Introduction: A cost-efficient alternative to measuring working postures directly could be to build statistical models for predicting results of such measurements from cheaper data, and apply these models to samples in which only the latter data are available. The present study aimed to build and assess the performance of statistical models predicting inclinometer-assessed trunk and arm posture among paper mill workers. Separate models were built using administrative data, workers’ ratings of their exposure, and observations of the work from video recordings as predictors.

Methods: Trunk and upper arm postures were measured using inclinometry on 28 paper mill workers during three work shifts each. Simultaneously, the workers were video filmed, and their postures were assessed by observation of the videos afterwards. Workers’ ratings of exposure, and administrative data on staff and production during the shifts were also collected. Linear mixed models were fitted for predicting inclinometer-assessed exposure variables (median trunk and upper arm angle, proportion of time with neutral trunk and upper arm posture, and frequency of periods in neutral trunk and upper arm inclination) from administrative data, workers’ ratings, and observations, respectively. Performance was evaluated in terms of Akaike information criterion, proportion of variance explained (R²), and standard error of the model estimate (SE). For models performing well, validity was assessed by bootstrap resampling.

Results: Models based on administrative data performed poorly (R²≤15%) and would not be useful for assessing posture in this population. Models using workers’ ratings of exposure performed slightly better (8%≤R²≤27% for trunk posture; 14%≤R²≤36% for arm posture). The best model was obtained when using observational data for predicting frequency of periods with neutral arm inclination. It explained 56% of the variance in the postural exposure, and its SE was 5.6. Bootstrap validation of this model showed similar expected performance in other samples (5^th-95^th percentile: R²=45-63%; SE=5.1-6.2).

Conclusions: Observational data had a better ability to predict inclinometer-assessed upper arm exposures than workers’ ratings or administrative data, but they are typically more expensive to obtain. The results encourage comparisons of the cost-efficiency of modeling based on administrative data, workers’ ratings, and observation.