Hi,
Can I fit a GEE model and adjust for clustering in MlwiN?
I understand it is possible by fitting a single level model, only fixed effects, and adjusting for clustering would be performed by adding clusters as dummy variables in the equation. Is this correct?
Happens to be that if I fit in my second level as countrys I only have 2, and if i set city I have 23 clusters, but on three cities I have just one participant, and in two cities two. A solution would be setting GEE model, but can I fit this model with Mlwin?
Thank you,
GEE approach adjust for clustering in MlwiN
Re: GEE approach adjust for clustering in MlwiN
MLwin does not fit a GEE model; it fits a random effects model, that is it estimates an explicit between-group variance You can fit a two level model with people at level 1 nested in cities at level 2 and have a random intercepts model for between city variation. You can then put in a dummy that distinguishes the two countries. Finally you can request robust (sandwich) standard errors for this model both in the fixed part (and if you want them in the random part)
The commands which you type in the command window after specifying the model and before start are
FSDErrors type <value>
Set the type of standard error computation for fixed parameters.
If <value> = 0, use standard, uncorrected, IGLS or RIGLS computation.
If <value> = 2, compute robust or 'sandwich' standard errors based on raw residuals.
RSDErrors type <value>
Set the type of standard error computation for random parameters.
If <value> = 0, use standard, uncorrected, IGLS or RIGLS estimates.
If <value> = 2, compute robust or 'sandwich' standard errors based on raw residuals.
Given the small sample size in some of the units you will clearly not be able to say much about the specific cities but you may be able to say something about the within and between city variation. If you put in a dummy for each and every city you will not be e able to include any city level variables, as you will have exhausted the degrees of freedom at that level.
The commands which you type in the command window after specifying the model and before start are
FSDErrors type <value>
Set the type of standard error computation for fixed parameters.
If <value> = 0, use standard, uncorrected, IGLS or RIGLS computation.
If <value> = 2, compute robust or 'sandwich' standard errors based on raw residuals.
RSDErrors type <value>
Set the type of standard error computation for random parameters.
If <value> = 0, use standard, uncorrected, IGLS or RIGLS estimates.
If <value> = 2, compute robust or 'sandwich' standard errors based on raw residuals.
Given the small sample size in some of the units you will clearly not be able to say much about the specific cities but you may be able to say something about the within and between city variation. If you put in a dummy for each and every city you will not be e able to include any city level variables, as you will have exhausted the degrees of freedom at that level.