Research/Medicine
Registry-posted trial results for two alcohol use disorder drugs
Pooling only the results sponsors posted on ClinicalTrials.gov, topiramate reduced heavy drinking at about three-quarters of its published effect size, and gabapentin's signal was real but less certain than in the journals.
Oct 14, 2026 · 6 min read

Meta-analyses of drugs for alcohol use disorder pool what was published in journals. We pooled something else: the results that trial sponsors posted on ClinicalTrials.gov, which include small, terminated and never-published trials. For topiramate, nine registry-posted trials give a standardized effect on heavy drinking of g = -0.30, about three-quarters of the roughly -0.40 the published meta-analyses report. For gabapentin, the registry signal points the same way but is smaller and less certain than the journal estimate. Both pools were re-checked number by number against the registry. This is an independent check on published evidence, not guidance for anyone's treatment.
-0.30
topiramate, pooled g on heavy drinking across 9 registry-posted trials
-0.40
approximate published estimate the registry pool is compared with
-0.44
gabapentin, pooled g across 5 trials after a correction
0 of 3
pre-registered tests that reproduced the post hoc subgroup split
The question
Journal meta-analyses can lean heavily on a few well-known trials. The two topiramate meta-analyses we compared against (Blodgett et al. 2014; Fluyau et al. 2023) both rest substantially on two Johnson trials, and neither of those has results posted on the registry. A pool built only from registry postings therefore shares almost none of its inputs with the published estimates. If the drug still looks effective there, that is evidence from a largely separate set of trials.
What we did
We searched ClinicalTrials.gov for randomized, placebo-controlled trials of each drug in alcohol use disorder with a posted drinking outcome, and transcribed each trial's numbers by hand. Each trial contributes one Hedges' g, signed so that a negative value favours the drug. We pooled with DerSimonian-Laird random effects and also report Hartung-Knapp-Sidik-Jonkman (HKSJ) intervals, which are more cautious when trials are few. A sealed job then re-downloaded every record and checked each transcribed mean, spread and sample size against the posted values. All matched.
Topiramate
Nine trials with 1,048 people in the arms used met the rules. Eight favour topiramate; one, in adolescents and young adults, does not.
Topiramate: effect on heavy drinking in each registry-posted trial (Hedges' g)
The pooled effect is clear of zero on both intervals, with real heterogeneity (I² 60%). That heterogeneity lines up with who was enrolled. In the four trials of people whose main problem is alcohol, the effect is moderate and consistent. In the five trials that also selected for smoking, PTSD, cocaine dependence or young age, it is small and uncertain. This split was noticed after the data had been seen, so it is a post hoc subgroup on four and five trials.
Topiramate registry pools
| Pool | Trials | g | 95% CI | HKSJ 95% CI | I² |
|---|---|---|---|---|---|
| All trials | 9 | -0.301 | -0.506 to -0.097 | -0.562 to -0.041 | 60% |
| Posted arm data only | 7 | -0.268 | -0.527 to -0.008 | -0.628 to +0.093 | 64% |
| General AUD samples | 4 | -0.574 | -0.780 to -0.368 | -0.895 to -0.253 | 0% |
| Samples with another condition | 5 | -0.103 | -0.265 to +0.059 | -0.333 to +0.127 | 8% |
Gabapentin, and a correction
The first gabapentin pool combined six trials and gave g = -0.36. Two days later, during a routine reference check, we found that one of those trials gave gabapentin for about ten nights and measured drinking in the twelve weeks after the drug stopped. It did not belong in a pool of treatment effects. Removing it, the job's five-trial pool gives g = -0.437, with a conventional interval clear of zero but an HKSJ interval that narrowly crosses it. A second trial's outcome window runs mostly after gabapentin ended. We now read each trial's dosing schedule and outcome window before it enters any pool, and we checked every row in all of our alcohol use disorder lists the same way: of 49 rows, 46 were on treatment.
Gabapentin registry pools
| Pool | Trials | g | 95% CI | HKSJ 95% CI |
|---|---|---|---|---|
| As first published: all generic-gabapentin trials | 6 | -0.360 | -0.659 to -0.062 | -0.757 to +0.037 |
| Corrected: 10-night trial removed | 5 | -0.437 | -0.754 to -0.119 | -0.883 to +0.010 |
| General AUD samples, as first published | 4 | -0.511 | -0.836 to -0.186 | -1.015 to -0.006 |
The published meta-analysis for gabapentin (Kranzler et al. 2019) found g = -0.64 on percent heavy drinking days. The registry's gaps also run in an unusual direction: three small trials reported as positive in journals have no posted results, so a registry-only pool here leaves out positive evidence rather than hiding negative trials. After the correction, the gabapentin "other condition" group rests on a single 30-person pilot, so the subgroup split for this drug is not well supported.
Testing the split in advance
A gap between general and comorbid samples, seen after the fact in two drugs, could be a general feature of alcohol use disorder trials. We fixed that hypothesis in advance and tested it on naltrexone, acamprosate and varenicline. None reproduced a split of that size: naltrexone went the other way (-0.13), and the average for acamprosate and varenicline was +0.16 (p ≈ 0.18). These three pools are preliminary: their numbers were computed by hand and are queued for the same sealed registry re-check that topiramate and gabapentin passed. We do not report their overall drug effects here. What survives so far is, at most, a small tendency for drugs to look weaker in samples selected for a second condition.
What it means
On registry evidence that barely overlaps the journals, topiramate still reduces heavy drinking, at a somewhat smaller size, and most clearly in people whose primary problem is alcohol. Gabapentin's registry signal favours less drinking but is less certain. Registry postings also carry errors: we found SDs on the wrong scale, model-based standard errors that cannot become SDs, and placebo arms that list the active drug. Each conversion we made is flagged in the study files.


