Debian Codegraph @

tests/testthat/test_lme_3level.R

raw
1 context("3-level lme objects")
2 set.seed(20190513)
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 library(mlmRev, quietly=TRUE, warn.conflicts=FALSE)
11
12 school_subset <- levels(egsingle$schoolid)
13 school_subset <- sample(school_subset, size = 15)
14 egsingle <- droplevels(subset(egsingle, schoolid %in% school_subset))
15
16
17 obj_A1 <- lme(math ~ year * size + female + black + hispanic,
18 random = list(~ year | schoolid, ~ 1 | childid),
19 data = egsingle,
20 control = lmeControl(tolerance = 1e-4, opt = "optim"))
21 obj_A2 <- update(obj_A1, weights = varIdent(form = ~ 1 | female))
22 obj_A3 <- update(obj_A1, correlation = corExp(form = ~ year))
23 obj_A4 <- update(obj_A2, correlation = corExp(form = ~ year))
24 objects <- list(A1 = obj_A1, A2 = obj_A2, A3 = obj_A3, A4 = obj_A4)
25
26 CR2_mats <- lapply(objects, vcovCR, type = "CR2")
27
28 test_that("bread works", {
29 bread_checks <- lapply(objects, check_bread, cluster = egsingle$schoolid, y = egsingle$math)
30 expect_true(all(unlist(bread_checks)))
31
32 obj_vcovs <- lapply(objects, vcov)
33 obj_bread <- lapply(objects, function(obj) obj$sigma^2 * sandwich::bread(obj) / v_scale(obj))
34 expect_equal(obj_vcovs, obj_bread)
35 })
36
37
38 test_that("vcovCR options work for CR2", {
39 skip_on_cran()
40
41 expect_equal(vcovCR(obj_A1, cluster = egsingle$schoolid, type = "CR2"), CR2_mats[["A1"]])
42 expect_equal(vcovCR(obj_A1, type = "CR2", inverse_var = TRUE), CR2_mats[["A1"]])
43 expect_false(identical(vcovCR(obj_A1, type = "CR2", inverse_var = FALSE), CR2_mats[["A1"]]))
44 target <- targetVariance(obj_A1)
45 expect_equal(vcovCR(obj_A1, type = "CR2", target = target, inverse_var = TRUE), CR2_mats[["A1"]])
46 attr(CR2_mats[["A1"]], "inverse_var") <- FALSE
47 expect_equal(vcovCR(obj_A1, type = "CR2", target = target, inverse_var = FALSE), CR2_mats[["A1"]])
48
49 expect_equal(vcovCR(obj_A2, cluster = egsingle$schoolid, type = "CR2"), CR2_mats[["A2"]])
50 expect_equal(vcovCR(obj_A2, type = "CR2", inverse_var = TRUE), CR2_mats[["A2"]])
51 expect_false(identical(vcovCR(obj_A2, type = "CR2", inverse_var = FALSE), CR2_mats[["A2"]]))
52 target <- targetVariance(obj_A2)
53 expect_equal(vcovCR(obj_A2, type = "CR2", target = target, inverse_var = TRUE), CR2_mats[["A2"]])
54 attr(CR2_mats[["A2"]], "inverse_var") <- FALSE
55 expect_equal(vcovCR(obj_A2, type = "CR2", target = target, inverse_var = FALSE), CR2_mats[["A2"]])
56
57 expect_equal(vcovCR(obj_A3, cluster = egsingle$schoolid, type = "CR2"), CR2_mats[["A3"]])
58 expect_equal(vcovCR(obj_A3, type = "CR2", inverse_var = TRUE), CR2_mats[["A3"]])
59 expect_false(identical(vcovCR(obj_A3, type = "CR2", inverse_var = FALSE), CR2_mats[["A3"]]))
60 target <- targetVariance(obj_A3)
61 expect_equal(vcovCR(obj_A3, type = "CR2", target = target, inverse_var = TRUE), CR2_mats[["A3"]])
62 attr(CR2_mats[["A3"]], "inverse_var") <- FALSE
63 expect_equal(vcovCR(obj_A3, type = "CR2", target = target, inverse_var = FALSE), CR2_mats[["A3"]])
64
65 expect_equal(vcovCR(obj_A4, cluster = egsingle$schoolid, type = "CR2"), CR2_mats[["A4"]])
66 expect_equal(vcovCR(obj_A4, type = "CR2", inverse_var = TRUE), CR2_mats[["A4"]])
67 expect_false(identical(vcovCR(obj_A4, type = "CR2", inverse_var = FALSE), CR2_mats[["A4"]]))
68 target <- targetVariance(obj_A4)
69 expect_equal(vcovCR(obj_A4, type = "CR2", target = target, inverse_var = TRUE), CR2_mats[["A4"]])
70 attr(CR2_mats[["A4"]], "inverse_var") <- FALSE
71 expect_equal(vcovCR(obj_A4, type = "CR2", target = target, inverse_var = FALSE), CR2_mats[["A4"]])
72
73 })
74
75 test_that("vcovCR options work for CR4", {
76 skip_on_cran()
77 skip("Not worrying about CR4 for now.")
78 CR4_mats <- lapply(objects, vcovCR, type = "CR4")
79
80 expect_equal(vcovCR(obj_A1, cluster = egsingle$schoolid, type = "CR4"), CR4_mats[["A1"]])
81 expect_equal(vcovCR(obj_A1, type = "CR4", inverse_var = TRUE), CR4_mats[["A1"]])
82 expect_false(identical(vcovCR(obj_A1, type = "CR4", inverse_var = FALSE), CR4_mats[["A1"]]))
83 target <- targetVariance(obj_A1)
84 expect_equal(vcovCR(obj_A1, type = "CR4", target = target, inverse_var = TRUE), CR4_mats[["A1"]])
85 attr(CR4_mats[["A1"]], "inverse_var") <- FALSE
86 expect_equal(vcovCR(obj_A1, type = "CR4", target = target, inverse_var = FALSE), CR4_mats[["A1"]])
87
88 expect_equal(vcovCR(obj_A2, cluster = egsingle$schoolid, type = "CR4"), CR4_mats[["A2"]])
89 expect_equal(vcovCR(obj_A2, type = "CR4", inverse_var = TRUE), CR4_mats[["A2"]])
90 expect_false(identical(vcovCR(obj_A2, type = "CR4", inverse_var = FALSE), CR4_mats[["A2"]]))
91 target <- targetVariance(obj_A2)
92 expect_equal(vcovCR(obj_A2, type = "CR4", target = target, inverse_var = TRUE), CR4_mats[["A2"]])
93 attr(CR4_mats[["A2"]], "inverse_var") <- FALSE
94 expect_equal(vcovCR(obj_A2, type = "CR4", target = target, inverse_var = FALSE), CR4_mats[["A2"]])
95
96 expect_equal(vcovCR(obj_A3, cluster = egsingle$schoolid, type = "CR4"), CR4_mats[["A3"]])
97 expect_equal(vcovCR(obj_A3, type = "CR4", inverse_var = TRUE), CR4_mats[["A3"]])
98 expect_false(identical(vcovCR(obj_A3, type = "CR4", inverse_var = FALSE), CR4_mats[["A3"]]))
99 target <- targetVariance(obj_A3)
100 expect_equal(vcovCR(obj_A3, type = "CR4", target = target, inverse_var = TRUE), CR4_mats[["A3"]])
101 attr(CR4_mats[["A3"]], "inverse_var") <- FALSE
102 expect_equal(vcovCR(obj_A3, type = "CR4", target = target, inverse_var = FALSE), CR4_mats[["A3"]])
103
104 expect_equal(vcovCR(obj_A4, cluster = egsingle$schoolid, type = "CR4"), CR4_mats[["A4"]])
105 expect_equal(vcovCR(obj_A4, type = "CR4", inverse_var = TRUE), CR4_mats[["A4"]])
106 expect_false(identical(vcovCR(obj_A4, type = "CR4", inverse_var = FALSE), CR4_mats[["A4"]]))
107 target <- targetVariance(obj_A4)
108 expect_equal(vcovCR(obj_A4, type = "CR4", target = target, inverse_var = TRUE), CR4_mats[["A4"]])
109 attr(CR4_mats[["A4"]], "inverse_var") <- FALSE
110 expect_equal(vcovCR(obj_A4, type = "CR4", target = target, inverse_var = FALSE), CR4_mats[["A4"]])
111
112 })
113
114
115 test_that("CR2 is target-unbiased", {
116 skip_on_cran()
117 CR2_checks <- mapply(check_CR, obj = objects, vcov = CR2_mats)
118 expect_true(all(CR2_checks))
119 # CR4_checks <- mapply(check_CR, obj = objects, vcov = CR4_mats)
120 # expect_true(all(CR4_checks))
121 })
122
123
124 CR_types <- paste0("CR",0:3)
125
126 test_that("Order doesn't matter.", {
127 skip_on_cran()
128
129 check_sort_order(obj_A1, egsingle)
130 check_sort_order(obj_A2, egsingle)
131 check_sort_order(obj_A3, egsingle)
132 check_sort_order(obj_A4, egsingle)
133
134 })
135
136
137 test_that("clubSandwich works with dropped observations", {
138 skip_on_cran()
139 dat_miss <- egsingle
140 dat_miss$math[sample.int(nrow(egsingle), size = round(nrow(egsingle) / 10))] <- NA
141 obj_dropped <- update(obj_A4, data = dat_miss, na.action = na.omit)
142 obj_complete <- update(obj_A4, data = dat_miss, subset = !is.na(math))
143
144 # obj <- obj_dropped
145 # cluster <- nlme::getGroups(obj, level = 1)
146 # target <- NULL
147 # inverse_var <- TRUE
148 # type <- "CR2"
149 # form <- "sandwich"
150 #
151 # full_grps <- get_cor_grouping(obj)
152 # R_list <- nlme::corMatrix(obj$modelStruct$corStruct)
153 # levels <- names(R_list)
154 # grps <- get_cor_grouping(obj, levels = names(R_list))
155 #
156 # V_list <- build_var_cor_mats(obj)
157 # V_grps <- attr(V_list, "groups")
158 # ZDZ_list <- build_RE_mats(obj)
159 # ZDZ_grps <- attr(ZDZ_list, "groups")
160 #
161 # V_dim <- sapply(V_list, nrow)
162 # identical(names(V_dim), names(table(V_grps)))
163 # data.frame(dim = V_dim, grps = table(V_grps))
164 # table(V_dim == table(V_grps))
165 # dat_miss$x <- NA
166 # dat_miss$x[!is.na(dat_miss$math)] <- V_grps
167
168 CR_drop <- lapply(CR_types, function(x) vcovCR(obj_dropped, type = x))
169 CR_complete <- lapply(CR_types, function(x) vcovCR(obj_complete, type = x))
170 expect_equal(CR_drop, CR_complete)
171
172 test_drop <- lapply(CR_drop, function(x) coef_test(obj_dropped, vcov = x, test = "All", p_values = FALSE))
173 test_complete <- lapply(CR_complete, function(x) coef_test(obj_complete, vcov = x, test = "All", p_values = FALSE))
174 expect_equal(test_drop, test_complete)
175 })
176
177
178 test_that("Possible to cluster at higher level than random effects", {
179 skip_on_cran()
180
181 # fit two-level model
182 obj_2level <- lme(math ~ year * size + female + black + hispanic,
183 random = ~ year | childid,
184 data = egsingle)
185
186 # cluster at level 3
187 V <- vcovCR(obj_2level, type = "CR2", cluster = egsingle$schoolid)
188 expect_is(V, "vcovCR")
189
190 # create 4th level
191 n_districts <- nlevels(egsingle$schoolid) / 3
192 districtid <- rep(1:n_districts, each = 3)[egsingle$schoolid]
193
194 # cluster at level 4
195 expect_is(vcovCR(obj_2level, type = "CR2", cluster = districtid), "vcovCR")
196 expect_is(vcovCR(obj_A1, type = "CR2", cluster = districtid), "vcovCR")
197 expect_is(vcovCR(obj_A2, type = "CR2", cluster = districtid), "vcovCR")
198 expect_is(vcovCR(obj_A3, type = "CR2", cluster = districtid), "vcovCR")
199 expect_is(vcovCR(obj_A4, type = "CR2", cluster = districtid), "vcovCR")
200 })