Real-world data (RWD) can provide insight into benefits, risks and usage of medical products. Unlike controlled randomized trials, RWD can capture patient populations and situations that may not be feasible to address in a clinical trial. This makes RWD a useful source for better understanding patients. However, when the goal is to understand how a medical product performs association can be misleading. Discovering that a drug is associated with improved outcomes, does not prove the drug caused the improvement. Correlation alone can mislead. This blog explores why correlation alone is insufficient in real-world evidence (RWE), how causal inference can strengthen study design, and practical tools for generating credible insights from observational data.
Why Association Falls Short for Decision-Making
Decision-makers such as regulators, payers, and clinicians need more than correlations. They need to understand how interventions change outcomes. Without a causal framework, RWE can be misleading. It can lead to distorted results, making treatments appear more or less effective.
Causal Questions Require Explicit Assumptions
Causal inference is fundamentally about assumptions. [1] Ideally, we would be able to estimate individual causal effects, however it not possible observe the same person both treated and untreated at the same time. Instead, we estimate treatment effects by comparing groups in a way that simulates this ideal counterfactual scenario. This requires clear assumptions about exchangeability, positivity, and consistency. Not all assumptions can be verified empirically but making them explicit helps identify the strengths and limitations of a study. [2]
Exchangeability and the Role of Confounding
Exchangeability means that the treated and untreated groups would be comparable if not for the treatment. In practice, this is rarely the case in observational data. Confounding variables, those that cause both treatment and outcome, can bias estimates if not accounted for. Simply adjusting for more variables is not a solution; including the wrong variables, like colliders, can introduce bias rather than remove it.
Why More Adjustment Is Not Always Better
Does coffee cause happiness? Imagine we have data on sleep, coffee, weightlifting and happiness.
Sleep affects both coffee consumption and happiness, making it a confounder we want to adjust for. Weightlifting, on the other hand, influences both sleep and coffee but is a collider on the pathway from happiness to coffee. Adjusting for weightlifting opens a path that creates bias rather than reducing it. Without carefully considering causal relationships, additional adjustments can distort rather than clarify the effect.
Causal Diagrams as a Design Tool
Directed acyclic graphs, or DAGs, are an indispensable tool for causal reasoning. By mapping relationships between variables, DAGs help identify confounders, mediators, and colliders, guiding which adjustments are necessary. They are especially valuable when dealing with time-varying confounding, where treatment affects a confounder that later influences outcomes. Using DAGs systematically ensures that causal questions are approached with transparency and rigor.
From Evidence Generation to Credible Results
Causal inference is not about labeling one study design as good and another as bad. Observational studies, randomized trials, or hybrid designs can all provide meaningful insights when assumptions are explicit and carefully considered. Thoughtful study design, clear reasoning, and an understanding of the underlying causal structure allow real-world evidence to support decisions with confidence, moving beyond correlation to credible, actionable insights.
References
- Hernán MA, Robins JM (2020). Causal Inference: What If. Boca Raton: Chapman & Hall/CRC
2.Haber, N. A., Wieten, S. E., Rohrer, J. M., Arah, O. A., Tennant, P. W. G., Stuart, E. A., Murray, E. J., Pilleron, S., Lam, S. T., Riederer, E., Howcutt, S. J., Simmons, A. E., Leyrat, C., Schoenegger, P., Booman, A., Kang Dufour, M.-S., O’Donoghue, A. L., Baglini, R., Do, S., … Fox, M. P. (2022). Causal and associational language in observational health research: A systematic evaluation. American Journal of Epidemiology, 191(12), 2084–2097. https://doi.org/10.1093/aje/kwac137
