I'm modeling longitudinal patient level health records data which will have the following structure:

\begin{equation*} \text{Prob}(Y_{it}=1 \text{ at time } t \text{ and member } i)=\beta_{i0} +\beta_{i1}x_{i,t-1} + \beta_{i2}x_{i,t-2} + ... + \beta_{ik}x_{i,t-k} + \text{additional covariates for patient age, diagnoses, etc.} \end{equation*}

where Y is an inpatient hospitalization indicator variable for time period t and patient i, and the x's measure frequency of home nursing visits during prior time periods.

The predictor variables are significantly correlated at the patient level. I'm considering a logistic regression, however the strong correlation between the x's makes things a bit more complicated.

Is a time series model appropriate here, or would a multilevel repeated measures model be better (say using proc mixed in SAS with a random intercept for each member)?

  • $\begingroup$ It seems like the random effects model is more practicable here and is probably a more appropriate choice. Whether or not the collinearity presents a problem depends on what your goals are with the analysis. $\endgroup$ – Macro Feb 22 '13 at 18:19
  • $\begingroup$ Thank you, I've decided on a random effects model as well. I reasoned frequency of visits are similar more due to member (or possibly provider) level characteristics rather than freq. of visits on day t actually being influenced by day t-1. $\endgroup$ – RobertF Feb 25 '13 at 15:45

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Browse other questions tagged or ask your own question.