issue with meologit

[andin anindya]

I'm new in Stata and I'm having a problem with meologit commands. I wanted to examine estimates of the state (level 2) effects in my data and I only want one record of each state in order to prduce a caterpillar plot. these are my syntax:

predict u, reffects reses(use)
(calculating posterior means of random effects)
(using 15 quadrature points)

egen pickone=tag(state)

egen urank=rank(u) if pickone==1
(7282 missing values generated)

I don't know why there are missing values generated and more than 50% of my empirical bayes means for _cons[state] is zero. anyone have an answer or suggestion?

Thanks in advance

[Clyde Schechter]

I don't know why there are missing values generated

Well, your command -egen urank=rank(u) if pickone==1- tells Stata to only create a value of urank in those observations where pickone == 1 (which amounts to one observation for each state). For all other observations (presumably the majority) you have told Stata to leave a missing value there. (And that is quite appropriate if you are going on to produce a caterpillar plot.)

If you mean that Stata is produceing missing values of urank even in observations where pickone == 1, that is a separate issue. But, like your second question about zero estimates for _cons[state], one that I don't see any way to help out with unless you show: both the exact -meologit- command you used and the complete and exact -meologit- output you got from Stata. Even with all of that, it may not be clear what the source of the problem is: it could be some peculiarity of your data. But at least with those two, some possibilities could be ruled out, and perhaps some considered probable.

[andin anindya]

Well, your command -egen urank=rank(u) if pickone==1- tells Stata to only create a value of urank in those observations where pickone == 1 (which amounts to one observation for each state). For all other observations (presumably the majority) you have told Stata to leave a missing value there. (And that is quite appropriate if you are going on to produce a caterpillar plot.)

If you mean that Stata is produceing missing values of urank even in observations where pickone == 1, that is a separate issue. But, like your second question about zero estimates for _cons[state], one that I don't see any way to help out with unless you show: both the exact -meologit- command you used and the complete and exact -meologit- output you got from Stata. Even with all of that, it may not be clear what the source of the problem is: it could be some peculiarity of your data. But at least with those two, some possibilities could be ruled out, and perhaps some considered probable.

Yes, I mean Stata producing missing values of urank. well, here is my command and output for meologit ((kode_kab in command below means state))

. meologit y || kode_kab:,intpoints(15)

Fitting fixed-effects model:

Iteration 0: log likelihood = -5752.0839
Iteration 1: log likelihood = -5752.0839


Refining starting values:

Grid node 0: log likelihood = -5741.42

[andin anindya]

Log likelihood = -5709.5501 Prob > chi2 = .
------------------------------------------------------------------------------
y | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
/cut1 | -2.087391 .0450602 -46.32 0.000 -2.175707 -1.999074
/cut2 | -.6080492 .0327438 -18.57 0.000 -.6722259 -.5438725
-------------+----------------------------------------------------------------
kode_kab |
var(_cons)| .4120288 .298579 .0995634 1.705122
------------------------------------------------------------------------------
LR test vs. ologit model: chibar2(01) = 85.07 Prob >= chibar2 = 0.0000

[Richard Williams]

As a sidelight this would be much easier to read if you used code tags; see pt 12 of the FAQ. Otherwise everything that is 2 spaces or more gets reduced to 1 space which totally screws up the spacing.

[andin anindya]

As a sidelight this would be much easier to read if you used code tags; see pt 12 of the FAQ. Otherwise everything that is 2 spaces or more gets reduced to 1 space which totally screws up the spacing.

I'm sorry Sir, I've edit it so it can be much easier to read

[Clyde Schechter]

OK. You are getting exactly the number of missing values of urank that you should be getting. The -meologit- output shows you start with 7,290 observations in the estimation sample. And they are in 8 groups (kode_kab). So your command generating the variable pickone will select 8 observations to set pickone = 1, and for the others pickone = 0. So there will be 7290-8 = 7282 observations that don't have pickone = 1. All of those observations will not get a value of urank because your -egen urank...if pickone == 1- command explicitly says to evaluate urank only in those 8 observations. In the remaining 7,282 observations urank is set to missing value.

As for 50% of your empirical Bayes estimates being zero, this isn't terribly surprising either. You have only 8 groups. So 50% of your empirical Bayes estimates is exactly four of them. With only 8 different groups, you are not really getting a very good sample of the distribution of effects at the kode_kab level. So pretty much anything can happen.

In all honesty, I don't think I would even do a random effects model with just 8 groups. There just isn't enough information there to really get a decent estimate of the parameters of that distribution.

[andin anindya]

OK. You are getting exactly the number of missing values of urank that you should be getting. The -meologit- output shows you start with 7,290 observations in the estimation sample. And they are in 8 groups (kode_kab). So your command generating the variable pickone will select 8 observations to set pickone = 1, and for the others pickone = 0. So there will be 7290-8 = 7282 observations that don't have pickone = 1. All of those observations will not get a value of urank because your -egen urank...if pickone == 1- command explicitly says to evaluate urank only in those 8 observations. In the remaining 7,282 observations urank is set to missing value.

As for 50% of your empirical Bayes estimates being zero, this isn't terribly surprising either. You have only 8 groups. So 50% of your empirical Bayes estimates is exactly four of them. With only 8 different groups, you are not really getting a very good sample of the distribution of effects at the kode_kab level. So pretty much anything can happen.

In all honesty, I don't think I would even do a random effects model with just 8 groups. There just isn't enough information there to really get a decent estimate of the parameters of that distribution.

thank you so much for your answer Sir, I get it now

[Liting Cai]

deleted post as i figured out the issue.