psychiatric measures of gambling in the general population: a reconsideration
TRANSCRIPT
Psychiatric Measures of Gambling in theGeneral Population: A Reconsideration
Glenn Harrison
Center for the Economic Analysis of Risk
Auckland, February 2016
Joint with Morten Lau and Don Ross
Motivating questions
> What do we mean by “problem gambling”?o People that are “bad for business”o People that (should) clinically present for treatmento People that suffer a welfare loss from gambling choices
> What is the prevalence of problem gambling …o In the general populationo In self-selected populations, such as active gamblers
> Can surveys be used to assess this prevalence?
Our overall design
Surveys
Lab Experiments
Field Experiments
Three issues with surveys
> I. They tend to reflect existing gambling, not the latent propensity to gambleo Likely to imply understatement of gambling problems
> II. They use “trigger questions” which lead to the possibility of sample selection bias in inferences about general population prevalenceo Likely to imply understatement of gambling problems
> III. They are statistically analyzed in a way that suggests massive co-morbidities with many other psychiatric disorders
Surveys of general prevalence, I
> Generally “reflective” of a history of gamblingo Does it lead to “disruptions in life,” such as bankruptcy, lying,
divorce, criminal activity? DSM-IV: “persistent and recurrent maladaptive gambling behavior” (p.615)
o Does it lead people to “clinically present” for treatment?> Not well designed to detect latent, “formative” propensity
to gamble (whether or not it leads to problem gambling)o Some surveys take these into account, all in our Danish work
reviewed by Morten and Don in the next session… Focal Adult Gambling Screen (FLAGS) The Gambling Craving Scale (GACS) The Gambling Related Cognition Scale (GRCS) The Gambling Urge Screen (GUS)
Surveys of general prevalence, II
> The use of trigger questionso Various forms, but things like “Have you ever lost $100 from
gambling?”o Only if this is answered affirmatively are the diagnostic questions
asked> Should be classified as “no detectable risk,” as in FLAGS
o But they can never, by definition, show up as problem gamblers> To an economist, this is simply sample selection bias
o Some process generates the observed sample in a way that could lead to biased inferences about the population
o Standard statistical corrections
Existing surveys
> Objective has been to mimic the criteria that psychiatrists would use to diagnose “problem gambling” for clinical purposeso People only ever present clinically if they have had a gambling
problem causing them to be concerned, so reflective constructs are therefore natural
> Dominance of criteria from the Diagnostic and Statistical Manual of Mental Disorders (DSM)
Existing surveys
> Objective has been to mimic the criteria that psychiatrists would use to diagnose “problem gambling” for clinical purposeso People only ever present clinically if they have had a gambling
problem causing them to be concerned, so reflective constructs are therefore natural
> Dominance of criteria from the Diagnostic and Statistical Manual of Mental Disorders (DSM) Important changes in
these, which have generated massive controversies (see Allen Frances books)
EveryQuestionassumes
someone has done some gambling
Major prevalence surveys
Survey Country Year Sample
NESARC Wave 1 USA 2000‐2001 43,093
NCS‐R USA 2001‐2003 9,282
CCHS Mental Health and Well‐Being Canada 2002 34,770
BGPS U.K. 2010 7,756
Legend:NESARC – National Epidemiologic Survey on Alcohol and Related ConditionsNCS-R – National Comorbidity Survey ReplicationCCHS – Canadian Community Health SurveyBGPS – British Gambling Prevalence Survey
By the way…
> Gambling disorders now completed droppedo from later NESARC waves in the USo from later CCHS waves in Canada
> Why?
> Extremely low prevalence estimates?
NESARC
Response N Fraction
Yes 11,153 26%
No (or invalid) 31,940 74%
NESARC, II
So 0.45% pathological gambling in the general population.Less than half of 1 percentage point.
Why use trigger questions?
> Saves time on long surveyso Yes, we get thiso But with popular instruments that only have a few questions?
Why use trigger questions?
> Saves time on long surveyso Yes, we get thiso But with popular instruments that only have a few questions?
> Many of the follow-on questions sound contrived or odd if someone has never gambled or lost moneyo Yes, but then also use questions getting at formative constructs
Why use trigger questions?
> Saves time on long surveyso Yes, we get thiso But with popular instruments that only have a few questions?
> Many of the follow-on questions sound contrived or odd if someone has never gambled or lost moneyo Yes, but then also use questions getting at formative constructs
> Everybody else does it
> Didn’t think of it before today
Why use trigger questions?
> Saves time on long surveyso Yes, we get thiso But with popular instruments that only have a few questions?
> Many of the follow-on questions sound contrived or odd if someone has never gambled or lost moneyo Yes, but then also use questions getting at formative constructs
> Everybody else does it
> Didn’t think of it before today
> Want to generate low estimates of gambling prevalenceo Very convenient for gambling industry
Not just gambling…
> In NESARC there are trigger questions for every psychiatric disorder
> My favorite: Specific Social Phobias
> After a trigger question they are then asked if they have strong fear or avoidance of being interviewed
Solutions
> Ask the threshold gambling question after asking the prevalence questionso We do this in Denmark with FLAGS, PGSI & DSM – Morten and
Don to discuss in the next session> Correct statistically using sample selection bias methods
o Due to James Heckman: Nobel Prize in Economics for 2000o Basic logic is to jointly model the sample-generating process and
the process explaining extent of gambling problems Probit model of participation (needs data on non-participants) Then see if errors in that participation model are correlated with the
errors of the process of interest, the extent of gambling problems
Surveys of general prevalence, III
> Gambling as a psychiatric disorder seems highly correlated with lots of other psychiatric disorderso Implications for treatment and therapy
Surveys of general prevalence, III
> Gambling as a psychiatric disorder seems highly correlated with lots of other psychiatric disorderso Implications for treatment and therapy
Surveys of general prevalence, III
> Gambling as a psychiatric disorder seems highly correlated with lots of other psychiatric disorderso Implications for treatment and therapy
Surveys of general prevalence, III
> Gambling as a psychiatric disorder seems highly correlated with lots of other psychiatric disorderso Implications for treatment and therapy
> But this is statistically detected by estimating the unconditional correlation of gambling with other disorders – the “total effect”
Surveys of general prevalence, III
> Gambling as a psychiatric disorder seems highly correlated with lots of other psychiatric disorderso Implications for treatment and therapy
> But this is statistically detected by estimating the unconditional correlation of gambling with other disorders – the “total effect”
> A different question is answered by the conditionalcorrelation of gambling with other disorders – the “marginal effect”
> Both questions are interesting, but only the first is ever answered
Solution
> Model the correlation and also control for other psychiatric disorderso Ordered probit rather than binary probito Infer the marginal effect of each psychiatric disorder on gambling
disorder, to measure conditional correlation
> Again, measuring unconditional correlation is not an erroro It is just not the only type of correlation we are interested ino I would argue that unconditional comorbidity is not that interesting
What we do
> Evaluate comorbidity of gambling disorders and other disorders using major national epidemiological surveyso US (NESARC and NCS-R), Canada (CCHS) and Britain (BGPS)o Just show results for NESARC here
Correct and replicate Petry et al. [2005] Same qualitative results for NCS-R, CCHS and BGPS
> Show marginal effect and total effect to compare
> Then correct estimates of comorbidities for sample selection bias
> Then show predicted gambling hierarchy with sample selection correction
Conclusions and Limitations
> Trigger questions can generate massive sample selection bias in gamblers at risko Have we been significantly underestimating the “at risk” fraction
of the population?> Comorbidities of gambling should be evaluated
conditionally and unconditionally (total and marginal)o Dramatic overstatement of comorbidity if unconditional
comorbidity is interpreted as a conditional comorbidity> Avoid trigger questions and do more econometrics
> Limitationso Data on the unwashed and unsampled?o Statistical assumptions are needed
Our overall design
Surveys
Lab Experiments
Field Experiments
1. Interpreting existing surveys2. Designing better surveys
Our overall design
Surveys
Lab Experiments
Field Experiments