Causal inference begins with clarity about the research question. What population, treatment, comparator, and outcome are you evaluating? Defining these elements upfront allows you to plan how real-world data can emulate the conditions of an ideal randomized trial, rather than trying to fit your question to the data you have. This blog challenges the idea that real-world data (RWD) automatically produces real-world answers and explains what it actually takes to make causal claims credible.
Aligning Time Zero to Avoid Immortal Time Bias
One of the most common pitfalls in observational studies is immortal time bias, misclassifying the start of follow-up so patients are “immortal” to the outcome for a period. [1] A simple way to prevent this is to ensure that the index date, or time zero, aligns with treatment initiation. This seemingly small step can dramatically reduce bias and strengthen causal claims.
Designing for Causality Before Choosing Methods
Causal inference is a design-first discipline. [2] Statistical methods like propensity score matching, inverse probability weighting, or doubly robust estimators are powerful tools, but their validity depends on well-considered study design. Determining eligibility criteria, treatment definition, follow-up rules, and censoring strategies before analysis ensures that the data can justify causal conclusions.
Using Causal Diagrams to Guide Variable Selection
Directed acyclic graphs (DAGs) remain invaluable for guiding adjustment decisions. They help distinguish confounders from colliders and mediators, clarifying which variables to adjust for and which to leave out. Time-varying confounders can also be identified, revealing how past treatment decisions influence future confounders and outcomes, and informing the choice of appropriate analytical methods.
When Unmeasured Confounding Is Inevitable
Even with the best planning, some confounders are unmeasured. Sensitivity analyses can quantify how unmeasured confounders would impact the result, providing insight into the robustness of results. [3] By explicitly addressing these uncertainties, researchers can maintain credibility and provide decision-makers with realistic interpretations.
The Challenge of Time-Varying Confounding
In real-world data, exposures and confounders often change over time. For example, a lab value may affect treatment choice, which in turn influences subsequent lab values and outcomes. DAGs and careful longitudinal design can help identify these feedback loops and guide appropriate statistical methods to account for them.
Choosing Data That Can Answer the Question
Not every dataset can answer every question. . Electronic health records, registries, or claims data may vary in sample size, completeness, and granularity. Considering the causal diagram when comparing data sources allows you to choose the dataset, or combination of datasets, that can best answer the research question while avoiding unnecessary compromises. [4, 5]
For Causal Claims, We Need Causal Inference
Real-world data can support causal claims, but only when study design, variable selection, and statistical methods are aligned with the research question. Through careful study design, explicit assumptions, and rigorous analysis, researchers can use real-world evidence to generate causal insights, bridge gaps in clinical knowledge, and inform decisions with confidence.
References
- Fu, E. L. (2023). Target trial emulation to improve causal inference from observational data: What, why, and how? Journal of the American Society of Nephrology, 34(8), 1305–1314. https://doi.org/10.1681/ASN.0000000000000152
2.Hernán MA, Robins JM (2020). Causal Inference: What If. Boca Raton: Chapman & Hall/CRC
3.D’Agostino McGowan, L. (2022). Sensitivity analyses for unmeasured confounders. Current Epidemiology Reports, 9, 361–375. https://doi.org/10.1007/s40471-022-00308-6
4.Wang, S. V., Pinheiro, S., Hua, W., Arlett, P., Uyama, Y., Berlin, J. A., Bartels, D. B., Kahler, K. H., Bessette, L. G., & Schneeweiss, S. (2021). STaRT-RWE: Structured template for planning and reporting on the implementation of real-world evidence studies. BMJ, 372, m4856. https://doi.org/10.1136/bmj.m4856
5.Gatto, N. M., Campbell, U. B., Rubinstein, E., Jaksa, A., Mattox, P., Mo, J., & Reynolds, R. F. (2022). The structured process to identify fit-for-purpose data: A data feasibility assessment framework. Clinical Pharmacology & Therapeutics, 111(1), 122–134. https://doi.org/10.1002/cpt.2466
