New package cooksd2 available on SSC

[David Vincent]

Thanks to Kit Baum, a new package cooksd2 is now available on SSC and can be installed by typing:

ssc install cooksd2

cooksd2 generates Cook's distance measures after regress or xtreg, which summarize the effect of deleting an observation, or an entire subject, on the estimated regression coefficients. The procedure uses efficient updating formulas which are described in the accompanying slides. Any comments are most welcome. Contact details are included in the help file.


Thanks,
David.

[Bruce Weaver]

Hi David. I believe the slides you mentioned can be viewed here:

• https://www.stata.com/meeting/uk22/s...22_Vincent.pdf

Cheers,
Bruce

[David Vincent]

Thanks Bruce. The hyperlink in the post (repeated below) is to the slightly amended version:

https://ideas.repec.org/p/boc/lsug22/03.html

Thanks,
David.

[Bruce Weaver]

Okay, fair enough David. If anyone is struggling to find the slides on that page (as I did), here is the direct link:

• http://repec.org/lsug2022/uk2022_vincent.pdf

[Amaa Ahmed]

Hi there. @David Vincent @Bruce Weaver I used the cooksd2 command and i got three columns; Chi, F, and the distances. so which one is to be compared with what? if it is an outlier, for which variable it would be ? Thanx

[David Vincent]

Hi Amaa, sorry for the delayed response (I have only just seen your message). The Cooks distance is usually compared to the percentiles of an F distribution, but I included the Chi2 percentiles as an additional reference (this represents the large sample distribution). The F and Chi2 values refer to the confidence levels that are attained by the distance. For example, if the distance of an observation corresponds to an F or Chi2 of around 0.5, then the removal of this observation, moves the estimated coefficients to the edge of a 50% confidence region based on the full sample estimates. Rather than working with specific cut-offs for these values, I would recommend looking for large relative differences (in the percentiles / distances). Or could you add the parms(newvar) option which adds the jackknifed coefficient estimates to the dataset and look at these directly (useful if you are interested in one specific variable say) . I hope this helps. Thanks, David.