Base rate in diagnosis: why the rare seems common

Flock of sheep grazing in a field

Anyone who spends the night studying pheochromocytoma wakes up thinking that every palpitation with sweating is an adrenal tumor. The classic picture is fresh in memory, and the simplest question gets set aside: how many people with that picture actually have this disease? The answer to that question has a name: base rate.

NEXO is a daily medical diagnosis game: one clinical case a day, six progressive clues, and ICD-10 proximity feedback across five levels.

This guide explains what the base rate is, why students tend to forget it, what studies show about doctors and prevalence, and how to train reasoning to put the common in the right place on the list.

Key points

  • Base rate is the frequency of the disease in the population the patient comes from, before any data from the case.
  • In a study with medical students, epidemiology came last among the components they consider important in diagnosis.
  • Residents who thought a disease was more common judged it more likely in the cases, which shows use of prevalence.
  • Most family doctors realized that the same test is worth more in the patient with symptoms, but underestimated the size of the difference.
  • Presenting the data in natural frequencies improved students' Bayesian reasoning.

Summary

What base rate is

Base rate is the frequency of a disease in the group the patient comes from, before looking at their symptoms. It is the starting point of any hypothesis. A headache in the emergency department is much more likely to be migraine or tension headache than subarachnoid hemorrhage, even if hemorrhage is the one that scares people most.

Neglecting the base rate means judging probability only by how much the picture resembles the textbook, without asking how common each disease is there. The guide on heuristics in clinical reasoning shows representativeness, the heuristic behind this error. Here the focus is prevalence itself.

Why students think first of the rare

Epidemiology comes last

In a cross-sectional study with medical students, participants answered diagnostic scenarios built in the classic format of base rate problems (Clarkson and Paruk, 2026). The correct rate was 48.3% in the scenario that asked to distinguish stroke from migraine and 33.3% in the one that asked to distinguish pheochromocytoma from panic disorder (Clarkson and Paruk, 2026). In the pure Bayesian task, 41.1% got it right, and students ranked epidemiology as the least important component of diagnosis (Clarkson and Paruk, 2026).

Medical student syndrome

The same study found that 56.3% of participants reported having already found in themselves the disease they were studying (Clarkson and Paruk, 2026). The authors propose that part of this phenomenon comes from the same failure: weighing the picture and forgetting how rare the disease is (Clarkson and Paruk, 2026).

The curriculum reinforces the rare

In medical school, each disease gets a class, and the class on the rare disease is usually the most memorable. The result is a memory in which the rare takes up as much space as the common. Correcting this requires remembering the frequency along with the picture.

Herd of brown and black horses grazing in a green field, with trees and hills in the background.
When you hear hooves, think first of horses: the saying sums up the base rate. Credit: Izvora, public domain, via Wikimedia Commons (file page).

What doctors do with prevalence

Residents use what they think about prevalence

In a study with internal medicine residents, each participant judged the probability of several hypotheses in ambiguous cases and then estimated the prevalence of those diseases on the ward (Rottman et al., 2016). Those who thought a disease was more prevalent tended to judge it more likely in the cases, a sign that prevalence entered the reasoning (Rottman et al., 2016). Over the three years of residency, prevalence estimates became more consistent, but not more accurate (Rottman et al., 2016).

Hypotheses follow incidence and remembered cases

In a classic experiment, physicians generated hypotheses for clinical stories in which two diseases were plausible, one more common and less severe, the other rarer and more severe (Weber et al., 1993). The order and frequency of the hypotheses followed the incidence of the diseases, with little evidence of base rate neglect (Weber et al., 1993). Having diagnosed a similar case before made the hypothesis appear earlier, and experience acted mainly through that path (Weber et al., 1993).

Reading them together is encouraging: the experienced physician uses prevalence, largely because they have already seen many common cases. The student does not yet have that repertoire, and so needs to remember the base rate on purpose.

The same test in different populations

In a study with family physicians and business students, each participant estimated the chance of cancer after an abnormal mammogram in two women: one with symptoms and one asymptomatic, in screening (Austin, 2019). Among the physicians, 88% took the base rate into account and estimated a higher predictive value for the woman with symptoms (Austin, 2019). On average, however, the physicians estimated a difference of 27 points, when the real difference was 55 points or more (Austin, 2019).

That is, the direction was right and the size was not. The guide to pretest probability shows the calculation that leads from the base rate to the predictive value.

Group of zebras standing in tall green grass, with trees in the background, on a savanna.
On the savanna, zebras are the common thing: the base rate depends on where you are. Credit: Sandra Aceng, CC BY-SA 4.0 (license), via Wikimedia Commons (file page).

Natural frequencies: the tool that helps

Counting people, not percentages

A natural frequency is writing the problem as a count of people, for example "out of every thousand people, ten have the disease", instead of conditional percentages. In a study with 577 medical students, accuracy on a Bayesian task was 34% with conditional probabilities and 68% with natural frequencies, without training (Feufel et al., 2023). After a short training, accuracy reached 89% when the data came in natural frequencies (Feufel et al., 2023).

The frequency tree

In a study with 111 medical students, the version with a frequency tree had the shortest median time to a correct answer, 2 minutes and 55 seconds, against 9 minutes and 59 seconds in the text with probabilities (Binder et al., 2021). The authors conclude that the frequency tree allows more accurate and faster judgments (Binder et al., 2021).

How to train prevalence thinking

Ask "where am I?" before "what is it?"

Before the first hypothesis, place the patient: outpatient clinic or emergency department, child or older adult, city or endemic region. The same chest pain has a different likely list in a young person's outpatient clinic and in an older adult's emergency department.

Order the list by frequency and by severity

Write two columns: the most common hypotheses for that scenario and the severe ones that cannot be missed. Investigate the severe ones, but do not put them at the top of probability without data that support them. The guide to differential diagnosis shows how to build that list.

Study the frequency along with the presentation

When reviewing a disease, note how common it is in the setting where you will be seeing patients. One line is enough.

Use natural frequencies

When facing a test, rewrite the numbers as a count of people. The calculation becomes visible and the scale error decreases (Feufel et al., 2023).

Change the scenario and watch the list change

Take a presentation, such as fever in a traveler, and redo the list for two different regions. The exercise shows how the base rate changes the order of the hypotheses.

Epidemiological scenarios in NEXO

In NEXO, the case arrives in up to six clues, and with each guess the proximity in ICD-10 across five levels shows in which chapter, group or category the hypothesis landed. That way you can notice when you jumped to a rare disease from another chapter while the case pointed to something common.

Continents mode gathers cases by regional epidemiology, which allows you to train the exercise of changing the scenario and watching the list of hypotheses change along with it. The Archive, on the Modes screen, gathers the previous daily cases. How each mode is unlocked is on the NEXO economy page.

If you coordinate a college, a league or a residency and want to discuss base rates with the class from the same case, get to know NEXO Closer. Researchers who study Bayesian reasoning can propose studies through NEXO Researcher. Doctors who review the cases in their specialty take part through NEXO Curator, and those who create health content on social media can take part through NEXO Creators.

Frequently asked questions

What is base rate neglect?

It is judging the probability of a disease by how similar the picture is and forgetting how common it is. In a study with students, epidemiology was rated as the least important component of diagnosis (Clarkson and Paruk, 2026).

Do experienced doctors also forget the base rate?

Studies show use of prevalence: residents judged as more likely what they thought was more common (Rottman et al., 2016), and doctors generated hypotheses in the order of incidence (Weber et al., 1993).

Why is the same test worth more in the patient with symptoms?

Because the chance of disease before the test is higher. Family doctors got the direction of this difference right, but underestimated the size (Austin, 2019).

What is the easiest way to do the math?

Natural frequencies. With them, students' accuracy on a Bayesian task doubled compared with conditional probabilities (Feufel et al., 2023).

How does NEXO help train the base rate?

Proximity in ICD-10 shows where the guess landed, and Continents gathers cases by regional epidemiology. To get to know the game, read what NEXO is.

Base rate checklist

  1. Place the patient before the first guess.
  2. List the common guesses for that setting.
  3. List the serious ones separately, the ones that cannot be missed.
  4. Ask for data that supports the rare guess before putting it at the top.
  5. Write down the frequency of each disease you study.
  6. Rewrite tests in natural frequencies.
  7. Redo the list in another setting from time to time.

Want to train with different settings? Play the case of the day on NEXO.

References

  1. Clarkson RC, Paruk N. Medical student syndrome: a bayesian reasoning failure. BMC Med Educ. 2026. https://pubmed.ncbi.nlm.nih.gov/41840426/
  2. Rottman BM, Prochaska MT, Deaño RC. Bayesian reasoning in residents' preliminary diagnoses. Cogn Res Princ Implic. 2016. https://pubmed.ncbi.nlm.nih.gov/28180156/
  3. Weber EU, Böckenholt U, Hilton DJ, Wallace B. Determinants of diagnostic hypothesis generation: effects of information, base rates, and experience. J Exp Psychol Learn Mem Cogn. 1993. https://pubmed.ncbi.nlm.nih.gov/8409853/
  4. Austin LC. Physician and Nonphysician Estimates of Positive Predictive Value in Diagnostic v. Mass Screening Mammography: An Examination of Bayesian Reasoning. Med Decis Making. 2019. https://pubmed.ncbi.nlm.nih.gov/30678607/
  5. Feufel MA, Keller N, Kendel F, Spies CD. Boosting for insight and/or boosting for agency? How to maximize accurate test interpretation with natural frequencies. BMC Med Educ. 2023. https://pubmed.ncbi.nlm.nih.gov/36747214/
  6. Binder K, Krauss S, Schmidmaier R, Braun LT. Natural frequency trees improve diagnostic efficiency in Bayesian reasoning. Adv Health Sci Educ Theory Pract. 2021. https://pubmed.ncbi.nlm.nih.gov/33599875/

Back to the blog