Does non-response bias affect the validity of a statistical model? I want to build a (generalised linear mixed) model on some survey data. The PROC GENMOD command in SAS doesn't admit weighting in the sense of survey weights. I am not interested in population-level frequencies, but do want to draw valid conclusions about how covariates interact with each other and the dependent variable in the real world. Does the lack of weights matter?
 A: If you are interested in making inferences on a group larger than your sample, I suspect that the answer is "yes," but how much bias is introduced?
The typical argument is that those who do not answer the survey are categorically different than those who do. Chapter 6 of the textbook Survey Methodology by Groves et al and related papers, partition non-response into three groups:


*

*refusal  

*non-contact or miss (never available)

*inability to respond (eg, language barrier)


If a cause behind one or more of these non-response categories is correlated with your outcome of interest, then there is a danger that your estimates  will be biased.
In some instances, you may be able to control for the non-response causes through modeling or weighting. When this is not a possibility, AAPOR recommends that, as a minimum, you report the response rate (various formulas might be found on their website).
The textbook chapter mentioned above as well as other research papers on this topic have found that well thought out and executed surveys can produce unbiased results even with a fairly high response rate, but these are typically surveys conducted with the great care.
My recommendation is to report the response rate (or a couple response rates based on different formulas) and discuss how this might affect your estimates based on the literature.
