meologit: null model without feasible initial values

[Theresa Markefke]

Dear Statalists,

I am analysing data from the European Social Survey and am trying to run a (three-level) multilevel ordinal logistic regression. The dependent variable are individual preferences for redistribution which is an ordered response variable with five categories. I ran these null models, which indicate that individuals are nested within country x waves interactions, which are nested within countries;

meologit prefredist || country: || year:,

meglm prefredist || country: || year, family(ordinal) link(logit)

and both times got the error that initial values are not feasible, r(1400). I have searched the forums and usually the advice is to reduce the modell in terms of variables and then use the solutions of the less complicated ones as starting values but since this is the null model, that is not an option. I tried running just

meologit prefredist || country:, but got the same error.

Do you have any idea what the problem might be?

Comments would be much appreciated.

[Clyde Schechter]

Based on nothing but a hunch, my next step would be to try a one-level model, just -ologit prefredist- and try using those results as starting values. I don't know if it will work or not.

Also I wonder if there is something peculiar about the distribution of prefredist in your data that makes it difficult to fit models to. Have you looked at -tab prefredist-? If, for example, it is practically a point mass at one value with only very rare selections of other values, I would expect considerable difficulty fitting a model to it. The same would be true if the marginal distribution looked OK but within each country the responses were almost all bunched up into a single category.

[Weiwen Ng]

Just to add a bit to what Clyde said, if the -ologit- model converges, then the next thing to do would be to add one level to the model and see if that converges, e.g.

meologit prefredist || year:,

That got me thinking about the nesting structure, though. Is it actually true that wave is nested within country? I assume each wave was administered simultaneously across several countries. Is that not more like a crossed random effect? In the mixed model manual, perhaps example 8 is parallel to your goal.

Alternatively, depending on how many years there are, it could be that you could use a fixed effect for year.

[Erik Lamontagne]

I would like to reactivate this issue as I am facing the same issue.
It is a two-level (individual, country) ordonated logistic regression with 100k participants from 160 countries.
The null model
meologit hivtest || country:,

gives: initial values not feasible r(1400);

The dependent variable (hivtest) does not have a point mass or any particular element that makes me think what could make it difficult to fit a model to it (see below)

hivtest Freq. Percent Cum.

never 23,858 23.77 23.77
more 12 months 17,998 17.93 41.69
within 12 months 13,706 13.65 55.35
within 6 months 44,828 44.65 100.00

Total 100,390 100.00


I tried what Clyde suggested, and ran a ologit hivtest successfully, but then, how can I see whether making a multilevel is adding anything ?

thanks in advance,
Erik

[Clyde Schechter]

How many countries are there? How many observations per country per level of hivtest are there (-tab country hivtest-)? If there are some small cells, that might be causing the problem. That's just a hunch. These initial values not feasible situations can be difficult to resolve.

[Richard Williams]

One of my generic suggestions is to add the -difficult- option. Sometimes it works miracles.

You could try a different xt or me command and see if it works, e.g. xtmlogit (if you have Stata 17).

[Erik Lamontagne]

Dear Clyde and Richard, I thought for so long nobody answered my post. I was obviously doing someting wrong when looking at Statalist. I gave up after two weeks.
There was indeed variables with small number of participants. hivtest is a 4 likert, including "never"; ">12months"; "6-12 months"; "<6months". The issue got resolved when looked at two part of this hivtest variable: nevertested (with those "never") and recencytest >12; 6-12; <6months)
Thank you once again.

[Erik Lamontagne]

Dear Clyde and Richard, I thought for so long nobody answered my post. I was obviously doing something wrong when looking at Statalist. I gave up after two weeks.
There was indeed variables with small number of participants. hivtest is a 4 likert, including "never"; ">12months"; "6-12 months"; "<6months". The issue got resolved when looked at two part of this hivtest variable: nevertested (with those "never") and recencytest >12; 6-12; <6months)
Thank you once again.