Lin Liu, Harvard University
Room 306, No.5 Science Building
In the era of big data and AI, applications of black-box machine learning in causal analysis will become the rule rather than the exception. Before we have a completely understanding on the theory of black-box machine learning, even the state-of-the-art causal effect estimates – the double machine learning (DRML) estimators – may have bias so large that prohibits valid inference. Invalid inference of causal effect estimates can have severe consequence when causal analysis may eventually change how treatment is prescribed to patients or what kind of policy is made to affect the society. In this talk, we describe a nearly assumption-free procedure which can either detect mis-coverage of the confidence interval associated with the DRML estimators of some causal effect of interest or falsify the certificates (i.e. the mathematical conditions) that, if otherwise to be true, could ensure valid inference. This work shows how higher-order influence functions can be used in modern causal data analysis.