error message - multinomial logit model

[scarpent]

Hi,

I try to run a multinomial logit discrete-time event history analysis with Mlwin 2.26, but I get the error message 'group 18 contains no columns(calc)'.
I examine the determinants of exit from welfare to three mutually exclusive states (work, work with supplementary welfare benefit, other benefit), while controlling for unobserved heterogeneity. I have a person-period data file (for 28052 persons) with time clustered within individuals.

I use the following syntax:

sort time_uniqueid
xi: runmlwin exitstates4_unordered cons i.kwartaal i.duration_lastspell i.geslacht_m leeftijdtv_c leeftijd2_c i.natio_tv_cat4 i.wetg_t1 i.householdtype_zm WItwst_c i.gewest_t1, level2(insz: cons) level1(time_uniqueid) discrete(distribution(multinomial) link(mlogit) denominator(cons) basecategory(1)) nopause maxiterations(500) mlwinsettings(size(100000))

xi: runmlwin exitstates4_unordered cons i.kwartaal i.duration_lastspell i.geslacht_m leeftijdtv_c leeftijd2_c i.natio_tv_cat4 i.wetg_t1 i.householdtype_zm WItwst_c i.gewest_t1, level2(insz: cons, residuals(u2)) level1(time_uniqueid) discrete(distribution(multinomial) link(mlogit) denominator(cons) basecategory(1)) mcmc (burnin(2000) chain(20000) orthogonal) initsprevious nopause mlwinsettings(size(100000))

It is not clear to me what the error message means. There is no personal id 18 in the dataset. The 18th individual does not has any particularities, and doing the analysis without 18th individual gives the same error message.

Do you have any suggestions?

Thank you in advance,

Sarah

[GeorgeLeckie]

Hi Sarah,

You describe a MLwiN side error message. Which model do you get the error message with? The MQL model or the MCMC model? You might simplify your model to the simplest model which replicates the same error message this often sheds light on what might be causing the error message.

In passing, I am concerned with your use of maxiterations(500), this is very high and suggests that you are having convergence problems with your MQL1 model. You might be providing your MCMC model with rather poor starting values

xi: runmlwin exitstates4_unordered cons i.kwartaal i.duration_lastspell i.geslacht_m leeftijdtv_c leeftijd2_c i.natio_tv_cat4 i.wetg_t1 i.householdtype_zm WItwst_c i.gewest_t1, level2(insz: cons) level1(time_uniqueid) discrete(distribution(multinomial) link(mlogit) denominator(cons) basecategory(1)) nopause maxiterations(500) mlwinsettings(size(100000))

The best way to ensure convergence when fitting by MQL1 is to fit a sequenced of increasingly complex models each time using the parameter estimates from the previous model as starting values for the current model. I strongly encourage that you do this. Put the most important sets of predictors in first.

Best wishes

George

[ChrisCharlton]

This does indeed look like MLwiN is having a problem, although I don't know exactly what the root cause of it is. When running discrete response models MLwiN has to transform the original data in order to fit it. So that it doesn't damage you original data it makes a copy of your ID, response and explanatory variables before each iteration and copies them back afterwards. This is done by making a reference to these variables in group 19 (g19) and assigning an equivalent number of columns to group 18 (g18). At the start of each iteration it issues a CALC g18=g19 command to copy the data and at the end of the iteration it does CALC g19=18 to copy the variables back. You can see which columns are in each group by looking at the Data Manipulation->Groups window. It appears that in your case it has failed to copy the variables into g18 and so when it attempts to restore them they cannot be found and hence the error. I can't say why it is failing in this case, although it is possible that there isn't enough memory assigned to make the copy, or that there aren't enough free columns. One thing that you could try is looking at the groups window to see which columns correspond to g18 and g19 and then see whether the number of columns match and contain data.

[scarpent]

Hi George and Chris,

Thank you for your comments and suggestions!

The IGLS-model gives the MLwiN error message.

I found out that I have indeed a convergence problem, even with the empty model. It seems to be a problem of non-identification of a dependent competing risks model.
However, the mlogit model (which does not model unobserved heterogeneity) converges very quickly.

I guess I need to turn to the literature on discrete-time competing risks. Any suggestions?

Sarah

[GeorgeLeckie]

Hi Sarah,

In terms of

"I guess I need to turn to the literature on discrete-time competing risks."

See the training materials by Fiona Steele and the references contained within them. These training materials are located at

http://www.bristol.ac.uk/cmm/software/s ... materials/

Best wishes

George