Modeling Longitudinal Gambling Data: Challenges and 1 Opportunities 2 Kristoffer Magnusson 3 Centre for Psychiatry Research, Department of Clinical Neuroscience, 4 Karolinska Institutet 5 Anders Nilsson 6 Centre for Psychiatry Research, Department of Clinical Neuroscience, 7 Karolinska Institutet 8 Per Carlbring 9 Department of Psychology, 10 Stockholm University 11 2019-09-11 12 Abstract 13 Clinical studies investigating treatments for problem gambling or gambling disorder frequently use gambling expenditure, such as gambling losses, as a treatment outcome. Gambling losses frequently vary substantially between participants; some report no losses, and some report substantial losses. In this article, we review how gambling losses are commonly analyzed in treat- ment studies, and show that frequently used methods, such as a log(y + 1) transformation or assuming a normal distribution, often perform poorly for these types of data. We propose that a marginalized longitudinal two-part model is a more attractive option. The models are compared using real data from a trial including 136 persons with gambling disorder. Additionally, different performance metrics are further evaluated in a Monte Carlo simu- lation study. We conclude that gambling researchers should consider using the longitudinal two-part model as it offers a flexible and powerful way of modeling gambling outcomes. The log(y + 1) transformation can be highly misleading in typical gambling data, as a difference in the number of zeros leads to biased estimates of the treatment effects. Keywords: semicontinuous data, two-part models, longitudinal gambling data, power analysis 14 Gambling disorder is recognized as an addictive disorder (Petry, Blanco, Stinchfield, 15 Correspondence concerning this article should be sent to Kristoffer Magnusson, Centrum för psykiatri- forskning, Norra Stationsgatan 69, 113 64 Stockholm, Sweden. E-mail: kristoffer.magnusson@ki.se