How to obtain robust standard error in -xtlogit fe-

[Alex Mai]

Dear statalist,

In the Stata manual of -xtlogit-, the FE option can be followed by only -vce(jackknife)- or -vce(bootstrap)-, while the RE option can be followed by -vce(robust)-. So is there any method to obtain robust standard error for -xtlogit fe-?

Thank you!

[Andrew Musau]

-vce(bootstrap)- will give you very similar standard errors to -vce(robust)-. However, to your question, you can use clogit which allows you to specify the robust option

clogit y x1 x2 ... xn, group(id) robust

[Alex Mai]

-vce(bootstrap)- will give you very similar standard errors to -vce(robust)-. However, to your question, you can use clogit which allows you to specify the robust option

clogit y x1 x2 ... xn, group(id) robust

Thanks a lot! Btw, you mentioned that -vce(bootstrap)- gives very similar standard errors to -vce(robust)- for -xtlogit fe-. Does this hold for other commands?

For example, unlike -qreg-, -sqreg- (simultaneous quantile regression) only has the option of bootstrap -reps(#)-, but not -vce(robust)-.

[Carlo Lazzaro]

Alex:
whenever Stata command do not allow something, there are sound theorethical reasons about that.

[Alex Mai]

Alex:
whenever Stata command do not allow something, there are sound theorethical reasons about that.

Dear Carlo,

Thank you! I think -xtlogit, fe- and -clogit- may be used interchangeably, which is my understanding after reading the note by Richard Williamson and the book by Paul Allison. But unlike -clogit-, -xtlogit fe- does not allow -vce(robust)-. Does it imply some (perhaps subtle) differences between the two commands?

[Carlo Lazzaro]

Alex:
you might be interested in the following thread: https://www.stata.com/statalist/arch.../msg00068.html

[Richard Williams]

whenever Stata command do not allow something, there are sound theoretical reasons about that.

I always tell my students that whenever Stata does not do something I expect it to, it is either because (a) Stata is smarter than me and knows it would be wrong, or (b) nobody's gotten around to programming it yet.

In this case, it seems odd to me that clogit allows vce(robust) but xtlogit, fe does not. In fact, I think it says somewhere that xtlogit, fe calls clogit, which makes the omission even more puzzling to me.

[Andrew Musau]

Thanks a lot! Btw, you mentioned that -vce(bootstrap)- gives very similar standard errors to -vce(robust)- for -xtlogit fe-. Does this hold for other commands?

For example, unlike -qreg-, -sqreg- (simultaneous quantile regression) only has the option of bootstrap -reps(#)-, but not -vce(robust)-.

The reference here is to xtlogit, see p. 623 of Cameron and Trivedi (Microeconomics using Stata, 2010) where they note that panel robust standard errors are obtained using the -vce(bootstrap)- option. I also share Richard's puzzlement in #7, it would be beneficial for StataCorp to be more explicit in the manual entry of xtlogit as to why -vce(robust)- is not allowed. The only thing that I can note is that clogit is not a panel data estimator in the Stata sense (i.e. those that require you to xtset your data) and therefore it may include other additional options beyond those available in a panel estimator.

[Alex Mai]

I always tell my students that whenever Stata does not do something I expect it to, it is either because (a) Stata is smarter than me and knows it would be wrong, or (b) nobody's gotten around to programming it yet.

In this case, it seems odd to me that clogit allows vce(robust) but xtlogit, fe does not. In fact, I think it says somewhere that xtlogit, fe calls clogit, which makes the omission even more puzzling to me.

Many thanks! I have read your note on Conditional Logit/Fixed-effects Logit in which you points out that "In fact, I believe xtlogit, fe actually calls clogit. "

Btw, may I ask two questions about the Hausman test for xtlogit?

First, I have tried adding -sigmamore- to deal with the error of "V_b-V_B is not positive definite". However, Stata gives the red error that "Estimators do not save e(sigma) or e(rmse), sigma option not allowed". I have checked the Stata manual. It seems that -sigmamore- may be valid only after -xtreg-. So may I ask if there is any solution to this problem for -xtlogit-?

Second, since the traditional Hausman test requires homoskedasticity or the absence of clustered errors, -xtoverid- is used. However, -xtoverid- is not valid after -xtlogit-. So is there command to do robust Hausman test after -xtlogit-?

Thank you!

[Amy Fang]

Many thanks! I have read your note on Conditional Logit/Fixed-effects Logit in which you points out that "In fact, I believe xtlogit, fe actually calls clogit. "

Btw, may I ask two questions about the Hausman test for xtlogit?

First, I have tried adding -sigmamore- to deal with the error of "V_b-V_B is not positive definite". However, Stata gives the red error that "Estimators do not save e(sigma) or e(rmse), sigma option not allowed". I have checked the Stata manual. It seems that -sigmamore- may be valid only after -xtreg-. So may I ask if there is any solution to this problem for -xtlogit-?

Second, since the traditional Hausman test requires homoskedasticity or the absence of clustered errors, -xtoverid- is used. However, -xtoverid- is not valid after -xtlogit-. So is there command to do robust Hausman test after -xtlogit-?

Thank you!

Hi Alex, have you solved this problem? I have encountered the same problem.

Besides, I have one problem with -clogit-. "Warning: variance matrix is nonsymmetric or highly singular" when I use the -robust- option.

Thank you! Look forward to your reply!

Amy

[Philipp Wegelin]

Hi Alex, have you solved this problem? I have encountered the same problem.

Besides, I have one problem with -clogit-. "Warning: variance matrix is nonsymmetric or highly singular" when I use the -robust- option.

Thank you! Look forward to your reply!

Amy

Hi Amy / Alex,

have you found a solution to your problem? (the Hausman test problem)?