A Unified Framework for Dynamic Causal Inference

• In Event: Methodological Pluralism

Abstract

The growing interest in causal inference has been one of the most noteworthy trends in political methodology over the last twenty years. Causal inference encompasses a wide array of estimation strategies but there are important commonalities among the various methods. Most approaches reduce the causal inference problem to the evaluation of a single-shot treatment on an outcome of interest at a single point in time. Through clever data manipulation and design strategies, the analyst attempts to recover an exact causal effect of a treatment on an outcome, as opposed to satisficing for mere evidence of a treatment effect. While this turn in political methodology has produced a number of important insights, it is not clear that the standard framework for causal inference is well suited to the empirical problems we face today.
While many of the popular data strategies deployed for causal inference tend to ignore temporal variation, time is one of the most important dimensions of the new social science data. The big data revolution is producing large volumes of data collected from commercial transactions, social media platforms, and continuous monitoring sensors. This has produced a veritable explosion of high-resolution time series data. As access to longitudinal social media, text, and event data continues to progress; the continued utility of single-shot causal identification strategies is more and more in question.
This chapter elucidates a general framework for dynamic causal inference. This endeavor has three components. First, innovation in this space has been hindered by differences in the complex nomenclature used in both time series analysis and causal inference. Some concepts, like Granger causality and causal identification, seem to be related when they are not. Some concepts, like weak exogeneity and exchangeability, are related in important ways that are not readily obvious. For scholars that have spent more time and energy focusing on either time series problems or causal inference problems, these differences in nomenclature represent important barriers to progress and for political scientists that are unfamiliar with both of these areas of research, these nomenclatures represent significant barriers to entry. Our first goal is to resolve confusion around critical concepts in an attempt to lay a foundation for causal inquiry.
Our second task is to outline the contours of the framework. Most of this work has already been done for us. Time series econometricians have been wrestling with the problems created by dynamic systems for decades. The novel element of this exposition is our effort to place existing causal identification strategies in their appropriate context. Most human phenomena of interest exist is dynamic systems where they cause and are caused by other human phenomena. In some circumstances, causal sequences within these systems can be identified. Single shot causal identification strategies can be effectively deployed in some of these contexts. In many circumstances, however, these causal identification strategies are inappropriate and efforts to apply these methods to dynamic systems produce misleading inferences. The framework we present describes when and how these dynamic systems can be restricted for causal identification, outlines the limits of traditional approaches to causal inference, and highlights new estimands of interest.
Our final task is to prescribe existing methods for causal inference within this framework. Where sequential cuts are plausible, many of the popular tools for causal inference can be applied. Where they are not, time series analysts have developed an alternative class of inferential strategies which can be theoretically informative. Tests for Granger causality, Sims causality, Structural causality, and other procedures that have been developed for policy analysis provide a set of tools that can be deployed in these contexts. Some facilitate what proponents of causal identification would call causal inference and some do not. The inferential limitations of some of these tools do not reflect an indolence among time series analysts or an ignorance of the importance of causal identification, but an awareness that some causal relationships cannot be easily reduced to simple cause and effect connections. Understanding which methods can be applied in which circumstances is critical to ensuring that the causal inference movement does not come to a sudden halt as it begins to come to terms with new challenges presented by the ever-growing volumes of dynamic data.

Authors

• Suzanna Linn, Pennsylvania State University

• Clayton McLaughlin Webb, University of Kansas