Deep Learning of Potential Outcomes

This review systematizes the emerging literature for causal inference using deep neural networks under the potential outcomes framework. It provides an intuitive introduction on how deep learning can be used to estimate/predict heterogeneous treatment effects and extend causal inference to settings where confounding is non-linear, time varying, or encoded in text, networks, and images. To maximize accessibility, we also introduce prerequisite concepts from causal inference and deep learning. The survey differs from other treatments of deep learning and causal inference in its sharp focus on observational causal estimation, its extended exposition of key algorithms, and its detailed tutorials for implementing, training, and selecting among deep estimators in Tensorflow 2 available at github.com/kochbj/Deep-Learning-for-Causal-Inference.

Experiments in IPE Research

Over the last decade, experiments have developed from a marginally employed to an increasingly standard method in the study of International Political Economy (IPE). After a short discussion of causal inference and the potential outcomes framework, this chapter outlines different kinds of experiments used to study questions concerning trade, migration, foreign aid, or investment. The chapter thereby not only strives to provide an overview of the state-of-the-art in experimental IPE research but also to outline the different roles or functions experimental studies can fulfill to further our knowledge on IPE. The chapter concludes by critically discussing both advantages and disadvantages of the experimental turn in IPE.

Emulating Target Trials to Improve Causal Inference from Agent-Based Models

Agent-based models are a key tool for investigating the emergent properties of population health settings, such as infectious disease transmission, where the exposure often violates the key ‘no interference’ assumption of traditional causal inference under the potential outcomes framework. Agent-based models and other simulation-based modeling approaches have generally been viewed as a separate knowledge-generating paradigm from the potential outcomes framework, but this can lead to confusion about how to interpret the results of these models in real-world settings. By explicitly incorporating the target trial framework into the development of an agent-based or other simulation model, we can clarify the causal parameters of interest, as well as make explicit the assumptions required for valid causal effect estimation within or between populations. In this paper, we describe the use of the target trial framework for designing agent-based models when the goal is estimation of causal effects in the presence of interference, or spillover.

Causal mediation in developmental science: A primer

Mediation has played a critical role in developmental theory and research. Yet, developmentalists rarely discuss the methodological challenges of establishing causality in mediation analysis or potential strategies to improve the identification of causal mediation effects. In this article, we discuss the potential outcomes framework from statistics as a means for highlighting several fundamental challenges of establishing causality in mediation analysis, including the difficulty of meeting the key assumption of sequential ignorability, even in experimental studies. We argue that this framework—which, although commonplace in other fields, has not yet been taken up in developmental science—can inform solutions to these challenges. Based on the framework, we offer a series of recommendations for improving causal inference in mediation analysis, including an overview of best practices in both study design and analysis, as well as resources for conducting analysis. In doing so, our overall objective in this article is to support the use of rigorous methods for understanding questions of mechanism in developmental science.

Panel experiments and dynamic causal effects: A finite population perspective

In panel experiments, we randomly assign units to different interventions, measuring their outcomes, and repeating the procedure in several periods. Using the potential outcomes framework, we define finite population dynamic causal effects that capture the relative effectiveness of alternative treatment paths. For a rich class of dynamic causal effects, we provide a nonparametric estimator that is unbiased over the randomization distribution and derive its finite population limiting distribution as either the sample size or the duration of the experiment increases. We develop two methods for inference: a conservative test for weak null hypotheses and an exact randomization test for sharp null hypotheses. We further analyze the finite population probability limit of linear fixed effects estimators. These commonly‐used estimators do not recover a causally interpretable estimand if there are dynamic causal effects and serial correlation in the assignments, highlighting the value of our proposed estimator.

Assessing causality in epidemiology: revisiting Bradford Hill to incorporate developments in causal thinking

The nine Bradford Hill (BH) viewpoints (sometimes referred to as criteria) are commonly used to assess causality within epidemiology. However, causal thinking has since developed, with three of the most prominent approaches implicitly or explicitly building on the potential outcomes framework: directed acyclic graphs (DAGs), sufficient-component cause models (SCC models, also referred to as ‘causal pies’) and the grading of recommendations, assessment, development and evaluation (GRADE) methodology. This paper explores how these approaches relate to BH’s viewpoints and considers implications for improving causal assessment. We mapped the three approaches above against each BH viewpoint. We found overlap across the approaches and BH viewpoints, underscoring BH viewpoints’ enduring importance. Mapping the approaches helped elucidate the theoretical underpinning of each viewpoint and articulate the conditions when the viewpoint would be relevant. Our comparisons identified commonality on four viewpoints: strength of association (including analysis of plausible confounding); temporality; plausibility (encoded by DAGs or SCC models to articulate mediation and interaction, respectively); and experiments (including implications of study design on exchangeability). Consistency may be more usefully operationalised by considering an effect size’s transportability to a different population or unexplained inconsistency in effect sizes (statistical heterogeneity). Because specificity rarely occurs, falsification exposures or outcomes (i.e., negative controls) may be more useful. The presence of a dose-response relationship may be less than widely perceived as it can easily arise from confounding. We found limited utility for coherence and analogy. This study highlights a need for greater clarity on BH viewpoints to improve causal assessment.

Population attributable fractions for continuously distributed exposures

When estimating population attributable fractions (PAF), it is common to partition a naturally continuous exposure into a categorical risk factor. While prior risk factor categorization can help estimation and interpretation, it can result in underestimation of the disease burden attributable to the exposure as well as biased comparisons across different exposures and risk factors. Here, we propose sensible PAF estimands for continuous exposures under a potential outcomes framework. In contrast to previous approaches, we incorporate estimation of the minimum risk exposure value (MREV) into our procedures. While for exposures such as tobacco usage, a sensible value of the MREV is known, often it is unknown and needs to be estimated. Second, in the setting that the MREV value is an extreme-value of the exposure lying in the distributional tail, we argue that the natural estimator of PAF may be both statistically biased and highly volatile; instead, we consider a family of modified PAFs which include the natural estimate of PAF as a limit. A graphical comparison of this set of modified PAF for differing risk factors may be a better way to rank risk factors as intervention targets, compared to the standard PAF calculation. Finally, we analyse the bias that may ensue from prior risk factor categorization, examining whether categorization is ever a good idea, and suggest interpretations of categorized-estimands within a causal inference setting.

Disseminated Effects in Agent Based Models: A Potential Outcomes Framework and Application to Inform Pre-Exposure Prophylaxis Coverage Levels for HIV Prevention

Pre-exposure prophylaxis (PrEP) for HIV prevention may not only benefit the individual who uses it, but also their uninfected sexual risk contacts. We developed an agent-based model using a novel trial emulation approach to quantify disseminated effects of PrEP use among men who have sex with men in Atlanta, USA from 2015 to 2017. Components (subsets of agents connected through partnerships in a sexual network, but not sharing partnerships with any other agents) were first randomized to an intervention coverage level or control, then within intervention components, eligible agents were randomized to PrEP. We estimated direct and disseminated (indirect) effects using randomization-based estimators and reported corresponding 95% simulation intervals across scenarios ranging from 10% to 90% coverage in the intervention components. A population of 11,245 agents was simulated with an average of 1,551 components identified. Comparing agents randomized to PrEP in 70% coverage components to control agents, there was a 15% disseminated risk reduction in HIV incidence (95% simulation intervals = 0.65, 1.05). Individuals not on PrEP may receive a protective benefit by being in a sexual network with higher PrEP coverage. Agent-based models are useful to evaluate possible direct and disseminated effects of HIV prevention modalities in sexual networks.

Population attributable fraction for continuously distributed exposures

When estimating population attributable fractions (PAF), it is common to partition a naturally continuous exposure into a categorical risk factor. While prior risk factor categorization can help estimation and interpretation, it can result in underestimation of the disease burden attributable to the exposure as well as biased comparisons across different exposures and risk factors. Here, we propose sensible PAF estimands for continuous exposures under a potential outcomes framework. In contrast to previous approaches, we incorporate estimation of the minimum risk exposure value (MREV) into our procedures. While for exposures such as tobacco usage, a sensible value of the MREV is known, often it is unknown and needs to be estimated. Second, in the setting that the MREV value is an extreme-value of the exposure lying in the distributional tail, we argue that the natural estimator of PAF may be both statistically biased and highly volatile; instead, we consider a family of modified PAFs which include the natural estimate of PAF as a limit. A graphical comparison of this set of modified PAF for differing risk factors may be a better way to rank risk factors as intervention targets, compared to the standard PAF calculation. Finally, we analyse the bias that may ensue from prior risk factor categorization, examining whether categorization is ever a good idea, and suggest interpretations of categorized-estimands within a causal inference setting.