R McDowell posted on Thursday, June 11, 2015 - 9:15 am
I have a dataset where the primary outcome is a count variable, and the length of follow-up time has been measured. I have a confounding variable I would like to include in the model which is time-varying. I am aware of fitting time-varying covariates with longitudinal models and in survival analysis, and was interested to know if it was possible or indeed correct to fit time-varying covariates in a Poisson model in Mplus, and if so under what conditions. Any advice would be appreciated!
Is your count outcome measured at several time points for each subject? Such as at the same times as the time-varying covariate?
R McDowell posted on Thursday, June 11, 2015 - 1:10 pm
The count outcome is number of deaths over the followup period-as such each person has either or 0 or 1 deaths. A typical exposure might be a form of medication-we know when they were on/off the medication, and hence the total time they were on/total time off the medication.
How is the outcome a count variable if it takes on only 0 or 1 values?
R McDowell posted on Friday, June 12, 2015 - 3:28 am
I apologise, that was a poor example. Consider an example where the outcome is number of doctor visits over a year, each of which has been recorded per month. The clinical context is such that we don't need to worry about lots of subjects having zero visits over the year, and everyone was monitored for the full year. There is an exposure variable, medication, detailing how much of the exposure each participant had per month. The primary outcome is the total number of visits the person had over 12 months. If I include whether the person ever had the medication over the 12 months as a predictor, I am assuming the effect is constant, whereas I know people commonly go on/off the medication of interest. Some form of longitudinal/growth model could accomodate this modelling situation I believe. My thinking is that if the effect of exposure varies per month, and the outcome is the total number of visits over 12 months, to break the effects of the exposure up by month say, and fit these effects separately, at least in the first instance, as an appropriate way of accomodating a time varying exposure with this particular outcome. I don't think this is necessarily a good way to do the analysis as a whole, however I wondered if in principle, this would be how one would go about it.