A custom AI reviewer for health equity reporting

PerfectPaper’s health equity agent checks how a health disparities manuscript treats race, ethnicity and socioeconomic position: whether race is framed as social or biological, how categories and the reference group were formed, whether an area-level measure is read as an individual one, and whether a disparity is explained or only documented. Copy the brief below into a custom agent slot.

Equity reporting sits in an awkward gap. A general reviewer asks whether the model is defensible; a disparities reviewer asks whether the variables inside it mean what the paper says they mean. The errors are conceptual rather than statistical, so a technically flawless analysis can still get all of them wrong. They are also checkable rather than a matter of taste: the ICMJE Recommendations, STROBE, the AMA Manual of Style and CONSORT-Equity 2017 each state, in publishable sentences, what a paper reporting race, ethnicity or socioeconomic position owes its reader.

When to use this agent

• Your manuscript reports differences in outcomes by race, ethnicity, or socioeconomic position
• Race or ethnicity came from administrative records rather than self-report
• You use a neighbourhood deprivation index as a measure of individual disadvantage
• Your discussion attributes a disparity to a biological or genetic difference
• You are submitting to a journal with an explicit equity reporting policy
• Your cohort came from a registry or claims database whose race field you did not collect yourself
• You collapsed one or more groups into an “other” category to keep cell sizes usable
• A referee on a previous submission asked why race was in the model at all

The agent is scoped to how the equity variables are handled. It does not check survival methods, multiplicity, or the adequacy of the adjustment set in general — the first of those belongs to competing risks and time-related bias and the rest to the standing team. Run it on original research reporting a group difference, not on a commentary or a narrative review.

What the reporting standards actually require

Four published standards cover this ground, and a disparities reviewer will quote at least one of them. The ICMJE Recommendations, under “Selection and Description of Participants”, state: “Authors should define how they determined race or ethnicity and justify their relevance. In the case where race or ethnicity was not collected, explain why it was not collected. Race and ethnicity are social and not biological constructs; authors should interpret results associated with race and ethnicity in that context.” Three obligations sit in those three sentences — method, relevance, and interpretation — and most manuscripts satisfy none of them.

STROBE is more specific than authors expect. Item 7 asks you to “clearly define all outcomes, exposures, predictors, potential confounders, and effect modifiers”, which includes saying what role race plays in your model. Item 8 asks that “for each variable of interest give sources of data and details of methods of assessment (measurement)” — the sentence that a registry-derived race field almost never satisfies. Item 11 asks you to “explain how quantitative variables were handled in the analyses” and, “if applicable, describe which groupings were chosen, and why”, which is precisely the collapsing question. See the STROBE checklist for the full instrument and the CONSORT checklist for its randomised-trial equivalent.

Beyond those: Flanagin, Frey and Christiansen set out the AMA Manual of Style position in JAMA (2021;326:621–627). CONSORT-Equity 2017 (Welch et al., BMJ 2017;359:j5085) and PRISMA-Equity 2012 (Welch et al., PLoS Medicine 2012;9:e1001333) extend the trial and review guidelines to equity-relevant reporting. And the category vocabulary itself changed. The US Office of Management and Budget published revisions to Statistical Policy Directive No. 15 in the Federal Register on 29 March 2024, replacing the 1997 standards. Under the revision, “a combined race and ethnicity question is required for both self-response and proxy data collection”, and agencies are to “add MENA as a new minimum reporting category distinct from all other reporting categories” — Middle Eastern or North African respondents having previously been counted as White. Existing collections are to be made consistent with the new standard “not later than March 28, 2029”. A dataset assembled under the 1997 standard and a manuscript written in 2026 may therefore use different category systems, and the paper has to say which. A study that pools two federal sources collected on either side of that transition needs to name, source by source, which category system produced each variable — a manuscript should not assume the newer combined question applies retroactively to older records.

Common questions about reporting race and ethnicity

A reviewer asked why race is even in my model — what’s the answer supposed to look like?

The ICMJE Recommendations require exactly this answer: state how race or ethnicity was determined and justify its relevance to the study, or say why it was not collected. “It’s a standard demographic covariate” is not a justification a disparities reviewer accepts; say whether race functions as the exposure, an effect modifier, or a confounder in your model, and why that role fits your question.

Do I have to explain how race was collected, or is reporting the categories enough?

Categories are not enough. STROBE item 8 requires the source and method of assessment for every variable, and a registry-supplied field, a self-reported questionnaire item, and a name-based imputation algorithm carry different, unstated error properties. Naming the categories without naming the collection method leaves a reader unable to judge how much to trust the numbers built on top of it.

My race variable came from a registry, not from asking patients directly — do I need to disclose that?

Yes, and it is disclosable rather than disqualifying. State that race was administratively assigned rather than self-reported, note that this method is known to under-detect specific groups rather than others uniformly, and say which direction that misclassification would push your estimate. A stated direction, even an approximate one, is a specification a reviewer can evaluate; silence is not.

The agent brief

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

Name: Health equity reporting

You are reviewing a health disparities manuscript for how it measures,
justifies and interprets race, ethnicity and socioeconomic position. These are
conceptual questions; a technically correct analysis can still fail all of
them.

Determine and report:

1. Measurement. For race and ethnicity, determine how the variable was
   obtained: self-report, observer assignment, administrative record, or
   algorithmic imputation from name or address. Report where this is unstated.
   Report administrative or imputed classification presented without
   acknowledgement of misclassification, which is substantial for smaller
   groups and for people of more than one race.

2. Categories. Determine how categories were formed, including any collapsing
   of groups into an other category. Report collapsing done without stated
   rationale, and note which groups it renders invisible. Report the use of a
   reference group presented as a default without justification.

3. Construct. Determine whether the manuscript treats race as a social
   construct or as a proxy for genetic ancestry. Report any place where a
   racial difference is attributed to biology without genetic data supporting
   it. Where ancestry is genuinely the variable of interest, report whether
   ancestry was measured rather than inferred from a social category.

4. Level of measurement. Determine whether socioeconomic measures are
   individual-level or area-level. Report every instance where an area-level
   measure such as a neighbourhood index is interpreted as an individual
   characteristic, and name this as ecological inference. Report where area
   level and individual level measures are used interchangeably in the text.

5. Confounding versus mediation. Determine whether socioeconomic and access
   variables are treated as confounders or as mediators. Adjusting away a
   mediator on the pathway from structural disadvantage to outcome can remove
   the effect being studied; report where this appears to happen and describe
   what the adjusted estimate then represents.

6. Mechanism. Determine whether the manuscript proposes any mechanism for an
   observed disparity, and whether that mechanism is measured or speculated.
   Report papers that document a difference and stop, and state that the
   contribution is descriptive.

7. Framing. Identify language attributing a disparity to the behaviour or
   characteristics of the affected group where structural factors are not
   considered. Quote the sentence and propose a rewording. Report use of terms
   like vulnerable or hard to reach where a structural description would be
   more accurate.

8. Limitations. Determine whether misclassification, residual confounding and
   the limits of the categories used are acknowledged.

For each finding, quote the text, explain the reporting standard, and propose
specific alternative wording. Do not soften a finding to avoid awkwardness.

The brief fits comfortably inside the 6,000-character instruction field that applies to every custom agent; the version above runs to roughly 3,000 characters, which leaves room to add the convention your target journal or your last reviewer asked for. Guidance on writing your own variant is in how to write a custom reviewer agent.

Worked example: how a race variable reaches the dataset

Administrative race codes are measurably wrong, and the error is concentrated in exactly the groups a disparities paper is about. Eicheldinger and Bonito, working with Medicare enrolment data, reported in Health Care Financing Review (2008;29(3):27–42) that the administrative race and ethnicity code, compared with self-report, had a sensitivity of 29.5% for Hispanic beneficiaries and 54.7% for Asian and Pacific Islander beneficiaries. Their RTI imputation algorithm raised those to 76.6% and 79.2% respectively, recoding 2,245,792 beneficiaries to Hispanic and 336,363 to Asian or Pacific Islander. A pre-2008 Medicare analysis that reported “no difference for Hispanic patients” was in part reporting on a group that was seven-tenths missing.

Three consequences follow, and the agent reports each separately. First, misclassification with more than two categories does not reliably attenuate an estimate toward the null. People move between specific categories — Hispanic to White, Asian to White — so the direction depends on which groups are confused with which and on the outcome rates in each. “Non-differential misclassification biases toward the null” is a two-category result and does not license the sentence it is usually used to license.

Second, an imputed variable is a model output, not an observation. Bayesian Improved Surname Geocoding, the standard surname-and-address method, produces a probability, and analysing the modal category discards that uncertainty. Some analyses instead carry the full probability distribution through multiple imputations rather than collapsing each person to a single modal race — heavier computation, but it keeps the uncertainty visible in the final variance rather than discarding it at the input step. State the method by name and report its validation in your population, not the vendor’s.

Third, the denominator changes when the numerator does. If reclassification moves people into a group, that group’s rates change even with no change in outcomes — which is a selection problem as much as a measurement one. Registry and claims analyses carry this class throughout; see registry and claims data validity.

Why collapsing categories changes the finding

Collapsing groups is a modelling decision that changes the answer, and STROBE item 11 requires you to say why you made it. The usual reason is a small cell, and the usual result is an “other” category holding American Indian and Alaska Native, Native Hawaiian and Pacific Islander, multiracial and unclassified respondents together. That category has no interpretation. An estimate for it describes no population, and its confidence interval is typically wide enough that “no significant difference” is the expected result whatever the truth — a power problem misreported as a null finding. Say which groups entered it and how many people each contributed. Where the cell is small but not negligible, exact logistic regression or a penalised-likelihood method such as Firth’s can sometimes model the category directly instead of folding it into “other” — worth naming as a route you considered and rejected, if you rejected it, rather than leaving “small cell” as the only stated reason.

Aggregation hides in the other direction too. A single “Asian” category pools populations whose outcome rates differ substantially by national origin; a single “Hispanic” category pools across origin, nativity and time in country. The revised SPD 15 requires the collection of detailed race and ethnicity data beyond the minimum categories unless an agency requests and receives an exemption from OIRA, which means newer federal data sources will carry disaggregated fields that older analyses did not have. If your source has them and you aggregated, that is a choice to state.

Reference-group choice is the third decision, and it is the one authors most often make without noticing. Whichever group is the referent gets the tightest interval and the implicit role of normal; every other estimate is a contrast against it. Choosing the largest group is a defensible statistical reason and it is still a reason you have to write down — and the largest group is rarely as homogeneous as its single label suggests, so a “White” referent pooling several distinct ancestries and immigration histories is subject to the same aggregation critique as an over-broad “Asian” or “Hispanic” category, just less often named as one. Boyd, Lindo, Weeks and McLemore, writing in Health Affairs Forefront (2 July 2020, “On Racism: A New Standard For Publishing On Racial Health Inequities”), ask authors to define race during the experimental design and specify the reason for its use in the study; a silently chosen referent is one of the places that reason goes unstated. Where a small-cell suppression rule in a data-use agreement forces the collapse, name the threshold and the agreement rather than presenting the grouping as analytic judgement.

Race as a social construct, in the analysis and not only the discussion

Writing “race is a social construct” in the limitations does not satisfy the standard if the analysis treats it as biology. The National Academies report Using Population Descriptors in Genetics and Genomics Research: A New Framework for an Evolving Field (National Academies Press, 2023) sets out the current position for genomics: race, ethnicity and ancestry are distinct descriptors used inconsistently, and a social category is not a measurement of genetic variation. Where ancestry is the actual variable of interest, it can be measured with genotype data; inferring it from a race checkbox is a substitution the agent reports. The distinction runs deeper than “measure genotype instead”: genetic ancestry is continuous and admixed within essentially every socially defined racial category in the United States, so a genotyped ancestry variable answers a different question from a discrete race category even when both appear in the same model, and a manuscript that swaps one for the other without saying so has changed its exposure rather than merely sharpened its measurement of the same one.

Clinical algorithms are the clearest worked case. Inker and colleagues, developing equations that omit race for estimated glomerular filtration rate in the New England Journal of Medicine (2021;385:1737–1749), stated the premise plainly: “race in eGFR equations is a social and not a biologic construct.” Their validation data showed the then-current creatinine equation overestimated measured GFR in Black participants by a median of 3.7 mL/min/1.73 m² (95% CI 1.8 to 5.4). A manuscript that applies a race coefficient without saying which equation version it used, and why, is making a biological claim by default.

Obermeyer and colleagues showed the same failure in a variable that never mentions race at all. In Science (2019;366:447–453) they analysed a widely used commercial risk algorithm and found that at a given risk score, Black patients were considerably sicker than White patients, because the algorithm predicted health care spending as a proxy for illness. Remedying the disparity, they calculated, would raise the share of Black patients referred for additional help from 17.7% to 46.5%. Any variable built from utilisation — visit counts, prior spending, documented comorbidity burden — encodes access, and using it as a measure of need imports the disparity into the covariate. The agent flags proxy variables of this shape and asks what the proxy is standing for.

Area-level measures and the individual conclusion

An area-level measure describes an area, and a sentence that turns it into a statement about a person is ecological inference. Robinson set the problem out in “Ecological Correlations and the Behavior of Individuals” (American Sociological Review 1950;15:351–357, reprinted in the International Journal of Epidemiology 2009;38:337–341): a correlation computed across geographic units need not hold, in size or in sign, among the individuals inside them. Seventy-six years later, the sentence “patients in deprived neighbourhoods were poorer” still appears in manuscripts whose data contain no individual income.

The composition of the index matters as much as its level. Petterson, in Health Affairs Scholar (2023;1(5):qxad063), recalculated the widely used Area Deprivation Index from the Neighborhood Atlas and found that because its component measures are not standardised before the index is formed, the published ADI “is reducible to a weighted average of just 2 measures — income and home values”. His unstandardised reconstruction matched the published index at r > 0.9999, while a properly standardised version correlated with it at only r = 0.7245. An ADI decile is therefore not the multidimensional deprivation summary its name implies, and urban areas with high home values rank in ways that surprise readers who assume it is. A second, separate pitfall sits alongside this one: what geographers call the modifiable areal unit problem, where the same underlying population can show a different — sometimes reversed — area-level association purely from how the data are aggregated into boundaries. A tract-level finding and a block-group finding drawn from the same city are not the same result at two resolutions; they can disagree.

Practical consequences for a manuscript: name the index, its version, its geographic unit (census block group and census tract are not interchangeable), and the year of the underlying estimates. Say whether the index enters as a decile, a percentile or a continuous score, and whether the ranking is national or within-state, because the Neighborhood Atlas publishes both and they order neighbourhoods differently. If you also use a segregation or polarisation measure such as the Index of Concentration at the Extremes, state which two extremes you contrasted — income, race, or the joint measure — since the choice defines the exposure. Then write conclusions at the level you measured. An area-level exposure supports an area-level claim, and pairing it with an individual-level claim requires individual data you either have or do not.

Questions about neighbourhood and socioeconomic measures

Can I use a neighborhood deprivation index as a stand-in for individual poverty?

No. An area-level index describes the area, not the people living in it, and treating it as an individual measure is ecological inference: a correlation computed across census tracts need not hold, in size or even in sign, among the individuals inside them. Report the index as an area-level exposure in every sentence, including the abstract, and pair it only with an area-level claim.

Does it matter whether I report the Area Deprivation Index nationally or within-state?

Yes. The Neighborhood Atlas publishes both a national and a within-state ranking of the same ADI, and they order neighbourhoods differently, so a decile reported without saying which ranking was used cannot be reproduced from your Methods. Name the ranking, the geographic unit — block group or tract — and the year of the underlying estimates in the same sentence as the index itself.

Is a higher ADI score always “worse,” no matter which components are driving it?

Not automatically. Because the index’s component measures are not standardised before it is formed, the published ADI reduces largely to income and home values, so an urban area with high home values but low income can rank as deprived for reasons a reader assumes are broader. State which components seem to drive a surprising result, rather than treating the composite score as self-explanatory.

Confounder or mediator: what the adjusted estimate then means

Insurance status, usual source of care and treatment receipt are usually mediators in a disparities analysis, not confounders, and adjusting for them removes the pathway the paper exists to describe. The direction of the arrow decides this, and no model diagnostic reveals it — the distinction is worked through in mediator or confounder and in confounding. Structural disadvantage causes lower insurance coverage, which causes later diagnosis, which causes worse survival. Put insurance in the model and the racial coefficient shrinks; the paper then reports that the disparity “was explained by insurance”, when what it has estimated is a direct effect conditional on a downstream variable.

That conditional quantity can be a legitimate target. Reporting it honestly means three things: name it as a direct effect, state the mediator you conditioned on, and state the assumptions a mediation estimate requires, including no unmeasured confounding of the mediator-outcome relationship. Where a coefficient shrinks after adjustment and no such account is given, see my effect disappeared after adjusting.

The reporting failure that follows is the Table 2 fallacy — presenting every coefficient in the adjusted model as though each were a causal effect estimated with the same care as the exposure of interest. Westreich and Greenland set it out in the American Journal of Epidemiology (2013;177:292–298), and disparities tables are unusually prone to it because readers scan the race row and the income row as if both were exposures. One adjustment set can be correct for one exposure and wrong for every other variable in the same table. State which coefficient your adjustment set was built for, and present the rest as what they are — model terms, not findings. The related conditioning trap is collider bias.

A related caution holds even when the adjustment set is entirely correct. Adding a covariate to a logistic or Cox model can move the exposure’s odds ratio or hazard ratio on its own, with no confounding and no mediation involved, because neither measure is collapsible: once the outcome is not rare, the marginal and the covariate-adjusted odds ratio differ by construction whenever the added covariate predicts the outcome, and Mansournia and Greenland, in Epidemiology (2015;26(4):466–472), show that this holds “whether or not the risk factor is associated with exposure” — that is, whether or not it confounds anything. The condition is that the covariate predict the outcome: one that is independent of the outcome given exposure leaves the odds ratio where it was, so this is not a blanket excuse for any coefficient that moves. A risk difference computed on a linear scale does not show the artifact at all. Before reading a shrinking race coefficient as evidence of a mediator or a suppressed confounder, check whether the same shrinkage appears when the outcome is expressed as a risk difference rather than an odds ratio or a hazard ratio.

Common questions about confounding and mediation

Is insurance status a confounder or a mediator in a health disparities study?

Usually a mediator, not a confounder. Structural disadvantage plausibly causes lower insurance coverage, which then causes worse outcomes, so insurance sits on the pathway the disparity is about, not upstream of race and outcome alike. Adjusting for it as though it were a confounder removes part of the effect you set out to measure, then reports the remainder as “the true disparity” — a mislabelled quantity, not a smaller one.

Why did my racial disparity shrink or disappear after I adjusted for insurance?

Most likely because you conditioned on a mediator rather than a confounder, which estimates a direct effect holding insurance status fixed instead of the total disparity. That can be a legitimate quantity, but only if you name it as a direct effect, state insurance as the mediator you conditioned on, and note that a valid mediation estimate additionally requires no unmeasured confounding of the insurance-outcome relationship itself.

If adjusting for a covariate is standard practice, why doesn’t its own coefficient count as a finding too?

Because one adjustment set is built to isolate a single exposure’s effect, and every other coefficient in the same model reflects that choice rather than its own clean estimate — a pattern epidemiologists call the Table 2 fallacy. State which coefficient your model was built for, and present the rest as model terms rather than as findings about income, insurance, or any other row in the table.

From documenting a disparity to explaining one

Documenting a disparity is a contribution; describing it as an explanation is the error. Kaufman and Cooper, in a commentary in the American Journal of Epidemiology (2001;154(4):291–298), set out considerations for variable definition, study design and statistical adjustment when racial and ethnic classification is used in etiologic research, and argued that interpretation should focus “on racism (i.e., social relations) rather than race (i.e., innate biologic predisposition)”. A coefficient on a race indicator summarises everything that differs between the groups being contrasted; it is not an estimate of an effect of being in one of them. The agent reports the point at which a descriptive result acquires explanatory verbs, and proposes the descriptive wording instead — a boundary also policed by causal language discipline.

A measured mechanism changes the claim. Segregation indices, area-level policy exposures, documented treatment differences, insurance transitions and discrimination measures collected from participants are all measurable; each supports a mechanistic sentence that a race coefficient alone does not. Where no mechanism was measured, the honest discussion says which candidate mechanisms exist, which the data could not distinguish, and what would distinguish them.

Framing is the last pass, and it is where a competent analysis most often loses a referee. “Non-adherence in this population” and “delayed presentation among Black patients” locate a cause in the patients; “shorter appointment availability at the participating clinics” and “longer median distance to a treating facility” locate it in a structure and are testable. Boyd and colleagues make the same point about naming racism rather than race as the exposure when the mechanism is discriminatory treatment. The agent quotes the offending sentence and returns a rewrite, which is what makes the finding actionable rather than an argument about tone.

When the obvious remedy is unavailable

Most of these findings have a remedy the author cannot apply, and the agent is briefed to say what to do instead rather than to demand data that does not exist. Secondary analysis of a registry cannot recover self-reported race that was never collected. The available move is to state the source field, the collection method used by the source, any published validation of that field in a comparable population, and the direction the resulting misclassification would push your estimate. A direction stated with a reason is a stronger limitation than an acknowledgement; see how to write a limitations section.

Small groups cannot be made large. Where a cell is too small to analyse, the alternatives are to report it descriptively with its count and no inferential test, to present it in a table while excluding it from the model with the exclusion stated in the Methods, or to name the group in the title of the limitation so it is not silently absorbed. What the agent flags is the third path being taken without any of the first two.

Individual socioeconomic position is frequently absent from clinical data. The available move is to keep the area-level measure, describe it as area-level in every sentence including the abstract, and add a sensitivity analysis under an alternative index or geographic unit to show how much the conclusion depends on the operationalisation. Where a race coefficient is required by a funder or a reporting requirement but is not the paper’s causal target, present it in a separate descriptive table rather than inside the primary adjusted model.

And where the reviewer’s demand is a mechanism you did not measure, the response is to reclassify the contribution rather than to speculate. A paper that says “this analysis establishes the magnitude and precision of the difference and does not identify its mechanism” is complete. A paper that offers an unmeasured biological mechanism to fill the gap has traded a clean descriptive finding for a claim it cannot defend, which is the exchange this agent exists to prevent.

What this agent catches

What a disparities reviewer writes

Referees on disparities manuscripts write specific sentences, and these are the ones that decide papers. “The Methods do not state how race and ethnicity were ascertained.” “Race is included in the model without a stated rationale for its role — confounder, effect modifier, or exposure.” “The ‘other’ category is not defined and cannot be interpreted.” “The authors adjust for insurance and then describe the residual difference as unexplained; insurance is plausibly a mediator here.” “The neighbourhood deprivation index is described in the Discussion as individual poverty.” “The abstract states that Black patients had worse survival due to more aggressive tumour biology; no genetic or molecular data are presented.” “Please justify the choice of reference group.” “The limitations acknowledge misclassification but do not state its likely direction.”

The last of those is the comment most often triggered by a limitations paragraph written to pre-empt it, and it is the one the agent is most useful against, because a general reviewer will read the acknowledgement as sufficient.

How this differs from the built-in review

The standing team includes causal inference and confounding, survey and instrument design, and conclusion validity specialists. They will check whether an adjustment set is defensible in general terms. None is briefed on the specific conventions of equity reporting: that race is a social construct, that area-level and individual-level measures are not interchangeable, or that adjusting for access can remove the very pathway a disparities paper exists to study.

The practical difference shows up in what gets quoted back. A general reviewer objects that a covariate is unjustified; this agent objects that ICMJE requires you to define how race was determined and justify its relevance, quotes your sentence, and proposes the replacement. That is the difference between a comment an editor can weigh and a comment an author can act on. How custom agents attach to a review is described in custom AI reviewer agents.

Pairs with causal language discipline. Full set: AI peer review for epidemiology.

Review my manuscript

Frequently asked questions

Can I say a racial difference in outcomes is caused by biology or genetics?

Only with genetic or molecular data supporting it — otherwise no. Race is a social category, not a measurement of ancestry, and the ICMJE Recommendations require interpreting racial and ethnic differences in that light. The Inker eGFR equations and the Obermeyer commercial-algorithm study both illustrate the pattern: a race-linked estimate that reads as biological turns out to trace to socially patterned differences in measurement or access instead.

What do I do with a race or ethnicity category too small to analyze on its own?

Do not silently fold it into “other.” Report the group descriptively with its count and no inferential test, present it in a table while excluding it from the regression with the exclusion stated in Methods, or name the specific group and its size in the limitations. Any of these three satisfies STROBE item 11’s “why” requirement better than an unlabelled residual category does.

Why do I need a separate agent for this if the regular review already checks confounding?

The standing team checks whether an adjustment set is defensible in general terms. This agent checks equity conventions layered on top: that race is a social construct, that area-level and individual measures are not interchangeable, and that adjusting for access can erase the pathway being studied. It occupies one custom-agent slot and quotes the standard back at your sentence.

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.

• 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.
• Causal language discipline — use it when an observational finding is written with causal verbs.
• Qualitative and mixed-methods rigour — use it for interview and focus group work, which the standing team does not cover.