I am using the PDL model to analyze the correlation between 8-days insect abundances and weather conditions. I also want to adjust for seasonality and first order autoregressive term in the model. Here is my models.
DATA SF_MODEL; SET PDL.SF_ALL_F;
PROC PDLREG data=SF_MODEL;
MODEL tarsalis=tavg(4,3) rh(4,3) prec(4,3) sea / nlag=1 partial;
My questions is:
1. Is it right to use "sea" to adjust the seasonality in the model? (46 is the total observation numbers of each year).
2. Does the statement "nlag=1" in the model option indicate that adjustment of the first order autoregressive term?
You mentioned 8 days, but your model is for 4 lags (time window of 5: current and previous 4). Is that what you want. You specified a cubic polynomial for the parameter constraints, which is just one degree less than the maximum (i.e., with no constraints). I think you are not gaining the advantages of PDL regression. With only 46 observations, I think you have an over-parameterized model. I have used PDL regression a lot, and you need a lot of data points to reliably estimate the parameters and determine the time length and polynomial order.
Your nlag option is fine for a 1-st order AR residual term.
Your seasonality term is probably OK, but it doesn't give you much flexibility. You might have to look into using a smoothing spline, but this can't be done within PDLREG. You would have to do this in another procedure, store the residuals, and then use PDLREG on the residuals. I suggest you check out the article by Schwartz (Epidemiology 11: 320-326 ). Also check out Madden and Paul (Phytopathology 100: 1015-1029 ).