Inferenza con Bias di Selezione Non Ignorabile: Un Approccio con Reti Neurali per i Sondaggi Elettorali Online

Non-probability samples, such as online samples, are ”almost everywhere” in the social sciences. Estimation with such ubiquitous datasets can be influenced by non-ignorable selection bias, which cannot be easily adjusted for. Therefore, methods able to re- duce bias and improve estimation for non-random samples are especially precious. In this contribution, methods making use of additional meta-information parameters in estimation emerge as useful to this aim. Meta-information transmit the belief of the researchers regarding the validity of the selection process into the estimation process. Often, the researcher is a field expert who might have information on the defects and problems of the employed selection methods, especially in the case of online panels. A novel method is presented, directional Rockafellar-Uryasev regression, which can be used successfully to improve the estimation through the addition of meta-information. The present contribution displays advantages over existing methods both in simulated scenarios and in an application in political polling. These advantages are most salient when the employed post-stratification variables are not strongly predictive of the target variable or when the selection mechanism is non-linear. Finally, the present contribution makes recommendations on which method to apply in each case.