Modeling data containing both repeated and non-repeated observations

[stumptowner]

Hello,

I’m in the early stages of developing a model to identify patient, provider, and encounter-level factors significantly associated with issuance of a specific type of prescription. Here’s some background:

• The outcome is binary (patient received prescription within 2 days of encounter/did not receive prescription).
• This is a longitudinal dataset with roughly 14000 observations (distinct encounters) that occurred over a 4 year period among ~4350 patients.
• The outcome is positive (patient received a prescription) in about 15% of encounters.
• For patients with multiple encounters, time between encounters is not evenly spaced.
• I anticipate developing a GLMM model using PROC GLIMMIX to account for repeated measures among patients.
• ~40% of patients had only 1 encounter in the study period, ~20% had 2 encounters, ~10% had 3 encounters, ~7% had 4 encounters, and the remaining  patients had between 5 and 60 encounters (patient counts get into single digits at 17 encounters and above).

It is this last point (the distribution of repeat encounters) that I’m wondering about. That is, I’m wondering if the high proportion of patients with only a single encounter will be an issue in a model designed for repeated observations. I’m also wondering if the skewed distribution in repeat observations might cause problems. To date, I’ve been unable to find any examples/commentary on modelling data with such characteristics.

This may be a non-issue. However, I thought I’d ask before I started building the model in earnest, so I can avoid any pitfalls at the outset. Any thoughts/advice/comments are appreciated.

Thanks,

Stuart

[SteveDenham]

My first thought was to fit the repeated measures covariance structure with a spline, but that may give those single digit encounters on the right tail too much emphasis.  Perhaps binning into some reasonable sizes would help, unless it becomes obvious that those patients who are doctor shopping until they get a response are in this category.  To me the substantive question isn't whether a prescription is obtained on the 12th or the 40th encounter--those are all the same.

The other thing is to maybe turn this "inside out" and deal with it as a time to event analysis, with those that never receive a prescription viewed as censored after their last encounter.  If you have a hierarchical model, though, this may not be so easily done.

Steve Denham

[stumptowner]

Thanks for your input Steve (and apologies for the tardy reply).

After consulting with colleagues, I ended up just using a pretty straightforward GEE model using PROC GENMOD. The model was not at all sensitive to the repeated measures covariance structure, I'm guessing due to the large number of observations compared to the number of repeated measures.

Thanks again,

Stuart