predict() fails for split fixest models with non-regular FEs (“factor level [...] is duplicated”)

[francescodelprato]

Hey Laurent,

thanks for the great package. I am using fixest 0.13.2 under R 4.5.1.

Here's the problem: predict() on a split model (fsplit=~female) with multi-way FEs errors with “factor level [...] is duplicated” when the FEs are non-regular (after singleton removal). The main (all-gender, in my case) submodel predicts fine; the gender submodels fail because predict.fixest rebuilds factors from duplicated fixef_names.

Here's a minimal reproducible example:

library(fixest)

packageVersion("fixest")  # 0.13.2 for me

set.seed(0)
N <- 200; G <- 10
dt <- data.frame(
  id     = sample(1:50, N, TRUE),
  firm   = sample(1:G, N, TRUE),
  year   = sample(1:5, N, TRUE),
  female = sample(0:1, N, TRUE),
  x      = rnorm(N)
)
dt$y <- rnorm(N)

mod <- feols(y ~ x | id + firm + year, dt, fsplit = ~ female)
# NOTE: 3 fixed-effect singletons were removed; FEs flagged as not regular.

predict(mod[[1]], newdata = dt, fixef = FALSE)  # works
predict(mod[[2]], newdata = dt, fixef = FALSE)  # error: factor level [...] is duplicated

This raises

Error in `levels<-`(`*tmp*`, value = as.character(levels)) : 
  factor level [2] is duplicated

while predict(mod[[2]], ...) should return the linear predictor (or NA for new levels) like the full-sample model.

A minimal workaround is to compute manual Xb's to bypass predict’s factor rebuild:

compute_xb <- function(model, data) {
  mm <- model.matrix(model, data = data, type = "rhs")
  as.numeric(mm %*% coef(model))
}

xb_all   <- compute_xb(mod[[1]], dt)
xb_male  <- compute_xb(mod[[2]], dt[dt$female == 0, ])
xb_fem   <- compute_xb(mod[[3]], dt[dt$female == 1, ])

Thanks for your work!