Real-world evidence (RWE) studies offer valuable insights into the effects of health interventions in real-world settings. However, these studies face a significant challenge of confounding. Confounding occurs when an external factor is associated with both the exposure (intervention) and the outcome of interest. If not properly addressed, confounding can lead to biased estimates of the exposure-outcome relationship.

To address confounding, researchers often use methods such as stratification, restrictions, or regression adjustment to estimate the effect of the exposure on the outcome while controlling for potential confounding variables. While adjusting for confounders is crucial to obtain unbiased effect estimates, it is essential to choose the variables for adjustment carefully. Including unnecessary variables in the analysis can introduce bias instead of correcting it.

In this blog, we will explore necessary and unnecessary covariate adjustments and discuss the concept of overadjustment bias in RWE studies. Schisterman et al. provide valuable insights to clarify the general term “overadjustment”, which has been inconsistently defined. They make the following distinction between unnecessary adjustments and overadjustment bias:¹

• Necessary adjustment is controlling for variables that are known or suspected to be confounders in the exposure-outcome relationship.
• Unnecessary adjustment is controlling for variables that do not affect the expected (mean) estimate of the total causal effect between the exposure and outcome. These variables may be related to the exposure or outcome alone or may have no relationship with either one. Including such unnecessary variables in the analysis does not contribute to obtaining a more accurate estimate of the exposure-outcome relationship but may affect the precision of the estimate.
• Overadjustment bias arises when one adjusts for intermediate variables (or their descending proxies) that lie on the causal pathway between the exposure and outcome. Intermediate variables are those influenced by the exposure and, in turn, influence the outcome. Adjusting for these variables can result in biased estimates of the exposure-outcome relationship because it obscures the true causal relationship between the exposure and outcome.

These concepts are illustrated in the figure below depicting different categories of variables in a casual system. Consider an example where the exposure of interest (E) is a cholesterol-lowering medication, and the outcome of interest (O) is a heart attack. Assume the intermediate variable (M) is blood cholesterol levels, and the potential covariates are C1-C7.

According to this illustration, to estimate the total effect of the medication (E) on heart attacks (O), it is necessary to control for C5 since it qualifies as a confounder (it is related to both the treatment and the outcome). However, controlling for variables C1, C2, C3, and C6 would be unnecessary since they have no effect on the estimated total effect of the medication, although they may impact the precision of the estimate. Previous research suggests that controlling for C7 may also affect precision but could introduce bias.

Controlling for the blood cholesterol levels (M) or its descending proxy (C4) when estimating the total effect of the medication (E) is not recommended since it leads to overadjustment bias. The coefficient for medication in a multivariable model that includes cholesterol levels would only estimate the direct effect of the medication (E) that does not go through cholesterol levels (M), thereby decreasing the medication’s total effect. If the medication affects the risk of heart attack only through lowering cholesterol levels, then the coefficient for medication in a model that includes cholesterol levels would be null.

 How can you avoid overadjustment bias?

1. Identify the relevant confounding variables that need to be adjusted for in the analysis, either through prior knowledge or assumptions being made in the study.
2. Use causal diagrams, such as directed acyclic graphs (DAGs), to graphically help identify the minimum set of confounding variables that need to be adjusted for while avoiding overadjustment bias on the effect of the exposure of interest.

One thing to note is that the minimum set of confounding variables identified is specific to the exposure under consideration. Focusing on the effect of a different variable in the system requires researchers to identify a different set of confounding variables.

 The impact of overadjustment bias

Adjusting for confounding is an essential aspect of epidemiologic research. Necessary adjustment improves the accuracy in estimating the size and direction of the exposure-outcome relationship (e.g., between a medication and health outcome). However, researchers must carefully consider the causal pathways between the exposure and outcome, as well as the potential confounding variables when selecting variables for adjustment. Causal diagrams such as directed acyclic graphs (DAGs) can help in identifying a minimum set of confounding variables that need to be adjusted for. Arbitrarily dumping every available variable into a regression model can lead to biased estimates and incorrect conclusions about the effectiveness of a health intervention, its safety profile, and its downstream effects.

 Why choose Medlior for your RWE project?

• Medlior has access to comprehensive real-world data sources,
• Medlior has a team of highly experienced epidemiologists, biostatisticians, and data scientists,
• Medlior applies advanced methodologies and will help you design your real-world study.

Contact us at  to learn more about our services and how we can support your RWE Strategy in 2023 and beyond.

 References:

1. Schisterman EF, Cole SR, Platt RW. Overadjustment bias and unnecessary adjustment in epidemiologic studies. Epidemiology 2009; 20(4): 488-95.