Background. Arm posture is a recognized risk factor for occupational upper extremity musculoskeletal disorders and thus often assessed in research and practice. Posture assessment methods differ in cost, feasibility and, perhaps, bias. An attractive approach could be to build statistical models for predicting results of expensive direct measurements of arm posture from cheaper or more accessible data, and apply them to large samples in which only the latter data are available. We aimed to build and assess the performance of such prediction models in a random sample of paper mill workers.

Methods. 28 workers were recruited to the study, and their upper arm postures were measured during three full work shifts using inclinometers. Simultaneously, the workers were video filmed, and their arm posture and gross body posture were assessed by observing the video afterwards. Models for predicting the inclinometer-assessed duration (proportion of time) and frequency (number/min) of periods spent in neutral right arm posture (<20°) were fitted using subject and observer as random factors, measured shift (1, 2 or 3) as fixed factor, and either observed time in neutral right arm angle or observed gross body posture as predictor.

Results. For the proportion of time spent in neutral arm posture, the best performance was achieved by using observed gross body posture as predictor (explained variance: R2=26%; standard error: SE=9.8). For the frequency of periods spent in neutral arm posture, the corresponding model fit was R2=60% and SE=5.6. Bootstrap resample validation of the latter model showed an expected performance in other samples of R2=59-60% and SE=5.5-5.6 (5th-95th percentile).

Discussion. Surprisingly, we found that observed gross body posture was a better predictor of variation in arm posture than observed arm angles. The findings suggest that arm posture during paper mill work can be cost-efficiently assessed using simple observations.