Abstract
OBJECTIVES: Despite the possible large number of missing values on the 25-question Geriatric Locomotive Function Scale (GLFS-25), how we should treat them is unknown. In a simulation study, we investigated how to handle missing values in the GLFS-25. DESIGN, SETTING AND PARTICIPANTS: We used three datasets with different participant characteristics: community dwellers who could walk by themselves, outpatients of orthopaedics owing to pain, and patients who required surgery for total knee replacement or lumbar spinal canal stenosis. OUTCOME MEASURES: The missing items of the datasets were artificially created, and four statistical methods, complete case analysis, multiple imputation, single imputation using individual mean, and single imputation using individual domain average, were compared in terms of bias and mean squared error. Simulation studies were conducted to compare them under varying numbers of participants with missing values (5%-40%) and under varying numbers of missing items of GLFS-25 (4-16). RESULTS: Multiple imputation had the lowest root mean squared error. Complete case analysis showed the largest bias, and the performances of the single imputation were between those methods. The relative performances were similar across the three datasets. The absolute bias of the single imputation was<0.1. The bias and mean squared error of multiple imputation and single imputation were comparable when the number of missing items was less than or equal to eight. CONCLUSIONS: Multiple imputation is preferable, although single imputation using subject average/subject domain average can be used with practically negligible bias as long as the number of missing items is up to 8 out of 25 items in each individual of the population.