Causal Inference and Data Fusion in Econometrics (1912.09104v4)

Abstract: Learning about cause and effect is arguably the main goal in applied econometrics. In practice, the validity of these causal inferences is contingent on a number of critical assumptions regarding the type of data that has been collected and the substantive knowledge that is available. For instance, unobserved confounding factors threaten the internal validity of estimates, data availability is often limited to non-random, selection-biased samples, causal effects need to be learned from surrogate experiments with imperfect compliance, and causal knowledge has to be extrapolated across structurally heterogeneous populations. A powerful causal inference framework is required to tackle these challenges, which plague most data analysis to varying degrees. Building on the structural approach to causality introduced by Haavelmo (1943) and the graph-theoretic framework proposed by Pearl (1995), the AI literature has developed a wide array of techniques for causal learning that allow to leverage information from various imperfect, heterogeneous, and biased data sources (Bareinboim and Pearl, 2016). In this paper, we discuss recent advances in this literature that have the potential to contribute to econometric methodology along three dimensions. First, they provide a unified and comprehensive framework for causal inference, in which the aforementioned problems can be addressed in full generality. Second, due to their origin in AI, they come together with sound, efficient, and complete algorithmic criteria for automatization of the corresponding identification task. And third, because of the nonparametric description of structural models that graph-theoretic approaches build on, they combine the strengths of both structural econometrics as well as the potential outcomes framework, and thus offer an effective middle ground between these two literature streams.

Overview of "Causal Inference and Data Fusion in Econometrics"

The paper "Causal Inference and Data Fusion in Econometrics" by Paul Hünermund and Elias Bareinboim provides a comprehensive review of the application of causal inference techniques, which have been developed in the field of artificial intelligence, within the field of econometrics. The primary objective is to extend the methodological repertoire available to econometric practitioners by integrating graph-theoretic causal inference tools. These tools promise to address the challenges of confounding bias, sample selection bias, and external validity, which frequently impair causal inference in empirical economic research.

Key Contributions

The authors delineate the theoretical advancements that enable the automation of the identification task—a crucial step in causal inference wherein causal queries are transformed into expressions that can be estimated from observable quantities. They argue that leveraging directed acyclic graphs (DAGs) allows researchers to articulate the assumptions underlying their causal models clearly and check the plausibility of critical identifying conditions, such as conditional independence, using graphical criteria.

1. Treatment of Confounding Bias:
   • The paper discusses the utilization of graphical models to identify backdoor and frontdoor adjustment strategies, addressing confounding bias comprehensively. The implementation of do-calculus offers a robust mechanism to transform causal queries into do-free expressions, thus facilitating the estimation of causal effects even when perfect randomization is unavailable.
2. Sample Selection Bias:
   • Sample selection bias is explored through the introduction of selection nodes that explicate the sampling mechanism. The authors provide a complete condition for the recoverability of conditional probabilities and demonstrate how causal effects can be estimated from a mixture of biased and unbiased datasets, thus extending the analytical horizon of researchers in settings fraught with selection issues.
3. Transportability:
   • The paper tackles the challenge of extrapolating causal knowledge across different populations. By introducing selection diagrams, the authors provide a formal framework to determine the transportability of causal effects, allowing researchers to project findings from experimental settings to new domains in a principled manner.
4. Integration with Surrogate Experiments:
   • An extension to causal inference is presented through the concept of $\mathpzc{z}$-identification and $\mathpzc{z}$-transportability, which facilitate the transfer of causal effects from surrogate experiments. This addresses scenarios where direct manipulation of a treatment is infeasible, thus broadening the applicability of experimental findings.
5. Meta-Transportability:
   • The discussion on meta-transportability explores the synthesis of causal knowledge from multiple heterogeneous sources. This framework empowers researchers to leverage a broad empirical base, exploiting shared mechanisms across different studies to ascertain causal effects more effectively than with isolated datasets.

Implications and Future Directions

The frameworks introduced in the paper have significant implications for empirical economics. The graph-theoretic approaches allow for greater transparency in causal modeling and provide formal solutions to many identification problems that have traditionally been addressed through parametric assumptions. This work paves the way for more robust and efficient causal inference techniques in econometrics, fundamentally grounded in the structural underpinnings of the data generating processes economists wish to understand.

Furthermore, the integration of artificial intelligence methodologies into econometric practice holds promise for future developments in automated causal inference. As these methods become more widespread, there is potential for more collaboration between AI researchers and economists to refine these techniques, ensure their practical applicability, and explore new territories in causal data fusion, which could be particularly beneficial in complex economic models and large-scale data integration scenarios.