Residual Plot

A residual plot is a diagnostic chart you draw after fitting a regression to see whether the model deserves your trust. A residual is the gap between what actually happened and what the model predicted — observed minus predicted — for each data point. The plot scatters those residuals on the vertical axis against the fitted values (or against a single predictor) on the horizontal axis. The whole test rests on one idea: a well-specified model should have squeezed all the structure out of the data, leaving residuals that look like a formless, random cloud centered on zero. If the residuals still show a shape, the model has missed something, and that leftover pattern is the clue to what.

Reading the plot is a matter of recognizing a few tell-tale shapes. Residuals that bow into a curve mean the true relationship is nonlinear and the straight line is mis-fit — a term is missing. A funnel or fan, narrow at one end and wide at the other, means the spread of errors changes across the range, called heteroscedasticity or non-constant variance, which undermines the model's standard errors. A drift or wave when points are ordered in time signals correlated errors, a violation of independence. Lone points sitting far from the cloud flag outliers or influential observations. A residual plot is a fast, visual gut-check you run before you trust a single coefficient, p-value, or prediction.

A residual plot is often paired with, and sometimes confused for, a Q-Q plot, but they check different things. The residual plot examines the structure of the errors against the fitted values or a predictor: is the relationship linear, is the variance constant, are the errors independent. It is the tool for spotting curvature, funnels, and time patterns. A Q-Q (quantile-quantile) plot examines the distribution of the residuals instead. It sorts the residuals and plots their quantiles against the quantiles of a theoretical normal distribution; if the residuals are roughly normal, the points fall along the diagonal line, while S-shapes or heavy tails reveal skew or fat tails. One asks whether the errors have the right shape across the data, the other whether they follow the right distribution.

Because they answer different questions, they are complements, not substitutes. Validating an ordinary linear regression usually means looking at both: the residual plot to confirm linearity, equal variance, and independence, and the Q-Q plot to confirm approximate normality of the errors, which matters for the reliability of confidence intervals and tests. Skipping the Q-Q plot can leave a distributional problem hidden even when the residual plot looks clean, and skipping the residual plot can leave a curvature or variance problem hidden even when the Q-Q plot looks straight. Treating either one as the whole diagnostic story is the mistake. Together they cover the main assumptions a linear model rests on, and reading them side by side is standard practice.

Using residual plots well means making them a habit, not an afterthought. Plot the residuals before you present an R-squared or celebrate a low p-value, because those headline numbers can look fine while the assumptions behind them quietly fail. Scan for the three classic problems — curvature, a funnel, and time-ordered drift — and act on what you find: add a nonlinear term or transform a variable for curvature, use robust or weighted methods for changing variance, and model the time structure when errors are correlated. Then re-plot and confirm the residuals have collapsed into a shapeless cloud. Diagnostics are iterative; the goal is a plot with nothing left to see.

The failures are mostly about trusting summary statistics over the picture. People report a strong R-squared without ever looking at the residuals, and so ship a model that is systematically wrong at the extremes. Others overreact, reading ordinary random scatter as a defect and chasing noise. Some notice a clear funnel or curve and ignore it because the overall fit looks acceptable. And many rely on a single diagnostic when normality, constant variance, and independence each need their own check. The discipline is simple: always plot the residuals, know what each pattern means, fix what the plot reveals, and pair the residual plot with a Q-Q plot so the model's assumptions are actually tested rather than assumed.

A residual plot charts residuals — from the Latin residuum, what is left over — the leftover error between observed and predicted values in a regression.

Etymology: source.

Search interest for this term over the last five years:

View interest-over-time on Google Trends →

Areas of marketing where residual plot is a core concern: