Changes in severity of problem gambling and subsequent suicide attempts: a longitudinal survey of young adults in Great Britain, 2018–20

Summary Background Cross-sectional studies identify problem gambling as a risk factor for suicidality. Using an online longitudinal survey, we aimed to examine the association between changes in severity of gambling behaviour and attempted suicide. Methods The Emerging Adults Gambling Survey is a longitudinal survey of people in England, Scotland, and Wales, aged 16–24 years interviewed online between June 25 and Aug 16, 2019 (wave 1) and 1 year later between July 13 and Oct 8, 2020 (wave 2). The Problem Gambling Severity Index (PGSI) was administered at both waves. Multivariable logistic regression models examined wave 1 PGSI score and between-wave change in PGSI score as risk factors for suicide attempts at wave 2, unadjusted and with adjustment for wellbeing, anxiousness, impulsivity, perceived loneliness, and suicide attempts at wave 1. Findings 3549 participants were interviewed in wave 1 and 2094 were interviewed in wave 2, of whom 1941 were included in this analysis (749 [39%] men; 1192 women [61%]). Prevalence of attempted suicide did not change between waves (wave 1: 3·7% [95% CI 2·9–4·8], n=75; wave 2: 3·3% [2·5–4·3], n=65). 78·9% (95% CI 76·7–80·9, n=1575) of participants had stable PGSI scores between the two waves, 13·7% (11·9–15·6, n=233) of participants had a decrease in PGSI score by 1 or more, and 7·5% (6·2–8·9, n=133) had an increase in PGSI score by 1 or more. An increase in PGSI scores over time was associated with suicide attempt at wave 2, even with adjustment for baseline PGSI score and other factors (adjusted odds ratio 2·74 [95% CI 1·20–6·27]). Wave 1 PGSI score alone was not associated with suicide attempt at wave 2 in fully adjusted models. Interpretation Repeated routine screening for changes in gambling harm could be embedded in health, social care, and public service settings to allow effective identification and suicide prevention activities among young adults. Funding Wellcome Trust.


Appendix B: generation of longitudinal weights between wave 1 and wave 2 of the Emerging Adults Gambling Survey Methodology
To generate the longitudinal weights, YouGov built a model that predicts which people are likely to drop out between wave 1 and wave 2. They then upweight people who were more likely to have dropped out (based on the predicted value), but did not actually dropout. Below are the steps followed to generate the attrition weights used in this study: 1. Take the wave 1 data, and add an attrition variable that tells us whether this respondent answered wave 2 (coded 0 = answered, 1 = did not answer) 2. Run a logistic regression with "attrition variable" as the dependent variable, and the "candidate variables" as predictor variables (see below for details). We then generate predicted probabilities from the regression modelling 3. Drop cases from the datafile who did not respond at wave 2 4. Divide each predicted value by the mean predicted value after dropping wave 2 cases (for example, if respondent has a predicted probability of 0.375, and the mean predicted probability is 0.25, then 0.375 / 0.25 = 1.5. This is the attrition weight 5. Multiply the wave 1 weight (which matched the responding sample in wave 1 to the age, gender, deprivation and regional profile of young people living in Great Britain) by the attrition weight. This will be the wave 2 longitudinal weight. This is the weighting variable used in this analysis.

Summary
The logistic modelling conducted in this study is based on these attrition predictor variables collected at wave 1:

W1sex
Gender, coded men/women. W1AgeBands Age (grouped), coded 16-18, 19-21, 22-24 years. W1ethnicg Self-reported ethnic identity (grouped), coded White/White British; Mixed; Asian/Asian British; Black/Black British; Other. W1neet Whether in Employment, Education or Training, coded yes/no. W1anystudent Whether currently in higher education, coded, yes/no. W1qimd Area Deprivation Decile, coded from least to most deprived. W1gamfreq2 Frequency of gambling on any activity, coded more than weekly; about weekly; about fortnightly; about monthly; a couple of times a year; did not gamble. W1anyacty Whether gambled on any activity in the past year or no, coded yes/no. W1ngamyr Number of gambling activities undertaken in the past year, range from 0 to 15. W1pgsiprob Problem Gambling Severity Index status, coded 0 = nongambler/non-problem gambler; 1-2 low risk gambling; 3-7 moderate risk gambling; 8+ problem gambling.
Four variables were significant in the model result: W1sex (Gender) , W1ethnicg (Ethnicity), W1qimd (Area-deprivation) and W1gamfreq2 (Gambling). The model shows that women were less likely to drop out than men. Participants of Black/Black British origin were less likely to drop out than those from White/White British backgrounds. Those whose area deprivation status was unknown were more likely to drop out than those where this was categorized. Those who gambled were more likely to drop out than those who did not gamble. This was particularly evident for those gambling fortnightly.

Attrition by suicide attempt status
For the analysis presented in this paper, we examined attrition rates by suicide attempt status at wave 1 (see Table B1 below). This shows that those who reported attempting suicide at wave 1 were no more likely than those who had not reported this to drop out of the survey. Respective attrition rates were 43.5% for those who had attempted suicide at wave 1 and 41.5% for those who had not. Because of small base sizes, we were not able to look at whether the profile of those who had attempted suicide in wave 1 (n=64) but dropped out was systematically different from those who had attempted suicide in wave 1 (n=83) but did not drop out. However, we note that our attrition weights do adjust for age, gender and area deprivation. We have noted this as a potential limitation of our article.

Appendix C: Treatment of missing values in analysis
This document sets out the number of missing values for each variable included in the analysis and the treatment applied to those values within the modelling. Analysis is based on those who took part in both waves (n=2080), thus missing values are presented for this group. This data uses in this analysis is from the Emerging Adults Gambling Survey. That original survey was designed to examine individual gambling trajectories over time. Sample size calculations were based on being sufficient to estimate change in gambling prevalence between waves, which was not the endpoint of the current analysis. Assuming a between wave correlation of 0.5, the study was designed to be able to detect changes in problem gambling behaviours of +/-0.3 percentage points (at 80% power).

Appendix E: Analysis focusing on suicide attempts
In supplementary tables 1 and 2, we repeat the analytical procedures documented in the main manuscript using past year suicidal thoughts as the outcome measure. Following the approach developed for the Adult Psychiatric Morbidity Survey 2007, the following question was asked at both wave 1 and wave 2: "In the last 12 months, have you ever thought of taking your life, even if you would not actually do it?". This was binary coded Yes (1) and No (0). As with suicide attempts, 139 participants did not answer this question and were excluded from analysis. Tables S1 show the unadjusted relationship between PGSI change scores and PGSI score at wave 1 with experience of suicidal thoughts in year prior to the wave 2 interview. Table S2 shows the adjusted relationship, using the same set of controls used in the main analysis, which were also all significantly associated with suicidal thoughts at wave 2, and additionally controlling for gender, which was also significant. Other analyses 17 Report other analyses done-eg analyses of subgroups and interactions, and sensitivity analyses