I am working with 3 waves of survey data. Many people have been surveyed multiple times, but many others are present only in 1 or 2 waves due to drop-off/resampling. Therefore, my design is multiple cross-sectional taking into account clusters of respondents across time.
My outcome is access to a specific service over the previous year (count variable, not very skewed, it goes from 0 to 16). I have a set of covariates (age, sex, income, education, etc), my main variable of interest is 'year of survery'.
Using the survey weigts I computer the standardized prevalence over the three survey years. However, i want to model the prevalence variation over time (controlling for covariates), therefore, I fitted a Poisson panel regression model using the cout outcome as dependent variable, time - as continuous variable - as main variable of interest and controlling for covariates. I accounted for clustering when computing the SE.
The IRR for time is 0.94 0.93-0.95
I have couple of questions:
what does the IRR mean in this context? In my case I am modelling change of a count variable, without setting offset/expected ratio as option (i am not using rates). Respondents might drop off, others recruited later. To me it seems that I am modeling the change in prevalence ratio, do you agree?
is it correct in this case to use survey year as continuous variable? I believe that the IRR in this case gives me the slope (the yearly change - considering the the survey is taken every year). I wouldnt consider to use survey year as categorical but just double checking.
shall i apply the survey weights (cross-sectional) when running the Poisson regression model? Results are almost identical but I just want to be accurate. I know there is no consensus but I would like to hear more