Please verify you are human to use the visualiser.
Each point pairs an ordered observation with the value a normal sample of the same size would have put in that position. Agreement is a straight line.
Skew bends the line, heavy tails put an S through it, and a mixture of two populations flattens its middle. The shape names the problem.
The band is simulated at the current sample size, so it shows how much departure to expect from a normal population before reading anything into it.
Skewness, excess kurtosis and Jarque-Bera are printed next to the plot, so the number and the picture can be read against each other.
The five samples are chosen so that each draws a different shape. The normal sample sits on the line with the ends wandering, which is what a passing plot looks like and is worth learning before judging a failing one. The lognormal sample curves up at the right, the most common departure in real measurements because so many quantities are bounded below by zero and unbounded above. Student t on three degrees of freedom is symmetric but heavy-tailed, so both ends turn away from the line while skewness stays near zero. The uniform sample turns the S the other way up. The two-group sample is the one that catches people out: the histogram gives it away immediately and the summary statistics can miss it entirely.
Sample size deserves as much attention as shape. Move the slider down to 20 and press "Draw again" a few times on the normal preset, and the plot produces curves and kinks that would be read as evidence at a glance. Move it to 400 and the same population draws a visibly straighter line. This is the reason a normality test reported without its sample size says very little, and the reason the band here is simulated at the current sample size.
For the formal test on the same question, the Shapiro-Wilk normality test calculator takes a sample and returns a p-value, and the distribution explorer plots the densities the presets here are drawn from. When a plot says the residuals are not normal, our tutorial on linear regression five ways covers which regression assumptions that threatens, and maximum likelihood and Bayesian inference covers fitting a distribution that fits the data instead.
In production the assumption that breaks first is rarely normality of the data. It is that the distribution seen in training still holds at serving time, which is a monitoring question rather than a modelling one. Building that check in before a model ships is part of our AI pilot design work.
Interactive probability distribution calculators. Visualise Normal, Binomial, Poisson, Beta, and 12 more distributions with live PDF/CDF charts.
Try itFree calculators for t-tests, chi-square, ANOVA, Mann-Whitney U, confidence intervals, sample size, power analysis, and effect size.
Try itInteractive linear and polynomial regression calculator. Enter data points, fit models, and visualise residuals with live R² and coefficient estimates.
Try itFrequentist and Bayesian A/B test analysis. Z-test for proportions, Beta posterior distributions, probability of winning, and expected loss.
Try itGet notified when new tutorials and deep-dives are published. No spam, unsubscribe anytime.
Common questions about quantile plots, normality and tail behaviour