Cohort studyindependent cross-checkfull text checked★ key study I-PPI-02 in-depth review
Why PPI–death links may be misleading
Proton pump inhibitors and risk of all-cause and cause-specific mortality: A cohort study.
Overview
Brown JP et al.. Br J Clin Pharmacol 2021 · PMID 33393677 · DOI 10.1111/bcp.14728
The link between PPIs and death shrank with better adjustment and appeared even for causes PPIs could not plausibly cause.
Does the study support the specific claim? No · Methodological quality: high
Plain-language summary
Study explained simply
- What did the researchers want to find out?
Confounding in PPI mortality.
- How was the study done?
UK cohort.
- How many people, animals or samples were studied?
>2 million.
- What was actually found?
Association attenuates; implausible associations.
- How large was the observed effect?
HR 1.65 → 1.31.
- What do the results mean in everyday life?
Sicker people get PPIs.
- What are the study's limitations?
Observational.
- What does it say about Elawady's specific claim?
Limits C-PPI-01.
- Do the results apply to Germany and Europe?
UK.
- What can explicitly not be concluded from this study?
True causal effect.
Scientific in-depth review
1. Full reference
- Authors
- Brown JP et al.
- Journal
- Br J Clin Pharmacol
- Year
- 2021
- Identifier
- PMID 33393677 DOI 10.1111/bcp.14728
- Verification
- verified – PubMed efetch 2026-10-25
- Retracted
- no
- Corrections
- none
- Extent checked
- full text
- Conflicts of interest and funding
No specific funding for the study. One author funded by a GlaxoSmithKline PhD studentship and another by an unrestricted GSK grant; others by MRC DTP and Wellcome/Royal Society. Authors report additional grants (NIHR, MRC, BHF, Diabetes UK) and personal fees from Grunenthal outside the work.
2. Research question
Association between PPIs and all-cause and cause-specific mortality; role of confounding
3. Study design and population
- Study design
- Cohort study – Retrospective new-user cohort study in UK primary-care electronic health records (CPRD GOLD) linked to ONS deaths, HES and IMD, with active-comparator and matched non-user comparisons and negative-control outcomes
- Population
Adults newly prescribed a PPI or H2RA in CPRD GOLD practices eligible for linkage; secondary cohort of PPI users and age/sex/year/practice-matched non-users of acid suppression
- Country
UK
- Period
Cohort entry 1998-2015; follow-up to 17 April 2017; median follow-up 4.1 years (PPI) and 3.0 years (H2RA)
- Sample size
Primary: 733,885 PPI vs 124,410 H2RA new users (95,489 and 8800 deaths); secondary: 689,602 PPI users vs 1,361,245 matched non-users
- Inclusion and exclusion criteria
First PPI/H2RA prescription after age 18 and >=1 year of up-to-standard registration; linkage-eligible practices
4. Methods and endpoints
- Exposure / intervention
New PPI prescription (omeprazole, lansoprazole, pantoprazole, rabeprazole, esomeprazole)
- Comparator
New H2RA users (main); matched non-users (secondary)
- Endpoints
All-cause mortality; cause-specific mortality (GBD categories; causes with plausible or implausible short-term causal links; control outcomes accidental trauma and pulmonary embolism)
- Statistical methods
Propensity-score ATT-weighted Cox models; high-dimensional PS (500 covariates); time windows 0-6 months, 1 year, 10 years; multiple sensitivity analyses; E-values
5. Actual results
- Results
Crude death rates 26.5 vs 16.1 per 1000 person-years (PPI vs H2RA). All-cause HR unadjusted 1.65 (1.62-1.69), PS-weighted 1.38 (1.33-1.44), hdPS 1.31 (1.26-1.37); first 6 months 1.34 (1.25-1.43). Implausible short-term associations, e.g. lung cancer mortality at 6 months 1.77 (1.39-2.25); neoplasm mortality 1.74 (1.63-1.86). Injury/external causes 1.00 (0.78-1.26); accidental trauma control 1.05 (0.69-1.59). Versus non-users: all-cause 1.96 (1.94-1.99), neoplasms 3.74, gastric cancer 14.59. Censoring at treatment discontinuation: 1.12 (1.04-1.20). E-value: unmeasured confounder with RR >=1.33 with both or >=1.99 with either could explain the lower CI bound.
- Absolute risks
Crude mortality 26.5 vs 16.1 per 1000 person-years (PPI vs H2RA) and 24.8 vs 11.5 per 1000 person-years (PPI vs matched non-users); these unadjusted differences are not causal estimates
- Authors' conclusion
PPI prescription was strongly associated with all-cause and cause-specific mortality, but the shifts in HRs with adjustment and comparator and the implausible early associations indicate residual confounding; fully accounting for it with observational data may not be possible.
6. Strengths
- Very large, representative UK population with linked death registration
- New-user active-comparator design plus non-user comparison
- Negative-control outcomes and time-window analyses to probe confounding
- hdPS and E-value quantitative bias analysis
7. Weaknesses
- Observational; unmeasured frailty/disease severity likely
- Prescription-based exposure; OTC use and non-adherence unrecorded
- Many cause-specific outcomes (multiplicity)
- Cause-of-death misclassification possible
- Industry studentship funding (GSK) for some authors
8. Risk of bias
- Orienting tool
- ROBINS-E
- Type of assessment
- structured assessment based on the full text by a single reviewer (no formal double review)
- Full formal assessment
- no
- Overall risk of bias
- high
- Rationale
Single-reviewer structured ROBINS-E judgement from full text; not formal. Judged 'high' for estimating a causal PPI effect, which supports - not undermines - the study's use here as evidence that observational PPI-mortality associations are confounded. Bias rating raised from some_concerns to high for the causal question.
| Domain | Judgement | Rationale |
|---|---|---|
| Confounding | high | Authors' own analyses demonstrate substantial residual confounding (HR depends on comparator and adjustment; implausible early associations). This is the study's key finding rather than a flaw of the analysis. |
| Exposure measurement | some concerns | Primary-care prescriptions only; OTC use and adherence unknown. |
| Selection of participants | low | New-user design with clear eligibility; 6% of PPI users unmatched in secondary analysis. |
| Post-exposure interventions | some concerns | Switching and discontinuation; sensitivity analyses performed. |
| Missing data | some concerns | Missing-indicator approach for BMI, smoking, alcohol. |
| Outcome measurement | low | National death registration with certified causes. |
| Selection of reported result | low | Prespecified outcome sets including control outcomes; full reporting in supplements. |
Full-text review (25 October 2026): Added crude death rates, time-window and control-outcome results, discontinuation-censored HR 1.12, E-values, comparator-dependence (1.38 vs 1.96), study period and funding detail; ROBINS-E domains with overall set to high for the causal question (from some_concerns), noting this supports the study's role as evidence of confounding.
9. Assessment of the statistical interpretation
Large, precise associations that shrink with better adjustment, depend strongly on comparator choice and appear for causes PPIs could not plausibly cause within 6 months - a pattern typical of confounding by indication/frailty. Not evidence of a causal mortality effect.
10. Relevance to the specific claim
Methodological counterpoint to A-PPI-12/13 (C-PPI-01).
11. Independent contrary findings
12. Transferability to Germany and Europe
UK data; high.
13. Overall methodological judgement
Strong methodological evidence that observational PPI-mortality associations (such as the '45 extra deaths per 1,000' estimate) are largely confounded; it does not prove the absence of any harm.
| Multidimensional assessment | |
|---|---|
| Authenticity of the source | verified |
| Methodological quality | high |
| Risk of bias | high |
| Precision of results | high |
| Directness | direct |
| Reproducibility | unknown |
| Consistency with the overall evidence | unknown |
| Regional transferability | high |
| Adequacy of the interpretation (by the archive) | not applicable |
14. Open questions
Long-term randomised data on mortality beyond 3-5 years.