I have fit two generalized estimating equation (GEE) models to my data:

1) Model 1: Outcome is longitudinal Yes/No variable (A) (year 1,2,3,4,5) with longitudinal continuous predictor (B) for years 1,2,3,4,5.

2) Model 2: Outcome is the same longitudinal Yes/No variable (A), but now with my predictor fixed at its year 1 value i.e. forced to be time invariant (B).

Due to missing measurements in my longitudinal predictor at a few time points for different cases, the number of data points in model 2 is higher than in model 1.

I would like to know about what comparisons I can validly make between the odds ratios, p-values and fit of the two models e.g.:

  • If the OR for predictor B is bigger in model 1, can I validly say that the association between A and B is stronger in model1?

  • How can I assess which is the better model for my data. am I correct in thinking that QIC/AIC pseudo R squareds should not be compared across models if the number of observations is not the same?

Any help would be greatly appreciated.

  • $\begingroup$ Because Model 2 is not really considered "nested" from Model 1, I don't know how valid using QIC to assess comparative fit would be. One thought I had was to use multiple imputation techniques to equalize the number of observations, and then one could arguably compare the QIC values for those models. However, some literature, e.g. "Applied Longitudinal Data Analysis for Epidemiology" by Twisk, showed really inconsistent results by using MI techniques on models that have dichotomous response variables. I wish I could help more. $\endgroup$
    – Iris Tsui
    Commented May 9, 2011 at 20:01
  • 1
    $\begingroup$ Why are the values missing? Is their missingness systematic in a way that makes missing values fundamentally different from non-missing values? $\endgroup$
    – Macro
    Commented Jun 26, 2012 at 1:10

1 Answer 1


I would definitely try multiple imputation (eg with mice or Amelia in R), possibly with several alternative methods to impute missing values.

In the worst case scenario you can consider it a sensitivity analysis.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.