Letter to the Editor
(Re: Rajamäki et al., Acta Orthop 2026; 97: 423-9. doi: 10.2340/17453674.2026.45965)
Citation: Acta Orthopaedica 2026; 97: 629–630. DOI: https://doi.org/10.2340/17453674.2026.46698.
Copyright: © 2026 The Author(s). Published by MJS Publishing – Medical Journals Sweden, on behalf of the Nordic Orthopedic Federation. This is an Open Access article distributed under the terms of the Creative Commons Attribution 4.0 International License (https://creativecommons.org/licenses/by/4.0/)
Published: 2026-09-07.
Complete disclosure of interest forms according to ICMJE are available on the article page, doi: 10.2340/17453674.2026.46698
Sir—, Rajamäki and colleagues were interested in identifying preoperative predictors of persistent pain 1-year post-surgery in patients undergoing knee arthroplasty [1]. Given the clinical importance of persistent pain following this typically highly successful surgery, prognostic studies of poor pain outcome are common in the literature [2,3], including a few of our prior publications [4-6].
The investigators chose 3 pain outcomes of interest. The primary outcome was a dichotomized single-item pain severity measure of either moderate or severe pain, derived from the first question in the Oxford Knee Scale (OKS), the second was a dichotomized final score of < 25 for the full OKS (ranging from 0 to 48 with higher scores equating to better outcome), and third was a change of < 8 points for the OKS, less than the minimal clinically important difference (MCID) reported for the OKS [7]. For the MCID outcome, participants had to have a preoperative OKS score of < 28 to reduce ceiling-effect risk. Regarding choice of outcomes, all were derived from the OKS and were therefore not independent. Additionally, dichotomizing all outcomes using arbitrary cut points reduces variability and increases measurement error.
Notably, the extent of missing data for each outcome was also substantial, ranging from 52% to 69%. The investigators indicated they assumed the data was missing completely at random, an assumption that was unlikely given that postsurgical outcome data is more likely to be missing not at random [8]. For example, participants with worse mental health and greater sociodemographic challenges are less likely to provide follow-up measures.
Given the extensive variation in MCID estimates for knee arthroplasty [9,10], and the lack of validity of these approaches [11], we have proposed an alternative prognostic approach: the use of latent growth curve modeling to statistically identify good vs poor pain and function outcome trajectories following knee arthroplasty [4,12]. This method has been externally validated [6], does not rely on arbitrary cut points, accounts for random measurement error, is scientifically falsifiable, and allows for identification of predictors of poor outcome. The investigators did not address this more contemporary and, we believe, superior prognostic approach to estimating the probability of poor pain outcome for patients undergoing knee arthroplasty. Either the latent variable modeling approaches or the latent difference score model [13] is a viable alternative to arithmetic difference score-based anchor-driven MCID calculation with AUCs in the 0.77–0.90 range. The investigators’ reported AUC range of 0.64–0.68 [1] calls for improvements in machine learning algorithms by better handling measurement and methodological issues in predictive modeling.
Daniel L Riddle 1 and Levent Dumenci 2
1 Distinguished Career Emeritus Professor of Physical Therapy, Orthopaedic Surgery and Rheumatology, Virginia Commonwealth University, Richmond, VA;
2 College of Public Health, Department of Epidemiology and Biostatistics, Temple University, Philadelphia, PA, USA
Correspondence: dlriddle@vcu.edu