Patient-reported outcomes may be 'missing not at random' in hip and knee arthroplasty
Simon Kornvig1,2, Thomas H Jakobsen3, Kirill Gromov4
1Department of Orthopaedic Surgery, Lillebaelt Hospital, Vejle, Denmark.
Aims:
The reporting of patient-reported outcome measures (PROMs) has become essential to the assessment of the outcome of treatment in many branches of medicine and surgery. However, missing PROM scores often present practical and statistical difficulties. The aim of this study was to investigate the missingness mechanism of PROM scores in patients who underwent hip or knee arthroplasty at three fast-track centres in Denmark by comparing demographics, mortality, the risk of revision, and PROM scores between responders and non-responders.
Methods:
This population-based cohort study included 6,300 primary hip and 4,964 primary knee arthroplasties from three fast-track centres in Denmark between January 2016 and September 2021 using the Danish arthroplasty registries. The Oxford Hip/Knee Score (OHS/OKS), EuroQol five-dimension three-level/five-level questionnaire, and EuroQol visual analogue scale scores were collected before and one year after surgery. Unadjusted relative risks (RRs) of death and revision were estimated with 95% CIs using binary regression. Median differences (MDs) in PROM scores with 95% CIs were calculated using bootstrapping.
Results:
Patients with and without baseline and follow-up scores had similar age, sex, Charlson Comorbidity Index, and BMI data. However, hip and knee patients with missing baseline scores had significantly increased RRs of death within one year of 6.5 (95% CI 3.8 to 11.0; p < 0.001) and 6.2 (95% CI 3.7 to 10.5; p < 0.001), respectively. Yet, missing baseline was not significantly associated with the risk of revision. Finally, missing follow-up was significantly associated with lower baseline OHS and OKS with MDs of -3 (95% CI -4.0 to -2.0; p < 0.001) and -3 (95% CI -4.3 to -1.7; p < 0.001), respectively.
Conclusion:
Non-responders had significantly increased mortality and significantly worse baseline scores despite similar demographics and revision risk. Thus, PROM scores may have been 'missing not at random' given our covariates, indicating that multiple imputation and/or multilevel models may not be sufficient to overcome the problem of missing PROM scores.
