The growing availability of electronic health records (EHR) and registry data has led to a surge in studies using real-world data.1 Real-world datasets can provide insights that randomized trials cannot easily provide.2 In metastatic breast cancer, where patients are exposed to long-term continuous treatment, observational data may be better suited for studying frequent dose adjustments in response to tolerability, comorbidity, or perceived frailty. However, the same features that make EHR-based research valuable, such as the direct capture of routine care, flexible treatment pathways, and large heterogeneous patient populations, also introduce methodological vulnerabilities that can result in misleading causal conclusions if analytic rigor is not applied. A real-world study3 derived predictive indices from early treatment experience and compared outcomes between reduced and full starting doses. The study aimed to emulate a pragmatic trial based on observational data using the target trial emulation framework.4 The conceptual appeal is clear: if patients who are likely to tolerate therapy poorly could be identified upfront, clinicians may safely initiate lower doses without compromising efficacy. The target trial emulation framework encourages investigators to explicitly define eligibility criteria, treatment strategies, and follow-up windows, thereby promoting conceptual clarity. However, translating these principles into valid causal inferences still requires careful attention to temporal alignment, treatment assignment, and endpoint definition, which are areas where retrospective datasets often blur distinctions that are explicit in randomized trials.5 A central methodological challenge in real-world EHR research is the frequent conflation of baseline treatment decisions with processes that unfold after therapy initiation. Initial therapeutic choices are made at a defined time point and reflect clinician judgment, patient characteristics, and contextual factors. However, subsequent modifications arise in response to clinical treatment response, toxicity, or other time-varying patient factors and therefore represent time-dependent processes. When observational analyses collapse these temporal factors into a single exposure definition, periods of guaranteed survival can be inadvertently introduced, creating a risk of immortal time bias.6 Confounding by indication presents additional challenges in interpreting real-world evidence. In contrast to randomized trials, observational studies reflect real-world clinical decisions rather than randomized patient assignments. These decisions are often influenced by unmeasured factors such as frailty, risk perception, patient preferences, and clinical presentation, which are not captured in structured variables. Consequently, the apparent differences in outcomes across treatment strategies may reflect the underlying selection rather than the treatment effects. Residual confounding should not be regarded as a technical limitation but as an inherent characteristic of observational studies. Irrespective of whether results align with prior expectations or challenge established evidence, this underscores the need for enhanced transparency, detailed reporting, and presentation of the underlying data to allow readers to assess robustness. The interpretation of survival outcomes presents another recurring challenge. In oncology, progression-free survival and overall survival reflect distinct biological and therapeutic processes. Because EHR datasets depend on routine clinical recording rather than standardized assessment schedules, endpoints such as progression may be variably defined across and even within centers. Alternative measures, including treatment discontinuation or time to the next therapy, can provide complementary insights into tolerability and real-world effectiveness. However, they require careful interpretation within the observational context.7 Beyond the analytic design, data provenance and quality determine the credibility of real-world analyses. As datasets scale across sites with heterogeneous coding practices, measurement variability can amplify rather than dilute bias. Transparent reporting of data generation processes, variable construction, validation procedures, and missingness is therefore essential, as observational studies depend more heavily on transparency to support credible inferences than randomized trials. Frameworks such as the ESMO Guidance for Reporting Oncology Real-World Evidence (GROW) provide structured recommendations to improve transparency, reproducibility, and interpretability, complementing causal design approaches such as target trial emulation.8 Together, they help ensure that observational studies are not only methodologically sound but also clearly communicated to clinicians, reviewers, and policymakers. Increased access to electronic health record data is reshaping oncological research, enabling levels of observational granularity that were previously unattainable. This data can provide insights into patterns of care, identify unmet needs, and guide future trial designs. However, the path from real-world observations to clinical recommendations is narrow. Best practices require careful separation of baseline from post-baseline characteristics, explicit articulation of causal assumptions, attention to data quality and validation, and transparent reporting aligned with frameworks such as target trial emulation and ESMO GROW. The value of observational research lies not in replacing randomized trials but in supplementing them and transforming routine clinical data collection into a learning system that advances oncology while maintaining methodological rigor.
Andreas Bjerrum (Wed,) studied this question.