Benchmarking gambling screens to health-state utility: the PGSI and the SGHS estimate similar levels of population gambling-harm – BMC Public Health -…

Our analysis was based on a comparison of health utility scores between unharmed / non-problem gamblers, for the SGHS and PGSI, respectively, who had participated in gambling at least once in the last year (hereafter, the control group) and those experiencing some degree of harm or problems (hereafter, the affected group). It is important to note that the control group for the SGHS and PGSI analyses were slightly different, as some respondents may have scored 0 on the SGHS and therefore been in the control group for SGHS-based analyses, but scored more than 0 on the PGSI and therefore been in the affected group for PGSI-based analyses. Sampling was stratified with respect to group, age and gender. Cases were propensity weighted based on key risk factors, and regression-estimated coefficients were estimated with control variables for gambling comorbidities. Similar analyses were run using the PGSI and the SGHS to define the reference (score 0) and affected (score 1+) groups. Categorical, linear and non-linear utility functions of 1+ scores were compared.

Australian participants aged 18+ were recruited from a commercial panel provider during late 2020 and early 2021 as part of a broader project to study gamblers, non-gamblers, and concerned significant others. The commercial panel has their network of respondents who have signed up to take part in research opportunities. The panel invited respondents through email and all data was collected online. As compensation, participants received points which could be exchanged for rewards as per the panels internal points-accumulation system.

All eligible participants were required to be Australian residents, aged 18years or above, provide consent to participate in the study, and to have gambledFootnote 1 in the past 12months. Residents of the state of Victoria were excluded due to COVID lockdown at time of sampling. Using soft-quotas, we attempted to sample approximately equal groups with respect to age (18-29, 30-44, 45+) and gender with respect to control / affected group status. A total of 22,699 started the survey, however 16,061 were screened out for the following reasons: 5848 did not meet the residency or age criteria, 5922 provided incomplete responses, 441 provided poor quality data (such as straight lining through the survey), and 3850 were excluded due to quotas being full. A total of 6638 responses were retained, of which 2603 were gamblers and formed part of the present analysis, with 1193 (45.8%) scoring zero on both population screens. Table1 provides the demographic characteristics for gamblers and figures are presented separately for gamblers who scored zero and 1+ on each screen. For the SGHS 1546 gamblers (59%) scored 0 and 1057 (41%) scored 1+, and for the PGSI 1331 (51%) scored 0 and 1272 (49%) scored 1+. The most common forms gambled on included lotteries (82.1% of sample), electronic gaming machines (65.3%), scratch tickets (64.0%), race betting (63.8%), raffle tickets / competitions (62.9%), sports betting (43.6%), and Keno (41.3%). Less than one-third of participants gambled on all other forms (casino table games, informal private betting, prize draws, Bingo, eSports, fantasy sports, and other).

All participants completed the following measures. Problem gambling status was assessed using the PGSI. The PGSI uses nine items (e.g. have you bet more than you could really afford to lose?) with each item measured on a four-point scale (from 0=never to 3=almost always). Total scores are summed and risk categories are yielded (non-problem 0, LR 1-2, MR 3-7, PG 8+) [2]. Reliability for the PGSI was high in the current sample (=0.95).

Gambling harm was assessed using the SGHS. The SGHS comprises 10-items (e.g. had regrets that made me feel sorry about my gambling) each measured in a binary no/yes format. The SGHS captures financial, emotional/psychological, and relationship harms due to gambling and yields scores 0-10 [3] however the screen does not specify categories. Nonetheless, recent research assessing the SGHS using the Personal Wellbeing Index suggests that cut-offs of 1-2, 3-5, 6+ provide a reasonable categorisation of differing degrees of harm [16]. Reliability for the SGHS was high in the current sample (=0.90).

We measured health utility using the SF-6D (see [32] for a detailed description). The SF-6D is a preference-based measure derived from the SF-12 item self-report measure [33]. It captures physical functioning, role limitations, social functioning, pain, mental health, and vitality, and yields health utility coefficients between 0.345 to 1.000 [34].

Demographic characteristics identified as risk factors for gambling problems and harms [17] were considered for inclusion in the propensity model: gender, country of birth, personal and parents highest level of education achieved, selected work status flags (FT student, unemployed, being unable to work due to infirmity, labourer), marital status, household composition (e.g. single, couple with children), personal and household income, and metropolitan/regional/rural residential location. Psychological risk factors such as cognitive style or rash impulsivity were measured but excluded due to potential endogeneity, particularly with respect to gambling problems when considered as a mental health condition.

The following key co-morbidities that affect health were also measured: excessive alcohol consumption (AUDIT-C) [35], any recreational drug use, cigarette smoking frequency (single-item measures), and ever having been diagnosed with a mood disorder, anxiety disorder, personality disorder, or any other mental health disorder (separate binary indicators). The AUDIT-C is a three-item measure of hazardous drinking (e.g. how often do you have six or more standard drinks on one occasion) with each item measured on a five-point scale. Reliability for the AUDIT-C in the current sample was (=0.67).

The analyses took a multi-step approach, and all analyses were conducted for SGHS as well as PGSI. Because SGHS and PGSI are correlated, there is significant overlap between the affected and control groups for the two measures. The first step was to determine the required weights for the propensity score matching, which was based on initial logistic regressions predicting SGHS or PGSI (0 vs 1+; Propensity models in Table2). Based on these regression results, propensity score weights were used in subsequent analyses predicting SF-6D scores using SGHS (and PGSI separately) as independent variables. Known risk factors were included as covariates (Causal models in Table 2). A final set of models was run predicting SF-6D using SGHS (and PGSI separately) as independent variables, but without the risk factors as covariates, to determine the effect of the covariates on the estimated decrements.

The working for the weights in the causal models is based on the binomial logistic regression predicting harm (SGHS=1+) compared to not experiencing harm (SGHS=0), or the equivalent for PGSI (0 vs 1+). The predictors in the models were known risk factors for experiencing gambling harm or problems and were chosen for each model based on backwards stepwise elimination using the Akaike Information Criteria, to avoid redundancy and multicollinearity. The models for SGHS and PGSI therefore had slightly different predictors to each other. While stepwise variable elimination has limitations for interpreting covariates, they do not apply in this case because our objective was not to interpret these covariate effects, but to achieve statistical control.

From the logistic regressions, predicted probability of harm (or problems) was derived for each individual and then cases in each group (affected vs control) were inversely weighted with respect to these group propensities based on the standard propensity weighting method:

$${displaystyle begin{array}{c}mathrm{if}left(mathrm{affected}right):frac{1}{hat{P}left(mathrm{affected}right)}\ {}mathrm{if}left(mathrm{control}right):frac{1}{1-hat{P}left(mathrm{affected}right)}end{array}}$$

This weighting acts to remove some potential selection bias from confounders in estimating the direct effect of gambling harm on health. This is because people with different demographic and other characteristics differ in their propensity to experiencing harm or problems from gambling, and these same risk factors can also contribute directly to lower wellbeing. For example, from Table 2, younger people in the present study were more likely to experience some degree of harm. The propensity weighting balances the groups of who were actually affected / unaffected, with respect to their propensity for being affected by gambling harms or problems. For example, looking at Table 1, affected gamblers were more likely to be younger compared to controls. The process weights down younger respondents in the affected group, and weights up younger respondents in the control group, balancing the groups with respect to this particular risk factor. One issue with propensity weighting is that excessively large or small weights can lead to outside case influence. However, skew and outliers of weights were moderate (median~1.8, mean~2, max ~6), so no thresholding of excessively large weights was required.

In a supplementary analysis, shown in Table3, the empirically derived estimates were applied to population estimates of SGHS and PGSI score prevalence using a recent Victorian prevalence dataset [13], in order to estimate population aggregate impact. Finally, a standard Pearson correlation matrix (see Additionalfile1), was calculated for descriptive purposes.

Ethical approval for this study was received from the institutional Human Research Ethics Committee (#22341) and all methods were performed in accordance with the relevant guidelines and regulations. Participants provided informed consent before participating.

Read more here:

Benchmarking gambling screens to health-state utility: the PGSI and the SGHS estimate similar levels of population gambling-harm - BMC Public Health -...