Abstract
BACKGROUND: An increasing number of randomized controlled trials (RCTs) have measured the impact of interventions on work productivity loss. Productivity loss outcome is inflated at zero and max loss values. Our study was to compare the performance of five commonly used methods in analysis of productivity loss outcomes in RCTs. METHODS: We conducted a simulation study to compare Ordinary Least Squares (OLS), Negative Binominal (NB), two-part models (the non-zero part following truncated NB distribution or gamma distribution) and three-part model (the middle part between zero and max values following Beta distribution). The main number of observations each arm, N(obs), that we considered were 50, 100 and 200. Baseline productivity loss was included as a covariate. RESULTS: All models performed similarly well when baseline productivity loss was set at the mean value. When baseline productivity loss was set at other values and N(obs) = 50 with ≤5 subjects having max loss, two-part models performed best if the proportion of zero loss> 50% in at least one arm and otherwise, OLS performed best. When N(obs) = 100 or 200, the three-part model performed best if the two arms had equal scale parameters for their productivity loss outcome distributions between zero and max values. CONCLUSIONS: Our findings suggest that when treatment effect at any given values of one single covariate is of interest, the model selection depends on the sample size, the proportions of zero loss and max loss, and the scale parameter for the productivity loss outcome distribution between zero and max loss in each arm of RCTs.