Evidence-Tier Checklist for the Studies Patients Quote

Patients now arrive carrying citations. A screenshot of a systematic review, an abstract with a hazard ratio, a news headline over a registry number — the clinician's job is no longer to find the evidence but to tier it in thirty seconds and answer within what that tier supports. Most patients cannot tell a validated-scale trial from a billing-code cohort, and neither can most headlines. The four-point checklist below maps the most common failure modes in the studies patients quote, each with a worked example from the October 2026 literature, and ends with the evidence-tier label we recommend as a standing element of every study summary on this site.

One boundary up front: nothing here resolves whether GLP-1 receptor agonists help or harm psychiatric outcomes. That question is exactly what the current evidence cannot answer, and this brief is deliberately about how to read the evidence, not what it concludes.

Checklist Item 1: Were outcomes measured with validated scales, or pulled from diagnostic codes?

The fastest way to over-read a study is to ignore how its outcome was operationalized. Administrative codes exist for billing; they detect coded encounters, not symptoms. Validated instruments (PHQ-9, GAD-7, and their peers) measure what the label claims. Sample size does not repair a measurement gap.

Worked example: a 2026 systematic review of GLP-1 receptor agonist psychiatric safety (Cureus, PMID 42829595) screened 617 records and included 11 primary studies drawing on more than 4.9 million patients. Ten of the eleven studies found no increase in psychiatric risk or a protective association; one obesity-specific study reported increased risk, with the authors noting surveillance or detection bias as a possible partial explanation — while stating they did not test that hypothesis. The methodological crux: only 1 of the 11 studies measured mood with validated questionnaires. The rest rest on diagnostic codes, and the review was not prospectively registered. "4.9 million patients" is a true sentence that describes the denominator of coded data, not the strength of mood measurement.

Conversation language that stays within the evidence: "Large reviews show no consistent signal in coded diagnoses, but almost none of those studies measured mood directly with validated questionnaires — so we can say the alarm headlines aren't well supported, and we can also say genuine mood effects haven't been ruled in or out by direct measurement." That sentence is fully defensible at the point of care and does not require relitigating the safety question.

Checklist Item 2: In a claims-data cohort, can confounding plausibly explain the effect size?

Propensity matching balances measured covariates; unmeasured confounding is the residue. The E-value quantifies it: the minimum strength of association an unmeasured confounder would need with both treatment and outcome to explain away an observed hazard ratio (PMID 28693043, with applied guidance in PMID 30676631). E-values near 1 mean small confounders suffice; larger E-values demand stronger explanations.

Worked example: a retrospective TriNetX cohort compared prior GLP-1 receptor agonist versus SGLT2 inhibitor use in adults with type 2 diabetes who developed sepsis (Drug Des Devel Ther, PMID 42829714). After 1:1 propensity matching, 11,969 patients per arm. Primary outcome, 90-day all-cause mortality: 48.6 versus 60.6 per 100 person-years, HR 0.81 (95% CI 0.75–0.88), E-value 1.6. Secondary outcomes: MACE HR 0.77 (0.65–0.91, E-value 1.7), myocardial infarction HR 0.82 (0.67–0.99), ICU admission HR 0.89 (0.84–0.94, E-value 1.4), MACCE HR 0.86 (0.74–0.99).

Three appraisal moves on this single table:

- Read the E-value against the effect, not the CI. An E-value of 1.6 is modest: an unmeasured confounder of moderate strength — severity of illness at presentation, formulary-driven channeling, adherence itself — could account for the observed mortality association. The correct tier for this study is claims-data association, not treatment effect. - Count the secondary outcomes and ask what survived multiplicity correction. The authors applied Holm correction across the seven secondary outcomes; only MACE (adjusted P=0.012) and ICU admission (adjusted P<0.007) survived, while MACCE, myocardial infarction, MAKE, septic shock, and acute respiratory failure did not. Correction does not rescue an underpowered finding set: with E-values this modest, even the Holm-surviving associations should be read as hypothesis-generating, exactly as the authors themselves label them. - Look for the null result. The kidney outcome was null: MAKE HR 1.08 (0.92–1.27), and septic shock and acute respiratory failure showed no significant difference either. Null results are findings — and they are the part of a positive cohort study that downstream summaries will drop first.

Checklist Item 3: What class of document is this — trial, review, or editorial?

PubMed-indexed does not mean evidentiary. Editorials, narrative reviews, and opinion pieces pass through the same database and often the same news pipeline as trials, with none of the data.

Worked example: "Integrative men's health: testosterone, metabolic fitness, and preventive urology" (Curr Opin Urol 2026;36(6):605-606, PMID 42828776) is a two-page editorial with zero new data. It makes a direction-setting argument — testosterone management belongs inside an integrative men's health frame — from a credible specialty voice. That is a legitimate use. What it cannot do is carry an outcomes claim, and our own ingestion pipeline initially tagged it priority high, which is precisely the error class this checklist exists to catch: a signal-detection system, like a news feed, will promote a quotable editorial on priors alone.

The appraisal habit is one question: does this document contain its own data? If not, cite it as rationale, never as evidence.

Checklist Item 4: Is the endpoint a finding or a plan — and if measured, is it a hard outcome or a surrogate?

Two registry records illustrate the distinction from both directions.

SELECT-LIFE (NCT04972721) is the follow-up of participants from the SELECT cardiovascular outcomes trial, designed to answer what happens to the cardiovascular benefit after the intervention stops — a durability question SELECT could not answer by design. As of this writing the record posts no results. It is a question with a study attached, not an answer; any durability claim sourced from it is premature by construction.

NCT06557811 measures oral semaglutide against epicardial and pericoronary adipose tissue in patients with type 2 diabetes after myocardial infarction — also results-free. Those fat depots are imaging-derived, inflammation-linked markers sitting upstream of hard cardiovascular events. A surrogate endpoint is a legitimate, faster proxy that can support a biology claim (this drug changes the tissue) but not an outcomes claim (this drug changes events). The title reads literally: adipose tissue change is the endpoint, not cardiac benefit.

The combined rule: registry record without results = plan; surrogate endpoint result = biology, not outcomes; extension-study result = durability in a selected surviving cohort, not the parent trial's population.

The Evidence-Tier Label: Making the Checklist a Standing Format

We propose the checklist collapse into a one-line label carried on every study summary on this site — and cross-linked from the patient-facing guide at LuxeFit, which works through the same four examples in plain language (for the patient view, see "How to read a health headline," luxefitwellness.com):

| Tier | Definition | Worked example | |---|---|---| | Validated-scale trial | Outcome measured with a validated instrument, ideally prospectively registered | The single validated-questionnaire study inside PMID 42829595 | | Claims-data association | Administrative-data cohort; associations only; report E-value and nulls | TriNetX sepsis cohort, PMID 42829714 (E-value 1.6) | | Editorial or commentary | No new data; direction-setting rationale only | PMID 42828776, two-page Curr Opin Urol editorial | | Registry-only | Registered plan, no results posted | NCT04972721; NCT06557811 (surrogate endpoints, no results) |

The label costs one line and prevents the two failures that matter in practice: overclaiming from association, and treating a plan or an opinion as a result.

Frequently Asked Questions

No. It means an unmeasured confounder of moderate strength could explain the association, so the finding stays at the association tier until randomized evidence or much larger E-values exist.

Codes determine what was captured, N determines how precisely. Ten million coded encounters still measure coding behavior, not mood.

They are the right tool for biology questions and the wrong tool for outcomes claims. The failure is labeling, not the endpoint.

Acknowledge the direction, name the document class, and offer the trial evidence on the same question.

Further Reading

- Wound Care Biologics Comparison — the evaluation-and-reimbursement pillar this checklist extends.

References

1. PMID 42829595 — Cureus 2026 systematic review, GLP-1 RA psychiatric safety (codes-versus-scales example). 2. PMID 42829714 — Drug Des Devel Ther 2026, TriNetX sepsis propensity-matched cohort (E-value example). 3. PMID 42828776 — Curr Opin Urol 2026 editorial, integrative men's health (source-class example). 4. NCT04972721 — SELECT-LIFE extension study registry record (results-free). 5. NCT06557811 — Oral semaglutide and epicardial/pericoronary adipose tissue (surrogate endpoints, results-free). 6. PMID 28693043 — VanderWeele & Ding, Ann Intern Med 2017, "Sensitivity Analysis in Observational Research: Introducing the E-Value." 7. PMID 30676631 — Haneuse S, VanderWeele TJ, Arterburn D, "Using the E-Value to Assess the Potential Effect of Unmeasured Confounding in Observational Studies," JAMA 2019;321(6):602-603.