Proxy respondents in a case-control study: validity, reliability and impact

Abstract

In many research settings, the study subject is unable to provide responses, requiring researchers to find alternative respondents. This thesis explores the use of proxy respondents in a case control study of premature mortality among Russian men of working age (25-54 years). Data obtained from proxy respondents is explored in four ways. Firstly, proxy questionnaire responses are validated against external data sources which were routinely collected, blind to case-control status - the city alcohol treatment clinic (Narcology Dispensary), Social Security and Police records. Secondly, agreement between proxy and index (control) responses to questions about alcohol use, tobacco use, health and socioeconomic factors is explored. Thirdly, the effect of proxy type is explored by examination of proxy-proxy and index-proxy agreement in a subset of households in which two proxy interviews were obtained. Finally, the impact on analysis outcomes is explored by mortality analyses using proxy versus index data. Cohen's kappa coefficient was used to explore inter-respondent agreement. Differences in agreement between pairs of respondents were examined using Agresti's loglinear model, and the directionality of disagreements were evaluated using McNemar's test. Findings confirm some assertions in the literature. Questions about easily observable characteristics and behaviours, avoiding excessive detail, subjective or sensitive topics, elicit valid proxy responses. Proxies tend to over-report alcohol use, but provide particularly valid responses about tobacco use and socioeconomic factors. Validity was highest among proxies who were the index's spouse. However, further exploration suggested that men who have spouses differ in their behaviour from men who do not in ways which affect its reporting. There was little additional evidence that proxy characteristics affect validity. The use of proxy responses biases odds ratios in this case control study toward more conservative estimates. These findings are generalisable to study settings which employ a protocol to ensure selection of the best available proxy.