www.cmm.bristol.ac.uk/forum

Dear runmlwin-runners,

We are trying to run a Bayesian analyses on a three level dataset (individuals=57771, schools=1344 and countries=25). Because of complexity (more random slopes) and not a big number of countries they advised us a Bayes estimator. Simple models run well (M0, see syntax down) but when it becomes more complex (M1, see below, and further) it gives an error (error whiled obeying batch file C:\Document~1\HJonkman\Look~1\temp\st010001o.tmp at line number 188: west, mcmc error 0003: matrix must be positive definite for inversion

In stata it tells: The model did not run properly in MLwin. Return the model without the nopause option to debug the model in MLwin. (me: I removed nopause but this was not the problem).

In the usergroup mail I see it has to do with starting values for covariance parameters. But, how to change this.

I send you a part of syntax where it gives this error. Maybe you know what to do.

Big best, Harrie






*M0 : Three level model: intercept-only model(empty model)
order acountry schoolid id
sort acountry schoolid id

runmlwin probuse cons, level3(acountry: cons) level2(schoolid: cons) level1(id:) discrete(distribution(binomial) link(logit) denominator(cons)) nopause

runmlwin probuse cons, level3 (acountry: cons) level2(schoolid: cons)level1(id:)discrete(distribution(binomial) link(logit) denom(cons)) mcmc(on) initsprevious nopause






*M1 : with social demographic predictors
runmlwin probuse cons male_r grade8 grade9 native_r, level3(acountry: cons) level2(schoolid: cons) level1(id:) discrete(distribution(binomial) link(logit) denominator(cons)) nopause

runmlwin probuse cons male_r grade8 grade9 native_r, level3 (acountry: cons) level2(schoolid: cons)level1(id:)discrete(distribution(binomial) link(logit) denom(cons)) mcmc(on) initsprevious nopause

Hi Harrie,

Please post the log file output for this series of models (i.e. output as well as commands)

M1 is a three-level random intercept model, and it would be rare to get this message with that model (you are more likely to run into this error message when the model is a random-slope model). There might be something more seriously wrong with your model so it would be helpful to see the model output.

Best wishes

George