Critics will likely argue that Leger respondents are not a random sample of Canadians, and this is true. But they represent an important segment of young adults -- those who may have income precarity.

These results suggest that online gambling problems fall heavily on these young Canadians.

AND THANK YOU FOR READING!!

Find the paper here: doi.org/10.1007/s108...

Download my analysis scripts here: osf.io/uw94q/

This project was made possible by my excellent collaborators, Drs. Sheurich, Monson, French, and Kairouz at UConcordia and UdeSherbrooke, as well as by funding from CIHR and FRQ-SC.

So, how do we make models our gambling harm detection models better?

1 - We try to account for these subgroups in large-sample research, or
2 - We use a graded intervention scheme instead of a binary decision threshold, or
3 - We select a different outcome variable with better properties.

In turn, these subgroups within the moderate-risk range varied on the specific items endorsed and several sociodemographic dimensions.

This kind of within-groups variability is exactly why a ML classifier might fail to perform optimally.

We found that possibly episodic cases (62% of sample) reported several different problems mainly occurring just sometimes, while specific problem cases (9%) reported significantly fewer problems that occurred more frequently.

We interpreted these groups as exhibiting "possibly episodic" gambling, "intermediate cases" of responding, and "specific problems", respectively.

We applied K-means clustering to the variance of individual PGSI responses in a large dataset of > 18,000 online gamblers.

When we did this, we found a 3-cluster solution that separated groups of respondents with lower, intermediate, and higher response variance.

To explore these possible subgroups, and to improve future ML detection systems, we took a variance clustering approach to the moderate risk group of the PGSI.

Thus, there are actually several ways one might land in the Moderate-risk range, including:

-Escalating towards problem gambling
-Remitting from problem gambling
-Applying non-standard question interpretations
-Reporting 1-2 problems very persistently, or
-Reporting up to 7 problems episodically

When we look under the hood, we can easily see why the PGSI might have problems: it confounds (1) whether a problem is reported to occur, and (2) how often that problem occurs.

This is bad because clinically-defined gambling disorder can be either persistently occurring, or *episodic*.

It works reasonably well when it comes to separating the "non-problem" and "problem gambler" groups, but other research has shown some real issues with those in-between groups of low and moderate risk (Currie et al., 2013; Williams & Volberg, 2014).

The Problem Gambling Severity Index is a very popular questionnaire that has been used for more than 20 years to partition non-clinical samples into risk groups of "non-problem" (PGSI = 0), "low-risk" (1-2), "moderate risk" (3-7), or "problem gambler" (8-27) groups.