A custom AI reviewer for competing risks and time-related bias

PerfectPaper’s competing risks agent reads a survival analysis for the two failures that most reliably produce a wrong answer in cancer populations: using a method that treats death from other causes as censoring when it is a competing event, and defining exposure in a way that guarantees exposed patients survived longer. The agent also checks screening studies for lead-time and length-time bias.

Copy the brief below into a custom agent slot. A review takes up to three custom agents, and they run alongside the standing specialist team rather than replacing any part of it.

Cancer cohorts are older and have substantial non-cancer mortality, which makes competing risks the norm rather than an edge case. Treating a competing event as censoring overstates cumulative incidence, always in the same direction. Koller, Raatz, Steyerberg and Wolbers reviewed 50 clinical studies performed in populations susceptible to competing risks and published in high-impact clinical journals, and reported competing risks issues in 70% of them (Statistics in Medicine 2012;31:1089–1097). That is a majority defect in the applied literature, which is why an agent briefed to look for it specifically finds something in most oncology and geriatric cohorts.

When to use this agent

Use the competing risks agent when your manuscript matches any one of the eight patterns below; the first two are the ones that most often coincide in an oncology registry.

• Your manuscript reports cancer-specific mortality or recurrence in an older cohort
• You use Kaplan-Meier or Cox regression and death from other causes is common
• Exposure is defined by something that happens after the start of follow-up, such as receiving treatment
• You compare survival between screen-detected and clinically detected cases
• A reviewer has previously mentioned Fine-Gray or subdistribution hazards
• You report absolute risk at a fixed horizon from a prediction model, nomogram or risk calculator
• You report cumulative incidence of relapse alongside non-relapse mortality, as transplant and haematology cohorts do by convention
• Your cohort comes from a registry or claims source, where the treatment date and the cohort entry date are usually different dates

One condition is enough. The two most consequential of them — substantial other-cause mortality, and an exposure ascertained after the time origin — travel together in registry oncology, which is why this brief covers both rather than splitting them across two agents.

The agent brief

Paste this into a custom agent. Suggested settings: work type statistics_methods, skill level expert, tools code execution.

Name: Competing risks and time-related bias

You are reviewing a time-to-event analysis for competing risks and for biases
arising from how time and exposure are defined.

Determine and report:

1. Events and competing events. Identify the event of interest and every event
   that prevents it from occurring. In a cancer cohort, death from other
   causes is a competing event for cancer-specific death and for recurrence.
   Determine how the analysis handled competing events. Report treatment of a
   competing event as censoring, state that this overstates cumulative
   incidence, and note that the overstatement grows with follow-up and with
   the competing hazard.

2. Method. Determine whether cumulative incidence was estimated with a method
   appropriate to competing risks, such as the cumulative incidence function
   or a Fine-Gray subdistribution model, or with Kaplan-Meier. Where
   Kaplan-Meier is used with substantial competing mortality, report it.
   Where a cause-specific hazard model is used, determine whether the
   interpretation in the text matches what that model estimates, since
   cause-specific and subdistribution models answer different questions and
   are frequently interchanged in discussion.

3. Time origin. Determine the start of follow-up for each subject and whether
   it is the same for all. Report any analysis where follow-up begins at a
   point that could only be reached by surviving, such as date of surgery in a
   cohort defined at diagnosis.

4. Immortal time. Determine whether exposure is defined by an event occurring
   after the time origin, such as receiving a treatment, completing a course,
   or achieving a response. Where it is, determine whether the time before
   that event was assigned to the exposed group. Report this as immortal time
   bias, state that it guarantees an apparent benefit, and name the standard
   remedies: a time-varying exposure, or a landmark analysis at a stated time.

5. Screening biases. For any comparison involving screen-detected disease,
   determine whether lead-time bias is addressed, since earlier detection
   lengthens measured survival without changing the date of death. Determine
   whether length-time bias is addressed, since screening preferentially
   detects slower-growing disease. Report survival comparisons between
   screen-detected and clinically detected cases where neither is addressed,
   and note that mortality rather than survival is the appropriate endpoint.

6. Overdiagnosis. For screening studies reporting incidence or survival
   benefit, determine whether overdiagnosis is considered.

7. Follow-up. Determine whether median follow-up is reported and how it was
   computed, and whether the number at risk is shown on survival curves.

Use code execution to check internal consistency of reported rates, numbers at
risk and event counts.

For each finding, state the direction of the bias, not only its presence.

Settings, and the tools to leave off

The work type statistics_methods and skill level expert carry more weight on this brief than on most, because items 1 and 2 turn on a judgement about the estimand rather than on a calculation. Give the agent code execution, because the checks in the next section are arithmetic performed on the numbers your manuscript already reports.

Leave web search off. Everything this brief needs is inside the manuscript, and an available tool is a tool that gets used — the failure mode is an agent that retrieves a tutorial on subdistribution hazards instead of reconciling your event counts. Add vision where the numbers at risk appear only under a figure and nowhere in the text, which is common in trial reports. Keep the brief under the 6,000-character limit that applies to every custom agent; the version above runs to roughly 2,900 characters, leaving room to add the specific convention your target journal or your last reviewer asked for. The general case for tool selection is set out in how to write a custom AI reviewer agent, and the slot mechanics in the custom agent documentation.

Common questions about this agent

Can an AI reviewer actually catch competing risks problems in my manuscript?

Yes, when the brief names the checks explicitly. The competing risks agent reads the event definitions, the estimator, the time origin and the exposure definition, reports where a competing event was treated as censoring, and states the direction of the resulting error rather than only flagging that something looks wrong. It runs alongside the standing specialist team, not instead of it.

Which settings should I use for the competing risks agent?

Work type statistics_methods, skill level expert, and code execution turned on, so the agent can reconcile event counts, numbers at risk and cumulative incidence figures against each other. Leave web search off, because everything the check needs is already inside the manuscript. Add vision only where numbers at risk appear solely in a figure.

What this agent catches

The competing risks agent catches eight failure patterns, each reported with the direction the estimate moves rather than a bare flag that something is wrong.

Every row states a direction, and that is deliberate. A finding that says only “competing risks were not addressed” is dismissed in a response letter; a finding that says the reported 10-year cancer-specific mortality is overstated, by roughly half again at that horizon, and that the overstatement grows with follow-up, is a finding an author has to act on. The last three rows are reporting failures rather than estimation failures, and they matter because they are the rows an editor can act on without a statistical referee.

The arithmetic the agent runs

Given code execution, the competing risks agent computes what the correct estimator would have produced rather than asserting that Kaplan-Meier was wrong. Two calculations do most of the work.

The cumulative incidence function. The cumulative incidence of cause 1 by time t is the cause-specific hazard weighted, at every instant, by the probability of still being under observation and alive: CIF₁(t) = ∫₀ᵗ S(u)·h₁(u) du, where S is overall survival from all causes. One minus Kaplan-Meier drops the S(u) weight, so 1 − KM ≥ CIF always, with equality only when no competing event has occurred by t. The correct nonparametric estimator is the Aalen-Johansen cumulative incidence function; the corresponding regression is the Fine-Gray subdistribution hazard model (Fine and Gray, Journal of the American Statistical Association 1999;94:496–509), and the corresponding group comparison is Gray’s k-sample test (Annals of Statistics 1988;16).

Take constant cause-specific hazards of 0.05 per year for cancer death and 0.10 per year for other-cause death — an unremarkable ratio in a cohort with a median age in the seventies — and the size of the error is not a rounding matter.

Drop the competing hazard to 0.02 per year and the 10-year overstatement falls from 13.5 points to 3.4. The bias is a function of the competing hazard and the follow-up horizon, and of nothing else in that calculation — which is why “our cohort was elderly and followed for a decade” is the sentence that makes this finding severe.

The sum-to-one check. Cumulative incidence functions for all causes sum to exactly 1 − S(t), the all-cause event probability. Naive 1 − Kaplan-Meier estimates do not: in the example above, cancer death gives 39.3% and other-cause death gives 63.2% at ten years, summing to 102.6% against an all-cause event probability of 77.7%. Any paper reporting a naive cumulative incidence for each of two competing causes has published an arithmetic contradiction, and the agent computes the sum and reports it. Where the manuscript reports only one cause, the agent says which quantity is missing instead of assuming a value for it.

Questions about Kaplan-Meier, Fine-Gray and cumulative incidence

When should I use Fine-Gray instead of Kaplan-Meier?

Use Fine-Gray whenever competing events are common and the question is the probability of the event actually occurring, not just its rate among survivors. Kaplan-Meier treats a competing death as censoring, which assumes that patient could still have the event of interest later — an assumption that is false the moment they have died of something else.

Why do my two cumulative incidence curves add up to more than 100 percent?

Because they were computed as separate 1-minus-Kaplan-Meier curves rather than as a proper cumulative incidence function. Correctly estimated cumulative incidences for every cause sum to exactly the all-cause event probability; in the worked example above, naive curves summed to 102.6% against a true 77.7%. The gap is the arithmetic signature of a competing event miscoded as censoring.

How much does treating a competing event as censoring actually change the answer?

Enough to change a clinical conclusion. With a cancer-specific hazard of 0.05 per year against an other-cause hazard of 0.10, 1-minus-Kaplan-Meier gives 39.3% cumulative incidence at ten years where the correct estimator gives 25.9%, and the gap widens with follow-up. In the Rotterdam Study, this class of error classified 18% of women as high risk where Fine-Gray classified 8% (Epidemiology 2009;20:555–561).

Cause-specific and subdistribution hazards answer different questions

Cause-specific and subdistribution hazard models are not two routes to one number, and the choice between them is a choice of question. Andersen, Geskus, de Witte and Putter put the reason plainly: in competing risks “the one-to-one correspondence between cause-specific hazard and cumulative incidence, between rate and risk, is lost” (International Journal of Epidemiology 2012;41:861–870).

The cause-specific hazard describes the rate of the event of interest among those still alive and still at risk. It is the aetiological quantity: use it to ask whether an exposure influences the biological process. A cause-specific hazard ratio does not translate into a probability without also knowing the competing hazard.

The subdistribution hazard, from Fine and Gray, keeps patients who experienced a competing event in the risk set with declining weight, so the coefficient maps directly onto cumulative incidence. It is the prognostic quantity: use it to ask what proportion of this population will experience the event by a stated time. The same hazard ratio vocabulary covers both, which is exactly why the discussion section slides between them.

The consequence authors find hardest to accept is that a covariate can leave the cause-specific hazard of the event untouched and still change its cumulative incidence, purely through its effect on the competing hazard. A treatment that reduces non-cancer mortality raises the cumulative incidence of cancer death, because more patients survive long enough to reach it. Latouche, Allignol, Beyersmann, Labopin and Fine argue that a complete account requires both to be analysed side by side, and that this is the more rigorous scientific approach (Journal of Clinical Epidemiology 2013;66:648–653). The agent applies that standard: it reports where only one is given and the text claims the other.

Wolbers, Koller, Witteman and Steyerberg measured the size of that difference in a prediction setting. Predicting 10-year coronary heart disease risk in women aged 55 to 90 in the Rotterdam Study, a standard Cox model ignoring the competing risk of non-coronary death classified 18% of women as high risk, above a 20% threshold, against 8% under the Fine and Gray model (Epidemiology 2009;20:555–561). The Fine and Gray model and the cause-specific hazards model performed similarly; the naive Cox model did not. Any nomogram or risk calculator built on Kaplan-Meier in an older population inherits that inflation, and inherits it at the decision threshold, where it changes who is treated.

The subdistribution weights carry their own assumption, and it is worth stating rather than leaving implicit: Fine and Gray’s original estimator treats censoring as independent of the covariates in the model, which is usually reasonable in a single-site trial with a common closing date but is a weaker assumption in registry data, where administrative follow-up length tracks calendar year of diagnosis. Where censoring plausibly depends on a covariate already in the model, the cause-specific and subdistribution estimates can disagree for a reason that has nothing to do with competing risks, and the fix is to model the censoring distribution conditional on that covariate rather than pool it across the whole cohort.

A worked example: 10-year cancer-specific mortality

Consider a registry analysis of 4,000 patients with localised prostate cancer, median age 73, reporting “10-year cancer-specific mortality of 39%, Kaplan-Meier method, log-rank p < 0.001” for a comparison between two treatment strategies.

The agent reports four things about that sentence. The event of interest is cancer-specific death, and death from other causes is a competing event that occurs more often than the event of interest in a cohort of that age. The method reported is Kaplan-Meier, which treats those other-cause deaths as censored — that is, as patients who could still die of prostate cancer later, which is false. The consequence is directional and estimable: with an other-cause hazard twice the cancer-specific hazard, the correct 10-year cumulative incidence is closer to 26% than to 39%. And the remedy is specific: re-estimate with the Aalen-Johansen cumulative incidence function, compare groups with Gray’s test rather than the log-rank test, and report a Fine-Gray model if the claim in the abstract is about absolute risk.

The uncomfortable part comes next, and it is worth anticipating in the manuscript rather than in the response letter. Correcting the estimator usually shrinks the treatment contrast as well as the level, because both arms are deflated by the same competing hazard only when the arms have the same competing hazard — and they rarely do, since the arm receiving the more aggressive treatment is usually the younger and healthier one. Reporting the corrected absolute risks alongside a cause-specific hazard ratio, and saying which question each answers, is what survives review. Reporting both estimators and letting the reader choose does not.

Immortal time in an oncology cohort

Immortal time enters an oncology cohort through exposure definitions that look administrative rather than analytic: exposure is defined by an event occurring after the start of follow-up, which is why the review brief asks for the time origin and the exposure-defining event in separate sentences rather than assuming they coincide. A cohort is defined at diagnosis; exposure is “received adjuvant chemotherapy”; a patient who died six weeks after diagnosis is counted as unexposed because they never reached treatment. The exposed group is then made, by construction, of people who survived long enough to be treated.

The magnitude is not speculative. Suissa’s reanalysis of inhaled corticosteroids after a chronic obstructive pulmonary disease hospitalisation moved the adjusted rate ratio from 0.69 (95% CI 0.55–0.86) to 1.00 (95% CI 0.79–1.26) by changing one thing: how person-time before the first prescription was allocated (American Journal of Respiratory and Critical Care Medicine 2003;168:49–53). No covariate was added and no patient was removed. The oncology dialect of the same error is the comparison of responders with non-responders, where a patient must survive to the first response assessment — RECIST 1.1 recommends re-evaluation every 6 to 8 weeks, and many protocols assess later still — before they can be classified as a responder.

The agent names the remedy rather than the term. A time-varying exposure lets patients contribute unexposed person-time until the moment of exposure and exposed person-time afterwards. A landmark analysis fixes a time after the origin, classifies exposure as of that landmark, excludes anyone who did not survive to it, and starts follow-up there; the landmark must be pre-specified and justified, and reporting the estimate at two or three landmarks is the standard sensitivity analysis. The full taxonomy, including the distinction between misclassified and excluded immortal time, is in immortal time bias. Where the treated and untreated curves also meet or invert, read my survival curves cross before reporting a single hazard ratio, because a flat early segment in the treated arm is usually the immortal span rather than a rapid treatment effect.

Questions about immortal time bias

What is immortal time bias?

Bias created when exposure is defined by an event that takes time to reach — a treatment, a response, a diagnosis code — and the time before that event is credited to the exposed group as if they had already been exposed. Those patients could not have died, relapsed or been reclassified before becoming exposed, so the exposed group looks like it survived longer regardless of any real treatment effect.

My treated patients look like they survived longer just because they got the drug — could that be immortal time bias?

Possibly, and it is worth checking before trusting the result. If exposure is defined by receiving a treatment that takes weeks to start, and the follow-up clock for the exposed group starts at diagnosis rather than at treatment, every patient who died before treatment is silently counted as unexposed — which manufactures survival time the treatment never earned.

Landmark analysis or time-varying exposure — which one fixes immortal time bias in my cohort?

Either can, and they answer slightly different questions. A time-varying exposure lets each patient contribute unexposed person-time before treatment and exposed person-time after, using every patient and every day of follow-up. A landmark analysis instead fixes one time point, classifies exposure as of that date, and excludes anyone who did not survive to it — simpler, but it discards early events.

The screening block: lead time, length time and overdiagnosis

Lead-time bias, length-time bias and overdiagnosis together make survival an invalid endpoint for a screening comparison, and the fix is a different number rather than a caveat. Marcus and colleagues’ long-term follow-up of the Mayo Lung Project shows the gap: median survival among participants with resected early-stage lung cancer was 16.0 years in the screened arm and 5.0 years in the usual-care arm — an eleven-year difference — while lung cancer mortality was 4.4 deaths per 1,000 person-years screened against 3.9 unscreened (Journal of the National Cancer Institute 2000;92:1308–1316). Survival and mortality pointed in opposite directions in the same trial.

Three distortions stack in that comparison and each needs its own answer. Lead-time bias moves the diagnosis date earlier without moving the date of death, so measured survival lengthens by the lead time even with a treatment effect of exactly zero. Length-time bias changes which cases are detected at all, enriching the screen-detected group for slow-growing disease; subtracting an estimated lead time adjusts a date and leaves the mixture untouched. Overdiagnosis adds cases that would never have surfaced, and its competing-mortality route is a direct competing risks problem — the same lesion is overdiagnosis in an 82-year-old with heart failure and is not overdiagnosis in a 45-year-old. Overdiagnosis also cannot be corrected after the fact the way lead time can be modelled away: the counterfactual absence of disease is unobservable for any single patient, so an overdiagnosis estimate is only as good as the sojourn-time or excess-incidence model behind it, and a randomised trial with long follow-up after screening stops is the design that bounds it empirically rather than assuming it. The agent reports which of the three is addressed and which is asserted, and it reports the endpoint: disease-specific mortality in the screened population, with all-cause mortality alongside it.

Edge cases this brief is written to handle

The competing risks brief is written to survive six situations that break a naive reading of a survival methods section.

Recurrence has two competing events, not one. Death without recurrence prevents recurrence, and so does a second primary in many protocols. A recurrence analysis that censors deaths and says nothing about second primaries has an incomplete event structure, and the agent asks for the definition rather than assuming it. Where the question goes further, asking about death after relapse rather than only whether relapse occurred, a competing-risks model no longer describes the data-generating process at all; that second transition needs a multistate (illness-death) model, which chains a competing-risks step at each transition rather than treating the whole trajectory as one event.

The competing event is not always death. Transplantation in a dialysis cohort, treatment discontinuation in a toxicity analysis, and salvage therapy in a relapse analysis all prevent the event of interest from being observed as defined. Naming the competing event correctly is item 1’s whole purpose.

Cause of death may not be reliably recorded. Registry cause-of-death coding misattributes deaths in both directions, and a cause-specific analysis inherits that error. Where the manuscript relies on cause of death without addressing coding validity, the finding belongs with the registry and claims data validity agent as well as here.

Randomisation does not remove competing risks. A trial has no confounding and still has competing events; randomisation makes the cause-specific hazard contrast causal without making 1 − Kaplan-Meier a valid risk. The agent applies items 1, 2 and 7 to trials exactly as to cohorts, which is where it overlaps with the clinical trials reviewer.

A composite endpoint can be a way of avoiding the problem rather than solving it. Combining cancer death with all-cause death removes the competing risk by removing the question, and the component that drives the composite is usually the one nobody wanted to study — see what is a composite endpoint. The agent reports the composite’s components and their event counts rather than treating the composite as settled.

Cohort entry is left-truncated, not just right-censored. A registry linkage often starts follow-up at first claims contact rather than at diagnosis, so subjects who died before entering the linked data are structurally excluded from it. The standard cumulative incidence estimator, built for right-censoring alone, needs its left-truncation extension before early time points can be trusted; the agent asks whether cohort entry and diagnosis share a date, not only whether follow-up start is uniform across subjects.

When the obvious remedy is unavailable

Re-estimating with a cumulative incidence function is correct in most designs, and the agent names it first for that reason. Four situations require something else, and the brief asks for the remedy that fits.

Cause of death is unknown or untrustworthy. Relative survival methods estimate excess mortality against a matched general-population life table without using cause of death at all; the Pohar Perme estimator is the consistent net-survival estimator in that framework (Pohar Perme, Stare and Estève, Biometrics 2012;68:113–120), implemented in the R package relsurv. State that the estimand is excess mortality, not cause-specific mortality, because they are not interchangeable in the discussion.

Too few events for a subdistribution model. A Fine-Gray model is a regression and needs events to identify coefficients. Where the event count will not support the covariate set, report the nonparametric cumulative incidence functions with Gray’s test, say why the regression was not fitted, and give the event count that drove the decision rather than presenting an unstable model.

A landmark would empty the cohort. Where most of the mortality falls in the first months, a landmark late enough to capture exposure removes most of the events. Report the number of deaths before the landmark alongside the landmark estimate so a reader can see what was excluded, and give the time-varying analysis as the primary result.

The paper is already published or already in meta-analysis. Both errors here are directional — censoring a competing event overstates cumulative incidence, and immortal time inflates apparent benefit — so a meta-analyst can treat susceptible studies as bounds, stratify the pooled estimate by whether competing events were handled, and report the strata separately rather than pooling across them. A stated bound makes a stronger limitations section than an acknowledgement.

What reporting guidance already requires

Three published reporting requirements turn the choice of estimator and the calculation of follow-up time into compliance checks rather than a referee’s preference, and the agent cites them because an editor can act on a compliance failure without a statistical reviewer.

Pocock, Clayton and Altman set the convention that survival plots carry numbers at risk beneath the time axis and that curves are not extended into a period where too few patients remain (Lancet 2002;359:1686–1689). Schemper and Smith established that median follow-up should be quantified by the reverse Kaplan-Meier method, reversing the roles of event and censoring, rather than by taking the median observed follow-up time among all patients — which counts early deaths as short follow-up and understates the maturity of the data (Controlled Clinical Trials 1996;17:343–346). Austin and Fine reviewed how the Fine-Gray model was reported across the 2015 medical literature, found that many authors interpreted its coefficients unclearly or incorrectly, and set out how a subdistribution hazard ratio should be stated — as an effect on the probability of the event occurring over time, reported alongside the cumulative incidence estimates themselves rather than in place of them (Statistics in Medicine 2017;36:4391–4400).

A manuscript that reports a subdistribution hazard ratio with no cumulative incidence figure, or a median follow-up with no method named, fails a stated convention rather than a matter of taste, and saying which convention converts the comment from an opinion into a revision instruction.

What reviewers say when this is missed

Reviewers rarely write “competing risks” first. They write the specific version, and these comments decide papers.

“Death from other causes is a competing risk in this population; Kaplan-Meier is not the appropriate estimator, and cumulative incidence functions should be presented.” “The subdistribution hazard ratio is interpreted as an aetiological effect; please state which estimand is intended.” “Follow-up begins at the date of surgery in a cohort defined at diagnosis, which introduces immortal time.” “Patients had to survive to receive the treatment, so part of the reported survival advantage is guaranteed by design.” “Survival is compared between screen-detected and clinically detected cases; this comparison is not informative about screening benefit.” “Numbers at risk are not shown, so the reader cannot judge the tail of the curve.” “Median follow-up is reported without stating how it was computed.” “The reported cumulative incidences for the two causes exceed the all-cause event probability.”

Answering the first of those takes three moves and nothing else: re-estimate at the correct estimator, update every number that changed in the abstract, results, figures and supplement together, and say which question the reported quantity answers. Do not add a Fine-Gray model to the supplement while the abstract retains the Kaplan-Meier figure — reviewers who raise this check the revision for exactly that. See how to write a response to reviewers, and where the objection arrives as a test choice rather than an estimand choice, the wrong statistical test covers the log-rank comparison specifically.

Frequently asked questions

Reviewer 2 said my survival analysis ignores competing risks — what do I actually need to change before resubmitting?

Three moves, not a paragraph of reassurance. Re-estimate cumulative incidence with the Aalen-Johansen estimator or a Fine-Gray model instead of 1-minus-Kaplan-Meier, update every number that changes across the abstract, results, figures and supplement together, and state plainly which question the reported quantity now answers. A revision that adds Fine-Gray to a supplementary table while the abstract still quotes the Kaplan-Meier figure fails the same review a second time.

Why is survival the wrong endpoint for a screening study?

Because earlier detection lengthens the interval between diagnosis and death even when the date of death does not move — that is lead-time bias, and it inflates survival regardless of whether screening does anything useful. Disease-specific mortality in the screened population, reported alongside all-cause mortality, is the endpoint that lead time cannot distort by itself.

Should I report the cause-specific hazard ratio or the subdistribution hazard ratio in my results?

Report whichever answers your stated question, and both if the manuscript makes claims about each. The cause-specific hazard ratio is the aetiological quantity: whether an exposure affects the underlying process among those still at risk. The subdistribution hazard ratio is the prognostic quantity, mapping directly onto cumulative incidence; the two can even move in opposite directions for the same covariate.

How this differs from the built-in review

PerfectPaper’s standing team includes statistical models and causal inference specialists that read the analysis competently. This agent adds the population-specific knowledge that makes those readings sharp in oncology: that competing mortality is the default rather than an exception, that immortal time enters through exposure definitions that look innocuous, and that survival is the wrong endpoint for a screening comparison. Each is a well-known bias with a name, and each is easier to find when you are looking for it specifically.

What none of the standing specialists is briefed to do is reconstruct the event structure of the study — every event that prevents every other event — and compare it against the estimator named in the methods and the quantity claimed in the abstract. That reconstruction is a domain habit rather than a statistical technique, and it is what separates a finding an author can act on from an observation that the analysis “could consider competing risks”. It also puts this agent upstream of the wording checks: an estimand mismatch reads as a causal language problem in the discussion long before anyone questions the model, and the Kaplan-Meier curve itself is where the mismatch is usually visible first.

Pairs with registry and claims data validity, where these designs usually live. Full set: AI peer review for epidemiology.

Review my manuscript

Related agents

The four agents below share a review with the competing risks agent; a review takes up to three custom agents at once, so pick from this list and the standing team fills the rest. 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.
• 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.