Mixed models for longitudinal data

[Erika Gonzalez]

Dear all,
I have a dataset with 93 patients (variable CETIRID) who had a blood sample taken at two time points: on hospital arrival and after admission to the intensive care unit (variable time), thus I have 186 observations. What I am looking to answer is if there is any effect of total blood products (variable totalblood) transfused in the first 24-hours of hospital admission, and the change from one time point to the other in my biomarker of interest (variable Syn).

For these, I have to also include variables that might have an effect on the dependent variable such as SBP, ISS, Age, BE. I want to include a random effect: CETIRID

My idea was to use the following model:


mixed Syn totalblood time AGE SBP BE ISS time#totalblood || CETIRID: , residuals(unstructured, t(time)) ml

I am not sure though if this model and approach are appropriate to test this question.
I am using Stata 13.

Thank you in advance for your help.

Best regards,
E Gonzalez

[Joseph Coveney]

Why are you using the -residuals(unstructured)- option when you have only two time points? Do you expect heteroskedasticity? I don't know how values of your biomarker typically distribute themselves, but you might want to take a look at the residuals after fitting the model and make a decision about whether to use a generalized linear mixed model instead of a linear mixed model. (I don't recall when -meglm- was introduced, but if it's after Release 13, then you could either use the user-written -gllamm- or upgrade.)

You would typically set up a repeated-measures ANOVA / MANOVA using the -residuals()- option in conjunction with the -noconstant- option in the random effects equation.

I recommend being explicit about the factor variables, e.g., c.totalblood##i.time c.(age sbp) etc.

With only 93 patients, a test of the total blood products × time interaction term will have an inflated Type I error rate, but unless patients were randomly assigned to levels of total blood products you've got more important things to worry about.

[Joseph Coveney]

Oh, in addition to -gllamm- (or upgrading), you could also look into -xtgee-.

[Erika Gonzalez]

Joseph, thank you greatly for your prompt response. My data is unbalanced for some of the variables such as c.SBP (which has observations for each time point) and c.BE. Patients were not randomly assigned to levels of blood products, they received as much as they needed.
I am looking to find the model that fits my data best. I decided to try xtgee:

xtgee Syn c.AGE c.BE c.ISS c.totalblood c.totalblood#i.time i.time , family(gaussian) link(identity) corr(unstructured)

Is this code correct?
In addition, if I want to compare models to look for the best model, is it appropriate to use qic?

Thank you in advance for your help.