`predict` methods for aggregated Sun & Abraham (2020) estimates

[fnaeher]

For most applications, fixest and marginaleffects packages interact well. However, it looks as if fixest does not support the estimation of marginal aggregated cohort effects and att as according to Sun & Abraham (2020) with marginaleffects. Please also see the short discussion at https://github.com/vincentarelbundock/marginaleffects/issues/315#issue-1239702882. It is suggested that fixest does not supply a predict method for aggregated cohort and att estimates that accepts a newdata argument which can be passed on to marginaleffects. If this is the case indeed, I would suggest to add this functionality to fixest. Thank you very much!

[vincentarelbundock]

I don't know this model, so I can't tell if it actually makes sense to add support for this. I am just adding a replicable example for convenience, not as a comment on the feature request:

library(fixest)
data(base_stagg)

sunab.agg = feols(y ~ x1 + sunab(year_treated, year) | id + year, base_stagg)

predict(sunab.agg, newdata = base_stagg)
#> Error in mm_info_new[[i]]: subscript out of bounds

[lrberge]

The philosophy of the software is: either it works, either it's an exception properly handled. So that's definitely a bug! 😄

Thanks for reporting!

[grantmcdermott]

Edit: Sorry, ignore my comment. It’s not a dup.

[m-pilarski]

The problem is that the "model matrix info" for newdata is not the same as it is in the original model.

library(fixest)

base_stagg <- fixest::base_stagg

sunab.agg <- feols(y ~ x1 + sunab(year_treated, year) | id + year, base_stagg)

mm_info <- sunab.agg$model_matrix_info

rhs_fml <- fixest:::fml_split(sunab.agg$fml, 1)

mm_new <- fixest:::fixest_model_matrix_extra(
  object=sunab.agg, newdata=base_stagg, original_data=FALSE, fml=rhs_fml,
  i_noref=TRUE
)

mm_info_new = attr(mm_new, "model_matrix_info")

In predict.fixest there is a comment stating:

https://github.com/lrberge/fixest/blob/a4d1a9bea20aa7ab7ab0e0f1d2047d8097971ad7/R/Methods.R#L2593-L2594

As you can see below, this is not the case:

str(mm_info)
#> List of 2
#>  $      :List of 8
#>   ..$ coef_names_full: chr [1:18] "year::-9" "year::-8" "year::-7" "year::-6" ...
#>   ..$ items          : num [1:18] -9 -8 -7 -6 -5 -4 -3 -2 -1 0 ...
#>   ..$ ref_id         : int 9
#>   ..$ ref            : num -1
#>   ..$ f_name         : chr "year"
#>   ..$ is_num         : logi TRUE
#>   ..$ is_inter_fact  : logi FALSE
#>   ..$ is_inter_num   : logi FALSE
#>  $ sunab:List of 4
#>   ..$ agg       : chr "(\\Qyear\\E)::(-?[[:digit:]]+):cohort"
#>   .. ..- attr(*, "model_matrix_info")=List of 8
#>   .. .. ..$ coef_names_full: chr [1:18] "year::-9" "year::-8" "year::-7" "year::-6" ...
#>   .. .. ..$ items          : num [1:18] -9 -8 -7 -6 -5 -4 -3 -2 -1 0 ...
#>   .. .. ..$ ref_id         : int 9
#>   .. .. ..$ ref            : num -1
#>   .. .. ..$ f_name         : chr "year"
#>   .. .. ..$ is_num         : logi TRUE
#>   .. .. ..$ is_inter_fact  : logi FALSE
#>   .. .. ..$ is_inter_num   : logi FALSE
#>   ..$ agg_att   : Named chr "\\Qyear\\E::[[:digit:]]+:cohort"
#>   .. ..- attr(*, "names")= chr "ATT"
#>   ..$ agg_period: chr "(\\Qyear\\E)::(-?[[:digit:]]+):cohort"
#>   ..$ ref.p     : num -1

str(mm_info_new)
#> List of 1
#>  $ sunab:List of 4
#>   ..$ agg       : chr "(\\Qyear\\E)::(-?[[:digit:]]+):cohort"
#>   .. ..- attr(*, "model_matrix_info")=List of 8
#>   .. .. ..$ coef_names_full: chr [1:18] "year::-9" "year::-8" "year::-7" "year::-6" ...
#>   .. .. ..$ items          : num [1:18] -9 -8 -7 -6 -5 -4 -3 -2 -1 0 ...
#>   .. .. ..$ ref_id         : int 9
#>   .. .. ..$ ref            : num -1
#>   .. .. ..$ f_name         : chr "year"
#>   .. .. ..$ is_num         : logi TRUE
#>   .. .. ..$ is_inter_fact  : logi FALSE
#>   .. .. ..$ is_inter_num   : logi FALSE
#>   ..$ agg_att   : Named chr "\\Qyear\\E::[[:digit:]]+:cohort"
#>   .. ..- attr(*, "names")= chr "ATT"
#>   ..$ agg_period: chr "(\\Qyear\\E)::(-?[[:digit:]]+):cohort"
#>   ..$ ref.p     : num -1