Extending the marketing ban after self-exclusion: a 2548-player randomised trial with a null primary outcome
- Parallel-group randomised trial run inside a French online gambling operator, n = 2548 adults who had just self-excluded for three months or less. 1265 were assigned to a suspension of direct promotional SMS and email running 9 months in total from the self-exclusion, so that it extended well past the block; 1283 received the standard procedure, in which solicitations resume the day the self-exclusion expires. Account-based follow-up to 18 months. The trial was financed by a grant issued via the operator itself.
- The registered primary outcome was null. Change in last-4-week net loss at month 9 was a reduction of €308.73 in the extended-ban arm versus €250.81 in the standard arm, p = 0.232. The trial was powered to detect a €100 gap on this measure; the observed gap was €57.92.
- The positive result sits on a secondary measure that does not appear in the trial registration at all. Change in last-4-week deposits was €455.68 versus €319.65 at 6 months (p = 0.017), €451.92 versus €343.21 at 9 months (p = 0.040), €484.27 versus €370.02 at 12 months (p = 0.025), and €492.61 versus €404.73 at 18 months (p = 0.087, no longer significant).
- Almost nothing else separated the arms. Stakes, number of sessions, episodes of compulsive depositing and repeat self-exclusions at 18 months (0.20 versus 0.21, p = 0.661) were comparable; average session duration at month 9 was the one further cell to clear the threshold (p = 0.045). The results table reports 49 between-group comparisons with no correction for multiplicity, of which four reached p < 0.05, and no clinical, diagnostic or harm measure was collected at any timepoint.
Self-exclusion is the one harm-reduction step almost every clinician recommends to a gambling patient, and almost nobody can say what it does after the block expires. This trial isolates a single modifiable variable at that moment: whether the operator's marketing resumes. The answer is more interesting for what it fails to show than for what it shows.
What the procedure does, and what it moved
The mechanics are simple enough to describe in one clause, which is unusual for a tool piece. A gambler on the PMU platform triggers a self-exclusion, at which point access to the whole site is closed for the chosen period and, critically, the duration cannot be altered once set. In the intervention arm the operator additionally suspended direct commercial solicitations, SMS and email, for a total of 9 months. In the control arm those solicitations restarted the moment the block lapsed. There is no session, no therapist and no dose: this is a back-office switch, delivered by the operator, at scale. Recruitment ran May to November 2022; participants were 85.3% men, mean age 44.9 (SD 14.5) years, with a median chosen self-exclusion of 31 days.
The registered primary outcome was change in four-week net loss at month 9, and it did not move. The extended-ban arm reduced net loss by €308.73 (SD 1399.06), the standard arm by €250.81 (SD 1020.41), p = 0.232. Read that against the sample-size calculation: the trial was built on an assumed €400 versus €300 reduction, common SD €900, requiring 2544 participants. The observed difference was €57.92, roughly half the effect the design was sized for. This is a well-powered null on the outcome the investigators nominated in advance, and it deserves to be reported as the headline finding rather than buried.
What did separate the arms was deposits, a different construct: money moved onto the platform rather than money lost on it. The differences are real and internally coherent, and I reproduced all four p-values from the published means and standard deviations. But three cautions travel with them. Deposits are absent from the registered outcome list entirely. The comparison is one of 49 unadjusted between-group tests in the same table, four of which cleared p < 0.05, which is close to what chance alone would deliver. And the intervention arm started €115.09 higher on baseline deposits (€691.52 versus €576.43), an imbalance that is itself nominally significant on my own recomputation, and a gap of almost the same size as the €136.03 between-arm difference in reduction at month 6, with the analysis using unadjusted Student's t-tests on change scores and no baseline covariate.
Reading it against the registry, and what to do on Monday
The registry record, NCT05413564, sponsored by CHU de Nîmes, is worth pulling because it both vindicates and complicates the paper. The primary outcome matches exactly, which is the honest part: net loss at month 9, reported as null. But the registry lists net loss as the only outcome, primary and secondary alike. Deposits, stakes, compulsivity episodes, session counts and repeat self-exclusions are prespecified as secondary outcomes in the paper's own Methods, yet none of them appears in the registration. Three further mismatches: the registry logs the design as non-randomised and factorial with no masking, while the paper is titled and analysed as a parallel 1:1 randomised trial; registered actual enrolment is 2554 against 2548 analysed; and the registry carries explicit inclusion and exclusion criteria while the paper states there was no exclusion criterion. What the registry does not show is any sign of tidying after the fact: the record was submitted on 24 May 2022, six days before the trial began on 30 May, posted on 10 June after quality control, and all four versions carry the same design, enrolment and outcome fields. Registration was prospective and has not been edited. One cell in the baseline table fails to reconcile: the intervention arm's male proportion, printed as 82.82% where 1073 of 1265 is 84.82%.
The competing-interests statement is short and material. The senior author declares a data-sharing agreement for the "OSE" study through an academic-private convention with FDJ, and a previous study with Winamax, both commercial gambling operators, with independence of protocol, analysis and publication guaranteed by convention and no money changing hands. The other two authors declare none. The funding statement matters more than any of that, and it sits in a separate section: the study was financed by a grant issued via PMU, the same operator whose platform supplied the data and whose marketing was the thing under test, under a regulatory obligation to redistribute 0.002% of stakes, administered through the French observatory for drugs and addictive tendencies. An operator-funded test of restraining operator marketing is not disqualifying, and the null primary outcome is the opposite of a flattering result, but a reader deserves to know who paid. Participants were never told they had been randomised and never consented to take part, although on opening an account they had consented in the terms of use to the collection and analysis of their gambling data; the local ethics committee at Nîmes approved that arrangement explicitly, reference 25.01.03. No adverse events were recorded, because no adverse-event or harm measure existed in the protocol.
So what changes in the consulting room. Not a recommendation, because a null primary on an operator-run behavioural nudge is not a recommendation. What changes is a question you can now ask with a date attached: when your patient self-excludes, find out how long the block runs and mark the day the promotional messages are legally free to resume, because on the standard procedure that is the same day. Build that date into the relapse plan the way you would build in a discharge date. And keep the ceiling honest with the patient: the best available trial of extending that quiet period found more money staying out of the account for about a year, no detectable change in money lost, no change in compulsive depositing, and no fewer people coming back to self-exclude again.
The outcome this trial registered in advance, money lost, did not move; what moved was money deposited, a measure that appears nowhere in the registration.
Deposits at a single operator are not a clinical outcome, and no measure of gambling-related harm, functional impairment or diagnostic status was collected, so gambling on other sites was invisible to the dataset. The registered primary outcome was null and the reported benefit rests on a secondary measure absent from the registration, among 49 unadjusted comparisons, in arms that differed by €115.09 on baseline deposits, an imbalance itself at the edge of significance. The study was financed by a grant issued via PMU, the operator whose platform supplied the data and whose marketing was under test. Participants had consented in the terms of use to analysis of their gambling data, but were not told they had been randomised and did not consent to take part; the ethics committee approved that explicitly.