When is a confounder not a confounder?

Abstract

As much as we would like to be guided by robust evidence from randomized controlled trials, the fact is that most of what we know about the causes of respiratory disease comes from observational studies (such as the associations of smoking with lung cancer and chronic obstructive pulmonary disease (COPD) or between asbestos and mesothelioma). For many exposures, it is often unfeasible or unethical to conduct randomized controlled trials in humans and we have to rely on observing associations with outcomes in non-interventional studies. The problem, of course, is that an association does not necessarily mean causation: there will always be uncertainty about whether the association is confounded by other differences between the exposed and non-exposed groups. Researchers can minimize confounding either by designing a study to eliminate major confounders (e.g. by excluding people with the confounder) or, more often, by measuring multiple potential confounding factors and adjusting for them in the analyses. Both approaches rely on our ability to identify and measure the confounding factors. So how should we decide what is a confounder and what is not? We teach students that potential confounders are factors that are associated with the exposure and the outcome, but not on the causal pathway. However, not every factor that meets these criteria is a confounder and adjusting for multiple non-confounding factors can bias the findings. Guidance for authors of causal inference studies has recently been published in the Annals of the American Thoracic Society by the editors of 35 leading respiratory, sleep and critical care journals, including Respirology.1 This calls for a more critical approach to selecting confounders and more transparent reporting of the results of observational studies. The article points out that several commonly used methods of selecting confounding variables do not adequately control for confounding and may introduce bias. These inappropriate methods include relying on stepwise or other statistical techniques to select variables that are significantly associated with the outcome, improve model fit, or those that make a difference to the strength of the observed association. The recommended approach is to decide, a priori, on the plausible confounding factors based on what is known or hypothesized about the causal pathway. To achieve this, the article encourages the use of causal diagrams or directed acyclic graphs (DAGs). There is nothing new in using conceptual causal path diagrams to aid our thinking, but DAGs put these onto a more formal setting. Unfortunately DAGs come with unfamiliar and potentially confusing terminology (nodes, vertices, edges, back-door paths etc.) but the concept is simple enough and an excellent description and example was provided in an earlier article in Respirology.2 Variables are linked by arrows representing the causal direction: bidirectional arrows and feedback loops are not permitted (variables cannot cause themselves). Paths between variables can include one or more intervening variables and the arrows on these paths can point in either direction. One of the advantages of using DAGs is that they help us to differentiate between confounders, mediators, and colliders and therefore to select which variables need to be controlled. Mediators are usually easy to distinguish from confounders: they are intermediate variables on the causal path between exposure and outcome. Mediators may provide valuable insights into causal mechanisms and we rarely want to adjust for them unless we are interested in quantifying different causal pathways between the exposure and the outcome. Less easily distinguished are colliders—these are easily mistaken for confounders because they are associated with both the exposure and the outcome and are not on the causal pathway. Unlike confounders, colliders are caused by both the exposure and the outcome or indirectly caused by other factors associated with the exposure and the outcome. Hence, the directional arrows from both exposure and outcome ‘collide’ at the collider variable. Colliders should not be adjusted for—controlling for them can introduce confounding.1, 2 For example, if we were investigating whether having COPD was a causal risk factor for lung cancer, we would consider smoking to be a confounder because it is a cause of both COPD and lung cancer and not on the causal path (Fig. 1). It would be necessary to control for smoking history to prevent confounding of the causal path between COPD and lung cancer. Weight loss, however, is unlikely to be a confounder but could be a collider: both severe COPD and lung cancer can cause weight loss (the arrows ‘collide’ at weight loss). Hence, it would not be appropriate to adjust the analysis for weight loss. Airway oxidative stress could potentially be on the causal path and could mediate the association. It would not be appropriate to adjust for measures of airway oxidative stress either. Of course, this causal diagram is oversimplified: there many be many other plausible variables (such as occupational exposures) that could confound an association between COPD and lung cancer and these should be included. An important distinction is drawn between causal inference studies and prediction studies. For example, the 5-item CURB65 score and the 20-item Pneumonia Severity Index have both been shown to be useful predictors of death among patients with pneumonia.3, 4 It is not relevant to consider whether the various factors that comprise these scores are causes, confounders, colliders, or mediators. What is important is whether they are valid and clinically useful predictors of mortality. By contrast, causal inference studies usually investigate the association between one exposure and an outcome. Other variables (‘covariates’) are usually only of interest to the extent that they confound or mediate this association. The practice of publishing the associations between these multiple covariates and the outcome in causal inference studies is discouraged.1 This guidance provides examples of confounders, colliders and mediators, more detailed explanations about DAGs, and principles for choosing confounders, interpreting, and reporting observational studies. There are links to further reading and online resources to help draw causal diagrams.1 Following this guidance will improve the conduct and interpretation of observational research. We can anticipate that the editors of the world's leading respiratory journals will expect authors to follow it.