3-levels & time

[alzbetab]

Hi,

I'm modelling transition to birth amongst European women and have a little difficulty to interpret the random part of my model.

I got observations (or waves if you like) on the first level, individuals (women) on the second and country on the third. I was looking for some examples of interpretation in the literature but so far I haven't found anything satisfactory. So I'm primarily using Leckie & Charlton's article to try to understand as I am using Runmlwin in STATA. What really puts me off is the interpretation of time on the first level. I'm not really fussed by it since my main motivation behind using panel data was to deal with the fact that birth is quite a rare event and also that my sample sizes are not that amazing. So basically I just want to control for the fact that the observations are not independent of each other and want to focus primarily on the level 2 outcomes. Can I do that? Also how big a problem in this case can be the fact that I have on average 1.8 observations?

Anyway, I've run an empty model and this is what I've got.

intercept = -2.0828, predicted probability = 0.1108
level-2 variance = 0.5862 => VPC = 0.1512
level-3 variance = 0.1267 => VPC = 0.0371

I would crudely interpret it as the probability of giving birth is about 11%, about 15% of the variance in transition to birth is between women and about 4% is between countries. Does it therefore mean that the differences in transition to birth is due to differences in women's characteristics rather than country-specific characteristics? It might be a stretch but does it also mean that the differences in birth occurrence may be due to compositional effects rather than country-specific characteristics? And what happens with the time? Does it relate only to the time-varient women's characteristics or does it need special interpretation?

As you can see I'm a little bit confused...Thank you.

Alzbeta

[GeorgeLeckie]

Hi Alzbeta,

I think you would be best posting your query on the multilevel JISC list as it is about statistical/multilevel modelling in general rather than a query specific to the runmlwin command.

Best wishes

George

[alzbetab]

Will do, thank you. In that case I would have one technical question though. Is it a problem if my pql2 model fails to converge before I run mcmc? If I use maxiterations it runs for ages and so far I always had to interrupt the process.

[GeorgeLeckie]

Hi Alzbeta,

No need for PQL2 estimates. MQL1 estimates are fine as starting values for MCMC. Actually you could even specify plausible starting values manually if you preferred. This is after all what you would typically do if you were using WinBUGS. As usual with MCMC estimation you want to check that you specify a long enough burnin period so that the parameter chains escape the influence of the starting values.

Best wishes

George