Published
The SEER database is the most accessible cancer outcomes resource in the world. Registration takes minutes, the data cost nothing, and the analytic software is free. That accessibility explains both why the resource is so valuable and why oncology editors have grown sharp about it. Investigators in mainland China alone published 1,566 SEER papers by August 2020, a count that grew significantly year on year (Lin and colleagues, JMIR Medical Informatics, doi:10.2196/21931), and the same small set of methodological failures appears in submission after submission. Reviewers now look for those failures before they read the discussion.
This guide sets out what the resource actually holds, which registry grouping to use and why the one most papers still call SEER 22 no longer exists under that name, the questions the data can and cannot support, the five analytical errors that account for most rejections, and the reporting standard a good journal now expects. Every claim below carries a DOI or a link to the National Cancer Institute page it came from.
In Simple Words
A cancer registry is a running list of every cancer diagnosed in a defined area, kept by people whose job is to chase each case from diagnosis to death certificate. SEER is the American version of that list, covering a little under half the country. Because it records everyone in its catchment rather than only the patients who walked into one hospital, it can tell you how long people with a given cancer actually live, how that has changed over decades, and whether it differs by race or insurance. What it cannot tell you is why one patient got a particular treatment and another did not, because nobody wrote that down.
What the SEER Database Actually Contains
The National Cancer Institute states that SEER “currently collects and publishes cancer incidence and survival data from population-based cancer registries covering approximately 45.9 percent of the U.S. population”. The registries routinely record patient demographics, primary tumour site, tumour morphology, stage at diagnosis, first course of treatment, and follow-up for vital status. Diagnoses in the most recent public research file span 1975 to 2023.
That headline coverage figure hides something most papers never mention. Coverage varies substantially by group, reaching 39.6 percent of Whites, 43.5 percent of African Americans, 64.9 percent of Hispanics, 59.3 percent of American Indians and Alaska Natives, 68.2 percent of Asians, and 69.9 percent of Hawaiian and Pacific Islanders. Figure 1 makes the spread visible. SEER was designed to over-sample populations that national estimates would otherwise describe poorly, which is a strength for disparities research and a genuine limitation for any sentence beginning “in the United States”. A survival estimate drawn from SEER is an estimate for the SEER catchment, weighted toward urban and minority populations, and a reviewer who works in cancer epidemiology knows that.
Follow-up is the other feature of the SEER database worth understanding properly. Registrars track vital status actively and link to the National Death Index, so a SEER cohort has long, near-complete mortality follow-up of a quality no single institution can match. Cause of death is coded from the death certificate, which is where the accuracy question starts rather than ends.
Which Registry Grouping, and Why SEER 22 Is Now SEER 21
SEER releases its data in overlapping groupings of registries, and the grouping determines both the earliest diagnosis year available and which variables exist. The programme’s own registry groupings documentation lists SEER 8 and SEER 9 running from 1975, SEER 12 from 1992, and SEER 17 from 2000. The largest contemporary grouping also starts in 2000 and carries a more limited set of variables than the smaller ones, a trade many authors discover only after building a cohort around a variable that is absent from it.
Naming catches people out. SEER documentation now labels the largest grouping SEER 21, noting that it was formerly SEER 22, and records that starting with the 1975 to 2022 research data drawn from the November 2024 submission, Massachusetts is no longer included in any of the new databases. A methods section that says “we used SEER 22” while citing a recent data release is describing something that no longer matches the file it was built from. The practical rule is to read the grouping definition attached to the specific submission year you downloaded, rather than copying the sentence out of an earlier paper. Detroit carries the same caution, flagged in the documentation as a research support registry no longer under contract to submit data.
Choosing between groupings is a real design decision rather than a formality. Long trend work needs the deep history of SEER 8 or SEER 9 and pays for it with narrow geographic coverage. Contemporary population estimates need the breadth of SEER 21 and pay for it with a shorter window and fewer variables. Whichever you pick, state it and justify it in the methods, because the reviewer will otherwise assume you picked the biggest one without thinking.
Questions the SEER Database Answers Well
The resource is at its strongest where the question is about population-level outcome rather than clinical decision-making. Survival by stage and histology sits squarely in that space, as do incidence and mortality trends across decades, disparities by race, insurance status or geography, and the natural history of uncommon tumours where no trial exists and no single centre accumulates enough cases to say anything useful.
Rare disease is where the SEER database earns its reputation. A cancer with a few hundred cases a year nationally will never support a randomised trial, and a single-institution series of eleven patients answers nothing. A population-based registry with fifty years of accrual turns that same disease into an analysable cohort, which is why the descriptive epidemiology of rare tumours is dominated by registry work and reasonably so.
| Research question | Can SEER answer it? | What decides |
|---|---|---|
| Five-year survival by stage for a given histology | Yes | Stage and vital status are core registry fields with long follow-up |
| Incidence trend across four decades | Yes | Use SEER 8 or SEER 9 for depth, and check coding changes across the window |
| Survival disparity by race or insurance | Yes | Over-sampling of minority populations helps here, though it limits national claims |
| Does drug A beat drug B? | No | No regimen detail, no comorbidity, no performance status, confounding by indication |
| Time to recurrence | No | Recurrence is not collected at all |
| Effect of second-line therapy | No | Treatment is recorded for first course only |
| Treatment patterns and cost in older patients | Only with linkage | Needs SEER-Medicare claims, a separate and larger undertaking |
Questions the SEER Database Cannot Answer
The absent fields matter more than the present ones. There is no comorbidity index, no performance status, no detailed chemotherapy regimen, no recurrence, and no record of any treatment given after the first course. Those omissions are not gaps to be imputed around. They are the variables that determine which patients receive which treatment, which is precisely what a treatment comparison needs to hold constant.
Evidence on that point is unusually direct. Soni and colleagues screened 3,657 studies and identified 350 registry treatment comparisons drawn from SEER, SEER-Medicare or the National Cancer Database, then matched them against 121 randomised trials testing the same comparisons, and found no significant correlation between the hazard ratios the observational studies reported and those the trials produced (Journal of Clinical Oncology, doi:10.1200/JCO.18.01074). A registry comparison agreeing with a trial, on that evidence, is close to a coin toss. Propensity matching does not repair it, because matching can only balance what was measured.
Jones and colleagues reached the same conclusion from the clinical side, cataloguing the recurring pitfalls of head and neck research built on SEER and the National Cancer Database and concluding that registry design constrains what these data can support regardless of analytic sophistication (Otolaryngology Head and Neck Surgery 2019;161(2):245-250, doi:10.1177/0194599819838823). None of this makes SEER a weak resource. It makes it a resource for a particular class of question, and the studies that fail are the ones that ask a trial question of a registry. Our guide to levels of evidence in clinical research sets out where registry work sits relative to trial evidence, and when a Mendelian randomization design is the better route to a causal claim.
Five Errors That Get SEER Database Papers Rejected
These are the failures reviewers check for first, in roughly the order they check for them.
1. Kaplan-Meier applied where a competing risk exists
In an older cancer cohort, death from cardiovascular disease is not a censoring event. It is a competing event that permanently removes the possibility of dying of the cancer. Austin, Lee and Fine put the rule plainly, stating that when estimating the crude incidence of outcomes analysts should use the cumulative incidence function rather than the complement of the Kaplan-Meier survival function, because the Kaplan-Meier approach overestimates the event probability (Circulation 2016;133(6):601-609, doi:10.1161/CIRCULATIONAHA.115.017719). For regression, the Fine and Gray subdistribution hazard model remains the standard, with more than 13,000 citations since its publication (Journal of the American Statistical Association 1999;94(446):496-509, doi:10.1080/01621459.1999.10474144).
A credible paper reports both the cause-specific hazard and the subdistribution hazard, and says which question each one answers. They are not interchangeable, and a manuscript that presents one while claiming the interpretation of the other invites a rewrite. We simulated exactly how far apart they can drift, and found that Fine-Gray returns a significant result in up to 100 percent of studies where the two groups have identical cancer-specific hazard. The full test is in Kaplan-Meier overstates, Fine-Gray answers a different question.
2. A Cox model with no proportional hazards test
The Cox model assumes a covariate’s effect on the hazard is constant over time. Registry cohorts followed for twenty years routinely violate that, particularly for stage, where an early separation of curves can converge later. That check has existed since Schoenfeld introduced partial residuals for the proportional hazards regression model (Biometrika 1982;69(1):239-241, doi:10.1093/biomet/69.1.239), and it costs one line of code.
Where the assumption fails, the repair is time-varying coefficients, stratification, or a switch to restricted mean survival time, and the failure itself is usually interesting. A hazard ratio reported as a single number over a period during which the hazard ratio demonstrably changed is a summary of nothing.
3. Immortal time bias in anything defined after diagnosis
Any exposure classified by an event that happens after diagnosis, receiving surgery or completing radiation among them, creates a stretch of guaranteed survival before the patient can enter the exposed group. Patients who die early cannot be classified as treated, so the treated group inherits survival time it did not earn. Suissa demonstrated the mechanism and quantified its size across observational drug studies (American Journal of Epidemiology 2008;167(4):492-499, doi:10.1093/aje/kwm324), and the same structure appears constantly in registry oncology.
Landmark analysis and time-dependent covariates both fix it. Neither is difficult. Reviewers see the naive comparison constantly, and it manufactures a survival benefit out of arithmetic.
4. Stage compared across incompatible AJCC editions
AJCC staging has been revised repeatedly across the SEER window, and the derived staging variables are not interchangeable between editions. A patient coded stage IIIA under one edition may be a different stage under the next with no change in the tumour. Pooling across the transition without reconciliation produces a stage migration artefact that looks like a real change in prognosis, an effect sometimes described as the Will Rogers phenomenon. Reconcile Summary Stage and Derived AJCC explicitly across the study period, or restrict to a window with consistent coding, and say in the methods which you did.
5. Trusting registry codes over the classification they predate
This is the error that most often survives peer review, and a 2026 scoping review measured it. Hanson and colleagues charted 170 articles from 1,079 citations that used SEER or the National Cancer Database to study gastroenteropancreatic neuroendocrine tumours between 2012 and 2022, and found that of 141 assessable studies, 88 percent, or 124 studies, applied the World Health Organization nomenclature inaccurately. Among them, 82.3 percent included poorly differentiated or undifferentiated neoplasms inside neuroendocrine tumour cohorts, contaminating those cohorts by a mean of 18.7 percent, and 9.9 percent conflated the database differentiation field with the WHO proliferation grade (Cancers 2026;18(15):2508, doi:10.3390/cancers18152508).
Two further findings from that review are worth sitting with. Accuracy did not improve over time, with 89.4 percent of assessable studies published between 2017 and 2021 containing at least one discordance and no significant temporal trend at p equal to 0.14. And the errors propagated, with the discordant studies accumulating 7,323 citations across 5,145 works, 87.9 percent of them cited by at least one review across 1,075 distinct reviews, and six cited in NCCN, ESMO or ENETS guidelines. Nothing about that pattern is careless, because the cause is structural. Registries store legacy codes that predate the current classification, so a cohort built by selecting on the database field is not the cohort the WHO definition describes. Checking that your case definition matches the clinical entity, rather than matching a code, is a step no statistical method replaces.
Getting Access to the SEER Database
Access is simpler than most researchers expect and comes in two tiers, described on the National Cancer Institute’s data access page. SEER Research Data requires only registration with an email address in the Data Request System, completion of an application form, and acknowledgement of the SEER Research Data Use Agreement along with the statement on SEER treatment data limitations. SEER Research Plus, which carries the more detailed fields, requires an eRA Commons or HHS account tied to an institutional email, the same agreements, and the contact details of a supervisor or mentor at the same institution. Applications are processed within two business days.
No institutional signature, ethics approval or fee is described for either tier, which is a meaningful difference from trauma registry work. Anyone planning an NTDB or TQP trauma registry study holds that data use agreement themselves through their institution, whereas SEER access is individual and quick. If you are still choosing between data sources, our comparison of which health database your study should use works through the trade-offs, and the worked example in our analysis of what happens when NHANES survey weights are handled three different ways shows how much a design decision of this kind can move a published estimate.
Reporting Standards Reviewers Expect
STROBE is the baseline, and completing it properly catches several of the errors above before a reviewer does. Our STROBE tutorial works through what each item actually requires rather than restating the checklist.
Three additions have become close to expected in registry oncology. Where survival is reported relative to a general population, the Pohar Perme net survival estimator has displaced older relative survival methods for cross-population comparison, its authors having shown that the standard estimators do not provide information on cancer mortality independent of national general population mortality and are therefore unsuitable for comparison between countries (Biometrics 2012;68(1):113-120, doi:10.1111/j.1541-0420.2011.01640.x). Where any comparison between treated and untreated groups is presented, an E-value quantifies how strong an unmeasured confounder would have to be to explain the result away, a sensitivity analysis VanderWeele and Ding designed for exactly this situation (Annals of Internal Medicine 2017;167(4):268-274, doi:10.7326/M16-2607). And several journals now expect retrospective registry protocols to be registered, a shift we cover in our piece on why a retrospective study should still be registered.
On the practical side, the analysis itself is mostly data preparation. Our walkthroughs on preparing patient data for survival analysis in Stata and cleaning clinical registry data cover the steps where most silent errors enter. For designs where the causal question matters more than the descriptive one, the same logic that governs instrument quality in genetic studies applies here, and our analyses of what the F-statistic actually tells you and what each of the four estimators protects you from show what a sensitivity battery looks like when it is done properly.
Free Research Tools from Editverse
- GraphPro: build publication-ready charts, 25 types, with built-in t-test, ANOVA, correlation, and chi-square p-values. Free, no signup.
- StatPro: free statistical calculators and a test-selector tool covering group comparisons, categorical outcomes, time-to-event analysis, and diagnostic accuracy.
- Flowchart Pro: generate PRISMA 2020, PRISMA-ScR, CONSORT, STROBE, STARD, MOOSE, and ARRIVE flow diagrams directly in your browser.
Professional Support for a SEER Database Study
The standard described above is now the working editorial bar, and it is also why so many SEER submissions are returned without review. A survival curve is rarely the problem. What fails is everything around it, a Cox model with no proportional hazards test, cancer-specific death analysed as though nothing else could kill an eighty-year-old, a treatment comparison presented as causal, a cohort defined by a legacy code rather than the disease. Our biostatisticians run the full battery in R or Stata and deliver the analysis with a completed STROBE checklist and runnable code rather than leaving the reporting to you.
Feasibility Check
$450
Whether the question is already answered, whether the variables exist in your chosen registry grouping, and whether the event count supports the model. Go, modify or stop, in writing. Credited in full against any study you then commission. 5 business days.
SEER Outcomes Study
$2,100
Full survival study with competing risks handled by Fine-Gray and cumulative incidence, proportional hazards tested, stage harmonised across AJCC editions, plus manuscript, figures, STROBE checklist and runnable code. We hold access, or work from your extract. 15 business days.
Extended Designs
From $3,500
SEER-Medicare linkage, multi-registry pooling, competing-risk nomogram development with external validation, and joinpoint trend modelling. Fixed quote and written scope before any work begins.
If you are weighing a cancer registry analysis against a population survey, a trauma registry or a genetic causal design, start with our guide to which health database your study should use.
Conclusion
SEER rewards a narrow kind of question and punishes a broad one. Population-level survival, incidence trends, disparities and the natural history of rare tumours are what the data were built to describe, and they describe them better than any other freely available resource. Treatment effects are what the data were not built to describe, and the evidence that registry treatment comparisons fail to track randomised results is strong enough that a reviewer is entitled to open with that objection.
Methodological failures in the SEER database literature form a shorter list than most authors fear. Handle competing risks with cumulative incidence and Fine-Gray, test proportional hazards rather than assuming it, design immortal time out with a landmark or a time-dependent covariate, reconcile stage across AJCC editions, and define the cohort by the disease rather than by a legacy code. The Hanson review suggests the last of those is the one most likely to slip through peer review and into a guideline, which is reason enough to check it before anything else.
References
- Hanson TF, Hua M, Zarate Rodriguez J, Yaeger LH, Chilukuri SR, Trikalinos NA, Hammill CW. Mapping misconceptions in neuroendocrine tumor nomenclature: a scoping review of national database studies and clinical implications. Cancers. 2026;18(15):2508. doi:10.3390/cancers18152508
- Soni PD, Hartman HE, Dess RT, Abugharib A, Allen SG, Feng FY, Zietman AL, Jagsi R, Schipper MJ, Spratt DE. Comparison of population-based observational studies with randomized trials in oncology. J Clin Oncol. 2020;37(14):1209-1216. doi:10.1200/JCO.18.01074
- Austin PC, Lee DS, Fine JP. Introduction to the analysis of survival data in the presence of competing risks. Circulation. 2016;133(6):601-609. doi:10.1161/CIRCULATIONAHA.115.017719
- Fine JP, Gray RJ. A proportional hazards model for the subdistribution of a competing risk. J Am Stat Assoc. 1999;94(446):496-509. doi:10.1080/01621459.1999.10474144
- Schoenfeld D. Partial residuals for the proportional hazards regression model. Biometrika. 1982;69(1):239-241. doi:10.1093/biomet/69.1.239
- Suissa S. Immortal time bias in pharmacoepidemiology. Am J Epidemiol. 2008;167(4):492-499. doi:10.1093/aje/kwm324
- Pohar Perme M, Stare J, Estève J. On estimation in relative survival. Biometrics. 2012;68(1):113-120. doi:10.1111/j.1541-0420.2011.01640.x
- VanderWeele TJ, Ding P. Sensitivity analysis in observational research: introducing the E-value. Ann Intern Med. 2017;167(4):268-274. doi:10.7326/M16-2607
- Jones EA, Shuman AG, Egleston BL, Liu JC. Common pitfalls of head and neck research using cancer registries. Otolaryngol Head Neck Surg. 2019;161(2):245-250. doi:10.1177/0194599819838823
- Lin MQ, Lian CL, Zhou P, Lei J, Wang J, Hua L, Zhou J, Wu SG. Analysis of the trends in publications on clinical cancer research in mainland China from the SEER database: bibliometric study. JMIR Med Inform. 2020;8(11):e21931. doi:10.2196/21931
- National Cancer Institute. Overview of the SEER Program; SEER registry groupings; SEER data access. Accessed 14 September 2026.