A Case for Survival Analysis in Shoulder Stiffness Research: Letter to the Editor

Dear Editor:

Feng et al ² address an important clinical question: whether postoperative metformin reduces shoulder stiffness after arthroscopic rotator cuff repair. The randomized design is a strength, and the authors should be commended for their careful surgical standardization and blinding procedures. Several aspects of the analysis, however, make the causal interpretation of the findings more difficult. We offer the following comments constructively.

First, stiffness is analyzed as a simple proportion at 3, 6, and 12 months, with group differences evaluated using the chi‐square or Fisher tests. Yet stiffness is a time‐to‐event outcome: patients may develop it at different times, and—as shown in the CONSORT (Consolidated Standards of Reporting Trials) diagram—20 of the 146 randomized participants were lost to follow‐up.When follow‐up time varies across individuals, a simple proportion (ie, number of events/total number of participants at baseline) fails to accurately capture the cumulative incidence. ⁶ It assumes identical follow‐up duration and noninformative censoring—conditions not met here. For example, if 6 of 10 patients develop stiffness by 24 months, but 3 are lost to follow‐up earlier, calculating 6/10 = 60% misrepresents the true incidence, as we do not know what happened to the censored patients. Time-to-event methods, such as Kaplan-Meier curves or Cox regression, are designed precisely for this situation. ³ They account for differential follow‐up, administrative censoring, and loss to follow‐up, producing valid estimates of cumulative incidence when the assumption of noninformative censoring holds.

It is important to recognize that noninformative censoring is a strong assumption, particularly for participants who are lost to follow‐up.For time‐to‐event analysis to yield unbiased estimates, the probability of censoring must be independent of the probability of developing stiffness. This is more defensible for administrative censoring (eg, late enrollment) than for dropout. If participants drop out due to worsening symptoms or recovery (ie, informative censoring), survival analysis estimates may be biased even if time‐to‐event methods are used. In such cases, sensitivity analyses, inverse probability weighting, and marginal structural models can help assess robustness to censoring assumptions.^4,5

Although the study is described as intention‐to‐treat, exclusion of the 20 randomized participants who were lost to follow‐up results in a complete‐case analysis. Removing participants based on post‐randomization events undermines the balance created by randomization and introduces potential for selection bias, especially when the outcome is relatively uncommon and may occur early. A true intention‐to‐treat analysis ⁷ would include all randomized participants and address missing outcomes using imputation or sensitivity analyses to reflect the uncertainty introduced by loss to follow‐up.

In addition, the effects are presented only on the multiplicative scale (eg, a risk ratio of 0.40 at 3 months). While relative measures are helpful, absolute measures—such as the risk difference or number needed to treat—offer essential clinical insight. As stiffness largely resolves by 1 year in both groups, risk differences over time would clarify the duration and magnitude of benefit. These absolute measures can be derived from risk‐based regression models or survival curves and should complement relative effect estimates to avoid overinterpreting modest effects that appear large on a multiplicative scale. ¹

In sum, the study addresses an important clinical problem, and the randomized design offers a valuable opportunity to evaluate the effect of metformin on postoperative stiffness. An analysis that (1) uses survival methods to account for differing follow‐up time, (2) evaluates the assumption of noninformative censoring, (3) reports both relative and absolute effect measures, and (4) adheres fully to the intention‐to‐treat principle would more directly estimate the causal effect of treatment assignment and strengthen confidence in the study's conclusions.

Sincerely,

Lucas Mena, MD Jorge Henrique Assunção, MD, PhD Eduardo Angeli Malavolta, MD, PhD São Paulo, Brazil