What does a rising smoother in a regression's scale-location plot indicate about the errors?

## What is on the axes The scale-location plot — sometimes called the spread-location plot — shows, for each observation, the fitted value `yhat_i` horizontally and `sqrt(|standardised residual|)` vertically. A standardised residual is the raw residual rescaled by an estimate of its own standard deviation, so that residuals from different parts of the fitted range are on a comparable footing before you compare their sizes. The two transformations each do a job: - **Absolute value** discards the sign. The question is no longer 'is the model too high or too low here' — that is the residual-vs-fitted plot's job — but 'how big is the typical error here'. Folding the negative half onto the positive half also doubles the point density, making a trend easier to see. - **Square root** compresses the upper end. Absolute residuals are strongly right-skewed even under perfectly well-behaved errors, so a smoother through them is pulled around by a few large values. The square root tames that and makes the vertical scale readable. ## Reading it Under constant error variance the cloud is a horizontal band and the smoother is flat. Note that the flat level is a positive number, not zero: the plotted quantity is a magnitude and can never go below zero. Expecting the smoother to sit on zero is a common misreading. A **rising** smoother means the spread of the errors grows with the predicted level. A **falling** smoother means the opposite — most common when large fitted values correspond to well-measured or heavily aggregated units. A **bowed** smoother means the model is noisiest in the middle of its range, which often comes from mixing two populations. ## Why bother, when the residual-vs-fitted plot shows the same thing Because judging the thickness of a band by eye is unreliable. Two things fool people. Point density usually varies across the fitted range, and a denser region simply *looks* taller because more points reach the extremes, even with identical variance. And the human eye anchors on the most extreme points, which are the least representative. The scale-location plot removes both problems: it puts a summary line through the magnitudes so you are reading a trend rather than estimating a silhouette. A modest, real change in spread — a factor of two across the range, say — is often obvious here and invisible in the raw residual plot. ## Where the pattern comes from Growing spread with level is not exotic; it is the default for many quantities. - **Positive, right-skewed outcomes** such as revenue, duration or income: the error is naturally proportional to the level, so a customer with ten times the spend is off by ten times as much in absolute terms. - **Counts**: for many count processes the variance tracks the mean, so bigger expected counts come with bigger deviations by construction. - **Aggregates over unequal group sizes**: an average over 5 people is noisier than an average over 5,000, so if group size correlates with the level, spread will too. - **A multiplicative process modelled additively**: when the truth is a product of factors, errors on the raw scale scale with the prediction. ## What it does not mean - It is **not** a statement about the level of the residuals. If you want to know whether the model over- or under-predicts somewhere, look at the residual-vs-fitted plot; the sign has been thrown away here. - It is **not** a normality check. The shape of the error distribution is a separate question and needs a quantile plot. - It is **not** a claim that the model's fitted line is in the wrong place. The line can be exactly right on average while the noise around it changes size. ## Traps The biggest is **sparse extremes**. There are usually few observations at the far ends of the fitted range, and a smoother fitted there is being asked to summarise five points. A dramatic upturn in the last five percent of the plot is often nothing. Look at where the bulk of the data is and whether the trend is supported across it. The second is **direction**. Spread may depend on one particular predictor rather than on the fitted value as a whole. If the scale-location plot is flat but you have reason to suspect one variable, plot the same vertical quantity against that predictor. A flat scale-location plot is reassurance about one direction, not a guarantee in every direction. The third is **confusing this with curvature**. A model can have a perfectly flat scale-location plot and a badly curved mean function, and vice versa. They are independent findings, and reporting the wrong one is a common interview slip. Choosing a remedy is a separate topic. The job of this plot is to establish that the constant-variance assumption is broken, and to describe how the spread changes across the range so that whatever you do next is informed by the shape rather than by a yes/no verdict.