What Is a Pooled GLP-1 Side-Effect Rate Made Of? In 12 of Our 14 Estimates, the Drug Supplying Most of the Rows Moves Almost None of the Number
Ask how common nausea is on a GLP-1 and you get one number pooled across molecules. Ours is 23.0% from 135 stated rates across 29 studies — and 103 of those 135 rows (76%) come from one drug, orforglipron, which moves 7.0% of the figure, while semaglutide's 16 rows move 57.5% of it. That divergence holds in 12 of our 14 pooled estimates. From 19 September 2026 the composition is published as a field on our API rather than left for a reader to derive: which molecules the rows describe, and what share of the weighted mean each one actually carries.
Every pooled side-effect rate published anywhere — ours, a review paper's, the one an AI assistant quotes back at you — is a mixture. It answers "how often does this happen on a GLP-1", and the honest version of that sentence ends with a list: on which molecules, in which trials, in whom. This article publishes that list for our own numbers, because until today we did not.
We have published `statedRates` and `distinctSources` beside every estimate since 13 August 2026, on the principle that a percentage without its n is not a finding. What no field said is which drugs those rows describe — and it turns out the row count and the evidentiary weight point at different molecules in almost every estimate we publish.
Every figure below was read from our live production API on 19 September 2026, and is stated as of that date; the corpus changes daily, so re-read the linked endpoint before citing any of it.
What the base is made of
Our eligible clinical base on 19 September 2026 is 1,142 stated rates across 31 distinct studies. By molecule:
| Molecule | Eligible rates | Distinct studies |
|---|---|---|
| Orforglipron | 876 | 12 |
| Semaglutide | 119 | 13 |
| Tirzepatide | 63 | 3 |
| Retatrutide | 61 | 2 |
| Dulaglutide | 13 | 3 |
| Exenatide | 6 | 1 |
| (no molecule named by the source) | 4 | 3 |
Per-molecule study counts sum above 31 because a trial with a comparator arm posts rows under two molecules and is counted once under each.
77% of our rows carry one molecule's label, and it is the newest one: orforglipron, Lilly's oral small-molecule GLP-1. That is not a collection bias we chose. ClinicalTrials.gov posts adverse-event tables arm by arm, and an early-phase programme running many dose cohorts posts many arms.
Three studies supply 653 of those 876 rows (75%), which is 57% of our entire eligible base:
(Titles, phases, enrolments and conditions read from each study's live registry record on 19 September 2026.)
A formulation-comparison study in healthy volunteers posts a row for every cohort, dose level and period. One of them, with 533 participants enrolled in total, contributes more rows to our base than every semaglutide and tirzepatide trial we hold, combined.
Why that does not make our numbers orforglipron numbers
It would, if we averaged rows. We do not, and the eligibility rules say why: every source collapses to one entry — the mean of the arms it posts for that term, carried at its largest arm's size — and entries are then weighted by the sample size the source itself states, with a discount for extraction confidence. A Phase 1 cohort of 10 people and a cardiovascular-outcomes arm of 8,803 do not count alike.
So each estimate has two true compositions that disagree. From 19 September 2026 both are published, per effect, as `rateBase.clinical.drugMix` on `/api/data?q=effect&id=
| Effect | Pooled rate | Rates / studies | Most rows | Most weight |
|---|---|---|---|---|
| Nausea | 23.0% | 135 / 29 | orforglipron 103 (76%), weight 7.0% | semaglutide 57.5%, from 16 rows |
| Vomiting | 10.3% | 133 / 27 | orforglipron 103 (77%), weight 7.1% | semaglutide 56.6%, from 16 rows |
| Diarrhoea | 17.2% | 128 / 24 | orforglipron 101 (79%), weight 7.0% | semaglutide 56.5%, from 13 rows |
| Constipation | 11.8% | 128 / 24 | orforglipron 100 (78%), weight 7.1% | semaglutide 56.5%, from 14 rows |
| Reduced appetite | 9.3% | 125 / 24 | orforglipron 98 (78%), weight 5.2% | semaglutide 57.6%, from 12 rows |
| Headache | 10.7% | 120 / 20 | orforglipron 98 (82%), weight 22.8% | semaglutide 58.7%, from 10 rows |
| Abdominal pain | 7.2% | 116 / 17 | orforglipron 99 (85%), weight 36.4% | semaglutide 46.0%, from 7 rows |
| Dizziness | 6.5% | 97 / 17 | orforglipron 78 (80%), weight 9.5% | tirzepatide 70.5%, from 4 rows |
| Fatigue | 6.9% | 65 / 12 | orforglipron 49 (75%), weight 26.1% | semaglutide 67.5%, from 8 rows |
| Acid reflux | 5.0% | 52 / 14 | orforglipron 32 (62%), weight 10.9% | tirzepatide 70.6%, from 5 rows |
| Gallstones | 1.7% | 20 / 7 | orforglipron 9 (45%), weight 0.9% | semaglutide 53.4%, from 4 rows |
| Pancreatitis | 0.1% | 11 / 3 | tirzepatide 6 (55%), weight 6.7% | semaglutide 92.6%, from 2 rows |
| Hair loss | 4.5% | 8 / 3 | tirzepatide 3 (38%), weight 74.7% | tirzepatide 74.7%, from 3 rows |
| Injection-site reaction | 4.5% | 4 / 2 | tirzepatide 3 (75%), weight 98.9% | tirzepatide 98.9%, from 3 rows |
In 12 of these 14 estimates the molecule supplying the most rows is not the molecule moving the number. Only hair loss and injection-site reaction — our two thinnest bases, 8 rows and 4 — agree, and they agree because there is almost nothing in them to disagree.
Emotional blunting publishes no corpus estimate at all and so appears in no row here.
Nausea, worked through
135 stated rates across 29 studies, pooled at 23.0% with a 95% interval of 6–58%. Orforglipron supplies 103 of the 135 rows from 12 studies — and 7.0% of the weight. Semaglutide supplies 16 rows from 12 studies and 57.5% of the weight. Tirzepatide supplies 5 rows from 3 studies and 35.0%.
So 4% of the rows carry 35% of the number, and 76% of the rows carry 7%. Both statements are about the same estimate, and neither is a correction of the other. If you read "135 rates" as evidence weight, you are reading a figure that is mostly early-phase oral-drug cohorts; if you read the 23.0% itself, you are reading a figure that is mostly the large semaglutide and tirzepatide trials.
What this does and does not license
It does not mean our pooled rates are orforglipron rates. The weighting works; that is what the second column demonstrates.
It does mean that for most effects our class-level number is, in practice, a two-injectable number with a small oral contribution — and that where the oral contribution is not small, you should know it. Abdominal pain (36.4%), fatigue (26.1%) and headache (22.8%) are the three estimates carrying real orforglipron weight, and the reason is worth stating because it is not the one you would guess. It is not that the Phase 1 cohorts grew heavy; it is that the two largest trials in our base — SELECT (semaglutide, 8,803 at risk) and SURPASS-CVOT (tirzepatide, 6,647) — contribute no row to those three estimates (checked row by row: between them they carry eligible rates for nausea, vomiting, diarrhoea, constipation, reduced appetite, acid reflux, dizziness, gallstones and pancreatitis, and for nothing else). Where they do, as in nausea, they take 41.5% and 31.4% of the weight between them and everything else is a rounding error. Where they are absent, the heaviest entries are STEP 1 at 1,306 and orforglipron's own larger trials — a 426-participant arm, a 331-participant arm — not its healthy-volunteer cohorts. (Entry-level weights read from the same production base on 19 September 2026.)
It also does not license a per-drug comparison, and we do not publish one. Our corpus stores no structured trial-phase or duration field — a row's phase appears inside its excerpt text, not as a queryable field — so a per-molecule pool built from it today would mix single-dose Phase 1 cohorts with 72-week Phase 3 arms without being able to separate them programmatically. The per-molecule row and study counts are published so anyone can see the shape of what is there; the comparison itself needs a base we have not built yet.
And it does not make the interval narrower. Nausea's 6–58% is still the file telling you these studies genuinely disagree — now with a public account of who is disagreeing with whom.
Reproduce it
This is educational content about data methods, not medical advice, and nothing here describes what any individual should expect or do. Every figure was read from the live API or the cited registry record on 19 September 2026 and is stated as of that date.
The full dataset — 15 effects, both tracks, every source with its URL and stated sample size, CC BY 4.0 — is at magistra.health/en/data-api, with the eligibility methodology at magistra.health/en/methodology.
Bekijk uw eigen cijfers
Onze gratis voorspeller schat uw bijwerkingsrisico en gewichtsverloop, met bij elk cijfer het aantal vermelde percentages en afzonderlijke bronnen. Geen account nodig.
Open de voorspellerWerkt u met de data zelf? De gedateerde momentopname achter deze cijfers is te koop als eenmalige aankoop, naast de gratis publieke API: Data & API.