Many research questions are causal, but the causal goal of a research endeavour is seldom explicitly stated. Because observational studies can never definitely ‘prove’ causation, many editors, reviewers and thesis supervisors believe that all causal language should be banned from non-randomized research manuscripts. This has led to the ubiquitous use of cloaked causal statements such as “X was independently associated with Y after adjusting for potential confounders.”

While the injunction on causal language is rooted in justifiable caution about the overinterpretation of observational studies, it is ultimately unhelpful in the evaluation of methods and results. The problem with cloaked causal statements is that causal vs. non-causal questions require different analytical choices, even if the same pair of variables is being analysed.

Suppose we are interested in the relationship between pulmonary embolism thrombus load (as quantified using CT angiography) and the risk of death. We have collected data on these two variables and on the cardiac biomarker troponin. If our research question is prognostic (i.e.: variable A predicts outcome Β), we could build a model that includes thrombus load and troponin as predictors of death. The results would show us the added predictive risk per unit of thrombus load. “Thrombus load was not associated with mortality after adjustment for troponin” could be a meaningful result in this context. But if our research question is about the causal relationship (i.e.: variable A causes outcome Β) between thrombus load and risk of death, including troponin as covariable in the model would be wrong.

Because more troponin release (through right ventricular strain) is a consequence of higher thrombus load, it is on the causal path between thrombus load and death. Adjusting for the mediating variable troponin in a model would ‘capture’ part of the causal effect of thrombus load on risk of death. This leads to a biased estimate (toward null) of the full effect of thrombus load. Therefore, “thrombus load was not associated with mortality after adjustment for troponin” would be a meaningless and misleading statement if our aim is to quantify the total effect of thrombus load on the risk of death. Readers of a study report can only judge whether the choice to include troponin or not in the model was correct if the authors state explicitly whether their question was causal or predictive.

This so-called overadjustment bias is one of the pitfalls that will be further explored below. It shows that answering causal questions requires a clear causal hypothesis. This module provides an introduction to formal causal thinking using Directed Acyclic Graphs (DAGs), a standard method to communicate hypothesized causal pathways.