Inference from Biased Polls

Abstract

Poll respondents often attempt to present a positive image by overstating virtuous behaviors. We examine whether people account for this "socially desirable responding" (SDR) when drawing inferences from poll data. In an experiment, we incentivize "predictors" to guess others' choice behaviors for actions with varying social desirability. To aid their guesses, predictors observe random subsamples of either (i) actual choice behavior or (ii) hypothetical claims from unincentivized polls. Predictors show reasonable skepticism towards hypothetical claims, which exhibit predictable SDR. However, their skepticism is not appropriately tailored to the direction or magnitude of SDR. This under-correction for SDR occurs even though subjects can accurately predict which behaviors are most prone to SDR when explicitly asked.