context("ignoring absorbed fixed effects") set.seed(20190513) skip_if_not_installed("plm") library(plm) data(MortalityRates) MV_Mortality <- subset(MortalityRates, cause=="Motor Vehicle" & state %in% 1:8) table(MV_Mortality$state) MV_Mortality$state_fac <- factor(MV_Mortality$state) # MV_Mortality$pop <- with(MV_Mortality, 1 + rbinom(nlevels(state_fac), size = 4, prob = 0.5)[state_fac]) summary(MV_Mortality$pop) MV_Mortality$pop_scale <- with(MV_Mortality, pop / mean(pop)) summary(MV_Mortality$pop_scale) # model specification specification <- mrate ~ 0 + legal + beertaxa + beerpercap + winepercap + factor(state) #----------------------- # unweighted #----------------------- ols_LSDV <- lm(specification, data = MV_Mortality) ols_within <- plm(update(specification, . ~ . - 0 - factor(state)), data = MV_Mortality, effect = "individual", index = c("state","year")) test_that("Unweighted lsdv and within estimators are equivalent", { lsdv <- coef_test(ols_LSDV, vcov = "CR2", cluster = MV_Mortality$state, coefs = 1:4, p_values = FALSE) wthn <- coef_test(ols_within, vcov = "CR2", p_values = FALSE) expect_equal(lsdv, wthn) }) #----------------------- # iv-weights #----------------------- wls_LSDV <- lm(specification, weights = pop_scale, data = MV_Mortality) MV_Mortality_full <- model.frame(lm(specification, weights = pop_scale, data = MV_Mortality)) U_mat <- model.matrix(update(specification, . ~ . - factor(state)), data = MV_Mortality_full) T_mat <- model.matrix(~ factor(state), data = MV_Mortality_full) w <- MV_Mortality_full$"(weights)" state <- MV_Mortality_full$"factor(state)" U_absorb <- residuals(stats:::lm.wfit(x = T_mat, y = U_mat, w = w))[,-31] Y_absorb <- residuals(stats:::lm.wfit(x = T_mat, y = MV_Mortality_full$mrate, w = w)) wls_within <- lm(Y_absorb ~ 0 + U_absorb, weights = w) test_that("Inverse-variance weighted lsdv and within estimators are equivalent.", { lsdv <- coef_test(wls_LSDV, vcov = "CR2", cluster = MV_Mortality$state, inverse_var = TRUE, p_values = FALSE)[1:4,] wthn <- coef_test(wls_within, vcov = "CR2", cluster = state, inverse_var = TRUE, p_values = FALSE)[1:4,] expect_equal(lsdv$beta, wthn$beta, check.attributes = FALSE, tolerance = 10^-5) expect_equal(lsdv$SE, wthn$SE, check.attributes = FALSE, tolerance = 10^-2) expect_equal(lsdv$tstat, wthn$tstat, check.attributes = FALSE, tolerance = 10^-2) expect_equal(lsdv$df_Satt, wthn$df_Satt, check.attributes = FALSE, tolerance = 10^-2) }) #----------------------- # p-weights #----------------------- test_that("Probability-weighted lsdv and within estimators are not necessarily equivalent.", { lsdv <- coef_test(wls_LSDV, vcov = "CR2", cluster = MV_Mortality$state, inverse_var = FALSE, coefs = 1:4, p_values = FALSE) wthn <- coef_test(wls_within, vcov = "CR2", cluster = state, inverse_var = FALSE, p_values = FALSE) expect_equal(lsdv$beta, wthn$beta, check.attributes = FALSE, tolerance = 10^-5) expect_equal(lsdv$SE, wthn$SE, check.attributes = FALSE, tolerance = 10^-2) expect_equal(lsdv$tstat, wthn$tstat, check.attributes = FALSE, tolerance = 10^-2) expect_equal(lsdv$df_Satt, wthn$df_Satt, check.attributes = FALSE, tolerance = 10^-2) })