Hi
I am looking for an implementation of a time-varying logistic regression, actually for time varying coefficients, to explore the changes of the odds ratio over time.
I found a R package called dynamichazard that claims to implement such a model, however it is based on specific modelling and the code doesn’t run on my data anyway.
Of course I can try a naïve model in which I divide the time to short time frames, and then run a logistic model for the first week, the for the first two weeks and so on, but I still hope to find something more sophisticated.
And one more note: a time-varying cox PH model may be used as a sensitivity analysis, but I am not able to use it as a primary analysis (please don’t ask)
I will appreciate any advice.
A GEE approach with a cluster sandwich covariance estimator allows for much flexibility in that setting, even for overlapping time intervals .
1 Like
Are you thinking like you would run separate models here? Because otherwise this sounds like a pooled logistic model which would approximate cox.
Exactly, but I don’t like this approach
This would be pooled logistic regression, using a single model fit.
Yossi, can you describe what your data look like? Are you working in continuous time or discrete time?
Hi David
I work in discrete time (weeks). It’s a simple setup: I have a (large) cohort of patients, an intervention, a rare outcome, and a bunch of baseline covariates. For every subject I have the time of the intervention (if the intervention was administered), and the time of the outcome (if it occurred), so there 4 combinations. If the outcome occurred before an intervention was administered, then intervention is no longer possible. I thought that a time-varying PH model will do fine, but for reasons I don’t want to discuss this is not an option. After some research I found a paper by Tan, Shiyko, Li, Li & Dierker (2012) with an implementation in the tvem package in R, and got nice results. But of course I’ll be glad to hear about other approaches.