Proton pump inhibitors (PPIs — omeprazole, esomeprazole, lansoprazole, pantoprazole and relatives) are one of the most widely prescribed drug classes, and they are a named standard drug in the pharmacotherapy taper: a large share of the adult population is on one, they define major strata, and they are a default primary-care prescription. This page appraises the harm half of that standing object — what chronic PPI use is associated with, how much of that is likely causal, and what it means for a decision to continue, deprescribe, or replace the drug. Efficacy for acid-related disease is not in question here; selection, dosing, and titration are prescriber-zone and out of scope. (the scoping and pharmacotherapy-taper framing are the wiki’s own; the harm findings below are (Salvo et al., 2021)).
The one-line verdict: the PPI-harm literature is large, consistent in direction, and almost entirely observational — so it establishes association, not causation, and the single biggest reason to distrust the causal reading is confounding by indication (sicker people get PPIs). The individual effect sizes are modest, the events are rare, GRADE certainty is very low for most outcomes, and the one large randomized trial nullified nearly all of them. The decision therefore hinges not on the harm estimate but on whether there is a benefit to weigh it against.
What the umbrella found
Salvo et al pooled 42 systematic reviews with meta-analyses (predominantly observational, some spanning millions of patients), taking the most comprehensive meta-analysis per outcome to avoid double-counting, and comparing significance and direction only — effect estimates were explicitly not quantitatively pooled across the umbrella (Salvo et al., 2021). The significant associations in the most-comprehensive analyses:
| Outcome (most-comprehensive MA) | Effect (95% CI) | n |
|---|---|---|
| Hip fracture | RR 1.20 (1.14-1.28) | 2,103,800 |
| Acute kidney injury | RR 1.61 (1.16-2.22) | 2,396,640 |
| C. difficile infection | OR 1.99 (1.73-2.30) | 356,683 |
| Gastric cancer | OR 2.50 (1.74-3.85) | 943,070 |
| Fundic gland polyps | OR 2.46 (1.42-4.27) | 40,218 |
(Salvo et al., 2021) Significant associations were also reported for other fracture sites, osteoporosis, CKD/ESRD, pneumonia, enteric infections, hypomagnesaemia, SIBO, and fall risk. These are relative effects; the absolute risk is small. Only 8 of the 42 studies gave enough data to compute a number-needed-to-harm «for one or more adverse outcomes, which ranged from 7 to 240 patients» (Salvo et al., 2021); Salvo reads this as evidence that «most identified adverse outcomes were generally rare» (Salvo et al., 2021). Salvo’s own summary judgement on the associations that are statistically significant: «the effect estimates were generally small» (Salvo et al., 2021).
What is NULL — the negative space is a finding
(Salvo et al., 2021) The umbrella concluded no association for several of the most-feared harms: «No associations with non-gastric cancers, or neurological disease were concluded, with conflicting evidence for cardiovascular outcomes.» Specifically colorectal and pancreatic cancer were null, and — directly against the popular belief that long-term PPIs cause dementia — dementia and Alzheimer’s disease were not significantly associated, most of these nulls supported by meta-analyses of >100,000 patients. Gastric mucosal atrophy was also null. This is the no meaningful effect evidence state (well-powered nulls), kept distinct from insufficient evidence.
Why the causal reading is weak — confounding by indication is the load-bearing caveat
The umbrella’s central limitation is not analytic — it is structural. «The large amount of observational evidence increases the risk of unrecognised bias» (Salvo et al., 2021), and because the evidence is predominantly observational, «only association, rather than causation, can be established» (Salvo et al., 2021). Salvo makes the confounding argument explicit with a worked illustration of a spurious association driven by a common cause: «ice cream consumption is associated with warm weather, as is gun violence. Consequently, ice cream consumption is associated with gun violence. However, restricting ice cream consumption is unlikely to impact on gun violence.» (Salvo et al., 2021) The PPI-specific form of this is confounding by indication: the people prescribed chronic PPIs are sicker, poly-medicated, and older than non-users, so they carry more of every adverse outcome for reasons that have nothing to do with acid suppression (inferred from Salvo et al., 2021) (the named mechanism the umbrella’s observational base cannot exclude).
Certainty is graded accordingly. «Individual systematic reviews with meta-analyses were usually considered to have very low certainty of evidence in the GRADE assessment, due to the incorporation of predominantly observational evidence» (Salvo et al., 2021), with high heterogeneity and unassessed primary-study quality driving further downgrades. GRADE starts observational evidence at low certainty by construction -> Risk of Bias Assessment Tools, Upgrading Observational Evidence.
The mechanisms are mostly unconfirmed. Proposed pathways exist (impaired B12/calcium/magnesium absorption, reduced gastric acidity letting ingested pathogens survive, gastrin-driven ECL-cell hyperplasia for gastric cancer), but «apart from the increased occurrence of enteric infections, these proposed mechanisms for PPI-related adverse events are not supported with clinical evidence leaving the issue of causation unresolved» (Salvo et al., 2021). The infection channel is the exception — it has both a direct mechanism (acid suppression -> pathogen survival) and, as below, the one randomized confirmation.
The randomized test nullifies nearly all of it
The decisive counter-evidence, cited within the umbrella, is a large RCT. In the COMPASS sub-randomization (Moayyedi 2019), 17,598 patients were randomized 1:1 to pantoprazole or placebo and tracked ~3 years for a pre-specified adverse-event list. «PPI use was not associated with any of these adverse events with the exception of enteric infections» (Salvo et al., 2021). A second observational check pointed the same way on the bone mechanism: the Canadian CaMos cohort «found no association between PPI use and bone demineralisation over 10 years, emphasising the importance of investigating putative mechanisms of adverse events» (Salvo et al., 2021).
This is the sharp case of the observational-trial discordance, and PPI is a drug, not a food. Unlike a dietary exposure — where a blinded trial necessarily tests a different exposure (an isolate, a short swap) so a null does not refute the lifetime pattern — a PPI can be blinded and randomized, and COMPASS tested the same exposure the cohorts did. When the commensurable randomized test is null, the discordance resolves cleanly toward the trial: the observational harm signals are very likely confounding by indication, not drug effect (inferred from Salvo et al., 2021). The one signal both streams agree on — enteric/GI infection — is the one to believe -> The Observational-Trial Discordance.
Duration-response
(Salvo et al., 2021) Several meta-analyses reported non-significant increases in effect with longer PPI duration (hip fracture, osteoporosis, CKD, pneumonia, gastric cancer, cardiovascular events). Only two duration-response signals were statistically significant: end-stage renal disease (rising across 1-3, 6-12, and 12-24-month bands) and fundic gland polyps (rising when use exceeded 12 months). A duration-response gradient is a weak positive causal cue, but it too is read off observational data and inherits the same confounding-by-indication caveat (longer use co-travels with sicker, longer-followed patients).
Decision relevance — the balance turns on benefit, not on the harm estimate
The operative rule from the umbrella’s clinical-implications section: «Regardless of the estimated size of the risk or the likelihood of causation, risk will always dominate a risk-benefit assessment in the absence of benefit» (Salvo et al., 2021). This is the whole decision, and it splits by stratum:
- Where there is no genuine indication — and 25%-70% of PPI prescriptions are considered inappropriate (no formal diagnosis, or continued post-discharge) (Salvo et al., 2021) — even a small, uncertain harm tips the balance toward deprescribing, because there is no benefit on the other side of the scale. This is the large, decision-relevant stratum: the lever is removing an unneeded drug, which has structural leverage (removes the exposure entirely) at essentially no cost.
- Where there is a genuine indication (severe GORD, peptic ulcer, bleeding prophylaxis, Zollinger-Ellison) — PPIs are «the most effective options for certain patients» (Salvo et al., 2021), the events are rare, the harm is probably confounded, and the benefit is real — so PPIs should continue. The fear of fracture/kidney/dementia harm should not drive discontinuation against a real indication.
So the harm literature does not license blanket PPI avoidance; it licenses de-adoption of inappropriate use and honest counselling that the widely-feared harms are mostly small, uncertain, and probably not causal. The realistic alternatives to a PPI (H2-receptor antagonists, lifestyle/weight-loss for reflux, surgical anti-reflux procedures) are a separate comparator appraisal not held here.
A constituent MA (Islam 2018) — granularity, not independence
A second PPI-harm meta-analysis, Islam 2018 (SR of 43 + MA of 28 observational studies, search to July 2016), reaches the same direction on every shared outcome. It is not independent corroboration: Islam is one of the 42 SR/MAs the umbrella pooled — it is listed in Salvo’s included-studies table as «Islam et al, 2018» (ref 25), its hip-fracture-SR constituent (Salvo et al., 2021). Its agreement is therefore laundered-E (a re-count of studies the umbrella already holds), not a type-E robustness gain, and Salvo in fact superseded it — for hip fracture the umbrella chose a larger MA (RR 1.20) over Islam’s. What Islam adds is type-F granularity the umbrella’s significance-and-direction-only abstraction dropped, plus an internal worked demonstration of why the certainty grade is very low.
Parameter table (BLOCKING — matched, individually-quoted, same-quantity column). Islam’s OR must not be read as confirming Salvo’s estimate: they are different pooled analyses, different metrics, different n.
| Outcome | Salvo (most-comprehensive MA) | Islam (constituent MA) | Same quantity? |
|---|---|---|---|
| Hip fracture | RR 1.20 (1.14-1.28), n=2,103,800 (Salvo et al., 2021) | OR 1.42 (1.33-1.57), k=9, I2=81% (Islam et al., 2018) | NO — RR vs OR, different (smaller, superseded) pooled set |
| Acute kidney injury | RR 1.61 (1.16-2.22), n=2,396,640 (Salvo et al., 2021) | OR 2.61 (1.93-3.52), k=4, I2=76% (Islam et al., 2018) | NO — RR vs OR, different MA/n |
| Gastric cancer | OR 2.50 (1.74-3.85), n=943,070 (Salvo et al., 2021) | OR 1.78 (1.41-2.25), k=2, I2=67% (Islam et al., 2018) | PARTIAL — both OR but different, non-overlapping pooled sets; not one analysis |
| Pneumonia (CAP) | “significant” (no effect size carried) | OR 1.67 (1.04-2.67), k=7, I2=99% (Islam et al., 2018) | N/A — Salvo carries no number; Islam supplies the granularity |
| Colorectal cancer | null | OR 1.55 (0.88-2.73) NS, k=4, I2=97% (Islam et al., 2018) | CONCORDANT (both null) — laundered, Islam supplies the NS estimate |
| Pancreatic cancer | null | OR 3.53 (0.36-34.49) NS, I2=100% (Islam et al., 2018) | CONCORDANT (both null) — laundered; estimate uninterpretable |
What genuinely earns Islam its sources: slot (type-F, not E):
- Per-outcome CIs the umbrella abstracted away for pneumonia and the cancers — a decision-thin gain (these outcomes were among those the COMPASS RCT nullified), but distinct content Salvo does not carry.
- Subgroup / dose / duration structure on CAP an umbrella-per-outcome summary structurally cannot hold: male OR 2.40 (1.50-3.86) vs female 0.95 (0.82-1.10); low-dose 1.67 vs high-dose 2.40; duration bands all raised (Islam et al., 2018). Every one sits on I2=99% observational pooling, so it refines — but does not rescue — an association the trial evidence already discounts.
- A worked heterogeneity/instability demonstration — Islam pooled OR/HR/RR together (all assumed to approximate OR), ran no sensitivity analysis, and produced I2 up to 100% and a pancreatic-cancer CI spanning two orders of magnitude (driven by one outlier, Lai 2014, OR 11.24). This is internal evidence for the umbrella’s very-low-certainty grade: a constituent’s own numbers show how fragile the observational pooling is (Islam et al., 2018).
- Concordant caveats (laundered, logged not weighted): Islam itself warns «these results should be interpreted with caution owing to the significant statistical and clinical heterogeneity … and the inherent inability of observational studies to clarify whether the observed epidemiologic association is a causal effect» and that «Publication bias may have resulted in an overestimate» (Islam et al., 2018). Its 2016 search predates COMPASS — it notes «no clinical trial considering the potentially adverse effect of PPI use has yet been conducted» (Islam et al., 2018) — so it cannot make the observational-trial-discordance move; the umbrella does.
Provenance, certainty, and conflict of interest
This page rests on a single gold-tier umbrella review (Salvo 2021, 42 SR/MAs); the second
sources: entry (Islam 2018) is one of those 42 constituents, not independent backing, so it raises
no confidence and confidence: stays low pending a genuinely independent gold source. Two
appraisal caveats bound the umbrella:
- Umbrella double-counting. An umbrella re-counts the primary studies pooled across its constituent meta-analyses, so nominal precision is overstated; Salvo mitigates this by taking one most-comprehensive MA per outcome, but the underlying cohorts still overlap across outcomes.
- Funder conflict of interest — directional. The study was funded by Ethicon Inc (Johnson & Johnson), and authors are employees of EVERSANA (a market-access consultancy) and J&J; the senior author declares industry advisory roles (Salvo et al., 2021). Ethicon manufactures surgical devices including anti-reflux alternatives to PPIs, so the funder has a competing product — a COI that runs toward emphasising PPI harm (inferred from Salvo et al., 2021). Notably the umbrella’s actual conclusions run the other way (harms mostly small, uncertain, non-causal; the RCT nulls them), which is reassurance against the COI having distorted the read — but the caveat is logged -> Risk of Bias Assessment Tools (the funding-effect lens).
Open loop. This grades the evidence on PPI harm and its causal status; it does not close the loop against any patient’s realized outcome. A second independent gold source on PPI safety (and a deprescribing-outcomes SR) would upgrade it -> the deprescribing-effect question the umbrella points to but does not answer.