Generalized estimating equations OR mixed effects models?

[Ashmita Karki]

Hi team,

I want to assess the effect of a lifestyle intervention in a cluster randomized trial on health-related quality of life(measured by using EuroQol5D). My sample size is 481 and total number of clusters is 30. Should I use a GEE model or a multilevel mixed effect linear regression model in STATA to perform my data analyses? How do we know which model to use?

Also, if my missing values is less than 5%, do I need to impute missing values using MI or can I do complete case analysis ignoring the missing values as its less than 5%?

Any help/suggestion would be very helpful.

Thank you!

[Joseph Coveney]

Should I use a GEE model or a multilevel mixed effect linear regression model in STATA to perform my data analyses?

My understanding is that GEE is intended more for longitudinal models. Nevertheless, if you're fitting a linear model with an exchangeable residual covariance structure, then xtgee and mixed give you the same large-sample results. You can see this by running the following code.

sysuse bplong, clear
xtgee bp i.sex##i.when, i(patient) family(gaussian) link(identity) corr(exchangeable) nolog
mixed bp i.sex##i.when || patient: , nolrtest nolog

With a sample size of 30, though, I would opt for mixed, because it has adjustments available to accommodate small samples, namely, the reml dfmethod() options.

About your missing values, the bigger question is liable to be the mechanism of missingness, especially but not exclusively if you observe differential dropout. You might want to contemplate a secondary sensitivity analysis in which you impute the best scores for missing values in the control treatment group and the worst scores for those in the experimental treatment group.