Multilevel analysis

[lourdes gastearena]

Hello statalist;
I am trying to do a multilevel analysis and when introducing second level variables the analysis takes long and finally does not come off. I do not know which is the reason and would appreciate any help with this matter. I am using meoprobit command

Another question is that I do not have clear which is the difference between the use of normal meoprobit as indicated in the menu, or meoprobit ending with " ,vce (robust)

[Clyde Schechter]

I don't see how anybody can troubleshoot this without seeing the actual code you ran and the actual output (including any error messages) you got from Stata. Please read the Forum FAQ, with special attention to #12, for information on how to most helpfully show this information.

[lourdes gastearena]

Dear Clyde,
Thank you so much for your response. I will try to explain better. I am studying job satisfaction in the public sector in European countries.
My dependent variable is job satisfaction (ordered probit from 1 to4). The independent 1 level variable is public sector (dummy variable, 1public, 2private) the 2 level variable is power distance, (pdi, Hofstede). I use the command meoprobit like this:

meoprobit satisfaction_ordered public_sector1 || Country: pdi

and got this result:


Fitting fixed-effects model:

Iteration 0: log likelihood = -12166.284
Iteration 1: log likelihood = -12155.196
Iteration 2: log likelihood = -12155.196

Refining starting values:

Grid node 0: log likelihood = -47.64866

Fitting full model:

could not calculate numerical derivatives -- discontinuous region with missing values encountered
could not calculate numerical derivatives -- discontinuous region with missing values encountered
r(430);

Previously I have dropped missing values in the dataset. Does anyone know how to solve this problem?
Thank you in advance,
Lourdes.

[Clyde Schechter]

These are often difficult problems to solve. The first thing I would do is check that the data lend themselves to this kind of analysis. Check -tab satisfaction_ordered public_sector- to make sure that there are adequate numbers of observations in each cell of that table.

Assuming that is OK, the first thing I note is that the model is mis-specified, because you have listed pdi as a variable with random slope, but you did not include it in the fixed-level of the model. So you need to change it to:

meoprobit satisfaction_ordered public_sector1 pdi || Country: pdi

This might or might not solve your problem. There's a decent chance it will, because unless the actual mean slope for pdi is close to 0, the mis-specification is forcing Stata to try to work in a section of the likelihood that is really distant from a good fit and the likelihood might be very badly behaved there.. If this does not result in a convergent model, the next step is to try to simplify the model.

Random slopes add a lot of computational burden and make the likelihood function, which is a difficult one with -meoprobit- to start with, even more unwieldy, and are often not that important to your research goals. In fact, I worry from your description of the problem that you do not actually want random slopes on pdi and that you have mistakenly put it in the || Country: part of the model simply because it is defined at the country level. If you did that, then you misunderstand how the syntax of these models work. All model variables, regardless of the level at which they are defined, need to appear in the bottom level, i.e. the varlist after -meoprobit- itself, of the model. They get listed again in the higher level sections of the model only if you also need random slope estimation for those variables. In any case, the next step in trying to get the model running would be to give eliminate the random slopes and just run:

meoprobit staisfaction_ordered public_sector1 pdi || Country:

If the model runs to convergence without the random slopes, but not with it, then you will probably just have to live without the random slope variance estimates.

If none of these attempts lead to a convergent model, post back showing what output you got from each attempt, and there may be other things that can be tried.

[Klaus Steitzel]

I agree with Clyde, I sense that you want to do something like:

Are public sector workers more satisfied with their jobs?

meoprobit satisfaction_ordered i.public_sector1 || Country:

Is there variation in job satisfaction gap between public and private sector workers across countries?

meoprobit satisfaction_ordered i.public_sector1 || Country: public_sector1

Is the job satisfaction gap bigger or smaller when power distance is bigger or smaller?

meoprobit satisfaction_ordered i.public_sector1##c.pdi || Country: public_sector1

[lourdes gastearena]

Thank you so much Clyde and Klaus for your quick and valuable response,

The way Clyde explain does not work in my case. Thank you anyway for your help.
The way Klaus explain Works but with minor changes. Sincé it is the first time I am running this regression, could you, please, help me to interpret the results? I send attached.

Many thanks,
Lourdes.

[Marcos Almeida]

Please act according to advice given in #2.

[Charilaos Mertzanis]

Hi. I am new to the blog and also a moderately experienced user of Stata. I am using Stata 15. I would like your help on the following: I am doing cross-country research and my dependent variable Y is ordinal (1-4). My sample includes country / sector / firm variables. I am using year/country fixed effects. Now, I am being asked to do a multi-level analysis of my research question. I have been looking at the meoprobit command. I read the manual and as much information as I could find on the net. My first question is whether there is a test that can justify the use of the meoprobit command over the oprobit one. I cannot see it in the manual with respect to this command specifically. My second question is whether, as far as I understand, that when using meoprobit the regression equation should not include the i.year and i.country dummies. I am correct? Thank you.