Factors Correlated With Physical Function 1 Year After Total Knee Arthroplasty in Patients With Knee Osteoarthritis

This systematic review and meta-analysis investigates the correlation of preoperative and intraoperative factors with physical function 1 year after total knee arthroplasty in patients with knee osteoarthritis.

We imputed correlation coefficients from estimates of association expressed as odds ratios, risk ratios, and linear model coefficients (including differences) as described in our protocol's supplementary materials. Where it was necessary to impute odds ratios from risk ratios prior to imputing correlation, we assumed a prespecified baseline probability of reduced postsurgical function of 20%. We defined canonical directions for all outcomes and factors and inverted reported directions of association as appropriate to ensure consistent directions of association in meta-analysis.
If studies did not report confidence intervals or sampling variances, we imputed them as appropriate (Higgins 2019). If a study did not report exact statements of uncertainty but provided statements about "statistical significance", we used a conservative approach in which we imputed "worst case" standard errors. For example, we imputed P≤0.01 to mean P=0.01 and "not statistically significant" to mean P=0.99. We performed all metaanalyses on the scale of Fisher's z (hyperbolic arctangent, not Z-score; Borenstein 2009). We used the inverse transform (hyperbolic tangent) to report meta-analytical estimates as correlation coefficients.
We anticipated that factors may be correlated and that there may be important differences in the methods used to quantify associations. We therefore planned to perform multivariate random-effects meta-analysis for each outcome using White's (2009White's ( , 2011 multivariate extension to Riley's (2008) bivariate random-effects model, as implemented in the MVMETA add-on command for Stata. Unfortunately, it was not possible to fit this model given the sparsity of our data. We therefore used a frequentist version of the Bayesian multivariate model we developed for a meta-analysis of pain after total knee arthroplasty (Rose 2020). We had planned to identify factors likely to be most strongly associated with postoperative function by estimating the probability of superiority of each factor using the pbest option of MVMETA. Because that model could not be used, we assessed using P-scores (cf. p-values;Rücker and Schwarzer 2015), in which larger magnitudes were defined to be superior to those with smaller magnitudes. Unlike the probabilities we had planned to estimate, P-scores are not as heavily influenced by imprecisely estimated factors with small point estimates whose confidence intervals extend far beyond those of more precisely estimated factors with larger point estimates. This is particularly important for multivariate meta-analysis of correlations, in which the superiority of a factor is a function of the magnitude of its coefficient rather than its magnitude and direction, as is the case in multiple treatment comparison via network meta-analysis. P-scores are therefore likely to better identify good factors. Multivariate estimates of correlation are presented as forest plots, which also show I² statistics (the percentage of heterogeneity attributable to between-study differences rather than sampling error) and the numbers of studies that provided usable estimates for each factor. We also performed exploratory univariate meta-analyses for each factor and outcome, but which do not account for correlation. We compared estimates from the three approaches to identify possible inconsistency. We report 95% confidence intervals throughout. Statistical analyses were performed using Stata 16 (StataCorp LLC, College Station, Texas, USA).
We had planned to investigate non-reporting bias and small study effects for factor supported by at least 10 results. However, none of the factors met this criterion. Similarly, we had planned to perform subgroup analyses with respect to study design, type of outcome measurement, and intervention if at least five studies could be included in each subgroup. However, this criterion was not satisfied, and no subgroup analyses were performed.
We performed a sensitivity analysis for the primary outcome (function 12 months postsurgery). For each of the six QUIPS risk of bias domains, we excluded studies judged to be at high risk of bias, re-ran the multivariate meta-analysis, and compared the estimated correlations with those obtained when all studies are included. We had planned to do a leave-one-study-out sensitivity analysis to explore the influence of each study on the meta-analysis results. Unfortunately, this was not feasible. However, the effect of particular studies can be inferred by inspecting the univariate meta-analyses. The following plot shows the multivariate meta-analytical estimates of correlation at each postoperative follow-up time. Estimates for factor studied at only one postoperative time point are omitted.

eFigure 1. Sensitivity Analysis
The following plot shows the results of a sensitivity analysis in which the multivariate meta-analysis model was used to estimate correlations for each factor, omitting all estimates from studies judged to be at high risk of bias for each of the six QUIPS domains. Estimates from the full meta-analysis are also included for comparison (shown as "All Studies").

. Sensitivity Analysis
The following table shows estimates of correlations for each of the sensitivity analyses. Estimates for the main analysis (i.e., no studies omitted) are also shown for comparison. 28 (function* or stiffness or contracture*).tw,kw. 5229428 29 (muscle adj3 (strength* or weakness or fatigue or tonus)).tw,kw. 76148 30 (sitting or lying or standing or balance or posture or rising or neeling or bend* or walk* or gait or stair* or extension* or stability or contracture* or movement* or motion* or locomotion* or mobility or twisting or pivoting or straighten* or swelling or grinding or clicking or squatting or running or jumping).tw,kw.