We developed and evaluated calibration models predicting objectively measured sitting, standing and walking time from self-reported data using a compositional data analysis (CoDA) approach. A total of 98 office workers (48 women) at the Swedish Transport Administration participated. At baseline and three-months follow-up, time spent sitting, standing and walking at work was assessed for five working days using a thigh-worn accelerometer (Actigraph), as well as by self-report (IPAQ). Individual compositions of time spent in the three behaviors were expressed by isometric log-ratios (ILR). Calibration models predicting objectively measured ILRs from self-reported ILRs were constructed using baseline data, and then validated using follow-up data. Un-calibrated self-reports were inaccurate; root-mean-square (RMS) errors of ILRs for sitting, standing and walking were 1.21, 1.24 and 1.03, respectively. Calibration reduced these errors to 36% (sitting), 40% (standing), and 24% (walking) of those prior to calibration. Calibration models remained effective for follow-up data, reducing RMS errors to 33% (sitting), 51% (standing), and 31% (walking). Thus, compositional calibration models were effective in reducing errors in self-reported physical behaviors during office work. Calibration of self-reports may present a cost-effective method for obtaining physical behavior data with satisfying accuracy in large-scale cohort and intervention studies.