I am interested in developing a regression model. The data (about 1000 observations) that I have are not really random (i.e., spatially, the closest ones are correlated to some degree), so I drew a stratified random sample from my dataset at a 50% sampling rate and I ran a regression model using the random sample. I tried two additional sampling rates of 60% and 75% and fitted regression models with the same predictors as earlier. I see that some predictors are now significant with a larger sample. My problem is which sampling rate to select? I do not want to bias the analysis by selecting a sampling rate that gives me significance. Can members suggest any appropriate approach to my problem? Would it be better to run the analysis with the three sampling rates and then use the models developed to validate another set of independent data, and select the model that gives the best prediction.

  • $\begingroup$ Can you please give a bit more details about how you model your data. (Possibly a line or two of code will also help). I am somewhat perplexed as why you do no define, at least at first instance, a simple AR(1) error structure and then see where this takes you. $\endgroup$ – usεr11852 Jan 28 '16 at 22:05
  • $\begingroup$ I tried the ar(1) error structure, but in SAS an error pops up: "R-side random effects are not supported for the multinomial distribution" (my response variable is ordinal). The data are from several different localities and within each locality responses are expected to be correlated. The model that I am using is a random-intercept GLMM with random effect term being the locality. proc glimmix; class loc; model resp = pred/dist=multinomial link=cumlogit; /* random _residual_/subject=id type=ar(1); ERROR */ random int/subject=locality; run; $\endgroup$ – Kris Jan 29 '16 at 15:11
  • $\begingroup$ I am sorry I do not know/own SAS. $\endgroup$ – usεr11852 Jan 29 '16 at 18:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.