| 1 |
############################# |
| 2 |
|
| 3 |
context("ivreg::ivreg objects") |
| 4 |
set.seed(20190513) |
| 5 |
|
| 6 |
skip_if_not_installed("ivreg") |
| 7 |
|
| 8 |
library(ivreg, quietly=TRUE) |
| 9 |
|
| 10 |
data("CigarettesSW", package = "AER") |
| 11 |
|
| 12 |
Cigs <- within(CigarettesSW, { |
| 13 |
rprice <- price/cpi |
| 14 |
rincome <- income/population/cpi |
| 15 |
tdiff <- (taxs - tax)/cpi |
| 16 |
}) |
| 17 |
|
| 18 |
CR_types <- paste0("CR",0:4) |
| 19 |
|
| 20 |
obj_un <- ivreg::ivreg(log(packs) ~ log(rprice) + log(rincome) | log(rincome) + tdiff + I(tax/cpi), |
| 21 |
data = Cigs) |
| 22 |
obj_wt <- ivreg::ivreg(log(packs) ~ log(rprice) + log(rincome) | log(rincome) + tdiff + I(tax/cpi), |
| 23 |
data = Cigs, |
| 24 |
weights = population) |
| 25 |
|
| 26 |
X <- model.matrix(obj_wt, component = "regressors") |
| 27 |
Z <- model.matrix(obj_wt, component = "instruments") |
| 28 |
y <- log(CigarettesSW$packs) |
| 29 |
w <- weights(obj_wt) |
| 30 |
|
| 31 |
test_that("Basic calculations from ivreg agree for unweighted model.", { |
| 32 |
XZ <- model.matrix(obj_un, component = "projected") |
| 33 |
ZtZ_inv <- chol2inv(chol(t(Z) %*% Z)) |
| 34 |
XZ_check <- Z %*% ZtZ_inv %*% t(Z) %*% X |
| 35 |
|
| 36 |
expect_equal(XZ, XZ_check, check.attributes=FALSE) |
| 37 |
expect_equal(coef(obj_un), lm.fit(XZ, y)$coefficients) |
| 38 |
expect_equal (bread(obj_un), chol2inv(chol(t(XZ) %*% XZ)) * nobs(obj_un), check.attributes=FALSE) |
| 39 |
|
| 40 |
hii <- diag(XZ %*% chol2inv(chol(t(XZ) %*% XZ)) %*% t(XZ)) |
| 41 |
expect_equal(hatvalues(obj_un, type = "stage2"), hii) |
| 42 |
|
| 43 |
r <- as.vector(y - X %*% coef(obj_un)) |
| 44 |
expect_equal(r, as.vector(residuals_CS(obj_un))) |
| 45 |
}) |
| 46 |
|
| 47 |
test_that("Basic calculations from ivreg agree for weighted model.", { |
| 48 |
XZ <- model.matrix(obj_wt, component = "projected") |
| 49 |
ZwZ_inv <- chol2inv(chol(t(Z) %*% (w * Z))) |
| 50 |
XZ_check <- Z %*% ZwZ_inv %*% t(Z) %*% (w * X) |
| 51 |
|
| 52 |
expect_equal(XZ, XZ_check, check.attributes=FALSE) |
| 53 |
expect_equal(coef(obj_wt), lm.wfit(XZ, y, w)$coefficients) |
| 54 |
expect_equal(bread(obj_wt), chol2inv(chol(t(XZ) %*% (w * XZ))) * nobs(obj_wt), check.attributes=FALSE) |
| 55 |
|
| 56 |
hii <- diag(X%*% chol2inv(chol(t(XZ) %*% (w * XZ))) %*% t(w * XZ)) |
| 57 |
expect_false(all(hatvalues(obj_wt) == hii)) # does not agree because hatvalues doesn't work with weighting |
| 58 |
|
| 59 |
r <- as.vector(y - X %*% coef(obj_wt)) |
| 60 |
expect_equal(r, as.vector(residuals_CS(obj_wt))) |
| 61 |
}) |
| 62 |
|
| 63 |
test_that("bread works", { |
| 64 |
|
| 65 |
expect_true(check_bread(obj_un, cluster = Cigs$state, y = log(Cigs$packs))) |
| 66 |
tsls_vcov <- bread(obj_un) * summary(obj_un)$sigma^2 / v_scale(obj_un) |
| 67 |
expect_equal(vcov(obj_un), tsls_vcov) |
| 68 |
|
| 69 |
expect_true(check_bread(obj_wt, cluster = Cigs$state, y = log(Cigs$packs))) |
| 70 |
wtsls_vcov <- bread(obj_wt) * summary(obj_wt)$sigma^2 / v_scale(obj_wt) |
| 71 |
expect_equal(vcov(obj_wt), wtsls_vcov) |
| 72 |
}) |
| 73 |
|
| 74 |
|
| 75 |
test_that("vcovCR options don't matter for CR0", { |
| 76 |
expect_error(vcovCR(obj_un, type = "CR0")) |
| 77 |
CR0 <- vcovCR(obj_un, cluster = Cigs$state, type = "CR0") |
| 78 |
expect_output(print(CR0)) |
| 79 |
attr(CR0, "target") <- NULL |
| 80 |
attr(CR0, "inverse_var") <- NULL |
| 81 |
CR0_A <- vcovCR(obj_un, cluster = Cigs$state, type = "CR0", target = 1 / Cigs$population) |
| 82 |
attr(CR0_A, "target") <- NULL |
| 83 |
attr(CR0_A, "inverse_var") <- NULL |
| 84 |
expect_identical(CR0_A, CR0) |
| 85 |
CR0_B <- vcovCR(obj_un, cluster = Cigs$state, type = "CR0", target = 1 / Cigs$population, inverse_var = FALSE) |
| 86 |
attr(CR0_B, "target") <- NULL |
| 87 |
attr(CR0_B, "inverse_var") <- NULL |
| 88 |
expect_identical(CR0_A, CR0) |
| 89 |
|
| 90 |
expect_error(vcovCR(obj_un, cluster = Cigs$state, type = "CR0", target = 1 / Cigs$population, inverse_var = TRUE)) |
| 91 |
|
| 92 |
wCR0 <- vcovCR(obj_wt, cluster = Cigs$state, type = "CR0") |
| 93 |
attr(wCR0, "target") <- NULL |
| 94 |
attr(wCR0, "inverse_var") <- NULL |
| 95 |
wCR0_A <- vcovCR(obj_wt, cluster = Cigs$state, type = "CR0", target = 1 / Cigs$population) |
| 96 |
attr(wCR0_A, "target") <- NULL |
| 97 |
attr(wCR0_A, "inverse_var") <- NULL |
| 98 |
expect_identical(wCR0_A, wCR0) |
| 99 |
wCR0_B <- vcovCR(obj_wt, cluster = Cigs$state, type = "CR0", target = 1 / Cigs$population, inverse_var = FALSE) |
| 100 |
attr(wCR0_B, "target") <- NULL |
| 101 |
attr(wCR0_B, "inverse_var") <- NULL |
| 102 |
expect_identical(wCR0_B, wCR0) |
| 103 |
|
| 104 |
expect_error(vcovCR(obj_wt, cluster = Cigs$state, type = "CR0", target = 1 / Cigs$population, inverse_var = TRUE)) |
| 105 |
}) |
| 106 |
|
| 107 |
test_that("vcovCR options work for CR2", { |
| 108 |
|
| 109 |
CR2_iv <- vcovCR(obj_un, cluster = Cigs$state, type = "CR2") |
| 110 |
expect_equal(vcovCR(obj_un, cluster = Cigs$state, type = "CR2", target = rep(1, nobs(obj_un))), CR2_iv) |
| 111 |
expect_false(identical(vcovCR(obj_un, cluster = Cigs$state, type = "CR2", target = 1 / Cigs$population), CR2_iv)) |
| 112 |
|
| 113 |
wCR2_id <- vcovCR(obj_wt, cluster = Cigs$state, type = "CR2") |
| 114 |
expect_identical(vcovCR(obj_wt, cluster = Cigs$state, type = "CR2", inverse_var = FALSE), wCR2_id) |
| 115 |
expect_identical(vcovCR(obj_wt, cluster = Cigs$state, type = "CR2", target = rep(1, nobs(obj_un))), wCR2_id) |
| 116 |
expect_identical(vcovCR(obj_wt, cluster = Cigs$state, type = "CR2", target = rep(1, nobs(obj_un)), inverse_var = FALSE), wCR2_id) |
| 117 |
|
| 118 |
}) |
| 119 |
|
| 120 |
test_that("vcovCR options work for CR4", { |
| 121 |
|
| 122 |
CR4_not <- vcovCR(obj_un, cluster = Cigs$state, type = "CR4") |
| 123 |
expect_identical(vcovCR(obj_un, cluster = Cigs$state, type = "CR4", target = rep(1, nobs(obj_un))), CR4_not) |
| 124 |
expect_identical(vcovCR(obj_un, cluster = Cigs$state, type = "CR4", target = rep(1, nobs(obj_un)), inverse_var = FALSE), CR4_not) |
| 125 |
expect_false(identical(vcovCR(obj_un, cluster = Cigs$state, type = "CR4", target = 1 / Cigs$population), CR4_not)) |
| 126 |
|
| 127 |
wCR4_id <- vcovCR(obj_wt, cluster = Cigs$state, type = "CR4") |
| 128 |
expect_identical(vcovCR(obj_wt, cluster = Cigs$state, type = "CR4", inverse_var = FALSE), wCR4_id) |
| 129 |
expect_identical(vcovCR(obj_wt, cluster = Cigs$state, type = "CR4", target = rep(1, nobs(obj_wt))), wCR4_id) |
| 130 |
expect_identical(vcovCR(obj_wt, cluster = Cigs$state, type = "CR4", target = rep(1, nobs(obj_wt)), inverse_var = FALSE), wCR4_id) |
| 131 |
|
| 132 |
}) |
| 133 |
|
| 134 |
|
| 135 |
test_that("CR2 is target-unbiased", { |
| 136 |
expect_true(check_CR(obj_un, vcov = "CR2", cluster = Cigs$state)) |
| 137 |
expect_true(check_CR(obj_wt, vcov = "CR2", cluster = Cigs$state)) |
| 138 |
}) |
| 139 |
|
| 140 |
test_that("CR4 is target-unbiased", { |
| 141 |
skip("Need to understand target-unbiasedness for ivreg objects.") |
| 142 |
expect_true(check_CR(obj_un, vcov = "CR4", cluster = Cigs$state)) |
| 143 |
expect_true(check_CR(obj_wt, vcov = "CR4", cluster = Cigs$state)) |
| 144 |
}) |
| 145 |
|
| 146 |
test_that("vcovCR is equivalent to vcovHC (with HC0 or HC1) when clusters are all of size 1", { |
| 147 |
library(sandwich, quietly=TRUE) |
| 148 |
CR0 <- vcovCR(obj_un, cluster = 1:nobs(obj_un), type = "CR0") |
| 149 |
expect_equal(vcovHC(obj_un, type = "HC0"), as.matrix(CR0)) |
| 150 |
CR1 <- vcovCR(obj_un, cluster = 1:nobs(obj_un), type = "CR1S") |
| 151 |
expect_equal(vcovHC(obj_un, type = "HC1"), as.matrix(CR1)) |
| 152 |
CR2 <- vcovCR(obj_un, cluster = 1:nobs(obj_un), type = "CR2") |
| 153 |
expect_false(all(vcovHC(obj_un, type = "HC2") == as.matrix(CR2))) |
| 154 |
CR3 <- vcovCR(obj_un, cluster = 1:nobs(obj_un), type = "CR3") |
| 155 |
expect_false(all(vcovHC(obj_un, type = "HC3") == as.matrix(CR3))) |
| 156 |
}) |
| 157 |
|
| 158 |
test_that("Order doesn't matter.",{ |
| 159 |
|
| 160 |
check_sort_order(obj_wt, Cigs, "state") |
| 161 |
|
| 162 |
}) |
| 163 |
# Clustering variable must have length equal to the number of rows in the data used to fit obj. |
| 164 |
|
| 165 |
test_that("clubSandwich works with dropped observations", { |
| 166 |
dat_miss <- Cigs |
| 167 |
dat_miss$rincome[sample.int(nrow(Cigs), size = round(nrow(Cigs) / 10))] <- NA |
| 168 |
iv_dropped <- update(obj_un, data = dat_miss) |
| 169 |
dat_complete <- subset(dat_miss, !is.na(rincome)) |
| 170 |
iv_complete <- update(obj_un, data = dat_complete) |
| 171 |
|
| 172 |
CR_drop <- lapply(CR_types, function(x) vcovCR(iv_dropped, cluster = dat_miss$state, type = x)) |
| 173 |
CR_complete <- lapply(CR_types, function(x) vcovCR(iv_complete, cluster = dat_complete$state, type = x)) |
| 174 |
expect_equal(CR_drop, CR_complete) |
| 175 |
|
| 176 |
test_drop <- lapply(CR_types, function(x) coef_test(iv_dropped, vcov = x, cluster = dat_miss$state, test = "All", p_values = FALSE)) |
| 177 |
test_complete <- lapply(CR_types, function(x) coef_test(iv_complete, vcov = x, cluster = dat_complete$state, test = "All", p_values = FALSE)) |
| 178 |
expect_equal(test_drop, test_complete) |
| 179 |
}) |
| 180 |
|
| 181 |
|
| 182 |
|
| 183 |
test_that("weight scale doesn't matter", { |
| 184 |
|
| 185 |
iv_fit_w <- update(obj_un, weights = rep(4, nobs(obj_un))) |
| 186 |
|
| 187 |
unweighted_fit <- lapply(CR_types, function(x) vcovCR(obj_un, cluster = Cigs$state, type = x)) |
| 188 |
weighted_fit <- lapply(CR_types, function(x) vcovCR(iv_fit_w, cluster = Cigs$state, type = x)) |
| 189 |
|
| 190 |
expect_equal(lapply(unweighted_fit, as.matrix), |
| 191 |
lapply(weighted_fit, as.matrix), |
| 192 |
tol = 5 * 10^-7) |
| 193 |
|
| 194 |
target <- 1 + rpois(nrow(Cigs), lambda = 8) |
| 195 |
unweighted_fit <- lapply(CR_types, function(x) vcovCR(obj_un, cluster = Cigs$state, type = x, target = target)) |
| 196 |
weighted_fit <- lapply(CR_types, function(x) vcovCR(iv_fit_w, cluster = Cigs$state, type = x, target = target * 15)) |
| 197 |
|
| 198 |
expect_equal(lapply(unweighted_fit, as.matrix), |
| 199 |
lapply(weighted_fit, as.matrix), |
| 200 |
tol = 5 * 10^-7) |
| 201 |
|
| 202 |
}) |
| 203 |
|
| 204 |
test_that("clubSandwich works with weights of zero.", { |
| 205 |
|
| 206 |
n_Cigs <- nrow(Cigs) |
| 207 |
Cigs$awt <- rpois(n_Cigs, lambda = 1.4) |
| 208 |
table(Cigs$awt) |
| 209 |
|
| 210 |
iv_full <- update(obj_un, weights = awt) |
| 211 |
Cigs_sub <- subset(Cigs, awt > 0) |
| 212 |
iv_sub <- update(iv_full, data = Cigs_sub) |
| 213 |
|
| 214 |
CR_full <- lapply(CR_types, function(x) vcovCR(iv_full, cluster = Cigs$state, type = x)) |
| 215 |
CR_sub <- lapply(CR_types, function(x) vcovCR(iv_sub, cluster = Cigs_sub$state, type = x)) |
| 216 |
expect_equal(CR_full, CR_sub, check.attributes = FALSE) |
| 217 |
|
| 218 |
test_full <- lapply(CR_types, function(x) coef_test(iv_full, vcov = x, cluster = Cigs$state, test = c("z","naive-t","Satterthwaite"), p_values = TRUE)) |
| 219 |
test_sub <- lapply(CR_types, function(x) coef_test(iv_sub, vcov = x, cluster = Cigs_sub$state, test = c("z","naive-t","Satterthwaite"), p_values = TRUE)) |
| 220 |
expect_equal(test_full, test_sub, check.attributes = FALSE) |
| 221 |
|
| 222 |
dat_miss <- Cigs |
| 223 |
miss_indicator <- sample.int(n_Cigs, size = round(n_Cigs / 10)) |
| 224 |
dat_miss$rincome[miss_indicator] <- NA |
| 225 |
with(dat_miss, table(awt, is.na(rincome))) |
| 226 |
|
| 227 |
iv_dropped <- update(iv_full, data = dat_miss) |
| 228 |
dat_complete <- subset(dat_miss, !is.na(rincome)) |
| 229 |
iv_complete <- update(iv_full, data = dat_complete) |
| 230 |
|
| 231 |
CR_drop <- lapply(CR_types, function(x) vcovCR(iv_dropped, cluster = dat_miss$state, type = x)) |
| 232 |
CR_complete <- lapply(CR_types, function(x) vcovCR(iv_complete, cluster = dat_complete$state, type = x)) |
| 233 |
expect_equal(CR_drop, CR_complete) |
| 234 |
|
| 235 |
test_drop <- lapply(CR_types, function(x) coef_test(iv_dropped, vcov = x, cluster = dat_miss$state, test = "All", p_values = FALSE)) |
| 236 |
test_complete <- lapply(CR_types, function(x) coef_test(iv_complete, vcov = x, cluster = dat_complete$state, test = "All", p_values = FALSE)) |
| 237 |
expect_equal(test_drop, test_complete) |
| 238 |
|
| 239 |
}) |
| 240 |
|
| 241 |
|
| 242 |
#------------------------------------------------------------------------------- |
| 243 |
# Other estimation methods |
| 244 |
|
| 245 |
ols_un <- ivreg::ivreg(log(packs) ~ log(rprice) + log(rincome) | log(rincome) + tdiff + I(tax/cpi), |
| 246 |
data = Cigs, method = "OLS") |
| 247 |
ols_wt <- ivreg::ivreg(log(packs) ~ log(rprice) + log(rincome) | log(rincome) + tdiff + I(tax/cpi), |
| 248 |
data = Cigs, |
| 249 |
weights = population, method = "OLS") |
| 250 |
|
| 251 |
mom_un <- ivreg::ivreg(log(packs) ~ log(rprice) + log(rincome) | log(rincome) + tdiff + I(tax/cpi), |
| 252 |
data = Cigs, method = "M") |
| 253 |
mom_wt <- ivreg::ivreg(log(packs) ~ log(rprice) + log(rincome) | log(rincome) + tdiff + I(tax/cpi), |
| 254 |
data = Cigs, |
| 255 |
weights = population, method = "M") |
| 256 |
|
| 257 |
rob_un <- ivreg::ivreg(log(packs) ~ log(rprice) + log(rincome) | log(rincome) + tdiff + I(tax/cpi), |
| 258 |
data = Cigs, method = "MM") |
| 259 |
rob_wt <- ivreg::ivreg(log(packs) ~ log(rprice) + log(rincome) | log(rincome) + tdiff + I(tax/cpi), |
| 260 |
data = Cigs, |
| 261 |
weights = population, method = "MM") |
| 262 |
|
| 263 |
test_that("vcovCR does not currently support ivreg models estimated using method = 'M' or method = 'MM'", { |
| 264 |
expect_error(vcovCR(mom_un, cluster = Cigs$state, type = "CR2")) |
| 265 |
expect_error(vcovCR(mom_wt, cluster = Cigs$state, type = "CR2")) |
| 266 |
expect_error(vcovCR(rob_un, cluster = Cigs$state, type = "CR2")) |
| 267 |
expect_error(vcovCR(rob_wt, cluster = Cigs$state, type = "CR2")) |
| 268 |
}) |
| 269 |
|