Cognitive procedures for correcting proxy‐response biases in surveys

To reduce survey costs, major surveys rely on self‐ and proxy‐responses. The use of proxies can reduce data quality introducing biases in the survey estimates. This paper identifies one source of systematic differences between self‐ and proxy‐reports: proxies' higher reliance on inferences. Using data from the National Health Interview Survey on Disability (NHIS‐D), proxy‐response biases were modelled by independently collected measures of cognitive inferences. Conditional likelihood judgements about a number of disabilities (e.g. likelihood that a person has a disability given another disability) predicted the conditional disability reports for proxy‐ but not for self‐respondents (e.g. the proportion of respondents who reported difficulty learning after reporting difficulty communicating). A model of self/proxy differences was estimated on data from the 1994 NHIS‐D and tested against 1995 data. The correlation between predicted and actual differences was 0.76. The correlation between predicted and actual proxy‐reports was 0.95. Such research can be used to estimate and correct for systematic proxy‐response biases. Copyright © 2002 John Wiley & Sons, Ltd.