inst/tinytest/test-felm-extreme-values.R

In capybara: Fast and Memory Efficient Fitting of Linear Models with High-Dimensional Fixed Effects

# These tests are a formality but an important one to check NA/Inf/NaN handling. This is justified provided that I am not using base R's model.matrix

# srr_stats (tests)
# {RE5.5} Ensures accuracy of prediction methods with unseen data subsets, maintaining expected patterns of error.

local({
  # NAs on the lhs ----

  ross2004_subset <- ross2004[ross2004$year == 1999, ]
  ross2004_subset <- ross2004_subset[ross2004_subset$ltrade > quantile(ross2004_subset$ltrade, 0.75), ]

  ross2004_subset1 <- ross2004_subset
  ross2004_subset1$ltrade[c(1, 3, 5)] <- NA

  m1 <- felm(ltrade ~ ldist | ctry1, data = ross2004_subset1)
  m2 <- lm(ltrade ~ 0 + ldist + as.factor(ctry1), data = ross2004_subset1)

  expect_equal(coef(m1), coef(m2)[1])

  # NAs on the rhs ----

  ross2004_subset2 <- ross2004_subset
  ross2004_subset2$ctry1[c(1, 3, 5)] <- NA

  m1 <- felm(ltrade ~ ldist | ctry1, data = ross2004_subset2)
  m2 <- lm(ltrade ~ 0 + ldist + as.factor(ctry1), data = ross2004_subset2)

  expect_equal(coef(m1), coef(m2)[1])

  # 0+log on the rhs ----

  ross2004_subset3 <- ross2004_subset
  ross2004_subset3$ltrade[c(1, 3, 5)] <- 0

  # we need a subset to avoid breaking lm!
  # otherwise
  # Error in lm.fit(x, y, offset = offset, singular.ok = singular.ok, ...) :
  # NA/NaN/Inf in 'y'
  ross2004_subset4 <- ross2004_subset3[-c(1, 3, 5), ]

  m1 <- felm(ltrade ~ ldist | ctry1, data = ross2004_subset3)
  m2 <- lm(ltrade ~ 0 + ldist + as.factor(ctry1), data = ross2004_subset4)

  expect_equal(coef(m1), coef(m2)[1], tolerance = 1e-2)
})