Best practice recommendations to measure and estimate workplace accident underreporting

• This paper provides recommendations to measure and analyze underreporting data. • Multi-item and recognition-based scales should be used. • Linear, Poisson & negative binomial models return similar results in medium samples. • Estimation using small samples should be avoided. • More than one model should be estimated & compared to other reasonable alternatives. Despite research indicating work-related accidents and injuries are extensively underreported, measurement and analytical strategies to investigate this phenomenon have widely varied. Using the results from three studies, we provide six best practice recommendations for applied scholars in underreporting research based on a comprehensive review across several disciplines (Study 1) of measurement and analytical choices and investigate which of those are the most appropriate to model counts of underreported workplace accidents using both synthetic (Study 2) and real-world (Study 3) data. These recommendations are: (i) use multi-item recognition-based scales, (ii) use counts (as opposed to proportions) but incorporate exposure information, (iii) think about the meaning of zeroes in the distribution, (iv) acknowledge that small samples are problematic regardless of model used to estimate parameters, (v) recognize that general linear model with raw outcomes and generalized linear models with negative binomial and Poisson distributions seem to be substantially equivalent in medium and large samples (i.e., N>500), and (vi) estimate more than one model and compare them using average marginal effects before reaching conclusions. These recommendations can also be extended to other research domains using count data both within and outside of occupational health psychology.