WordGPT

Which probability distribution should I use?

Ask three questions: is the outcome a count or a measurement, what process produced it, and what shape does the data have. The table below maps common situations to the right distribution, and each links to an explanation with a simulator.

Counts or measurements?

A count is a whole number of things: emails received, defective items, goals scored. A measurement can take any value in a range, such as a height or a waiting time. Counts are modelled by discrete distributions (binomial, Poisson and their relatives) and measurements by continuous ones (normal, exponential, gamma and so on).

Why it matters

The distribution you choose decides which probabilities you calculate, which confidence intervals are correct and which test is valid. A normal curve fitted to skewed data underestimates how often extreme values occur; a Poisson fitted to clustered counts understates the variance.

A distribution is a model, not a fact. Pick the one whose story matches how the data were produced, then check the fit.

How to work it out

  1. 1

    Is the variable a count of things (0, 1, 2, …) or a measurement that can take any value in a range? Counts use a discrete distribution, measurements a continuous one.

  2. 2

    For counts, ask how the count arises. A fixed number of yes or no trials gives the binomial. Events happening at a steady rate over time or space give the Poisson. Waiting for the first or the r-th success gives the geometric or the negative binomial. Drawing without replacement from a small group gives the hypergeometric.

  3. 3

    For measurements, look at the shape. Symmetric and bell-shaped: normal. Positive and right-skewed: gamma, log-normal or Weibull. Bounded between 0 and 1: beta. Equally likely anywhere in a range: uniform. Time between events: exponential.

  4. 4

    For inference about means, variances and counts in categories use the sampling distributions: t, chi-square and F.

  5. 5

    Check your choice: compare a histogram of your data with the distribution's shape, and see whether the mean and variance behave as the distribution says (for a Poisson they should be about equal).

Situation and distribution

If you have…Use the…
A single yes or no outcomeBernoulli
Successes in a fixed number of independent yes or no trialsBinomial
Successes when drawing without replacement from a small populationHypergeometric
Trials until the first successGeometric
Trials until the r-th success, or overdispersed countsNegative binomial
Events per unit of time or space at a steady ratePoisson
A fair die or a random whole number in a rangeDiscrete uniform
A random value anywhere in an intervalUniform
A symmetric bell-shaped measurement, or an average of many valuesNormal
The time until the next event at a steady rateExponential
The time until the k-th event, or a positive skewed amountGamma
A proportion or probability between 0 and 1Beta
A positive amount made by multiplying many factors (income, prices)Log-normal
How long a part lasts, when the failure rate changes with ageWeibull
A mean with an unknown standard deviation and a small sampleStudent's t
A test or interval for a variance, or counts in categoriesChi-square
Comparing variances or several group means (ANOVA)F

Questions

How do I know which distribution my data follows?

Start from how the data arose (trials, events over time, measurements), plot a histogram, and compare its shape and its mean-to-variance ratio with the candidates. Formal goodness-of-fit tests and Q–Q plots help.

When should I use the normal distribution?

For symmetric, bell-shaped data, and for sample means and sums of many values by the central limit theorem. Avoid it for strongly skewed data or counts near zero.

What is the difference between the binomial and the Poisson?

The binomial counts successes in a fixed number of trials; the Poisson counts events in an interval with no fixed number of trials. A binomial with many trials and a small success probability is close to a Poisson.

What is the difference between the t and the normal distribution?

The t distribution has heavier tails and is used when the standard deviation is estimated from a small sample. It becomes the normal distribution as the sample grows.

All distributions

    Which probability distribution should I use? A decision guide | WordGPT