Extra option and explanatory variables in MM models

[Jacqueline]

Dear runmlwin users,

I have two questions with which I was hoping you could please help.

First, I am analysing data that could be described as count data and that exhibit over-dispersion. Thanks to Chris, I understand that the negative binomial distribution is not implemented currently in MLwiN when using MCMC estimation. I was wondering, therefore, whether an alternative to the negative binomial distribution would be estimating a Poisson model with the extra option. If so, I would appreciate any guidance/information on how to interpret the results of the model with the extra option. For instance, if the value of α is greater (less) than 1, does this indicate over- (under-) dispersion?

Second, I am attempting to estimate a multiple membership model where patients may be treated by more than one consultant, nested within hospitals. I have constructed the multiple membership unit identifiers (consultant1-consultant6) and the corresponding multiple membership weights. I would like to include consultant-level explanatory variables, but I’m not sure how this should be done. Should the order of the consultant-level explanatory variables be the same as that of the multiple membership unit identifiers? If so, some of the explanatory variables may then be repeated if, for example, consultant A is consultant1 for one patient, but consultant6 for another patient.

As always, any assistance would be greatly appreciated.

Many thanks,

Jacqueline

[GeorgeLeckie]

Good to hear from you Jacqueline,

Yes you can do poisson model with over-dispersion by specifying the extra option. Values greater than 1 indicate over-dispersion, values less than 1 indicate under-dispersion.

An alternative way of bringing in over-dispersion is to fit your level-1 units (patients) at both level-1 and level-2. A significant level-2 variance suggests over-dispersion,

You should find more information if you look in multilevel textbooks and google overdispersion in multilevel poisson models etc.

In terms of entering consultant-level variables into your multiple membership models simply construct weighted averages of your consultant variables using the multiple membership weights and enter only the weighted average variable into your model.

So if a patient is seen by 2 consultatnts and you want to enter consultant age into the model generate a variable equal to 0.5 time the age of the first consultant plus 0.5 times the age of the second consultant.

Best wishes

George

[Jacqueline]

Dear George,

Thank you very much for your very helpful response.

Kind regards,

Jacqueline