Negative confounding demystified: revisiting classic examples and proposing an updated teaching framework

Observed non-null effect estimates are routinely checked carefully for the possibility of positive confounding biasing away from the null and generating a false positive (type 1 error). Further investigation of the exposure of interest as a potential target of intervention, using more rigorous study designs, typically clarifies whether the observed association is fact or fiction. In contrast, observed null associations are not obviously sources of new targets of intervention, although they could be an artifact of negative confounding creating a false negative (type 2 error),1,2 causing potential interventions to be overlooked. Negative confounding may also create spurious protective effects, which also need to be investigated. Here, to address this gap, we provide graphical illustrations of negative confounding in contrast to positive confounding along with numerical examples, bearing in mind that whether an exposure-outcome pair is subject to positive or negative confounding depends on the direction of the confounders and of the hypothesized effect. Common teaching examples given for positive confounders include socioeconomic position (SEP), health status, and health-conscious behaviors.3 Given that negative confounding is easily overlooked, examples of negative confounding are pedagogically valuable. The most widely cited example of negative confounding, from a study published over 15 years ago (with over 160 citations), uses an outdated definition of a confounder as associated with the exposure rather than as a cause of the exposure as well as of the outcome.4 The paper describes fish consumption as a negative confounder when examining the effect of methylmercury exposure on the outcome of neurodevelopmental deficits in children or risk of cardiovascular disease (CVD) in adults.4 In this example, negative confounding is possible assuming fish consumption increases the intake of methylmercury4 and decreases the risk of cardiovascular disease and neurodevelopmental deficits (Figure 1A). However, the paper also considers the effect of fish intake on CVD as confounded by methylmercury. Methylmercury does not cause fish intake, so methylmercury is not a confounder (negative or positive) in this situation. Illustrations of positive and negative confounding using directed acyclic graphs. (A) Fish consumption as a negative confounder of methylmercury on adverse health outcomes. (B) Smoking as a positive and as a negative confounder of alcohol consumption on cardiovascular disease (CVD). (C) Socioeconomic position as a negative confounder of job stress on CVD. Apart from this example, another one from a study intended as a teaching tool for positive/negative confounding posited smoking as a negative confounder of the association of alcohol use with CVD.5 However, this example is in the context of alcohol having a protective effect on CVD, making smoking a negative confounder of alcohol on CVD, and potentially obscuring any benefits of alcohol on CVD. Conversely, if alcohol had been hypothesized as having a harmful effect on CVD, then smoking would have been a positive confounder of alcohol on CVD, potentially exaggerating the harmful effect of alcohol. Failure to comprehensively adjust for positive confounding overestimates the effect of alcohol consumption on CVD (Figure 1B). Neither of these two often-cited papers clearly present negative confounding. The lack of informative teaching examples may lead to underappreciation of the pervasiveness of negative confounding, resulting in the effects of harmful exposures being overlooked. Consequently, the opportunity to intervene on such exposures may not be investigated. To facilitate the detection of negative confounding, we recently published a Mendelian randomization (MR) study demonstrating how MR can be used to identify negative confounders,6 as long as the MR assumptions are met, particularly the selection of appropriate genetic instruments7 and avoidance of selection bias.8 We also showed that body mass index (BMI) can be a negative confounder in many scenarios, sometimes differing by sex. For example, BMI reduces low-density lipoprotein cholesterol and increases the risk of CVD, which means the true harms of low-density lipoprotein cholesterol may not be observed due to negative confounding by BMI. Here, we propose a revised example of negative confounding where the association of job stress (exposure) with CVD (outcome) is considered,9 with SEP as a negative confounder. People with higher SEP tend to experience greater job stress10,11 while simultaneously having lower CVD risk for many reasons. Crucially, SEP can create a spurious inverse association of job stress with CVD, because the effect of job stress on CVD is biased by SEP increasing job stress but reducing CVD risk (Figure 1C). In this case, if adequate adjustment is not made for SEP, then the association of job stress with CVD would be biased toward the null or reversed. This example also demonstrates that SEP, often perceived as a positive confounder, can also be a negative confounder for specific causal pathways. We also give corresponding 2 × 2 tables to illustrate this negative confounding by SEP (Table 1) alongside for completeness an example of positive confounding by SEP, also illustrating that whether an attribute is a negative or positive confounder depends on the research question. Examples where SEP is a negative confounder, a positive confounder, or not a confounder. Directed acyclic graphs (DAGs) are powerful tools for conceptualizing confounding, as demonstrated in Figure 1 and Figure S1. Open-access platforms like DAGitty (www.dagitty.net) offer user-friendly interfaces to create DAGs interactively, so as to identify bias from confounding. Importantly, our letter highlights a key, yet often underappreciated, implication: negative confounding can mask true causal effects, potentially obscuring important public health targets of intervention. By illustrating this phenomenon, particularly through our proposed example, we emphasize the necessity of actively considering negative confounding during study design and interpretation to avoid misleading conclusions about interventions, and hence facilitate identification of threats to public health and prioritization of effective interventions. We also emphasize that whether confounding is positive or negative depends on the confounders and the hypothesized direction of effect of the exposure on the outcome, that is, protective or harmful. In this letter, we used simplified examples with a single confounder to illustrate negative versus positive confounding. We also assumed that the effects of the confounder on exposure/outcome are monotonic. To reduce complexity, we did not consider time-varying confounding or nonlinear associations of confounders with exposure or outcome. Zhu Liduzi Jiesisibieke, C. Mary Schooling (Conceptualization), Zhu Liduzi Jiesisibieke, C. Mary Schooling, Hong Tang (Formal analysis), Zhu Liduzi Jiesisibieke, C. Mary Schooling, Hong Tang (Methodology), C. Mary Schooling (Project administration), C. Mary Schooling (Resources), C. Mary Schooling (Supervision), Zhu Liduzi Jiesisibieke, C. Mary Schooling, Hong Tang (Visualization), Zhu Liduzi Jiesisibieke, C. Mary Schooling (Writing—original draft), and Zhu Liduzi Jiesisibieke, C. Mary Schooling, Hong Tang (Writing—review & editing) Supplementary material is available at AJE Advances: Research in Epidemiology online. All authors declare no competing interests. This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Not applicable.