|
Abstract:
|
Survey data users frequently invoke the assumption of uncorrelated error terms. Scant research, however, has addressed its validity in surveys of human populations. The few existent studies have yielded two dominant findings: correlated errors are 1) most common between items pertaining to a shared topic and 2) not well predicted by demographic variables. This paper uses data from a validation study featuring randomized mode assignment (IVR, CATI, Web) to bolster the literature on this topic in two ways. First, we determine whether the previous findings replicate in a contemporary survey, noting that the survey environment has evolved in numerous ways during the past 20 years. Second, prior studies relied exclusively on data from in-person interviews, but we are able to test for interactions between the newer modes of data collection and the nature and prevalence of correlated errors.
|
