Level of Agreement Between Problem Gamblers’ and Collaterals’ Reports: A Bayesian Random-Effects Two-Part Model

Abstract

This study investigates the level of agreement between problem gamblers and their concerned significant others (CSOs) regarding the amount of money lost when gambling. Reported losses were analyzed from 266 participants (133 dyads) seeking treatment, which included different types of CSO–gambler dyads. The intraclass correlation coefficients (ICCs) concerning the money lost when gambling during the last 30 days were calculated based on the timeline followback. In order to model reports that were highly skewed and included zeros, a two-part generalized linear mixed-effects model was used. The results were compared from models assuming either a Gaussian, two-part gamma, or two-part lognormal response distribution. Overall, the results indicated a fair level of agreement, ICC = .57, 95% CI (.48, .64), between the gamblers and their CSOs. The partner CSOs tended to exhibit better agreement than the parent CSOs with regard to the amount of money lost, ICCdiff = .20, 95% CI (.03, .39). The difference became smaller and inconclusive when reports of no losses (zeros) were included, ICCdiff = .16, 95% CI (− .05, .36). A small simulation investigation indicated that the two-part model worked well under assumptions related to this study, and further, that calculating the ICCs under normal assumptions led to incorrect conclusions regarding the level of agreement for skewed reports (such as gambling losses). For gambling losses, the normal assumption is unlikely to hold and ICCs based on this assumption are likely to be highly unreliable.

Internet-delivered cognitive-behavioral therapy for significant others of treatment-refusing problem gamblers: A randomized wait-list controlled trial.

“…The secondary outcomes were 1) gambling-related outcomes, as measured by the timeline follow-back (TLFB; Weinstock, Whelan, & Meyers, 2004) and treatment engagement. Studies have found that treatment-seeking gamblers and CSO are in fair agreement (ICCs just below 0.6) regarding the amount of self-reported losses using TLFB (Hodgins & Makarchuk, 2003; Magnusson, Nilsson, Andersson, Hellner, & Carlbring, 2019); 2) CSO’s depressive feelings, as measured by the Patient Health Questionnaire (PHQ-9; Kroenke, Spitzer, & Williams, 2001), anxiety, as measured by the Generalized Anxiety Disorder Scale (GAD-7; Spitzer, Kroenke, Williams, & Löwe, 2006), and quality of life, as measured by the 26-item version of the World Health Organization’s Quality of Life assessment (WHOQOL-BREF; Skevington, Lotfy, O’Connell, & the WHOQOL Group, 2004); and 3) relationship satisfaction, as measured by the relationship assessment scale (RAS; Rask et al, 2010). Measurements were administered at pretest and posttest and 3, 6, and 12 months after posttest.…”

Assessing the importance of individual‐ and colony‐level variation when using seabird foraging ranges as impact assessment and conservation tools

“…Moreover, the choice of distribution and link function can influence the ICCs calculated (Magnusson et al . 2019). We compared three possible distributional models (see Appendix S3): a gamma GLMM (with log link), a Gaussian model and a log‐normal model.…”

“…foraging ranges > 1000 km in Kittiwake, Guillemot and Razorbill), far beyond that which we observed and which could influence the reliability of calculated ICCs (Magnusson et al . 2019). Moreover, consideration of which distribution best reflects observed foraging ranges per trip is important because it can, in turn, be used to generate predicted foraging range radii for untracked colonies to help understand species space use using posterior distributions (e.g.…”

“…for the gamma, and other standard distributional choices in GLMMs) be linked to the mean for distributions other than the Gaussian, so these models assume that there is an irreducible source of variation, which therefore implies that on the original scale, ICCs from GLMMs are not able to reach a theoretical value of 1.0 (de Villemereuil et al 2016, Nakagawa et al 2017. Moreover, the choice of distribution and link function can influence the ICCs calculated (Magnusson et al 2019). We compared three possible distributional models (see Appendix S3): a gamma GLMM (with log link), a Gaussian model and a log-normal model.…”