R Session Aborted

[da-zar]

Hi,

I've noticed that for some data sets, small once where the data is not sufficient to estimate the parameters, an error is thrown but for others the R session simply crashes. Below an example.

library(fixest)
packageVersion('fixest') # 0.10.3’

# gives an error that can be handled
df <- data.frame(id = 1, time = 2, y = 1, x = 1)
model <- fepois(
  y ~ x | id + time,
  df
)

# crashes the session
df <- data.frame(id = c(1, 1), time = c(1, 2), y = c(1, 0), x = c(1, 0))
model <- fepois(
  y ~ x | id + time,
  df
)

[etiennebacher]

I narrowed it down to this but then it happens in C++ (which I don't know). I'm putting this here if someone wants to explore:

y = 0.004401089
X = matrix(1, dimnames = list(NULL, "x"))
weights = 1.1
iterMax = 10000
diffMax = 1e-05
r_nb_id_Q = c(1L, 1L)
fe_id_list = list(1L, 1L)
table_id_I = c(1L, 1L)
slope_flag_Q = c(0L, 0L)
slope_vars_list = list(0)
r_init = 1
nthreads = 4

fixest:::cpp_demean(y, X, weights, iterMax, diffMax, r_nb_id_Q, fe_id_list, table_id_I,
           slope_flag_Q, slope_vars_list, r_init, nthreads)

[lrberge]

Hi, and sorry for the longuiish delay!
Thanks a lot @da-zar for reporting and providing a simple reproducible example, and to @etiennebacher for making my life so easy!

Now fixed.

[etiennebacher]

I confirm, thanks! One small mistake in the error message though, in the example below I suppose it should be "Some variables" instead of "All variables"?

library(fixest)
df <- data.frame(id = c(1, 1), time = c(1, 2), y = c(1, 0), x = c(1, 0), x2 = c(2, 1), x3 = c(0, 0))
model <- fepois(
  y ~ x + x2 + x3 | id + time,
  df
)
#> NOTE: 0/1 fixed-effect (1 observation) removed because of only 0 outcomes.
#> Error in fepois(y ~ x + x2 + x3 | id + time, df): All variables, 'x' and 'x2', are collinear with the fixed effects. In such
#> circumstances, the estimation is void.

[lrberge]

No, it's all , the three should be there...
:-/ brb

[lrberge]

Ah, it's because x3 is an edge case and was removed before because equal to a constant 0.
Works OK with non constants:

df <- data.frame(id = c(1, 1), time = c(1, 2), y = c(1, 0), x = c(1, 0), x2 = c(2, 1), x3 = c(6, 0))
model <- fepois(y ~ x + x2 + x3 | id + time, df)
#> NOTE: 0/1 fixed-effect (1 observation) removed because of only 0 outcomes.
#> Error in fepois(y ~ x + x2 + x3 | id + time, df) : 
#>  All variables, 'x', 'x2' and 'x3', are collinear with the fixed effects. In such circumstances, the estimation is void.

I remove 0-only variables beforehand. I do not give warnings since it can happen for factors when manipulating the data (hence there would be too much false positives).

Thanks for checking! :-)