Direct answer
Choose a probability distribution by defining what is random, which values are possible, how observations arise, and which assumptions connect the process to a model; visual resemblance alone is not enough.
Visual explanation
Choose from process and support
What this calculation tells you
A distribution is a mathematical model for possible values and their probabilities or density. Its parameters must have a coherent meaning for the process being studied.
Identify whether the variable is a count, category count, bounded proportion, positive duration, signed measurement, maximum, or another quantity. Then examine independence, event rate, replacement, bounds, and tail behavior.
Where it is used
Reliability
Distinguish event-time, count, and extreme-value processes.
Quality
Connect measurement support and process behavior to candidate models.
Research
Document model selection before estimation and testing.
Education
Organize distribution families by process rather than memorized formulas.
When this guide helps
- Choosing a model for counts.
- Modeling positive waiting times.
- Checking whether a bounded variable fits an unbounded model.
- Explaining why a candidate still needs diagnostics.
Start with the statistical question
Identify whether the variable is a count, category count, bounded proportion, positive duration, signed measurement, maximum, or another quantity. Then examine independence, event rate, replacement, bounds, and tail behavior.
A decision map branches from variable type and support to process questions, ending at candidate families rather than a single declared answer. Diagnostic checks form a loop back to model revision.
Worked example
A count of successes in a fixed number of independent equal-probability trials suggests binomial. A count of events over exposure may suggest Poisson only if the rate and event assumptions are plausible.
Assumptions that carry the result
Every named model carries assumptions beyond its formula. Parameterization conventions, truncation, mixtures, zero inflation, censoring, clustering, and time dependence must be considered explicitly.
Interpret the result without overreaching
The finder organizes candidate models; it does not fit parameters, prove goodness of fit, validate independence, or establish a physical mechanism.
- Selecting normal merely because a histogram looks bell-shaped.
- Ignoring impossible values implied by the model’s support.
- Using a familiar distribution without checking process assumptions.
Worked case: fixed trials
Count successes in 20 independent trials with constant success probability 0.3.
A binomial model matches fixed n, binary outcomes and constant p.
Mean would be np = 6 and variance np(1-p) = 4.2 under the model.
Dependence or changing p breaks the simple binomial assumptions.
Reproduce this worked caseOpen Probability Distribution Finder Calculator
Worked case: event count in time
Count arrivals during an hour under a constant-rate independent-event model.
A Poisson model may be considered with mean rate λ per hour.
Support is nonnegative whole counts with no fixed trial maximum.
Overdispersion, seasonality or clustering can require another model.
Reproduce this worked caseOpen Probability Distribution Finder Calculator
distribution selection: compare assumptions, not just answers
A distribution is a model, not a label discovered from shape alone. Validate support, sampling process, parameterization and fit.
| Case | Calculation focus | Interpretation |
|---|---|---|
| Successes in 20 trials | Fixed n; binary | Binomial candidate |
| Arrivals per hour | Count over exposure | Poisson candidate |
distribution selection: calculation checklist
- Outcome support matched
- Process assumptions stated
- Independence checked
- Parameterization named
- Diagnostics retained
Practical questions
Frequently asked questions
Can software choose the distribution automatically?
It can compare candidates numerically, but study design, support, mechanism, dependence, and consequences still require judgment.
Is the best-fitting distribution true?
No. Fit is relative to data, candidates, diagnostics, and sampling variation; several approximations may be useful.
Should every dataset be transformed to normal?
No. Transformations change interpretation and may not address dependence, mixtures, censoring, or the actual process.
Further reading
Authoritative sources
Use these primary and professional resources to check definitions, conventions, or requirements that may extend beyond this guide.