Introduction
Real-world data (RWD) can provide insight into treatment effects in a real-world setting. However, because treatment allocation in RWD is not random, analyses are susceptible to confounding, when a variable influences both treatment and outcome, potentially biasing results.
When the exposure of interest varies over time, there may be time-varying confounding. Key confounders can change over time, sometimes in response to earlier treatment (i.e., a result from a laboratory test) impacting subsequent treatment. This is known as time-varying confounding. Conventional methods such as generalized linear models, or propensity score matching provides biased effects. Time-varying confounding requires a different analytical approach.
This blog describes what time-varying confounding is, why traditional methods lead to biased results, and the advanced techniques available to address it. It explains when time-varying confounding is likely to occur, why it matters for study validity, and three approaches to adjust for it.
Why Time-Varying Confounding Matters in Clinical Trials
Time-varying confounding can occur in clinical trials with long-term follow-up or repeated measurements. Examples include:
- Disease progression markers (e.g., tumor size, biomarker levels) that change during the trial and may influence subsequent treatment decisions.
- Safety parameters such as blood pressure or liver function tests, which respond to treatment and can trigger dose modifications or discontinuations.
- Concomitant medication use, which can vary over time and impact both treatment exposure and trial outcomes.
- If time-varying confounding is not addressed, standard analyses can yield misleading estimates of treatment effect—potentially influencing key trial decisions, regulatory review, and post-approval commitments.
The Problem: How Time-Varying Confounding Differs
Time-varying confounding arises when both exposures and confounders change over time, and when treatment influences future confounders. [1]
For example, a biomarker might improve after initial treatment but still factor into decisions about whether to adjust therapy later on. If this time-varying confounder is simply included as a time-dependent covariate in a conventional model, such as a time-dependent Cox regression, random effects model, or generalized estimating equation the result may be biased. [1]
The Solution: Methods for Addressing Time-Varying Confounding
Several causal inference methods have been developed to address this problem, particularly when confounders are influenced by prior treatment. Three commonly used methods include:
1.Inverse-Probability Weighting (IPW) via Marginal Structural Models (MSMs)
Creates a pseudo-population in which treatment assignment at each time point is independent of measured confounders. [2]
Conceptually similar to propensity score weighting, but applied repeatedly across time.
2. Parametric G-Formula
Simulates the distribution of exposure, confounders and outcome that would have been observed under two scenarios: 1) where every participant received the exposure of interest, 2) where every participant did not receive the exposure of interest.
Averages the confounder-specific mean outcomes under a specific treatment over the joint distribution of confounders over time [1]
3.G-Estimation
Uses two models: a model for the exposure and a semiparametric structural mean model.
G-estimation uses both of these models together to find a situation that makes the exposure independent of the estimated outcome given previous treatment and confounder history [1]
Conclusion
When exposures and confounders both changes over time, and when past treatment can influence future confounders, traditional adjustment methods provide biased results. Methods such as MSMs, the parametric g-formula, and g-estimation are useful alternatives, helping to produce more accurate results for decision-making.
At Phastar, we apply advanced statistical approaches to tackle challenges like time-varying confounding, ensuring that real-world and longitudinal clinical trials generate trustworthy evidence for clinical and regulatory use.
References
1.Mansournia, M. A., Danaei, G., Forouzanfar, M. H., Mahmoodi, M., & Jamali, M. (2017). Handling time varying confounding in observational research. BMJ, 359, j4587. https://pubmed.ncbi.nlm.nih.gov/29038130
2.Hernán, M.A., & Robins, J.M. (2020). Causal Inference: What If. Chapman & Hall/CRC. Causal Inference: What If (the book) — Miguel Hernán
3. Hernán, M. A., Cole, S. R., Margolick, J., Cohen, M., & Robins, J. M. (2005). Structural accelerated failure time models for survival analysis in studies with time-varying treatments. Pharmacoepidemiology and Drug Safety, 14(7), 477–491. https://doi.org/10.1002/pds.1064
