Causal inference methods for real-world data

Professor of Machine Learning at Saarland University

Full Professor of Medical Statistics, University of Florence

Abstract

Correlation does not imply causation, but teaching applied researchers pretty much just that – and nothing else – about causal inference has had some peculiar consequences. For example, researchers have come up with very creative phrasings to imply causality while maintaining some degree of plausible deniability (such as the language of « unique » prediction), or have put undue faith in pseudo-solutions. In this webinar, I will share some observations that are mainly drawn from psychology but echo concerns raised in other fields (including other social and life sciences). I will discuss a major tension in the debate about causality: if researchers are taught how causal inference works in principle, will they end up drawing even more bad causal inferences? And I will also propose potential ways to make rigorous causal inference more mainstream.

Abstract

This lecture provides an introduction to causal machine learning methods for estimating treatment effects in observational and experimental studies with high-dimensional covariates. The focus is on modern approaches that combine the flexibility of machine learning with causal inference in order to estimate treatment effects conditional on covariates and uncover effect heterogeneity across subgroups in a data-driven manner. The lecture discusses the challenges of treatment or impact evaluation when the set of potential confounders is large and conventional parametric models may be too restrictive. It also introduces two prominent approaches that allow researchers to control for covariates in a data-adaptive way and estimate heterogeneous treatment effects: causal (random) forests and the double machine learning framework, the latter of which can be combined with a wide range of machine learning methods. Applications from the health sciences/economics illustrate how these methods can be used to estimate both average and subgroup-specific causal effects and to detect effect heterogeneity. The lecture also demonstrates how these methods can be implemented in R, making the session useful for applied researchers interested in integrating causal machine learning into their empirical work.

Abstract

We introduce methods for studying for causal effects in settings characterized by interference, arising when the outcome of a unit (e.g., patient’s infection) depends not only on their own treatment (e.g., vaccination) but also on the treatment of others (e.g., friends).

We first introduce estimands and explore identification and estimation strategies in experimental and observational settings. We then focus on important but nonstandard settings having two distinct sets of units:
units to which the intervention is applied and units on which the outcomes are measured, which is called bipartite interference: treatments applied to one intervention unit can affect multiple outcome units, and the outcome of a unit may depend on the treatments applied to multiple intervention units. We illustrate the methods with some case studies.

🇬🇧 Policy relevant effects in infectious disease studies

When facing infectious diseases, healthcare professionals must make decisions under uncertainty. Certain considerations are frequently recurring in this process. Knowing whether a vaccine’s effect wanes is important for determining the timing of seasonal flu shots or assigning COVID-19 boosters. Understanding whether a vaccine has heterogeneous effects on different subtypes, such as Malaria variants, has implications for the development of new treatments. Whether vaccinating a subgroup of individuals is sufficient to reach herd immunity, due to spill-over effects, shapes public health policies. Although the previous plain English sentences might seem clear, I will argue that terms like “waning”, “heterogeneous effects”, and “spill-over” are ambiguously interpreted. I will then propose ways to clarify these terms based on theory and methods for causal inference. Using Malaria, HIV, influenza, and COVID-19 as illustrative examples, I will show that these clarifications matter in practice. In particular, I will leverage the clarifications to define new treatment effects in vaccine settings, which sometimes justify existing analytic approaches and also motivate new methods that are feasible to implement.

🇬🇧 « Causal AI”: Is causal inference from healthcare data about to be automated?

Decisions about the treatment and prevention of disease are guided by causal inference and health researchers often make causal inferences using healthcare databases. The emergence of tools referred to as “AI” may transform the way in which those databases are used for causal inference. However, for “AI” to speed up the causal learning for healthcare databases, we need a better understanding of what both “AI” and causal inference are. This talk dissects the components of “Causal AI” and
discusses its potential to automate causal research in the health sciences.

🇬🇧 Causal inference with compositional data

Compositional data is a form of hierarchical data in which a whole (or a total) is the sum of its constituent components. Although compositional data can arise in any setting, they are particularly common in health research, with typical examples including dietary, physical activity, or microbiome data. This type of data can bring a range of analytical and interpretational challenges which risks results being misinterpreted.

In this talk, I will introduce compositional data using directed acrylic graphs (DAGs), outline the different types of causal effects that may be of interest in such data, and discuss suitable analytical approaches.

🇬🇧  Causal Inference Methods for Real-world data

Estimating causal effects in non-experimental data is a key aim of applied health and social science research. Unfortunately, it is also notoriously difficult.

Contemporary causal inference methods, including directed acyclic graphs, promise to revolutionise the analysis and interpretation
non-experimental data, not least by making our ambitions and assumptions far more explicit. In the health sciences, these methods are rapidly gaining popularity, but they are still yet to be adopted as widely as other entrenched
methods.

This talk offers a simple introduction to the new
science of causal inference, with a particular focus on directed acyclic graphs.