|
|
The association between gambling and financial, social and health outcomes in big financial data Naomi Muggleton, Paula Parpart, Philip Newall, David Leake, John Gathergood and Neil Stewart - - 2021 Gambling is an ordinary pastime for some people, but is associated with addiction and harmful outcomes for others. Evidence of these harms is limited to small-sample, cross-sectional self-reports, such as prevalence surveys. We examine the association between gambling as a proportion of monthly income and 31 financial, social and health outcomes using anonymous data provided by a UK retail bank, aggregated for up to 6.5 million individuals over up to 7 years. Gambling is associated with higher financial distress and lower financial inclusion and planning, and with negative lifestyle, health, well-being and leisure out- comes. Gambling is associated with higher rates of future unemployment and physical disability and, at the highest levels, with substantially increased mortality. Gambling is persistent over time, growing over the sample period, and has higher negative associations among the heaviest gamblers. Our findings inform the debate over the relationship between gambling and life experiences across the population.
We analyse gambling behaviour via detailed, anonymous, individual-level financial transaction data from millions of cus- tomers of the United Kingdom’s largest retail bank, Lloyds Banking Group (LBG). Our largest dataset tracks ~6.5 million people or around 10.6% of the population of the United Kingdom, over 7 years. Big financial transaction data provide a unique view of individual-level gambling behaviour, consisting of the full spread of electronic payments to gambling platforms, which allows us to identify the distribution (who, when and for how long) of gambling and its associated outcomes across a national population. The relationship of gambling with financial outcomes (for example, savings and debt) and non-financial outcomes (for example, spending on hobbies, social activities and night-time online spending), can all be inferred objectively and analysed alongside information on gambling behaviour. We also measure longer-term outcomes, including transitions into unemployment, disability and mortality. This view of individual outcomes is rivalled only by what a state monopolist could see—it cannot be seen in the data of gambling firms, self-reported survey data or the aggregated data reported by firms, industry groups and regulators. This observational study documents gambling in the United Kingdom with large-scale objective data. Previous approaches had to rely primarily on self-report surveys and smaller sample sizes13. For example, the United Kingdom ran three waves of the British Gambling Prevalence Survey in 1999, 2007 and 2010, considered by expert witnesses in a recent government select committee as the best national data on gambling in the United Kingdom14. The 2010 survey used a sample of 7,756 respondents or ~0.01% of the then population of the United Kingdom15. This survey estimated that 0.7–0.9% of the then population of the United Kingdom met diagnostic criteria for disordered gambling, although this estimate is based on less than 100 cases, as is typical in prevalence surveys given population base rates16. It has been argued that these base rates may be understated if gamblers hide or cover-up their gam- bling when filling out these surveys17. Prevalence surveys also ask respondents to self-report their gambling involvement and expen-diture. However, it has been demonstrated that disordered gamblers cannot self-report their gambling expenditure reliably18, that memory biases are an established feature of disordered gambling19 and that prevalence surveys may struggle to recruit sufficient disordered gamblers given population base rates16. Similar20, or smaller sample sizes21,22, have so far been used to examine the relationship between gambling and mortality. A further advantage is that transaction data take the form of individual-level panels which follow the same individual over time. To date, most gambling research is cross-sectional in nature, with a comparative lack of longitudinal studies23—which exhibit increased levels of attrition amongst disordered gam- blers24. By comparison, our big financial transaction data approach unobtrusively follows a random sample drawn from a substantial fraction of the banked population of the United Kingdom. The empirical gambling-related harm literature has added a focus on the negative consequences associated with gambling but is also limited by a focus on cross-sectional self-report sur- veys25–28. Thus far, there have been two main attempts to create conceptual frameworks to better understand the multidimensional nature of the relationship between gambling and individual out- comes29,30. Langham et al.29 derived a list of 72 distinct ‘harms’, covering financial, relationship, psychological, health, work, study and social deviance harms. Later research has shown that these harms differ markedly with respect to prevalence, with financial harms being the most prevalent and social deviance harms the least prevalent31. Wardle et al.30 conceptualized gambling-related harm as affecting economic resources, relationships and health, with harms potentially having persistent effects through time, and being felt beyond individuals and across wider communities. Moreover, there is a current debate about the extent to which gambling harms are concentrated amongst disordered gamblers32,33 versus the overall impact of harm felt amongst the larger group of lower-risk gamblers31.
We contribute
to this literature with a data-driven approach. Our analysis focuses on
quantifying the association between gambling and personal outcomes. The
evidence we present raises questions of causation
and the
mechanisms by
which associations
arise, which are topics
for future work. Levels of gambling. We used a random sample (sample 1) of 102,195 customers active in each month of 2018. The unit of analy- sis in this panel data sample is an account calendar month. To iden- tify gambling transactions, we relied on the pre-existing gambling category in the bank’s typography of transactions, which includes various forms of gambling such as offline and online bookmakers, casinos, lotteries and other providers. Cash gambling and gambling at other types of retailers (for example, a lottery ticket at the super- market) are not captured and thus we are conservative in estimating total gambling. Summary data in Table 1 reveal that 43% of individuals in the sample made at least one electronic gambling transaction in 2018. Among those who made at least one electronic transaction, the median number of transactions was 12 (mean = 56), with a median year spend of £125 (mean = £1,345), which is approximately a median of 0.5% of monthly spending (mean = 4%). The gap between the mean and median values highlights the highly skewed nature of gambling behaviour (Supplementary Table 1). We define spend as the sum of all gambling transactions that were processed via a debit card or credit card. The distribution of spending has a long right-tail, with the top 10% of gamblers spending >£1,800 on gambling in the calendar year, close to 8% of their total spending. Gambling and financial stress. Here, we describe how gambling is associated with financial distress, financial inclusion and finan- cial planning in a random sample of active customers (sample 1) (top rows of Fig. 1). The unit of analysis in this sample is a calendar month. The measures of financial distress are: using an unplanned overdraft, missing a credit card payment, taking a payday loan, missing a loan repayment and missing a mortgage repayment. Financial inclusion measures are: having a credit card, loan or mortgage, credit card use and making a payment to a debt recovery agency. Financial planning measures are: holding insurance, paying down a mortgage, saving money, saving money in a tax-preferred savings account (known as an individual savings account (ISA) in the United Kingdom) and paying into a self-invested pension. A detailed description of all the outcome variables is contained in Supplementary Table 2, with summary statistics reported in Supplementary Table 1. The set of outcome variables shown includes all outcomes that were analysed. In all of the binned scatterplots related to financial outcomes in Fig. 1 (rows 1–3) the unit of analysis is one account calendar month. For each account month, we calculated the percentage of the individual’s total spend in that month devoted to gambling. Total spend was calculated by summing all outflows of cash across a given month, and included credit card, debit card, direct debit and ATM transactions but not internal movements of money (for example, movement from a personal current account to savings account). The x axis shows the percentile rank of this variable. Each panel contains 101 dots. The dot at 0% on the x axis include account months in which the individual had zero gambling (not all indi- viduals who gamble do so in each month of the sample period). That is, if a gambler had an account month where they did not gamble, he or she would be captured in the dot at 0%. Each of the remaining 100 dots represent one percentile of account months (typically 150–3,000 account months, depending on the sample size; Supplementary Tables 4–6). Thus, the dot at 1% represents the 1% of observations where gambling was lowest (but not zero) and the dot at 100% represents the 1% of observations where gambling was highest. (The discontinuity between 0% and 1% results for technical reasons: selecting accounts with zero gambling selects accounts that were less likely to be active for other transactions.) The y axis shows the mean value of the dependent variable at each percentile. For this analysis, the dependent variable is measured 1 month forward, to avoid a mechanical relationship whereby higher gambling mecha- nistically reduces the value of outcome variables related to spending due to individuals having less net income to spend on other items in months when more is spent on gambling. The lines are penalized cubic regression splines estimated directly from the underlying data with 95% confidence intervals (CI). Higher gambling is associated with a higher rate of using an unplanned bank overdraft, missing a credit card, loan or mort- gage payment, and taking a payday loan. A 10% point increase in absolute gambling spend is associated with an increase in payday loan uptake by 51.5% (so, for example, 0.97% of those with 0% of spending on gambling have a payday loan but 1.47% of those with 10% of spending on gambling have a payday loan, an increase of 51.5%) and the likelihood of missing a mortgage payment by 97.5% (Supplementary Table 3). In all reported cases, the effect of a 10% point increase in absolute gambling spend are reported after con- trolling for age, gender and annual income. Gambling is associated with lower rates of holding a credit card, loan or mortgage, higher use of credit card balances and a higher likelihood of the individual being subject to debt collection by bailiffs. A 10% point increase in absolute gambling is associated with an increase in credit card use by 11.2% and bai- liff interaction by 8% (Supplementary Table 3). Conversely, higher gambling is associated with smaller spends on insurance and mort- gage repayments, smaller total savings and smaller pension contri- butions. For many of the outcome variables, the association with gambling is notably stronger at high percentile ranks approximately above the 75th percentile (which equates to ~3.6% of total monthly expenditure). This suggests that the relationship between gambling and many of the harmful outcomes is stronger when the individ- ual is devoting a relatively large share of total monthly spending to gambling. We conducted regression analyses, using an ordinary least squares regression estimator in a specification that controlled for age, gender and income in addition to gambling as a percentage of monthly spend (all variables entering linearly, together with a con- stant term). All statistical tests were two-sided. The coefficients on the gambling covariates, together with 95% CI and marginal R2, are reported in Supplementary Table 3 (with the full regression esti- mates reported in Supplementary Tables 4–6).
Gambling, lifestyle and well-being. Outcomes associated with gambling extend beyond the purely financial (bottom rows of Fig. 1). The wider themes are lifestyle (spend on fast food, gaming, bars, tobacco and off licences), health and well-being (spend on prescrip- tions, self-care, fitness and night-time spending between 1:00 and 5:00) and leisure and interests (spend on hobbies, social activities, education and travel), which are analysed in a random sample of active customers (sample 1), where the unit of analysis is a calendar month. Results show a negative association between gambling and self-care, fitness activities (for example, gym membership), social activities, and spending on education and hobbies. There is also an association between gambling, social isolation and night-time wakefulness—individuals spending more on gambling travel less and are more likely to spend at night. A 10% point increase in absolute gambling equates to an 11.5% increase in nights awake and 9% reduction in social activities (Supplementary Table 3). The relationship between gambling on reduced socialization is also seen in lower spend at bars and pubs. But higher levels of gambling are associated with lower off-licence spending. The relation with fast-food spend is more complex (see Supplementary Table 3 for regression coefficients, with the full regression estimates reported in Supplementary Tables 7–9). The coefficient estimates are precisely estimated and confirm the directional relations illustrated in Fig. 1, with the exception being tobacco spend, for which the coefficient is not precisely estimated. Gambling, unemployment, disability and mortality. Here, we describe medium-term associations with unemployment, disability and mortality using data from all 6.5 million active customers in each month in 2013 (sample 2). We tracked these individuals across the subsequent 5 years, 2014–2019. We find that higher gambling is associated with a higher risk of future unemployment and future physical disability. The panel ‘Disability payments’ in Fig. 1 restricts sample 2 to individuals who were not receiving disability payments in 2013 and plots the relationship between the percentile rank of gambling spend as a percentage of monthly income and the likeli- hood of subsequently claiming disability payments over the period January 2014 to July 2019. The plot reveals a positive association (Supplementary Table 10).
The panel
‘Unemployment’ in Fig.
1
restricts sample 2 to individuals
who
were
employed
in
2013
and
plots
the
relationship
between the percentile
rank of gambling spend as a percentage of monthly income and the likelihood
of subsequently experiencing at least one spell
of unemployment
over the
period January
2014 to
July 2019. The
positive relationship
is notably
stronger at
high levels
of gam- bling,
with employed
individuals in
the highest
percentiles of
gam- bling
having
a
6%
likelihood
of
experiencing
future
unemployment
(Supplementary Table 10). We examined the relationship between gambling spend and mortality. We model mortality using survival analysis in adult males and females drawn from sample 2. We fitted Cox proportional hazard models to the data, controlling for amount gambled, individual gender and individual age. The model censors individuals who left the sample for reasons other than mortality. Figure 2 plots the Cox model fits, showing the relationship between levels of gambling, where gambling is expressed as a proportion of monthly income of 0%, 10%, 20% or 30%. (Table 1 shows that the top 1% of gamblers gambled over 58% of their income in 2018.) The x axis plots time in years (from January 2014) and the y axis plots the survival probability. Plots are shown for men and women at three age points. For all groups, the survival probability is lower at higher levels of gambling. Information is not available on cause of mortality. The heaviest gamblers exhibit higher 5-year mortality rates. For example, among 44-year-old women, gambling 30% of annual expenditure (relative to 0%) is associated with an increased chance of death from 50 in 10,000 (95% CIs [50, 51]) to 69 in 10,000 (95% CIs [66, 72]) or by a factor of 1.37 (Supplementary Table 11). High levels of gambling are associated with a likelihood of mortality that is about one-third higher, for both men and women, younger and older. The time course of gambling. Gambling is also persistent over time, although individuals can transition into (and out of) high lev- els of gambling within a few months. We used a random sample of 101,151 customers active over all months from 2012 to 2018 (sample 3). The top panel of Fig. 3 illustrates the movement over time of individuals between levels of gambling. The analysis is cen- tred on the year 2015, showing the level of gambling that leads to and leads from 2015. Gambling is persistent but some small fractions of individuals move from no gambling in 2012 to the highest levels in 2015 and some small fractions gambling at the highest level in 2015 have stopped in 2018. The bottom panel zooms in on the highest-spending gamblers to see whether they have always gam- bled heavily in the past. The sample comprises a subset from sample 3 whose gambling was >10% of their total spending in Quarter 2 of 2015 (2,168 individuals). We find that, for example, 3 years earlier around half of the highest-spending gamblers were already gam- bling heavily, while only 6 months before, over 6.9% of these heavy gamblers were not gambling at all, highlighting the fast acceleration with which some individuals can transition into heavy gambling. In contrast, 6 months later 4.6% of heavy gamblers were not gambling at all. This asymmetry shows that gambling expenditure represents sticky behaviour. DiscussionThis paper demonstrates that financial transaction data can produce a view of gambling-related outcomes that is objective, longitudinal and mass-scale. By comparison, prevalence surveys, which have dominated the view that academics and policy-makers have of gambling for the last 30 years, are self-report, cross-sectional and largely small sample in nature13. We described the association between gambling and 31 outcome variables from the financial and wider social and health domains. Given that our data do not cover cash gambling transactions, or electronic transactions using third-party payment processors or another person’s account details, the estimated effects of gambling expenditure on gambling-related harm are probably conservative. Our evidence complements existing approaches, which draw upon self-report surveys, case studies or inferences from industry or aggregate-level statistics13,21,22,25–28,34–36 by relying on large-scale objective data. As such, the reported findings have implications for the future study of gambling epidemiology and public health.
This study contains some limitations that could be addressed with future
research. First, and similarly to gambling prevalence surveys, we do not
establish causality, which means that findings
demonstrate associations
that may
reflect causality
or comorbidity—both
of which
are of
concern. Causality
would indicate that
higher levels of gambling increase one’s risk of negative outcomes like
financial distress, social exclusion, disability and unemployment.
Comorbidity, however, would indicate that individuals
who are
susceptible to
these negative
outcomes due
to other
factors
are
more
likely
to
be
drawn
to
gambling.
In
reality,
the
observed effects
could result
from a
blend of
causality and
comorbidity, both of
which have
significant implications
for policy-makers
and public
health experts.
Nonetheless, a longitudinal financial transaction approach informs the current gambling policy debate. Some argue that associations between gambling and negative outcomes exist primarily among a small group of disordered gamblers, who should be the focus of mitigating gambling-related harm32,33,37. In support of this view, we find a number of negative outcomes such as nights |
| |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
|