Calculation of ICCs after meglm

[Christophe Haon]

Dear all,
I am looking for a way to calculate intraclass correlation coefficients after an meglm estimation.
estate icc is not an option after meglm.
Thanks,
Christophe

[Clyde Schechter]

You can always do it "by hand." The -meglm- output includes the variance components at all levels (or at the bottom level it is fixed at 1 for probit link and pi²/3 for logit) and you can just calculate the appropriate sums of those and take the ratio. If you are unsure how that works, see the Methods and formulas section of the estat icc chapter in the [ME] volume of the online help. (-help meglm-, click on the link near the top to get to the chapter, and there click on the ink to Methods and formulas.)

The other possibility is to re-run the model using the more specific -me- command, such as -mixed- or -melogit-, etc. as the case may be.

[Joseph Coveney]

What kind of generalized linear model are you trying to compute the ICC for?

ICC might not be handy for some models that meglm can fit.

webuse epilepsy
meglm seizures treat lbas lbas_trt lage v4 || subject: , family(poisson) link(log)

[Clyde Schechter]

Yes, Joseph Coveney is right, of course. I was presuming that the model used was one for which ICC is definable, namely those where a fixed residual variance is either implicit in the model or can be estimated. Poisson, gamma, and negative binomial are "off the table".

[Christophe Haon]

Thanks to both of you. In my case the "by hand" solution works (logit).

[Rafal Raciborski (StataCorp)]

We will add support for estat icc after meglm in cases where ICCs can be computed.

[Nanco Bonnevieri]

Finch, Bolin & Kelley Multilevel modeling using R (2014, p.154-155) estimate ICC for random intercept Poisson regression (see my red boxes in attachment). Is this wrong?
Throughout the book (e.g. p. 44-45), they seem to plug in standard deviations and not variances. Can anyone explain why?
Finch_multilevel-modeling-using-r_p44-45_154-155.pdf