Error in terms.formula(tmp, simplify = TRUE) : invalid model formula

[ccmullally]

I'm receiving the above error when trying to run boottest using a feols model. I don't have same issue with lm. I'm probably doing something wrong, but just in case this is a bug, I wanted to let you know.

library(dplyr)
library(fwildclusterboot)
library(fixest)
# 
data(mtcars)
feols_fit <- feols(disp ~  am + hp, weights = ~qsec, cluster = ~cyl, data = mtcars)

boot_feols <- boottest(feols_fit, clustid = "cyl", param = "am", B = 499)

lm_fit <- lm(disp ~  am + hp, mtcars, weights = qsec)

boot_lm <- boottest(lm_fit, clustid = "cyl", param = "am", B = 499)

[s3alfisc]

Hi, thanks - boottest() currently only works if the weights argument is passed to lm() and feols() as a vector. I was not aware of this (fixest has a lot of different options to specify models, which makes my life rather complicated), so thanks for letting me know about it! I'll try to fix this over the next days.

Until then, a workaround for you is to run

library(dplyr)
library(fwildclusterboot)
library(fixest)

# 
data(mtcars)

feols_fit <- feols(disp ~ am + hp, weights = mtcars$qsec, cluster = "cyl", data = mtcars)
boot_feols <- boottest(feols_fit, clustid = "cyl", param = "am", B = 499, seed = 986)

lm_fit <- lm(disp ~  am + hp, mtcars, weights = qsec)
boot_lm <- boottest(lm_fit, clustid = "cyl", param = "am", B = 499, seed = 986)

generics::tidy(boot_feols)
generics::tidy(boot_lm)
# > generics::tidy(boot_feols)
# term  estimate statistic p.value  conf.low conf.high
# 1 1*am = 0 -100.2779  -2.69509       0 -166.8028 -48.54865
# > generics::tidy(boot_lm)
# term  estimate statistic p.value  conf.low conf.high
# 1 1*am = 0 -100.2779  -2.69509       0 -166.8028 -48.54865

[s3alfisc]

Btw as you are new to R and use WLS - one difference between R and Stata is that R does not differentiate between frequency and sampling/probability weights. In R's lm() function as well as in feols(), it is my understanding that weights are always treated as sampling/probability weights. This does not matter for point estimation, but for inference via small sample corrections - with frequency weights, the number of rows of the data M no longer corresponds to the number of observations N, and in consequence, R implements an incorrect small sample adjustment m = (M-k) / (M-1). This does not matter for wild bootstrap inference, as the p-value is calculated as mean(abs(m x t_stat) <= abs(m x t_boot)) - both t_stat and t_boot are multiplied with the same factor.

[ccmullally]

Very interesting, thanks. I will add this to my problem set so the students know.

[s3alfisc]

With commit 874facb, boottest() should no longer fail when the weights argument in feols() is specified as a formula, or when the cluster argument in feols is specified as a column vector (with the caveat that the column vector needs to be specified in relation to the input data, e.g. feols(fml, data, cluster = data$cluster, weights = data$weights), I will add this option in the next days).

[ccmullally]

Fantastic, thanks. Can confirm it worked for me.

…

On Wed, Dec 29, 2021 at 12:33 PM Alexander Fischer ***@***.***> wrote: With commit 874facb <874facb>, boottest() should no longer fail when the weights argument in feols() is specified as a formula, or when the cluster argument in feols is specified as a column vector (with the caveat that the column vector needs to be specified in relation to the input data, e.g. feols(fml, data, cluster = data$cluster, weights = data$weights), I will add this option in the next days). — Reply to this email directly, view it on GitHub <#26 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AKYJV45JZ4XFIDSZG3OGBWLUTNA5LANCNFSM5K2NPB6A> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>. You are receiving this because you authored the thread.Message ID: ***@***.***>