Abstract

Bridging probability and nonprobability samples: inference and data integration under nonstandard conditions (P.L. Conti, D. Marella)

A major concern with non-probability samples (web-scraped data, online panels, voluntary web surveys) is selection bias: since the selection mechanism is unknown and uncontrolled, inclusion probabilities are not available and cannot be used to weight observations and to remove bias of population parameters estimates. In this talk we first discuss the uncertainty in estimating a discrete probability distribution arising from the non-identifiability of the selection mechanism, and show how it can be reduced by auxiliary information external to the sample. Next, due to the lack of identifiability, the concept of coherent estimator is introduced, an inverse probability weighting estimator is proposed, and its uniqueness in terms of coherence is discussed. Finally, estimators with asymptotically minimax bias are also discussed, as well as applications to the estimation of the population mean.