Is a probability plot the same as a histogram?

No. It compares ordered observations with theoretical quantiles rather than grouping observations into bins.

Must the reference line have slope one?

No. Its slope depends on the scaling and axis convention; a fitted location-scale comparison need not follow the identity line.

Can a small sample confirm a distribution?

No. It may provide little ability to distinguish plausible alternatives, even when the points appear straight.

AI Glossary / Statistics and probability / Intermediate

Probability plot

Definition

A probability plot compares ordered observations with corresponding theoretical quantiles from a proposed probability distribution. If the distribution describes the observations reasonably well, the plotted points should lie approximately along a straight line, subject to sampling variation. Systematic departures can reveal differences in shape or tail behavior. The plot is a diagnostic tool for assessing a distributional assumption, not proof that the assumption is exactly true.

What it means

Instead of comparing a histogram with a curved density, the plot transforms the comparison into a question about straightness. This can make departures in the tails easier to notice. Normal and Weibull probability plots are common examples, but the idea applies to other distributions with suitable quantile functions.

For a location-scale family, the line's intercept and slope can help describe location and scale under the chosen axis convention. These interpretations depend on the distribution and plotting method. Always inspect the axis labels rather than assuming that every software package uses the same orientation or reference line.

How it works

Sort the n observed values from smallest to largest. Assign a plotting probability to each rank using a stated convention. For example, one simple convention uses (i − 0.5)/n for rank i. Other methods, including approximations to order-statistic medians, use slightly different probabilities.

Apply the proposed distribution's inverse cumulative distribution function to those probabilities. Plot the resulting theoretical quantiles against the ordered observations and add an appropriate reference line. If the distribution has shape parameters, specify or estimate them before interpreting the comparison.

Look for consistent curvature, asymmetric departures, isolated extremes, and differences concentrated in the tails. Sampling variability is especially relevant with few observations. A plot does not reveal time dependence after the data have been sorted, so inspect time-order and residual diagnostics separately. For censored observations, use methods designed for censoring rather than treating the visible values as a complete sample.

Three practical examples

An analyst uses a normal probability plot to examine residuals from a response-time model.
A reliability team compares a Weibull model with observed component lifetimes using an appropriate probability plot.
A forecasting team checks whether large prediction errors are more extreme than its assumed error distribution suggests.

Think of it this way

Think of stretching the ruler so that observations from the proposed distribution should line up. A good match produces an approximately straight pattern. A bend suggests that the proposed ruler does not describe some part of the observed distribution, but random variation can also create small bends.

Worked example

For an original construction example, take five plotting probabilities using (i − 0.5)/5: 0.10, 0.30, 0.50, 0.70, and 0.90. Their standard normal quantiles are approximately −1.28155, −0.52440, 0, 0.52440, and 1.28155.

Now construct ordered values using y = 10 + 2z. They are approximately 7.43690, 8.95120, 10, 11.04880, and 12.56310.

Plotting z horizontally and y vertically places all five constructed points on a line with intercept 10 and slope 2, apart from rounding. These are the location and scale used to create the example.

If the last value is replaced with 16, that point lies about 3.43690 above the original line, drawing attention to the upper tail. This is a deliberate illustration of the plot's geometry. The constructed straight line is not evidence that five real observations would establish normality, and a line refitted after the replacement would differ.

Illustrative case study

Fictional case study: Cedar Analytics assumes normally distributed errors when summarizing an AI service's response times. A probability plot of model residuals shows that the largest positive residuals depart from the reference line.

The team investigates whether unusually long responses coincide with specific workloads or system events. It compares alternative error models and tests predictive interval coverage on later observations. This fictional case illustrates diagnostic investigation; it does not establish a particular distribution or justify deleting inconvenient tail observations merely to straighten the plot.

Common misconceptions and mistakes

A nearly straight probability plot does not prove independence, stationarity, or correct model specification. Sorting discards the original sequence, and different processes can have similar marginal distributions.

Do not confuse the fitted line's correlation with a universal pass-or-fail threshold. A high overall correlation can coexist with important tail departures. Another mistake is using an ordinary probability plot for heavily censored or truncated data without adjustment. Finally, slight differences across software may reflect plotting-position conventions rather than an error. Record the method when exact reproduction matters, especially with small samples.

Business use

Probability plots help businesses assess assumptions behind forecasts, quality analyses, and reliability estimates. They are useful when a decision depends on unusually high or low outcomes that a summary average may conceal.

Use the plot to guide further investigation and model comparisons. Combine it with knowledge of how data were collected and with checks relevant to the intended decision. If the business concern is an extreme percentile, assess performance in that tail rather than choosing a model solely because most central observations lie near a line.

Check your understanding

A normal probability plot is nearly straight, but a time-order plot shows long runs of positive residuals. Has the probability plot established independent errors?

Answer: No. It assesses a distributional comparison and does not preserve temporal order. The runs require separate investigation of dependence or changing conditions.

Related terms

Quantile
Cumulative distribution function
Goodness-of-fit
normal probability plot
quantile-quantile plot
plotting position
residual analysis
location parameter
scale parameter

Learn next

Practice constructing plots from ordered observations and theoretical quantiles. Compare plotting-position conventions, then study normal and Weibull examples. Learn how confidence envelopes, residual diagnostics, and censoring-aware methods can support interpretation without treating visual straightness as conclusive evidence.

Sources

NIST/SEMATECH, Engineering Statistics Handbook, explains probability plots, theoretical quantiles, reference lines, and distributional checking: https://www.itl.nist.gov/div898/handbook/eda/section3/probplot.htm

The five-point construction uses explicitly stated midpoint plotting positions, not NIST's order-statistic-median approximation. All calculations and the fictional case are original.

Previous: Probability of improvement Next: Probability proportional to size