Letter: Comparison of Lattice Flow Diverter and Pipeline Embolization Device in Unruptured Intracranial Aneurysms: A Real-World, Propensity Score Matching Study

To the Editor: We read with interest the outstanding work by Bao et al,1 who compared outcomes of lattice flow diverters and pipeline embolization devices for unruptured intracranial aneurysms using propensity score (PS) matching to estimate marginal treatment effects. Although in this commendable study, the authors attempted to address potential imbalance between groups and acknowledged the risk of residual bias classically stemming from nonrandomized studies, biased estimates may more plausibly arise from methodological decisions than from study design alone in this case. The inclusion of 22 covariates in the PS model despite a modest postmatching sample size increases the risk of model overfitting. When too many variables are included relative to the sample size, PS estimates can become unstable, and matching may paradoxically worsen imbalance rather than improve it.2 Indeed, Figure 5 shows that several covariates remain outside established thresholds even after matching, with some showing greater deviation post-adjustment—patterns consistent with overconditioning. Equally important, the study does not show whether the matched groups were truly comparable across the full PS range. Displaying histograms of PS distributions would allow readers to assess overlap and judge whether the devices were evaluated in similar patient populations.3,4 Limited overlap would indicate that groups were not entirely comparable across the covariates included in the PS model, reducing confidence in the estimated treatment effect. In addition, the inclusion of nonbaseline variables (such as stent distance and angiographic follow-up time) in the matching model violates temporal ordering, an essential requirement when utilizing PS matching to compare 2 treatment strategies. Conditioning on factors that occur after treatment can introduce collider bias and distort effect estimates, leading to misleading inferences about treatment efficacy.5 More broadly, these shortcomings underscore the importance of explicitly representing the causal structure of a comparison of 2 interventions. Mapping the assumed causal structure through a directed acyclic graph clarifies temporal relationships, distinguishes confounding variables from mediators or colliders, and guides selection of an appropriate covariate set for PS modeling.6 Doing so reduces the risk of inappropriate conditioning and improves the validity and reproducibility of studies that ask a causal question. Although not yet routine in the field, E-values can quantify how strongly an unmeasured confounder would need to relate to both treatment and outcome to explain away an observed effect, providing readers a clearer sense of the findings' robustness.7 This study contributes valuable early data comparing 2 novel approaches in cerebrovascular surgery. Nonetheless, the analytic characteristics described above may have yielded results that seem rigorously controlled yet remain biased. We recognize that these comments may seem overly technical to some readers; however, careful attention to model specification, overlap, and temporal assumptions is essential for ensuring that PS-based analyses yield valid conclusions.