Diehr P, Koepsell T D, Cheadle A, Psaty B M
Department of Biostatistics, University of Washington, Seattle 98195.
Stat Med. 1992 Jun 15;11(8):1009-21. doi: 10.1002/sim.4780110803.
Knowledge of the characteristics of survey non-respondents is important to determine generalizability to the population of interest. In a recent random-digit dialling survey of health behaviours only 73 per cent of the households contacted provided any information about household composition, and only 74 per cent of those actually completed the extended interview, for an overall response rate of 54 per cent. To identify possible biases we grouped all attempted phone numbers by their prefix, and looked for the association between the response rate for that prefix and other summary variables known about the prefix. A simulation study showed that the method can identify non-response biases if certain assumptions are correct. The analysis suggested that our survey data under-represent older people and those with a college education. We found no significant biases in health behaviours, possibly because the basic assumptions did not hold. This method may assist in identification of non-response bias in other studies.
