Abstract:
|
Robust causal inference with Instrumental Variables (IVs) is a powerful tool in observational settings, but is often limited to estimating the effect of intervening on all subjects at a specific level vs. on none. In many cases, this is not the effect that is most interesting to subject matter experts, or it is inappropriate given the data. In this paper, we develop nonparametric instrumental variables methods, which extend modern influence function-based doubly robust IV methods to more general settings by investigating the effects of arbitrary interventions on continuous instruments. These do not require the strong parametric modeling assumptions of other methods, although like all causal inference methods they require strong assumptions about the causal structure of the data. This new methodology was motivated by the need to investigate the effects of visitation on recidivism. We apply one version of these new estimators to this problem for prisoners released from Pennsylvania prisons in 2008, and find that policies which facilitate visitation may lead to a decrease in recidivism.
|