meqrlogit about ICC

[Jay Jang]

First of all, thank you for the advice. I ran models using stata command "meqrlogit" with 3 level.

The most of results show the icc of level 3 be larger than level 2.

Some result has case that the icc at level 3 be less than the icc at level 2

icc at level 3 is 0.03779

icc at level 2 is 0.22457

Have you ever seen this result before??

[Joseph Coveney]

Have you ever seen this result before??

Well, .. .

version 18.0

clear *

// seedem
set seed 1893061210

quietly set obs 35
generate byte hid = _n
generate double hid_u = rnormal(0, 0.25)

quietly expand 35
generate long pid = _n
generate double pid_u = rnormal()

quietly expand 20
generate byte out = rbinomial(1, invlogit(hid_u + pid_u))

meqrlogit out || hid: || pid: , nogroup nolog
estat icc

exit

Do-file and log file attached.

1. Why do you think that the results are inadmissible?

2. Why are you using meqrlogit instead of melogit?

[Jay Jang]

It seems I posted the wrong question. However, you still answered, and when I opened the .do file you sent, I understood it quickly. Thank you.

I initially used 'meqrlogit' and have been using it consistently. I am also using it with 'melogit'. I wanted to present the results consistently using the same command. However, I would like to use 'melogit' in the future."

[Leonardo Guizzetti]

Why do you use meqrlogit instead of the newer melogit? Very few applications need the former with modern Stata.

[Joseph Coveney]

. . . I understood it quickly. Thank you.

You're welcome.

I initially used 'meqrlogit' and have been using it consistently. . . . I wanted to present the results consistently using the same command. However, I would like to use 'melogit' in the future."

I recommend switching over to to melogit.

In working up the example that I posted above, there was a trial run (different sample size and seed) where meqrlogit had difficulties getting feasible starting values, but melogit just sailed through.

Even with the example dataset that I posted, meqrlogit takes considerably longer to converge than does melogit, the former taking fourfold to sixfold the time-to-convergence of the latter. (Log files attached. The first is AB and the second is BA, because there is about a two-second penalty for going first—there's some latency in setting up the regression machinery.)