Data Imputation Improves Sedentary Behavior and Physical Activity Estimates in Low Wear Time Accelerometer Data

Abstract

Missing accelerometer data from low participant wear time underestimates sedentary behavior (SB) and physical activity (PA) measurements. Yet, it remains unclear if imputing data for low participant wear time improves SB and PA estimates. PURPOSE: To determine if a data imputation technique improves SB and PA estimates in accelerometer data with low participant wear time. METHODS: One-hundred participants wore an accelerometer at the hip for ≥22.0 hours/day, at least 4 days including 1 weekend day, to capture habitual SB, light physical activity (LPA), and moderate-to-vigorous physical activity (MVPA) levels. After removing sleep time (RAW; 15.9±3.5 hours/day), random 60-minute blocks of data were removed from the RAW data set until participants had a unique data set with wear time adherence at 10 hours/day. A minute-by-minute, mean data imputation technique was used to impute estimates of SB, LPA, and MVPA in place of the missing data for the 10-hour adherence level. A series of paired t-tests with a Bonferroni correction (alpha level=0.006) compared the estimates of SB, LPA, and MVPA to the RAW data set at the 10-hour adherence level. Similarly, imputed estimates of SB, LPA, and MVPA were compared to the RAW data set at the 10-hour adherence level. RESULTS: SB, LPA, and MVPA were underestimated by 163.7 (95% confidence intervals [CI]: 156.0, 171.5; p<0.0001), 138.4 (CI: 129.1, 147.9; p<0.0001), and 27.2 (CI: 24.3, 30.1; p<0.0001) minutes/day at 10-hours of wear compared to the RAW data set, respectively. When utilizing the data imputation technique at the 10-hour adherence level, SB and MVPA were underestimated by 16.8 (CI: 8.7, 24.9; p<0.0001) and 17.1 (CI: 14.5, 19.6; p<0.0001) minutes/day compared to the RAW data set, respectively. LPA at the 10-hour adherence level was overestimated by 33.9 (CI: 25.9, 41.9; p<0.0001) minutes/day compared to the RAW data set after utilizing the data imputation technique. CONCLUSION: A minute-by-minute, mean data imputation technique improved SB, LPA, and MVPA estimates in accelerometer data with low wear time adherence. Future studies should examine the impact of data imputation techniques on accelerometer data with low participant wear time.