Design–analysis mismatch in the analysis of longitudinal HER2 data in breast cancer

This graphical abstract presents findings from a simulation study examining the impact of within-patient correlation on statistical inference in analyses of repeated binary clinical outcomes. Under conditions of strong intracluster correlation (200 patients, 10 observations per patient, 1000 simulated data sets), naive logistic regression that assumed independence substantially underestimated standard errors and led to marked inflation of the type I error rate (30.5%). In contrast, generalized estimating equations (GEE), which appropriately account for within-patient clustering, preserved nominal error control (5.7%). These results underscore the importance of using correlation-adjusted analytic approaches to ensure valid inference when evaluating patient-level exposures in longitudinal or clustered oncology data. The recent article by Qiu et al. on the evolution and prognostic impact of HER2-low, HER2-ultralow, and HER2-null status in HER2-negative early breast cancer addresses a clinically significant topic regarding dynamic HER2 expression in the neoadjuvant setting.1 However, the study suffers from a fundamental design–analysis mismatch that violates core principles of longitudinal data analysis. By treating inherently correlated, time-dependent observations as independent, static variables, the authors have generated results that are not statistically defensible. The study used a classic paired longitudinal design, in which HER2 status was measured repeatedly within the same patients before and after neoadjuvant chemotherapy (NAC). Established statistical theory makes clear that analyses must explicitly account for this correlation structure.2, 3 Nevertheless, the authors analyzed these data using standard logistic regression for pathologic complete response analysis and conventional Cox proportional hazards models for survival—both of which assume independence of observations. This assumption is mathematically incompatible with repeated measures designs. These issues are not matters of analytic preference or alternative modeling philosophies but represent departures from accepted statistical science in three critical domains. First, in the pathologic complete response analysis, pre-NAC and post-NAC HER2 status were entered simultaneously as independent covariates in a standard logistic regression model. This approach ignores the covariance induced by repeated measurement within patients and is known to produce underestimated standard errors, spuriously narrow confidence intervals, and inflated type I error rates.2, 3 Valid inference for paired binary outcomes requires methods that explicitly model within-subject correlation, such as generalized estimating equations or conditional logistic regression. Without such methods, the reported p values and confidence intervals lack statistical interpretability. Second, the survival analysis is affected by immortal time bias. Post-NAC HER2 status is determined at surgery and thus constitutes a postbaseline, time-dependent variable. Treating it as a fixed baseline covariate in a standard Cox model violates the temporal ordering between exposure and outcome. Patients must necessarily survive the neoadjuvant period to have post-NAC HER2 status observed, creating a guaranteed survival interval that biases hazard ratio estimates. Established survival analysis methodology requires time-dependent Cox models or appropriate landmark analyses in this setting.4 Failure to apply these methods renders the reported hazard ratios neither biologically nor statistically meaningful. Third, the analytical approach does not comply with international methodological and reporting standards. The STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) Statement explicitly requires authors to describe and apply statistical methods that account for clustering and repeated measurements (see Item 12a in von Elm et al.).5 Similarly, the International Committee of Medical Journal Editors recommendations state that statistical methods must be described and implemented with sufficient rigor to allow readers to assess the validity of the findings.6 The methods used in this study do not meet these requirements. Because the reported associations are derived from an analytically incompatible framework, they cannot reliably inform clinical interpretation or decision-making. A post-hoc limitation statement or minor correction would be insufficient to remedy these fundamental flaws. To safeguard the integrity of the scientific record, a comprehensive reanalysis using statistically appropriate methods—namely, clustered or marginal models for pathologic complete response and time-dependent modeling strategies for survival outcomes—is necessary to determine whether the study's conclusions remain tenable under valid statistical assumptions (see Supporting Methods and Figures S1). To illustrate the magnitude of this inferential distortion, I conducted a simple simulation study under a null association scenario (see Table S1). The naive logistic model that ignored within-subject correlation showed inflated type I error rates compared with generalized estimating equation–based analysis, confirming that the reported inference framework is statistically unstable. Ilene Chen: Conceptualization; methodology; investigation; writing—original draft; writing—review and editing; visualization. The author declares no conflicts of interest. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.