |
| 91 |
obj_dropped <- update(MVML_full, data = dat_miss, na.action = na.omit) |
| 92 |
obj_complete <- update(MVML_full, data = dat_miss, subset = !is.na(y)) |
| 93 |
|
| 94 |
CR_drop <- lapply(CR_types, function(x) vcovCR(obj_dropped, type = x)) |
| 95 |
CR_complete <- lapply(CR_types, function(x) vcovCR(obj_complete, type = x)) |
| 96 |
expect_identical(CR_drop, CR_complete) |
| 97 |
|
| 98 |
test_drop <- lapply(CR_drop, function(x) coef_test(obj_dropped, vcov = x, test = "All", p_values = FALSE)) |
| 99 |
test_complete <- lapply(CR_complete, function(x) coef_test(obj_complete, vcov = x, test = "All", p_values = FALSE)) |
| 100 |
expect_identical(test_drop, test_complete) |
| 101 |
}) |
| 102 |
|
| 103 |
|
| 104 |
test_that("Possible to cluster at higher level than random effects", { |
| 105 |
|
| 106 |
# create 4th level |
| 107 |
n_groups <- nlevels(factor(dat$id)) |
| 108 |
group_id <- rep(1:n_groups, each = 4)[dat$id] |
| 109 |
|
| 110 |
# cluster at level 4 |
| 111 |
expect_is(vcovCR(MVML_full, type = "CR2", cluster = group_id), "vcovCR") |
| 112 |
expect_is(vcovCR(MVML_diag, type = "CR2", cluster = group_id), "vcovCR") |
| 113 |
|
| 114 |
}) |