Stars in fixest

[vincentarelbundock]

Hello, I am running into similar issues when using modelsummary with the fixest package. The p-values in my fixest results do not match those printed by modelsummary.

• p-value for language variable from fixest estimate: 0.013087
• p-value for language variable printed in modelsummary: 0.03491

Code is below. I made sure to install the latest development versions of modelsummary and fixest, but these didn't fix the issue. Apologies if I'm missing something. Any help would be appreciated.

library(modelsummary)
library(fixest)
library(carData)
d <- Greene
d$dv <- ifelse(Greene$decision == 'yes', 1, 0)

mod <- feglm(dv ~ language | judge,
             data = d,
             cluster = c('judge'), family = 'logit')

mod
modelsummary(mod, fmt = '%.5f', statistic = "{std.error} ({p.value})")

Originally posted by @knrob in #264 (comment)

[vincentarelbundock]

Very sorry the delay. Work has been crazy, then vacation, etc.

The problem is that modelsummary uses parameters to extract p values. parameters uses the t distribution and believes that this particular model has only 9 degrees of freedom (it has more). So we get overly conservative results.

library(fixest)
library(carData)
d <- Greene
d$dv <- ifelse(Greene$decision == 'yes', 1, 0)

mod <- feglm(dv ~ language | judge,
             data = d,
             cluster = c('judge'), family = 'logit')

insight::get_df(mod)
# [1] 382

pnorm(-2.48139) * 2
# [1] 0.01308711
pt(-2.48139, df = 9) * 2
# [1] 0.03491173

This should be fixed upstream, so I opened an issue here: easystats/parameters#892

[vincentarelbundock]

tag @kbroman