www.cmm.bristol.ac.uk/forum

I am currently working on a paper to be submitted to a Special Issue of the Journal of Management on Bayesian Methods.

I have already conducted a thorough analysis of a 3-level model of venture capital investment rounds in Stata (ie. xtmixed) whereby 2083 VC investment occasions are nested within 966 firms nested within 16 regions. Building on Browne&Draper (2006) I am framing the paper with a view to compare Bayesian and likelihood-based estimation methods for my model.

I have decided to use MLwiN on Stata for mcmc estimation (runmlwin) and have just received my free MLwiN v2.26 licence earlier today. However, whilst useful results were obtained with likelihood-based estimation from Stata xtmixed (i.e. level3 random intercepts, level2 random slopes/intercepts, level1 random intercepts) I am not able to get runmlwin to produce a meaningful estimation of parameters based on IGLS, a preliminary step for estimating parameters (prior distribution) before running mcmc (see attached Stata output). Instead, I am getting the following warning: Caution. MLwiN was unable to calculate the log likelihood, there may be a problem with your model. There is no problem estimating the model with Stata xtmixed and thus, the MLwinN error message is in stark contrast with claims that runmlwin and xtmixed produce similar results. Simple 2-level models run well with runmlwin but 3-level models appears to be struggling. The Stata syntax I am (or planning to) using is as follows:

xtmixed lcromount royear ronum1 ronum2 ronum3 ronum4 patcount cclaims ccreceive ccmade cgeneral coriginal c1 c2 c3 c4 c5 c6 p foyear ipoyear || cometro: || coname: royear, cov (unstr)

runmlwin lcromount cons royear ronum1 ronum2 ronum3 ronum4 patcount cclaims ccreceive ccmade cgeneral coriginal c1 c2 c3 c4 c5 c6 p foyear ipoyear, level3(cometro: cons) level2(coname: cons royear) level1(cons) maxi(1000) nopause

runmlwin lcromount cons royear ronum1 ronum2 ronum3 ronum4 patcount cclaims ccreceive ccmade cgeneral coriginal c1 c2 c3 c4 c5 c6 p foyear ipoyear, level3(cometro: cons) level2(coname: cons royear) level1(cons) mcmc(on) initsprevious nopause


I am only just starting to use MLwiN on Stata and your advice would be much appreciated. Alternatively, I am prepared to travel to Bristol to sort this out if necessary.

Kind regards

Hi Pierre,

When using runmlwin you must explicity include the level-1 residual error term e_ij

You do this by adding the following option

...level1(level1id: cons)

where level1id is the name of the unique identifier which indexes the records in the dataset and cons is a variable which equals the value one.

This usage contrasts xtmixed where the level-1 residual error is included by default

Once you include the level-1 residual errors I expect that your runmlwin results should match your xtmixed results.

Best wishes

George

Dear George

Many thanks. This does the trick.

P