Analysis of a national gambling and violence study using a block-wise selection process
Janet E. Pearson1, Maria Bellringer2, Katie Palmer du Preez2, Denise Wilson3, Jane Koziol-McLain4, Nick Garrett1, and Max Abbott2
1. Department of Biostatistics and Epidemiology, Auckland University of Technology
2. Gambling and Addictions Research Centre, Auckland University of Technology
3. Taupua Waiora Centre for Maaori Health Research, Auckland University of Technology
4. Interdisciplinary Trauma Research Centre, Auckland University of Technology
We have data on a convenience sample of 164 adult (18+) gamblers recruited from New Zealand gambling treatment services, as part of a larger study to investigate the occurrence of family violence. Here, we are interested in the effect of gambler gender on the relationship between having dependent children at home, and the gambler perpetrating or being a victim of family violence. Further, we wish to adjust for the many inter-related potential confounder variables collected, and to do so in a way that illuminates the effect of adjusting for variables in the macro social domain, and then the more micro psychological and gambling domains. The identified process to manage this is described as follows. We sequentially add blocks of conceptually related and statistically correlated variables, to each of our two models (one predicting perpetration, and one predicting victimisation). Thus, in each model, we can see the effect on the primary predictor (the interaction of gambler gender and having dependent children) of adjusting for each block in turn. This enables us to investigate the effect of adjusting for later blocks, once earlier blocks have been taken into account. Before inclusion in the final multivariable model, only the most predictive variables are kept in each block (controlling for the primary predictor), enabling a more parsimonious final model, reduction in the number of redundant variables, reduction in multicollinearity, and better accuracy in prediction. With the primary predictor in place, we then adjust in sequence and cumulatively, for significant predictors from each block: socio-demographic, then psycho-social, and finally gambling factors, in the process removing variables that are no longer significant in the final multivariable logistic regression model. We show how this block-wise selection process uncovers interesting insights about the data.