Hi,
I'm working with colleagues on the design of a survey of primary schools. I
came across the idea of designing surveys for balance in the work of Yves
Tille (of the R sampling package fame). Roughly the survey is designed to
provide accurate estimates of known data, in our case things like enrolment,
gender ratios and so on, in the (reasonable) hope that this will improve
estimates of what you actually want the survey to measure. The idea appeals,
but I've hit one major conceptual snag.
We expect non-response. We expect this to be non-random. I know, more or
less, what this does to estimates from srs and pps samples. I can find no
papers on the impact of non-response on balanced samples. I think that it
might be quite severe, but I don't know. Before I embark on a vast
simulation study has anyone got a better idea, or (better again) can anyone
point me to previous work on this question?
Ta,
Anthony Staines
--
Professor of Health Systems Research,
School of Nursing, Dublin City University, Dublin 9,Ireland.
Tel:- +353 1 700 7807. Mobile:- +353 86 606 9713
|