Setting up a statistical analysis comparing two groups adjusted for age and twins I'm investigating if there is a difference in volume of a specific brain region (T1 weighted MR scans) when comparing Migraneurs with aura to healthy controls. I would like to adjust for age and the fact that i have twin pairs in the Migraneurs group. I've thought about creating a linear mixed effect model, but I lack experience using these models. Can anybody help me on the way? Would it be possible to solve this problem using simple t-tests?
Btw. I'm mainly using matlab and I have several other parameters such as blood pressure, smoking, number of migraine attacks and so forward that I would like to add to the model (next step).
 A: A great review on analysis of twin data can be found in Carlin 2005 https://www.ncbi.nlm.nih.gov/pubmed/16087687


*

*Twins as individuals: ignore the fact that twins are dependent and analyze them as-is. It is a fine approach to data analysis when either a) the sample is representative and happens to contain twins at a rate consistent with the general population or b) the sample is entirely twins but still is representative of the population. This can be enhanced by fitting a GEE. The GEE is a counterpart to mixed models that boasts much better stability when the correlation structure is irregular (such as in the case of having smaller, fewer, or imbalanced clusters such as is often the case in analysis of observational where there is accidentally a smattering of twinpairs). They are also easier to fit. You can specify an exchangeable correlation for intratwin correlation, but it is not necessary.

*Use Cotwin controls: if there are twins who are discordant in terms of their exposure, calculate a paired T-test for the difference in T1 weighted MR outcomes between the Migraneur twin and his or her cotwin control. You run an analysis on a much smaller dataset, but it controls for virtually every possible confounder and gives a much more precise analysis with far better balance.
A: You can indeed use a mixed effect model to account for the correlations within the twins. Typically this will involve just adding a random intercept for the grouping variable that identifies which pairs of subjects are twins.
You wrote that you compare migraneurs to healthy controls. But did you match these controls to the migraneurs using some characteristics, e.g., age and sex? Then you would need also to account for the correlations/clustering that is introduced because of the matching. That is, you could include an extra random intercept term for the grouping variable that identifies which subjects are matched.
The GEE approach versus mixed models:


*

*Mixed models are more efficient (i.e., make better use of the data and provide) than the GEE when you have complex correlations, such as taking into account for both the twins and matching.

*GEEs obtain a better estimate of the within-cluster correlations when you have balanced data, i.e., if you had all twins, and when you include few covariates that are not continuous.

*If your outcome is continuous and you assume a normal distribution, then the GEE and mixed models are almost equivalent.

*If your outcome is categorical and you need to using a nonlinear link function (e.g., the log or the logit), then there is a difference in the interpretation of the coefficients between mixed models and the GEE approach.

*If you have missing data in your outcome, the GEE will provide you correct results under the missing completely at random missing data mechanism whereas the mixed models under the weaker assumption of missing at random.

