I am running an event study design and get some unexpected behavior using OLS. One of my outcome variables is by definition zero before the treatment, i.e., whenever event time is weakly smaller than 0. I am convinced OLS shouldn't return estimates if there is no variance in the outcome (it's always zero for event_time <= 0). Below is a reproducible example. I get the expected behavior if I use fixed effects in a Poisson regression (which I intend to use for reasons unrelated to this issue), but not using fixed effects in feols.

Am I wrong, or is this a bug? Could it be some edge case I am unaware of? I am grateful for any pointers.

library(fixest)
library(dplyr)

set.seed(123)

n_units <- 500
n_periods <- 10 
n_rows <- n_units * n_periods

panel <- expand.grid(unit_id = as.character(1:n_units),
                     event_time = -4:5) %>%
  arrange(unit_id, event_time) %>%
  mutate(treated = ifelse(runif(n_rows) > 0.5, 1, 0),  
         outcome = ifelse(event_time <= 0, 0, rpois(n_rows, lambda = 1)),  
         time = sample(1:1000, n_rows, replace = TRUE),
         cluster_id = sample(1:100, n_rows, replace = TRUE))

# outcome is always 0 for event_time <= 0
panel %>% group_by(event_time <= 0) %>% distinct(outcome)

model_1 <- lm(outcome ~ i(event_time, treated, ref = 0),
              data = panel)

model_2 <- feols(outcome ~ i(event_time, treated, ref = 0) |
                   unit_id + event_time + time,
                 data = panel)

model_3 <- feglm(outcome ~ i(event_time, treated, ref = 0) |
                   unit_id + event_time + time,
                  family = "poisson",
                 data = panel)

summary(model_1)
etable(model_2, model_3)