Web Survey Bibliography
pdf (120 k)
Although non‐probability samples have been a part of statistical analysis from the beginning, there appear to have been only a handful of meaningful attempts to combine probability and non‐probability samples (Yoshimura, 2004). There are three likely reasons for the use of data from a non‐probability sample when a probability sample is available. First, only the non‐probability sample may contain the detailed outcomes of interest. Second, the non‐probability sample may be substantially larger than the probability sample, allowing the possibility of substantially improved estimators if the increase in precision is not overwhelmed by bias from the non‐probability sample. Finally, analysts will likely continue to use non-probability samples in lieu of probability samples in many settings. Non‐probability samples are likely increasing as Web surveys become increasingly entrenched in market research and other settings. Survey methodologists arguably should propose methods that can improve the quality of analyses obtained from these datasets, at least under clearly specified assumptions. This paper proposes a method to construct "pseudoweights" for a non‐probability sample that uses available data in both a probability and a nonprobability sample to estimate probabilities of selection for the non-probability sample, had it actually been sampled via a randomized mechanism.
Journal Homepage (abstract)/(full text)