Bayesian Pooling of Self- and Peer-Reports to Improve Measurement of Sensitive Behaviors

• In Event: New Approaches in Social Network Research: Theory, Methods, and Applications - Open Call Paper Session

Abstract

It is well known that self-reports, especially those of sensitive behaviors, tend to be biased. For example, respondents may over-report their family origin, voting activities, and volunteering experiences while under-reporting their drinking and smoking behaviors. In this paper I examine a network-based approach to addressing self-reporting bias, which is to ask peers (e.g. classmates or co-workers) to provide additional, alternative reports that can be used to cross-validate and supplement self-reports. In this paper, I argue it is important to understand there might be reporting bias in both self-reports and peer-reports and to provide a more accurate correction of self-reporting bias, it will be more fruitful to combine both self-reports and peer-reports and to consider the characteristics of both the egos and the peers and even the dyadic characteristics between them.

Specifically, I first formalize the data generation process statistically. I propose a model containing two sub-models, one for self-reports and the other for peer- reports. In each of the sub-models, I treat the observed report as a probabilistic realization of a truthful report and a false report. The model can incorporate both self- and peer-characteristics and also their dyadic characteristics in the formation of truthful reports and false reports. However, the model depends on knowing an ego’s behavior, which is unknown unfortunately. A Bayesian approach is useful here in that it solves this problem by imposing a prior on the ego’s behavior (i.e., making a guess about it). Then the Bayesian approach simulates self- and peer-reports according to the data generation process iteratively until the simulated reports emulate the observed self- and peer-reports. As a result, the Bayesian approach can show how multiple characteristics are related to reporting bias and incorporate them effectively in imputations of the behavior of interest. The method is illustrated with a case study.

Author

• Weihua An, Emory University