COOK, DIFBETA, DIFFIT

[Marcel Cabral]

Hi,

I´m working with a panel data and my results are stranges. I think that this is happening because some outiliers datas for my explanatory variables (epslag, reclag, retlag). I runned a xtreg command (fe and re) exlcluding the outliers after calculate they as Q1/Q3 -/+ 1,5*Interquartile, but I´m not confortable to do it. I know too, that for OLS I can run dfit, cook and dfbeta to exclude my influential observations (instead of the outliers). But I don´t know if is correct to use these for the panel data.

My doubt is if can I run the following command and interpret that as well as in the OLS?

"reg w epslag vcret vcsize vcpbook vcdesv vcturn
dfbeta
predict dfit, dfits
predict cook, cooksd
xtreg w epslag vcret vcsize vcpbook vcdesv vcturn,fe robust, if abs(_dfbeta_1)<1 & abs(dfit)<1 & abs(cook)<1
xtreg w epslag vcret vcsize vcpbook vcdesv vcturn,re robust, if abs(_dfbeta_1)<1 & abs(dfit)<1 & abs(cook)<1
drop _dfbeta_1 _dfbeta_2 _dfbeta_3 _dfbeta_4 _dfbeta_5 _dfbeta_6 dfit cook

reg w reclag vcret vcsize vcpbook vcdesv vcturn
dfbeta
predict dfit, dfits
predict cook, cooksd
xtreg w reclag vcret vcsize vcpbook vcdesv vcturn,fe robust, if abs(_dfbeta_1)<1 & abs(dfit)<1 & abs(cook)<1
xtreg w reclag vcret vcsize vcpbook vcdesv vcturn,re robust, if abs(_dfbeta_1)<1 & abs(dfit)<1 & abs(cook)<1
drop _dfbeta_1 _dfbeta_2 _dfbeta_3 _dfbeta_4 _dfbeta_5 _dfbeta_6 dfit cook

reg w retlag vcret vcsize vcpbook vcdesv vcturn
dfbeta
predict dfit, dfits
predict cook, cooksd
xtreg w retlag vcret vcsize vcpbook vcdesv vcturn,fe robust, if abs(_dfbeta_1)<1 & abs(dfit)<1 & abs(cook)<1
xtreg w retlag vcret vcsize vcpbook vcdesv vcturn,re robust, if abs(_dfbeta_1)<1 & abs(dfit)<1 & abs(cook)<1
drop _dfbeta_1 _dfbeta_2 _dfbeta_3 _dfbeta_4 _dfbeta_5 _dfbeta_6 dfit cook

reg w epslag reclag retlag vcret vcsize vcpbook vcdesv vcturn
dfbeta
predict dfit, dfits
predict cook, cooksd
xtreg w epslag reclag retlag vcret vcsize vcpbook vcdesv vcturn,fe robust, if abs(_dfbeta_1)<1 & abs(_dfbeta_2)<1 & abs(_dfbeta_3)<1 & abs(dfit)<1 & abs(cook)<1
xtreg w epslag reclag retlag vcret vcsize vcpbook vcdesv vcturn,re robust, if abs(_dfbeta_1)<1 & abs(_dfbeta_2)<1 & abs(_dfbeta_3)<1 & abs(dfit)<1 & abs(cook)<1
drop _dfbeta_1 _dfbeta_2 _dfbeta_3 _dfbeta_4 _dfbeta_5 _dfbeta_6 _dfbeta_7 _dfbeta_8 dfit cook"

Can someone help me with this?

My best regards.

[Steve Samuels]

dfbeta, cook's D, dffit and associated statistics are not robust to multiple outliers, as these can mask one another. To identify outliers and high leverage observations, download Verardi and Croux's mmregress (SSC) and read their Stata Journal article, which can be downloaded at no cost: http://www.stata-journal.com/article...article=st0173

[Phil Bromiley]

There seems to be even more disagreement on how to handle outliers than on most other statistical issues. I am not taking a position on whether cook's d etc. are "better" than robust regression or other approaches to outliers.

However, if you choose to go with the cook's d or similar approaches, I don't think you get the answers you're looking for if your regression ignores the panel controls. That is, the outliers with the fixed effects can differ greatly from the outliers without. So, you may want to run the regression with the fixed effects. Consequently, I'd put in panel dummies (i. plus whatever is your panel variable) in the regression before doing the dfbeta etc.

[hend elgamiel]

dfbeta, cook's D, dffit and associated statistics are not robust to multiple outliers, as these can mask one another. To identify outliers and high leverage observations, download Verardi and Croux's mmregress (SSC) and read their Stata Journal article, which can be downloaded at no cost: http://www.stata-journal.com/article...article=st0173

how i can read the results of mmregress and then remove outliers