Born out of the "causal revolution" at the end of the 20th Century and its subsequent introduction into the field of epidemiology, a clearer distinction between three domains in epidemiological research began to emerge: description, causal inference (or counterfactual prediction), and prediction. While description research is of utmost relevance to answering many epidemiological questions, this Research Topic in Frontiers in Epidemiology focuses on the intersection of prediction and causal inference.
Causal inference research aims to answer questions about cause and effect by contrasting between counterfactual worlds. To obtain an accurate answer to a causal question using observational data, one must identify the underlying data generation process and conduct analyses consistent with the identified process. Indispensable and inherent to this domain is a priori subject-matter knowledge and the use of graphical tools such as directed acyclic graphs.
On the other hand, prediction-focused research questions center around forecasting outcomes in “future” individuals using available information about predictors. In the context of health, this domain is indispensable in supporting (shared) clinical decision-making and in identifying individuals at particular risk of an outcome of interest. Methodologically, prediction methods, which often employ data-driven approaches that are becoming increasingly automated, such as machine learning, are generally agnostic to causal structures.
The contemporary strict separation of causal inference and prediction in epidemiology is useful in the prevention of common methodological fallacies, such as the misinterpretation of risk prediction regression model coefficients as causal effect estimates. Such confusions likely arise because these domains employ the same set of statistical tools. However, this practical division is ultimately a simplification, which masks the fact that when answering clinical prediction questions, data are always the product of a data generation mechanism and therefore, causal tools and thinking can be employed also in the domain of prediction. Recent work has underlined the importance of causal methods in clinical risk prediction; namely, showing how causal knowledge and causal inference tools can be used to select predictors, evaluate the transportability of a clinical prediction model, build clinical prediction models adjusting for treatment drop-in, build fairer models with respect to sensitive attributes such as race and gender, and create more interpretable and impactful models to better inform clinical decision-making. Within this Research Topic, our aim is to collate innovative approaches using causal methods and causal thinking to solve clinical prediction problems, in an effort to bring more attention to the overlap between these two domains.
We welcome both new methodological developments and applied research using causal methods in clinical risk prediction.
Possible submission topics include:
- Addressing the clinical predictor selection problem using causal methods
- Transportability assessment of clinical prediction models informed by causal inference tools
- Adjustment for treatment drop-in relying on counterfactuals in clinical prediction models
- Algorithmic fairness in clinical prediction models relying on causal-based notions of fairness
We also explicitly encourage authors of manuscripts focused on other applied topics at the intersection between causal inference and prediction domains or focused on method developments in this area to submit their work for consideration.
Born out of the "causal revolution" at the end of the 20th Century and its subsequent introduction into the field of epidemiology, a clearer distinction between three domains in epidemiological research began to emerge: description, causal inference (or counterfactual prediction), and prediction. While description research is of utmost relevance to answering many epidemiological questions, this Research Topic in Frontiers in Epidemiology focuses on the intersection of prediction and causal inference.
Causal inference research aims to answer questions about cause and effect by contrasting between counterfactual worlds. To obtain an accurate answer to a causal question using observational data, one must identify the underlying data generation process and conduct analyses consistent with the identified process. Indispensable and inherent to this domain is a priori subject-matter knowledge and the use of graphical tools such as directed acyclic graphs.
On the other hand, prediction-focused research questions center around forecasting outcomes in “future” individuals using available information about predictors. In the context of health, this domain is indispensable in supporting (shared) clinical decision-making and in identifying individuals at particular risk of an outcome of interest. Methodologically, prediction methods, which often employ data-driven approaches that are becoming increasingly automated, such as machine learning, are generally agnostic to causal structures.
The contemporary strict separation of causal inference and prediction in epidemiology is useful in the prevention of common methodological fallacies, such as the misinterpretation of risk prediction regression model coefficients as causal effect estimates. Such confusions likely arise because these domains employ the same set of statistical tools. However, this practical division is ultimately a simplification, which masks the fact that when answering clinical prediction questions, data are always the product of a data generation mechanism and therefore, causal tools and thinking can be employed also in the domain of prediction. Recent work has underlined the importance of causal methods in clinical risk prediction; namely, showing how causal knowledge and causal inference tools can be used to select predictors, evaluate the transportability of a clinical prediction model, build clinical prediction models adjusting for treatment drop-in, build fairer models with respect to sensitive attributes such as race and gender, and create more interpretable and impactful models to better inform clinical decision-making. Within this Research Topic, our aim is to collate innovative approaches using causal methods and causal thinking to solve clinical prediction problems, in an effort to bring more attention to the overlap between these two domains.
We welcome both new methodological developments and applied research using causal methods in clinical risk prediction.
Possible submission topics include:
- Addressing the clinical predictor selection problem using causal methods
- Transportability assessment of clinical prediction models informed by causal inference tools
- Adjustment for treatment drop-in relying on counterfactuals in clinical prediction models
- Algorithmic fairness in clinical prediction models relying on causal-based notions of fairness
We also explicitly encourage authors of manuscripts focused on other applied topics at the intersection between causal inference and prediction domains or focused on method developments in this area to submit their work for consideration.