RESEARCH

The widely employed regression discontinuity (RD) causal inference method is intended for scenarios where there is a known cutoff of a running variable such that the probability of receiving treatment changes abruptly at this cutoff. Examples of such settings are countless. The key assumption underlying these analyses is that outcomes of the untreated units just left to the cutoff are good representatives of the counterfactual outcomes of the treated units just right to the cutoff had they been untreated.

However, this assumption is often implausible when changes other than the intervention of interest occur at the cutoff (for example, other treatments or policies are implemented at the same cutoff). In these scenarios, researchers retreat to ad-hoc analyses that are not justified by any theory and yield results with unclear interpretations. These analyses seek to exploit additional data for which no intervention was applied for all units (regardless of their running variable value). This could be the case when data from multiple time periods and/or multiple comparison groups are available.

In this project, we (1) Develop formal theory and statistical methods for studying causal effects with generalized RD designs utilizing multiple time periods or comparison groups data and (2) Implement the proposed methods to real data and provide practical guidance on how to conduct RD analysis with multiple time points and/or multiple comparison groups.

Instrumental variable (IV) estimation is a widely used method that supports high stakes government policies and business decisions, when controlled experiments are not feasible or when a treatment is not successfully delivered to every unit in a randomized experiment.

In this research project, we will develop statistical methods for assessing the validity of such designs. IV designs rely on exclusion restriction assumptions that are not directly testable and that are therefore challenging to assess empirically. In this context, this project aims to formalize the logic of falsification tests, a set of tests that leverage contextual knowledge about the absence of causal links (for example, from future to past outcomes) to test the validity of candidate instruments.

We will establish that IV falsification tests can be mapped to a class of prediction problems that can leverage current machine learning methods. Based on this conceptualization, we will develop general methods that would both improve falsification test efficiency and help guide the construction of such tests.

These methods will be particularly applicable to research using large datasets with many candidate variables that can be used for falsification, an increasingly common situation for which no formal methods currently exist. Developing methods for evaluating and improving the validity of these common research designs entails clear societal benefits.

In the era of Big Data, exploratory data analysis (EDA) is gaining great importance. Yet, EDA is still a difficult process, especially for non-expert users, as it requires deep understanding of the investigated domain and the particular context. Consequently, users may skip significant analysis actions and overlook important aspects of the data. In this research, we aim to set the ground for employing Deep Reinforcement Learning (DRL) techniques in the context of EDA. We suggest an end-to-end framework architecture coupled with an implementation of each component. Our goal is the development of DRL models and techniques for facilitating a full-fledged, autonomous solution for EDA.