Binomial, Poisson and Exponential: Three Questions, Three Models
Learn the distributions by asking the right question—not by memorizing formulas. Use Binomial for yes-or-no trials, Poisson for event counts, and Exponential for waiting time or distance.
Learning objectives
By the end of this lesson, you will be able to:
- Identify whether a response is a binary-trial count, an event count, or a waiting-time measurement.
- Select Binomial, Poisson, or Exponential from the way the data are generated.
- Calculate exact, cumulative, tail, and interval probabilities.
- Explain how Poisson counts and Exponential waiting times arise from the same process.
- Choose between p, np, c, and u control-chart applications.
- Evaluate assumptions before trusting a probability or control-chart result.
1. Start with three different questions
Imagine inspecting steel pipes. A pipe may pass or fail, defects may appear along its surface, and the distance between defects may vary. Those are three different random variables, so they require three different models.
Answer: 0, 1, 2, …, 20
Use Binomial. There is a fixed number of trials, and each pipe has two possible outcomes: fail or pass.
Answer: 0, 1, 2, 3, …
Use Poisson. You are counting random event occurrences inside a fixed interval.
Answer: 4.2 m, 18.7 m, 53.1 m, …
Use Exponential. You are measuring a continuous waiting distance between events.
Key terminology
Plain meaning: one repeated opportunity.
Technical meaning: one Bernoulli experiment with two mutually exclusive outcomes.
Practical interpretation: one pipe, weld, order, or record that can be classified yes/no.
Plain meaning: the chance of the selected outcome.
Technical meaning: the constant Bernoulli probability assigned to each Binomial trial.
Practical interpretation: “success” may mean failure, rejection, or any outcome you choose to count.
Plain meaning: something that happens and is counted.
Technical meaning: an occurrence in a defined time, length, area, volume, or opportunity interval.
Practical interpretation: a defect, call, breakdown, inclusion, or arrival.
Plain meaning: how frequently events occur.
Technical meaning: expected events per unit of exposure in a stationary Poisson process.
Practical interpretation: 2 defects per 100 m, 5 calls per hour, or 0.01 failures per day.
Plain meaning: the long-run average count in the chosen interval.
Technical meaning: μ = λt for a Poisson count over exposure t.
Practical interpretation: changing the inspection length changes μ even when λ stays constant.
Plain meaning: the time or distance until the next event.
Technical meaning: a continuous random variable modelled by Exponential under a constant event rate.
Practical interpretation: metres until the next defect or hours until the next arrival.
2. Choose the model from the question
Read how the response is generated, identify the data type, and select the model. “Neither” is available because a familiar distribution should not be forced onto a process that violates its assumptions.
Incoming inspection
A supplier ships 40 coils. Each coil is classified as accepted or rejected. What models the number rejected?
Choose the model that matches how the response is generated.
3. Binomial: repeated yes-or-no trials
The Binomial distribution applies when you have a fixed number of approximately independent trials, each trial has only two outcomes, and the probability of the selected outcome is reasonably constant.
Pipe inspection example
You inspect 20 pipes. Each pipe either passes or fails.
- Trial: one pipe
- Selected outcome: fail
- Number of trials: n = 20 pipes
- Probability of failure: p
- Random variable: X = number of failed pipes
Why this is not Poisson
Each pipe contributes at most one selected outcome: fail or not fail. The maximum possible count is therefore 20.
Poisson is different: one pipe or one inspection interval may contain several defect events.
X = selected-outcome count; x = a particular whole-number count; n = number of trials; p = probability per trial. Counts and probabilities are dimensionless.
The combination term counts how many trial arrangements contain x selected outcomes. The probability terms weight each arrangement.
n is fixed, each trial is yes/no for the selected response, p is reasonably constant, and dependence is negligible.
multiple events can occur within one unit, the trial count is not fixed, or p changes materially across suppliers, shifts, products, or time.
The expected number of selected outcomes is E(X) = np, and the variance is Var(X) = np(1-p). In plain English, np is the long-run average count from repeated samples of n trials.
Live Binomial simulator
Simulation histogram showing the frequency of each selected-outcome count.
4. When Poisson can approximate Binomial
A Binomial process remains Binomial, but when n is large and p is small, a Poisson distribution with μ = np can closely approximate its probabilities. This is a computational approximation, not a change in the data-generating mechanism.
The chart compares exact Binomial probabilities with Poisson probabilities using μ = np.
5. Live Poisson–Exponential process laboratory
Adjust the event rate and inspection length. The same generated event locations are viewed as both counts and gaps.
Poisson view: events per window
Each bar answers: “How many events occurred in this fixed window?”
Exponential view: distance between events
Each space answers: “How far did we travel before the next event?”
Count all events inside equal windows.
Measure the distance between consecutive events.
6. Core formulas, symbols, and decision criteria
Poisson probability
X and x are event counts; λ is events per exposure unit; t is exposure; μ is expected events in that exposure. λ and t must use compatible units.
The model converts an average event rate into the probability of observing a particular whole-number count.
events are counted in a defined interval, multiple events are possible, events are approximately independent, and the rate is reasonably constant.
the response is pass/fail out of fixed n, events cluster strongly, exposure is ignored, or the rate changes materially across the interval.
Example: if λ is 2 defects per 100 m and the window is 300 m, express t as 3 units of 100 m. Then μ = 2 × 3 = 6 expected defects.
Exponential waiting probability
T is the random waiting time or distance; t is a chosen threshold in the same exposure units; λ is events per exposure unit.
P(T > t) is the chance of travelling or waiting beyond t without the next event. The mean gap is 1/λ.
you measure the continuous gap to the next event and the underlying Poisson-process rate is reasonably constant.
you need an event count, the hazard changes with age or time, or maintenance, clustering, seasonality, or wear-out materially changes the process.
The Poisson–Exponential bridge
Interpretation: waiting longer than t for the next event is equivalent to observing exactly zero events during the next t exposure units.
7. Exponential memorylessness—what it does and does not mean
The Exponential distribution is memoryless: under a constant event rate, time already waited does not change the probability of waiting an additional amount.
8. Worked examples: choose, calculate, and interpret
A steel quality team studies rejected pipes and surface defects. Each question uses the same operating context but a different random variable.
Example A: exactly 2 failures in 20 pipes
Problem setup: Each inspected pipe has a 10% probability of failure. Find the probability that exactly 2 of 20 pipes fail.
Data: n = 20 pipes, p = 0.10 failure per pipe, x = 2 failures.
Selected method: Binomial, because there are 20 fixed pass/fail trials and the response is the number of failed pipes.
- Substitution: Combination term: 20!/(2!18!) = (20 × 19)/(2 × 1) = 190.
- Failure term: (0.10)2 = 0.0100.
- Pass term: (0.90)18 ≈ 0.150095.
- Multiply: 190 × 0.0100 × 0.150095 ≈ 0.285180.
Final answer: P(X = 2) ≈ 0.2852, or 28.52%.
Practical interpretation: if many comparable samples of 20 pipes were inspected under the same conditions, about 28.5% of those samples would contain exactly two failures.
Limitations: this result is only defensible if the 10% failure probability is reasonably constant and dependence among pipes is negligible.
Example B: exactly 3 defects in 100 m
Problem setup: Surface defects occur at an average rate of 2 per 100 m. Find the probability of exactly 3 defects in the next 100 m.
Data: λ = 2 defects per 100 m, t = 1 unit of 100 m, so μ = λt = 2 × 1 = 2; x = 3.
Selected method: Poisson, because multiple defect events can occur inside a fixed length.
- Substitution: e−2 ≈ 0.135335.
- 23 = 8.
- 3! = 3 × 2 × 1 = 6.
- Substitute: (0.135335 × 8) ÷ 6 ≈ 0.180447.
Final answer: P(X = 3) ≈ 0.1804, or 18.04%.
Practical interpretation: in repeated comparable 100 m sections, roughly 18% would contain exactly three defects under this model.
Limitations: clustering, changing line conditions, or inconsistent inspection sensitivity can make a basic Poisson model underestimate variability.
Example C: more than 75 m without a defect
Problem setup: Using the same defect process, find the probability that the next defect is more than 75 m away.
Data: λ = 2/100 = 0.02 defect per metre; t = 75 m.
Selected method: Exponential survival probability, because the response is continuous distance until the next event.
- Substitution: Compute the exponent: −0.02 × 75 = −1.5.
- Evaluate the exponential: e−1.5 ≈ 0.223130.
Final answer: P(T > 75) ≈ 0.2231, or 22.31%.
Practical interpretation: there is about a 22% chance that the next 75 m contains no defect.
Limitations: this interpretation depends on a reasonably constant hazard; it is not appropriate if defects become more likely after a setup change or occur in bursts.
Try the live Poisson–Exponential calculator
Enter λ as events per base exposure unit and enter exposure in the same base units. For example, use λ = 2 events per 100 m and exposure = 1 for a 100 m window.
9. Unified probability workbench
Choose the model and the probability wording. The workbench highlights the relevant region, reports the calculation method, and translates the result into operational language.
Choose inputs and calculate.
The highlighted region represents the requested probability.
10. Why c and u charts are both Poisson
Both c charts and u charts are used for defect counts. A defect is a random event, and more than one defect can occur on the same unit. That is why both charts are based on the Poisson distribution.
c chart: equal opportunity
Use a c chart when every inspected unit is the same size and is inspected in the same way.
Example: every steel plate is 2 m × 2 m, and the full surface of every plate is inspected.
| Plate | Defects |
|---|---|
| 1 | 2 |
| 2 | 4 |
| 3 | 1 |
| 4 | 3 |
Because every plate provides the same inspection opportunity, the raw defect counts can be compared directly.
u chart: unequal opportunity
Use a u chart when the inspected sizes, lengths, areas, or sample quantities differ.
Example: plates have different inspected surface areas.
| Plate | Area, m² | Defects | Defects/m² |
|---|---|---|---|
| A | 2 | 2 | 1.00 |
| B | 6 | 5 | 0.83 |
| C | 10 | 7 | 0.70 |
Plate C has the largest raw count, but it has the lowest defect rate. A u chart corrects for unequal opportunity.
A comparison that makes it stick
Same-length exam
If every student answers the same number of questions, compare the raw number of mistakes.
This is like a c chart.
Different-length exams
If one student answers 20 questions and another answers 100, compare mistakes per question.
This is like a u chart.
Decision rule
| Chart or model | What it monitors | Underlying distribution |
|---|---|---|
| np chart | Number of defective units when sample size is constant | Binomial |
| p chart | Proportion defective when sample size varies | Binomial |
| c chart | Defect count when opportunity is constant | Poisson |
| u chart | Defects per unit when opportunity varies | Poisson |
| Exponential model | Time, distance, or area until the next event | Exponential |
11. Interactive c/u chart opportunity lab
Enter defect counts and inspection opportunities. The lab selects a c chart when opportunity is constant and a u chart when opportunity varies, then calculates the centre line and basic 3σ limits.
| Sample | Defects cᵢ | Opportunity nᵢ | Rate uᵢ = cᵢ/nᵢ |
|---|---|---|---|
| 1 | — | ||
| 2 | — | ||
| 3 | — | ||
| 4 | — | ||
| 5 | — | ||
| 6 | — | ||
| 7 | — | ||
| 8 | — |
Control chart with centre line and 3σ limits.
12. A practical selection procedure
- Define the observation. State whether one observation is a pipe, a 100 m length, an hour, a surface area, or the gap between events.
- Define the response. Is it a yes/no classification, a whole-number event count, or a continuous waiting time or distance?
- Check the data-generating mechanism. Fixed two-outcome trials point to Binomial; events in exposure point to Poisson; a gap to the next event points to Exponential.
- Align units. Convert λ and t to compatible exposure units before calculating μ = λt or e−λt.
- Check assumptions. Evaluate independence, constant p or λ, consistent exposure, clustering, aging, seasonality, and inspection consistency.
- Select the probability wording. Decide whether the question asks for exactly, at most, at least, between, by a time, or beyond a time.
- Calculate and validate. Show substitution and intermediate values; compare with software or a second method where practical.
- Interpret and document. State the result in operational language, identify limitations, and record the model, parameters, units, and assumptions.
13. When you see this, it means this
| When you see this | It means this |
|---|---|
| A fixed n with pass/fail outcomes | The selected-outcome count is bounded by n; consider a Binomial model. |
| Multiple events can occur in one interval | The response is an event count rather than a defective-unit count; consider Poisson. |
| A continuous time or distance until the next event | The response is a gap; consider Exponential if the event rate is stable. |
| Large λ | Events are frequent; expected counts are higher and average gaps are shorter. |
| Small λ | Events are rare; expected counts are lower and average gaps are longer. |
| μ = λt | Expected count grows in direct proportion to the observation opportunity. |
| 1/λ | The long-run average waiting time or distance between consecutive events. |
| P(T > t) | The probability of seeing zero events during the next t exposure units. |
| Poisson variance far above the mean | Clustering, mixed rates, or omitted exposure effects may be present. |
| Opportunity is equal across samples | A c chart can compare raw defect counts. |
| Opportunity varies across samples | A u chart should compare defects per unit with sample-specific limits. |
| Event rate changes over time | A single stationary Poisson/Exponential model may be inappropriate or require stratification. |
14. Common mistakes and exam traps
Calculation mistakes
Why it is wrong: λt is only meaningful when the units cancel correctly.
Recognize it: λ is entered per 100 m while t is entered in metres.
Avoid it: convert λ to events per metre or express t in 100 m units before multiplying.
Why it is wrong: P(X ≥ x) is not the same as P(X ≤ x).
Recognize it: the wording says “at least” or “more than,” but the cumulative lower tail is reported directly.
Avoid it: translate the wording before calculating and use complements carefully.
Interpretation mistakes
Why it is wrong: failed units out of fixed n are Binomial, while defects in exposure are Poisson.
Recognize it: each unit contributes at most one selected outcome.
Avoid it: ask how the count was generated, not merely whether it is a whole number.
Why it is wrong: a signal indicates unusual variation; it does not identify why it occurred.
Recognize it: the conclusion jumps directly from “outside limits” to a specific cause.
Avoid it: investigate process history, measurement, materials, methods, and operating conditions.
Assumption mistakes
Why it is wrong: basic Poisson variance may be too small for a bursty or mixed-rate process.
Recognize it: count variance greatly exceeds the mean or events occur in bursts.
Avoid it: stratify the process and consider an overdispersed model.
Why it is wrong: memorylessness follows from a constant hazard; real wear-out can increase hazard with age.
Recognize it: older components fail more frequently.
Avoid it: examine hazard behavior and consider Weibull or another reliability model.
Software-output mistakes
Why it is wrong: for a continuous Exponential variable, the density at one exact time is not the probability of that exact time.
Recognize it: a software “probability density” value is interpreted as an interval probability.
Avoid it: use cumulative or survival probabilities for practical time intervals and thresholds.
Why it is wrong: rate λ and scale 1/λ are reciprocals.
Recognize it: Minitab asks for scale but the event rate is entered unchanged.
Avoid it: calculate scale = 1/λ and verify the software parameterization.
Exam traps
Why it is wrong: a defective unit is a yes/no classification; a unit may contain multiple defects.
Recognize it: the question alternates between “units failing” and “nonconformities found.”
Avoid it: map defective to Binomial charts and defects to Poisson charts.
Why it is wrong: raw counts are not comparable when inspection opportunity varies.
Recognize it: sample length, area, or number of units changes.
Avoid it: equal opportunity → c; unequal opportunity → u.
15. Assumption checker
A formula can be mathematically correct and still be operationally misleading. Use this checklist as a screening tool before relying on a model.
| Diagnostic clue | What it may indicate | Next action |
|---|---|---|
| Poisson count variance far above the mean | Clustering, mixed rates, or omitted exposure effects | Stratify the data; consider negative-binomial or another overdispersed count model. |
| Poisson count variance far below the mean | Regular spacing, process constraints, or dependence | Review the event definition and whether occurrences inhibit one another. |
| Failure rate increases with age | Constant-hazard Exponential assumption is violated | Consider Weibull or a reliability model that allows changing hazard. |
| Binomial p changes by supplier, shift, or product | Trials do not share one constant probability | Stratify or model the probability using relevant predictors. |
16. Excel and Minitab guidance
Software should execute the probability calculation; it should not choose the model for you. Define the random variable, align units, and confirm assumptions first.
Excel
Binomial exact:
=BINOM.DIST(x, n, p, FALSE)Returns P(X = x).
Binomial cumulative:
=BINOM.DIST(x, n, p, TRUE)Returns P(X ≤ x). For P(X ≥ x), use 1-BINOM.DIST(x-1,n,p,TRUE).
Poisson exact:
=POISSON.DIST(x, mean, FALSE)Returns P(X = x), where mean = μ = λt.
Poisson cumulative:
=POISSON.DIST(x, mean, TRUE)Returns P(X ≤ x). For P(X ≥ x), use 1-POISSON.DIST(x-1,mean,TRUE).
Exponential cumulative:
=EXPON.DIST(t, lambda, TRUE)Returns P(T ≤ t). For P(T > t), use 1-EXPON.DIST(t,lambda,TRUE) or EXP(-lambda*t).
Minitab
Probability calculations: choose Calc → Probability Distributions, then select Binomial, Poisson, or Exponential.
- Select Probability for an exact discrete probability or a continuous density value.
- Select Cumulative probability for P(X ≤ x) or P(T ≤ t).
- Enter the correct parameter: n and p for Binomial; mean μ for Poisson; scale 1/λ for Exponential.
- For an upper tail, subtract the appropriate cumulative probability from 1 and check whether the boundary is inclusive.
- Store the result in a constant or column when documenting repeated calculations.
Output fields that matter: the entered parameters, input value, probability type, and returned probability. Record all four so another analyst can reproduce the result.
Attribute charts: choose Stat → Control Charts → Attribute Charts. Use C for constant opportunity and U when opportunity varies; supply subgroup sizes or opportunity values for the u chart.
17. Practice activity
Try it first: answer each question before revealing the worked solution.
A coating line inspects 30 panels. Each panel has a 5% chance of being rejected, while pinholes occur at an average rate of 4 per 200 m².
- Which distribution models the number of rejected panels?
- What is the expected number of rejected panels?
- Which distribution models the number of pinholes in the next 200 m²?
- Which distribution models the inspected area until the next pinhole?
- What is the mean area between pinholes?
1. Binomial. There are 30 fixed yes-or-no panel trials.
2. Expected rejected panels = 1.5. E(X) = np = 30 × 0.05 = 1.5.
3. Poisson. The response is the number of pinhole events in a fixed area.
4. Exponential. The response is the continuous area until the next pinhole.
5. Mean gap = 50 m². Rate = 4/200 = 0.02 per m², so E(T) = 1/λ = 1/0.02 = 50 m².
Extension: choose a chart and test a waiting-time claim
Eight inspected coil lengths are 100, 150, 80, 200, 120, 100, 250, and 100 m. Their defect counts are 2, 5, 1, 7, 3, 2, 9, and 2. The long-run rate is estimated at 0.03 defect per metre.
- Should the team use a c chart or u chart?
- What is the probability of travelling more than 50 m before the next defect under the Exponential model?
- State one assumption that must be checked.
1. Use a u chart. Inspected lengths vary, so raw counts have unequal opportunities. Plot defects per metre or per 100 m with limits that reflect each length.
2. P(T > 50) = e−0.03×50 = e−1.5 ≈ 0.2231. There is about a 22.31% chance of more than 50 m without the next defect.
3. Check that the defect rate is reasonably constant and events are not clustered. A bursty process can invalidate the basic Poisson–Exponential assumptions.
18. Mini quiz
19. Final summary
- Binomial models selected outcomes in a fixed number of yes-or-no trials.
- Poisson models random event counts in a defined interval or opportunity.
- Exponential models the continuous waiting time or distance until the next Poisson event.
- Binomial counts are bounded by n; Poisson counts are not bounded by a fixed trial count.
- For Binomial data, E(X) = np and Var(X) = np(1-p).
- For Poisson data, μ = λt and both the mean and variance equal μ under the ideal model.
- For Exponential data, the mean gap is 1/λ and P(T > t) = e−λt.
- p and np charts are Binomial; c and u charts are Poisson.
Remember this: Count selected trials, count random events, or measure the spaces.
Suggested next lesson: Attribute control charts—choosing among p, np, c, and u charts.
References
- Montgomery, D. C. Introduction to Statistical Quality Control. Wiley.
- Ross, S. M. Introduction to Probability Models. Academic Press.
- NIST/SEMATECH. e-Handbook of Statistical Methods, sections on probability distributions and control charts.
- Microsoft Support. Excel statistical distribution function documentation.
- Minitab Support. Probability distributions and attribute control chart documentation.