A custom AI reviewer for causal language in observational studies

PerfectPaper’s causal language agent reads an observational manuscript for the gap between what the design estimates and what the sentences claim. It identifies causal verbs applied to associational findings, checks whether the adjustment set is justified or merely assembled, and flags the recommendation that quietly assumes causation. Copy the brief below into a custom agent slot.

The pattern is consistent: methods are appropriately cautious, results are neutral, and the abstract and final discussion paragraph revert to causal claims. That is where this agent concentrates.

That pattern has been measured. Haber and colleagues rated 1,170 articles drawn from 18 high-profile medical, public health and epidemiology journals published between 2010 and 2019, 65 articles per journal, using three reviewers per article (“Causal and Associational Language in Observational Health Research: A Systematic Evaluation”, American Journal of Epidemiology 2022;191(12):2084–2097). The causal implication of the exposure–outcome linking language in the abstract was rated none in 13.8%, weak in 34.2%, moderate in 33.2% and strong in 18.7% of abstracts. Roughly one abstract in five in the most selective journals in the field already carries strongly causal wording. That rating required three independent reviewers per article, not one, because classifying implied causality from wording is a judgment call trained readers do not always resolve the same way — the sentence-level check below exists for the same reason a banned-word list does not.

When to use this agent

The causal language agent is the right default slot when several of these signals appear together in one manuscript:

• Your study is observational and the discussion recommends an intervention
• Your abstract uses verbs like reduces, improves, prevents or leads to
• Covariates were selected by stepwise procedures or by significance testing
• The paper has many co-authors and the discussion was written last
• A reviewer has previously asked you to soften your conclusions
• Your covariate list includes a variable measured after the exposure
• Your limitations paragraph mentions residual confounding and gives no number
• A press office will write a release from your abstract

Questions about spotting causal language before submission

How do I check my observational paper for causal language before submission?

Extract the abstract’s conclusion sentence, the first sentence of the discussion, the final discussion paragraph, and any recommendation, onto one page, then compare each against what the design can identify. Search the full text for causal verbs — reduces, prevents, leads to — and note which section each falls in. PerfectPaper’s causal language agent automates that comparison and returns a replacement for each mismatch.

Is my finding an association or a cause?

Ask what the design licenses, not how strong the number looks. A randomised trial licenses causal language for its primary comparison; an observational study licenses an association unless the analysis explicitly states a causal framework and the assumptions behind it — exchangeability, positivity and consistency. Absent that framing, the honest description is an association, however large the estimate.

Can an AI reviewer catch overstated causal claims in a manuscript?

Yes, at the sentence level, which is where the failure usually sits. PerfectPaper’s causal language agent quotes each causal verb applied to an associational finding, compares the caution of the methods against the strength of the abstract, and proposes specific replacement wording. It does not re-derive the identification strategy; a separate specialist on the standing team covers that.

How far causal language outruns the design

Causal language outruns the design most reliably in two places: the action recommendation and the choice of linking verb. In the Haber 2022 sample, the implied causality of the action recommendation was higher than the implied causality of the linking sentence in 44.5% of articles and merely commensurate in 40.3% — that is, the advice a paper gives is more causal than the relationship it reports in nearly half of published observational work.

The linking verb itself is a poor safeguard. “Associate” was the most common linking word in abstracts at 45.7%, and the same study found reviewers’ ratings of linking-word roots “highly heterogeneous”, with over half of reviewers rating “association” as carrying at least some causal implication. The word most authors reach for to signal caution is read as causal by more than half of trained readers.

That is the practical case for a sentence-level check rather than a vocabulary rule. The question is not whether a paper used the word “cause”. It is whether each claim, in the place a reader will encounter it, matches what the design can identify. The agent brief below is written to ask that question sentence by sentence, in the abstract, the first line of the discussion and the final paragraph, which is where the Haber ratings concentrate.

Why softening the verb is not the remedy

Softening the verb without changing the claim is a known failure mode with a named primary source. Miguel Hernán argued in “The C-Word: Scientific Euphemisms Do Not Improve Causal Inference From Observational Data” (American Journal of Public Health 2018;108(5):616–619) that “being explicit about the causal objective of a study reduces ambiguity in the scientific question, errors in the data analysis, and excesses in the interpretation of the results.”

The consequence for this agent is specific: it does not enforce euphemism, and it does not flag the word “cause”. Where a manuscript adopts a causal framework explicitly and states the identification assumptions — exchangeability, positivity and consistency — causal language is the correct language, and the brief instructs the agent to say so rather than flag it. What it flags is the third state, which is neither: associational vocabulary carrying a causal claim, or causal vocabulary with no stated assumptions behind it.

The cleanest way to move a paper out of that third state is to name the target trial the analysis is emulating. Hernán, Sauer, Hernández-Díaz, Platt and Shrier set out the framework in “Specifying a target trial prevents immortal time bias and other self-inflicted injuries in observational analyses” (Journal of Clinical Epidemiology 2016;79:70–75), where the biases follow from failing to align eligibility, treatment assignment and start of follow-up. A manuscript that states its eligibility criteria, its treatment strategies, its assignment procedure and its start of follow-up as a protocol has answered the causal-language objection structurally, and has usually also foreclosed immortal time bias on the way.

Target trial emulation earns its keep on a narrower claim than it is sometimes given credit for. Aligning eligibility, assignment and follow-up start removes design-induced biases — immortal time, prevalent-user bias, self-inflicted misalignment of the three — but it does not by itself remove confounding from an unmeasured common cause. Exchangeability is a separate, untestable assumption that the protocol states explicitly rather than one the protocol satisfies; a beautifully specified target trial can still be confounded.

The agent brief

Paste this into a custom agent. Suggested settings: work type causal_inference, skill level graduate, no tools required.

Name: Causal language discipline

You are reviewing an observational study for consistency between what its
design can estimate and what its sentences claim. Focus on wording and on the
justification of the adjustment set. Do not re-examine identification strategy
in depth; another reviewer covers that.

Determine and report:

1. Design and estimand. State what the design supports: an association, an
   adjusted association, or a causal effect under stated assumptions. Where the
   manuscript explicitly adopts a causal framework with stated assumptions,
   causal language is appropriate and should not be flagged; say so.

2. Causal verbs. Identify every sentence using causal verbs for an
   associational finding: causes, reduces, increases, improves, prevents,
   leads to, results in, protects against. Quote each with its location and
   propose a specific rewording that preserves the finding. Concentrate on the
   abstract, the first line of the discussion, and the final paragraph, where
   causal language most often appears even when the methods are careful.

3. Asymmetry. Compare the caution of the methods and results with the strength
   of the abstract and conclusions. Report where the abstract makes a stronger
   claim than the results support, and quote both.

4. Adjustment set. Determine how covariates were selected: from a stated
   causal model, from prior literature, by stepwise selection, or by
   significance in bivariate testing. Report selection by statistical
   criteria, and note that it can introduce bias by selecting on colliders or
   mediators. Report the absence of any stated rationale.

5. Colliders and mediators. Identify any adjustment variable plausibly on the
   pathway between exposure and outcome, or plausibly caused by both. Report
   these individually, naming the variable and the likely consequence.

6. Causal diagram. Determine whether a causal diagram or an explicit statement
   of assumed structure is provided. Report its absence as a gap rather than
   an error, and note what it would let the reader evaluate.

7. Residual confounding. Determine whether the manuscript acknowledges
   unmeasured confounding and whether any quantitative assessment is offered.
   Report a limitations paragraph that mentions it in passing and then
   proceeds as though it were absent.

8. Recommendations. Identify recommendations for practice or policy. For each,
   determine whether it requires a causal interpretation the design cannot
   support, and quote the sentence.

For every flagged sentence, supply a concrete replacement rather than advising
the author to be more cautious.

The brief follows the structure described in how to write a custom reviewer agent: a named scope, an explicit non-scope in the second paragraph, numbered determinations rather than themes, and a closing instruction that fixes the output format. The final line — supply a replacement, not an exhortation — is the one that changes what comes back. Without it, a model returns “consider more cautious phrasing” for every flagged sentence, which is the least useful sentence in peer review. Other prebuilt briefs in the same shape are listed under custom AI reviewer agents.

What this agent catches

The causal language agent catches the sentence patterns below by translating each into the weaker claim the design actually supports.

The linking-verb ladder

Linking verbs form a ladder of commitment, and a manuscript should sit on one rung and stay there. The rungs below are ordered from the weakest commitment to the strongest; the agent’s job is to check that the rung used in the abstract is not higher than the rung the methods earn.

Two rungs deserve particular care. “Independently associated” is the phrase that most often smuggles a causal claim past a methods section, because “independent” names a property of the model rather than of the world. “Predicts” is a discrimination claim: a coefficient with a small p-value in an explanatory model says nothing about out-of-sample prediction, and reviewers in clinical journals now say so routinely.

A third caveat cuts the other way on “predicts.” A covariate can raise a model’s discrimination substantially while having no causal effect on the outcome at all, because prediction and intervention answer different questions — a variable that improves how well the model separates cases from non-cases is not thereby a target for intervention, and a validated predictive model is not evidence for any of the causal claims further down the ladder.

Questions about linking verbs and wording

Can I say a treatment “is linked to” an outcome in my results section?

Avoid it there. Haber and colleagues found reviewers’ ratings of linking-word roots highly heterogeneous, and “is linked to” carries no defined statistical meaning of its own — it reads as causal to most readers without committing to a testable claim. Use “was associated with” for a plain relationship, or “was independently associated with” once the adjustment set is stated and justified.

Can an observational study use causal language without being flagged?

Yes, when it adopts a causal framework explicitly and states the assumptions required — exchangeability, positivity and consistency. The agent is instructed to recognise that case and say so rather than flag it. What it flags is causal wording without that framing, and associational wording carrying a causal claim.

How do I fix causal language a reviewer has already flagged?

Change the scope of the claim rather than the strength of the verb. Replace “the treatment reduced mortality” with “receipt of the treatment was associated with lower mortality after adjustment for the covariates listed, with an E-value of 2.37,” and replace “these findings support adopting X” with “these findings are consistent with a benefit of X worth testing in a randomised trial.”

How adjustment sets get assembled instead of justified

Adjustment sets get assembled when the selection procedure is statistical and justified when it is causal, and the distinction is visible in one sentence of the methods. Selection by bivariable screening — entering a variable into the multivariable model only if it reaches p < 0.05 alone — was shown to be inappropriate by Sun, Shook and Kay in the Journal of Clinical Epidemiology (1996;49(8):907–916), whose central point is that a genuinely important variable can be non-significant in bivariable analysis precisely because its relationship with the outcome is confounded by a variable not yet in the model. The screen discards the variables that most need controlling.

The causal alternative is stated compactly by VanderWeele in “Principles of confounder selection” (European Journal of Epidemiology 2019;34(3):211–219): control for each covariate that is a cause of the exposure, or of the outcome, or of both; exclude from that set any variable known to be an instrumental variable; and include any proxy for an unmeasured common cause of exposure and outcome. This is the modified disjunctive cause criterion, and it requires no complete causal diagram — only a judgement, for each candidate covariate, about whether it causes the exposure and whether it causes the outcome.

The criterion cuts the other way too: “adjust for everything measured” is not the safe default it looks like under it. A variable that is a proxy for an unmeasured cause of the outcome alone, with no path to the exposure, does not qualify for the set, and including it anyway can open a collider path if two such proxies happen to share a hidden common cause. The criterion is a reason to adjust for less, not more, once a covariate’s role is genuinely unclear.

Where the full structure is available, the diagram is better. Greenland, Pearl and Robins introduced causal diagrams to epidemiology in Epidemiology (1999;10(1):37–48) as “a starting point for identifying variables that must be measured and controlled to obtain unconfounded effect estimates”, and a published directed acyclic graph lets a reviewer dispute your assumptions rather than guess at them. STROBE item 16(a) already asks for the justification: it requires unadjusted and confounder-adjusted estimates with their precision, and asks authors to “make clear which confounders were adjusted for and why they were included”. The clause after the comma is the one manuscripts skip; the STROBE checklist is where a reviewer will point when they do.

Colliders, mediators and the variables that must leave the model

Colliders and mediators are the two covariate classes whose inclusion makes an estimate worse, and neither announces itself in a regression table. A collider is a variable caused by both the exposure and the outcome, or by causes of each; conditioning on one induces an association that does not otherwise exist, which is why collider bias is a consequence of adjusting rather than of failing to adjust. A mediator sits on the pathway from exposure to outcome; adjusting for one removes part of the effect you are trying to estimate, converting a total effect into a partial direct effect without saying so.

The operational test is temporal and causal, not statistical. For each covariate, establish when the process that produced its value occurred relative to the exposure, and state whether it is a cause of the exposure, a cause of the outcome, both, or neither. A variable measured after the exposure is a mediator or a collider until argued otherwise. That classification — not model fit, not a change-in-estimate threshold — decides membership of the adjustment set, and the distinction is worked through in mediator versus confounder.

One structure defeats the timing test on its own, and it needs only one covariate to do it. A variable measured before the exposure can have no causal effect on the exposure and none on the outcome, and still be a collider: one unmeasured variable causes both it and the exposure, a second unmeasured variable causes both it and the outcome. The covariate then sits at the point of the M in the path exposure ← U1 → covariate ← U2 → outcome, and adjusting for that single variable opens a path that was closed unconditionally. This shape, called M-bias in the causal-diagrams literature, is why “measured before the exposure” is a necessary check on a candidate confounder, not a sufficient one — pretreatment timing rules out mediators, not colliders of this shape.

Questions about adjustment sets and confounders

Why did my effect estimate change so much when I added one more covariate?

Check what kind of variable you added before reading the shift as confounding removed. If the new covariate is a mediator — on the pathway between exposure and outcome — adjusting for it removes part of the effect you are estimating, which produces the same attenuation as confounding control but means something different. Classify the variable by timing and causal role first; see mediator versus confounder.

How do I check whether my adjustment set supports the causal language in my abstract?

Classify each covariate by its causal role rather than by a statistical criterion: is it a cause of the exposure, of the outcome, of both, or neither, and was its value produced before or after the exposure. Selection by stepwise or bivariable screening can pull in a collider or a mediator instead of a confounder. The agent reports the selection procedure used and any covariate whose role is unstated.

Do I need a DAG before an AI reviewer can check my causal language?

No. A published directed acyclic graph makes the assumed structure explicit and lets a reader dispute it, and the agent reports its absence as a gap rather than an error. Where no full diagram exists, a judgement per covariate about whether it causes the exposure, the outcome, or both is enough for the agent to check the adjustment set against the claims.

Two downstream errors follow from getting this wrong. The first is reading the attenuation of a coefficient as evidence that the effect was confounded, when conditioning on a mediator produces the same attenuation for an entirely different reason — the case set out in my effect disappeared after adjusting. The second is the Table 2 fallacy: presenting every coefficient from one model as though each were an effect estimate for its own variable. Westreich and Greenland named and diagrammed this in the American Journal of Epidemiology (2013;177(4):292–298), showing that covariate coefficients from a model tuned for the exposure may be confounded even when the exposure coefficient is not, and that direct-effect estimates get read as total effects. The remedy — separate models or explicit labelling — is described in the Table 2 fallacy.

Residual confounding: bound it rather than acknowledge it

Residual confounding should be bounded with a number, not acknowledged with a sentence, and the tool for that is the E-value. VanderWeele and Ding defined it in the Annals of Internal Medicine (2017;167(4):268–274) as the minimum strength of association, on the risk ratio scale, that an unmeasured confounder would need with both the exposure and the outcome to fully explain away the observed association, conditional on the measured covariates. For an observed risk ratio above 1 it is RR + √(RR × (RR − 1)): an observed risk ratio of 1.5 has an E-value of 2.37, and a risk ratio of 2.0 has an E-value of 3.41. The informative version is the E-value for the confidence limit nearest the null, because it states what would be required to move the result to non-significance.

The formula takes an estimate that is already on the risk-ratio scale. An odds ratio or hazard ratio stands in for one only when the outcome is rare — the E-value authors’ own software draws that line at an outcome affecting under 15% of the cohort by the end of follow-up — so where the outcome is common the estimate has to be converted toward the risk-ratio scale first, which is why that conversion is built into the software rather than left to the analyst. Feeding a common-outcome odds ratio straight into the formula overstates how strong an unmeasured confounder would need to be, because a common-outcome odds ratio sits further from the null than the risk ratio it is standing in for. A reviewer checking a submitted E-value should first check what scale the reported estimate is actually on.

A second check is a negative control. Lipsitch, Tchetgen Tchetgen and Cohen set out negative control exposures and negative control outcomes in Epidemiology (2010;21(3):383–388) as a routine precaution borrowed from laboratory practice: a variable that could not plausibly be affected by the exposure, or an exposure that could not plausibly affect the outcome, should show a null association if the analysis is unconfounded. A non-null result on a negative control is evidence of confounding, selection or an analytic flaw that no adjustment list will fix.

The check runs in one direction more safely than the other. A null result is necessary, not sufficient — it rules out the bias structure the negative control shares with the main analysis, but a different, unshared source of bias can still be present. A non-null result does not automatically indict the main analysis either, since the chosen negative control may carry a small genuine effect of its own; Lipsitch and colleagues frame the pair as evidence to weigh, not proof, and recommend running it alongside the other checks in this section rather than instead of them.

This is what STROBE item 19 already asks for. Its wording — “Discuss limitations of the study, taking into account sources of potential bias or imprecision. Discuss both direction and magnitude of any potential bias” — makes direction and magnitude a checkable demand, and most published observational limitations paragraphs supply neither. Guidance on converting a prose worry into a stated quantity is in how to write a limitations section. Where the concern is who entered the study rather than what was measured inside it, the relevant structure is selection bias, which adjustment does not remove and a larger sample does not either.

Recommendations that quietly assume causation

Recommendations are where an observational paper commits itself, and they are the highest-yield place for this agent to look. The Haber 2022 finding that action recommendations carried higher implied causality than the linking sentence in 44.5% of articles means the recommendation is not a summary of the result — it is a separate, stronger claim, usually written last and reviewed least.

The consequence of that sentence is measurable outside the journal. Sumner and colleagues analysed 462 press releases on biomedical and health-related science issued by 20 leading UK universities in 2011, alongside the peer-reviewed papers and 668 associated news stories (BMJ 2014;349:g7015). They found 33% of press releases contained exaggerated causal claims, 95% confidence interval 26% to 40%. Where the release exaggerated, 81% of news stories carried a similar exaggeration (70% to 93%), against 18% (9% to 27%) where it did not — an odds ratio of 20, 95% confidence interval 7.6 to 51. The press office is not the origin of the distortion in most cases; the abstract is.

That confidence interval is wide enough that no single release, in either direction, is informative on its own — the pattern holds across the sample of 462 press releases, not as a claim about any one of them. A press office that got a specific story right did not thereby show the pipeline is generally safe, and one that got a story wrong did not thereby show every release from that university is unreliable.

The rewrite is mechanical once the estimand is fixed. “These findings support adopting X in routine practice” becomes “these findings are consistent with a benefit of X that would justify a randomised trial powered for the clinical endpoint”. “Clinicians should consider X for patients with Y” becomes “among patients with Y, receipt of X was associated with a lower rate of the outcome; whether offering X would reduce that rate has not been tested”. Each replacement preserves the finding and removes the assumption, which is exactly what the brief’s final instruction demands. Because the abstract conclusion is the sentence most often quoted onward, it is worth writing last and checking first — see how to write an abstract.

Common questions about causal language in review

Reviewer 2 says my results section sounds like I’m claiming causation — what do I write back?

State what the design supports, then reconcile the flagged sentence with it rather than arguing about tone. Rewrite the sentence to name the association explicitly, with the adjustment set and its precision, and check the abstract’s conclusion separately — Haber and colleagues found causal language concentrates there even when the flagged sentence sits in the results. Quote the STROBE item the rewrite now satisfies.

What does the agent return for each sentence it flags?

Each flagged sentence comes back quoted with its location, a statement of what the design actually supports, and a specific replacement that preserves the finding. The brief’s closing instruction exists to prevent the alternative, which is a model returning “consider more cautious phrasing” for every hit — the least useful sentence in peer review.

Running the check yourself before the agent does

Run the check yourself in about twenty minutes with four passes over the manuscript, in this order.

Pass one: extract the four sentences that matter. Copy the abstract’s conclusion sentence, the first sentence of the discussion, the final paragraph of the discussion, and any sentence in the results containing a recommendation, onto one page. Read them without the rest of the paper. If those four sentences read as a causal paper and the methods describe an associational one, you have the defect the agent reports, and you have it before submission.

Pass two: search for the verbs. Search the full text for cause, causes, caused, reduce, reduces, reduced, increase, improve, prevent, protect, leads to, results in, due to, attributable to, impact and effect of. Record the section each hit falls in. A distribution weighted toward the abstract and the last discussion paragraph is the asymmetry the brief asks about in point 3.

Pass three: classify every covariate. Write the adjustment set as a list and mark each variable C (cause of the exposure), O (cause of the outcome), B (both), M (plausible mediator), K (plausible collider) or U (unknown role). Any M or K needs an argument or needs to leave the model. Any U means the justification STROBE item 16(a) asks for does not yet exist. Where a matching or weighting design was used, the same classification governs which variables belong in the model that produces the weights.

Pass four: put one number on residual confounding. Compute the E-value for the point estimate and for the confidence limit nearest the null, and write one sentence naming the most plausible unmeasured confounder and whether an association of that strength with both exposure and outcome is credible in your field. That sentence replaces the paragraph a reviewer would otherwise call an acknowledgement.

What reviewers write when causal language is not controlled

Reviewers rarely write “causal language” as the heading of a comment. They write the specific version, and these are the sentences that appear on observational manuscripts.

“The design is cross-sectional; the conclusion that the intervention improves outcomes is not supported.” “The abstract states that treatment reduced mortality, while the discussion correctly describes an association — please reconcile.” “Covariates appear to have been selected by stepwise regression; please justify the adjustment set on causal grounds.” “The model adjusts for a variable measured after the exposure, which may be a mediator; please state its assumed causal role.” “No causal diagram is presented, so the reader cannot evaluate the adjustment set.” “Residual confounding is acknowledged but not quantified; please provide an E-value or equivalent sensitivity analysis.” “The recommendation for practice requires a causal interpretation the design cannot support.” “Confounding by indication is likely and is not addressed.” “Please report unadjusted alongside adjusted estimates.”

The comment authors find hardest to answer is the reconciliation request, because it cannot be met by softening one sentence: the abstract and the discussion have to agree on what was estimated. Responses to the broader objection, including the version phrased as correlation versus causation, are worked through in correlative, not causal, and the statistical framing of the same complaint sits in confounding.

What this agent does not do

The causal language agent reviews the sentences of an observational original research article and the justification of its adjustment set. It does not re-derive the identification strategy, refit models, recompute an E-value, or judge whether the effect size is clinically meaningful — it names where each of those belongs and quotes the sentence that needs it.

Four boundaries are worth stating plainly. The agent does not flag causal language in a randomised trial report, where the design licenses it. It does not review systematic reviews or meta-analyses, whose claims are about a body of evidence rather than about one estimand. It does not review qualitative or mixed-methods work, which has a separate agent. And it does not enforce a vocabulary rule: a paper that adopts a causal framework and states exchangeability, positivity and consistency is using causal language correctly, and the brief instructs the agent to say so in point 1 rather than flag it.

Reporting items a reviewer will cite

Four STROBE items carry the causal-language requirements, and quoting them in a response letter is more effective than arguing about tone. Item 16(a) asks for unadjusted and confounder-adjusted estimates with their precision and for a statement of “which confounders were adjusted for and why they were included”. Item 19 asks you to discuss limitations “taking into account sources of potential bias or imprecision” and to “discuss both direction and magnitude of any potential bias”. Item 20 asks for a cautious overall interpretation considering objectives, limitations, multiplicity of analyses and results from similar studies. Item 21 asks you to discuss generalisability.

Items 19 and 20 are the pair this agent maps onto directly, and they are complementary rather than duplicative: item 19 is about the bias, item 20 is about the sentence. A manuscript can satisfy item 19 with a quantified bias analysis and still fail item 20 because the abstract never absorbed the result. The checklists are published at strobe-statement.org, and the item-by-item walkthrough is in the STROBE checklist.

Item 16(a) is satisfied by stating a rationale, not by the rationale being a causal one. A methods section that reports “adjusted for age, sex and comorbidity count, selected by backward stepwise regression” has technically answered the item’s request for which confounders were adjusted for and why — while still describing a statistical selection procedure rather than a causal one. Ticking the STROBE box and closing the causal-language objection are different achievements, and a reviewer citing item 16(a) should check which one the manuscript actually did.

How this differs from the built-in review

The standing team includes a causal inference and confounding specialist that examines identification strategy, and a conclusion validity specialist. This agent is deliberately narrower and overlaps with both: it works at the sentence level, comparing each claim against the design, and supplying replacement wording. That scope is chosen because the failure is usually not that the authors misunderstand their design — the methods section normally shows they do — but that the abstract was written to be interesting.

Narrow scope and directly actionable findings make it a good default slot for any observational paper. Full set: AI peer review for epidemiology.

PerfectPaper compares every causal claim in an observational manuscript against what its design can identify, quotes both sentences where the abstract outruns the results, and returns a specific replacement rather than advice to be cautious.

Review my manuscript

Related agents

Four other population-science agents share a review with this one; you can run three at a time. See AI peer review for epidemiology for the full set of five population-science agents.

• Health equity reporting — use it for how race, ethnicity and socioeconomic position are measured and interpreted.
• Competing risks and time-related bias — use it when death from other causes is common, or exposure is defined after follow-up starts.
• Registry and claims data validity — use it for SEER, NCDB and claims analyses that may outrun what the source records.
• Qualitative and mixed-methods rigour — use it for interview and focus group work, which the standing team does not cover.