Denominators and conditional probability
In one sentence
A conditional probability describes how often an event occurs within a specified group, so its denominator must identify that group.
The intuition
Suppose someone says, “Half had a response.” Half of whom: everyone enrolled, everyone treated or everyone with a scan? The fraction may be calculated correctly and still answer a different question from the one you have in mind.
Read the denominator as a doorway into the group. Changing that doorway changes the question. It does not automatically reveal why the group had its outcome.
How it works
The numerator counts the observations meeting the stated event rule. The denominator counts the relevant group or amount of observation. A simple proportion divides the first by the second.
Write P(A | B) as “the probability of A among those satisfying B.” The vertical bar means “given,” not “caused by.” For events with P(B) > 0, the rule is P(A | B) = P(A and B) / P(B). A conditional probability is undefined when the conditioning event has probability zero. Hernán and Robins: association and conditioning.
The direction matters. The chance of a positive test among people with disease differs from the chance of disease among people with a positive test. Positive predictive value and base rates works through that reversal.
Name the observation unit too. People, lesions, blood draws and laboratory wells give different denominators. An event count per person-time is a rate, not a simple probability: one person can contribute multiple periods or events under the specified rule.
When follow-up is incomplete, an observed event count divided by everyone enrolled may not estimate risk at a chosen date. Kaplan–Meier estimation and competing-risk methods address particular time-to-event questions and assumptions.
A worked example
A fictional study enrolls twenty people. Six never receive the product. Of fourteen treated people, ten have an evaluable follow-up scan. Five of those ten meet the stated response rule.
| Description | Calculation | What the denominator says |
|---|---|---|
| Observed responders among enrolled people | 5/20 = 25% | Includes everyone enrolled |
| Observed responders among treated people | 5/14 ≈ 35.7% | Includes everyone treated |
| Responders among scan-evaluable people | 5/10 = 50% | Excludes treated people without evaluable scans |
These descriptions do not establish the missing participants' outcomes. The first two count observed responses; calling everyone else a confirmed non-responder would add an unsupported assumption. A protocol may define an analysis that treats missing assessments as non-response. Report that rule and its sensitivity analyses rather than silently improvising one. CONSORT 2025: participant flow and analysis counts.
Now take a separate, complete fictional cohort of one hundred people: twenty have recurrence, and five of those twenty have brain involvement. The proportion with brain involvement among those with recurrence is 5/20 = 25%. The proportion among the original cohort is 5/100 = 5%. Neither number is a personal forecast.
Try it
A fictional study has forty enrolled people, thirty treated people and twenty evaluable scans. Eight scans meet the response rule. Complete this sentence: “The observed response proportion among people with evaluable scans is ___.” Does that establish the outcomes of the other enrolled people?
Answer: Eight divided by twenty is 40%. It does not establish the other outcomes. Eight divided by forty is 20% observed responders among everyone enrolled; that denominator answers a different question.
Why it matters in cancer
- A safety table may count treated people, while an efficacy table counts a different analysis set.
- A biomarker subgroup or later disease-free landmark changes who contributes to the estimate.
- Several reports from the same participants do not create additional independent people.
Common confusions
- Conditional probability versus causation: a group defined by treatment response may also be selected by underlying biology.
- Evaluable versus enrolled: exclusion reasons can affect interpretation, especially when disease progression prevents assessment.
- Percentages versus percentage points: a changed proportion needs its reference proportion attached.
- Patient denominator versus sample denominator: repeated samples do not become independent patients.
Related concepts
Sources and scope
Source check: October 10, 2026. General probability and denominator education; all cohorts and numbers are fictional. Expert and learner review remain pending.
- Hernán and Robins, Causal Inference: What If — author-linked August 19, 2026 version, chapters 1–2; conditional associations and causal contrasts.
- CONSORT 2025 explanation and elaboration — distinguish assigned, treated, observed and analyzed participants.
- FDA: cancer-trial endpoint guidance — assessment and analysis rules belong to the endpoint.