Diagnosing Changing Vertical Spread

Introduction

A scatterplot can show a clear trend in the center of the response values while also showing that the values become more or less dispersed as the predictor changes. For example, outcomes might cluster tightly at low predictor values but occupy a much taller vertical band at high values. This change in vertical spread is a feature of the relationship in its own right; it is not the same as a change in direction or curvature.

The key is to compare response values at similar predictor values. The overall height of a cloud may reflect both how its center changes and how widely observations vary around that center. A visual comparison should therefore ask two separate questions: where is the middle of the response cloud, and how thick is the cloud vertically at different predictor values or within different groups?

What Changing Vertical Spread Means

Definition: The vertical spread at a predictor value, or within a group, describes how dispersed the observed response values are around their local center. Widening spread means that this vertical dispersion tends to increase across the predictor range or in one group relative to another. Narrowing spread means that it tends to decrease. These are descriptive patterns in the observed data, not conclusions about cause.

In a scatterplot, the predictor is on the horizontal axis and the response is on the vertical axis. To assess spread, imagine looking at a narrow vertical slice of the plot: how far apart are the response values in that slice? Move from left to right and compare the thickness of successive slices. A cloud that forms a wedge or fan may indicate increasing or decreasing spread. The middle of the cloud can move at the same time, so a sloping center line does not by itself tell you whether spread changes.

For a categorical predictor, the analogous comparison is between the vertical distributions of responses in the groups. Strip plots or box plots can make this comparison visible. The response must use the same measurement scale in every group, and the groups should be relevant to the comparison being made.

A numerical summary can help describe what the eye sees. One simple measure is the range, the largest response minus the smallest. Another is the interquartile range, or IQR, the third quartile minus the first quartile; it describes the width of the middle half of the observations and is less affected by the most extreme values than the range. Neither number replaces the plot: a summary can hide the shape of a distribution, and a range can be strongly affected by a single observation.

\text{Range} = \max(y)-\min(y) \qquad\qquad \text{IQR} = Q_3-Q_1

For a quantitative predictor, comparisons are usually made across narrow ranges, or bands, of predictor values. If a band covers a large horizontal span and the response center changes substantially within it, the responses may look vertically dispersed even when their spread at any one predictor value is modest. Narrower bands help distinguish this effect. Bins should still contain enough observations to make a comparison informative; a very thin slice with only one or two points cannot reliably reveal the shape of a cloud.

Worked Example: A Cloud That Widens

Suppose a hypothetical plot compares a machine setting, \(x\), with the number of units produced per hour, \(y\). The following response values were recorded at each setting.

Setting \(x\)	Responses \(y\)	Mean response	Response range
1	18, 20, 22	20	4
2	17, 20, 23	20	6
3	14, 20, 26	20	12
4	10, 20, 30	20	20

At setting 1, the range is \(22-18=4\) units per hour. At setting 4, it is \(30-10=20\) units per hour. The mean response stays at 20 at all four settings, while the range increases from 4 to 6 to 12 to 20. A scatterplot would show the response values fanning out around a level center. This is evidence of widening vertical spread in these hypothetical observations, not evidence that the average output rises with the setting.

The conclusion is descriptive. The increasing ranges identify the pattern in this small example, but range alone would not establish that the pattern continues outside the observed settings or in a larger population.

Worked Example: A Cloud That Narrows While Its Center Rises

Now consider a hypothetical study of daily water use, \(y\), at several temperature levels, \(x\). The response values (in liters per household) are:

Temperature level \(x\)	Responses \(y\)	Mean response	Response range
1	10, 20, 30	20	20
2	20, 26, 32	26	12
3	30, 34, 38	34	8
4	40, 42, 44	42	4

At the first level, the range is \(30-10=20\) liters. At the fourth level, it is \(44-40=4\) liters. The ranges decrease across the listed levels, while the mean responses increase from 20 to 42 liters. In a scatterplot, the center would move upward and the vertical cloud would tighten. Both features matter: saying only that water use rises would miss the narrowing spread.

The ranges provide a simple check on the visual comparison, but the plotted points should still be examined. A change in range can arise from one unusually high or low observation, so it is sensible to ask whether the broader or narrower band is visible across the observations rather than being driven by just one point.

Separate Changes in Center From Changes in Spread

A common source of confusion is to judge spread from the entire vertical extent of a plot. That extent can be large simply because the center of the response changes across predictor values. Imagine two narrow groups of points: one at a low predictor value with responses near 10, and one at a high predictor value with responses near 30. The combined plot spans a large vertical distance even if each group is individually tight. That total height is not, by itself, evidence that response spread changes.

The distinction can be stated precisely for observations partitioned into groups, such as predictor bands. Let \(y_{ij}\) be observation \(i\) in group \(j\), let \(\bar{y}_j\) be that group’s mean, and let \(\bar{y}\) be the mean across all observations. The total squared distance from the overall mean divides into within-group spread and differences between group centers:

\sum_j\sum_i (y_{ij}-\bar{y})^2 = \sum_j\sum_i (y_{ij}-\bar{y}_j)^2 + \sum_j n_j(\bar{y}_j-\bar{y})^2

Here \(n_j\) is the number of observations in group \(j\). To see why the identity holds, write each difference as \(y_{ij}-\bar{y}=(y_{ij}-\bar{y}_j)+(\bar{y}_j-\bar{y})\), then square and sum. The cross-term for group \(j\) is zero because \(\sum_i(y_{ij}-\bar{y}_j)=0\). The first term on the right describes variation within groups; the second describes shifts in their centers. Thus, a wide overall cloud can result from either or both. When diagnosing changing spread, the within-band or within-group part is the relevant visual comparison.

This identity is a justification for looking locally, not a requirement to calculate sums of squares every time you view a plot. In practice, examine bands narrow enough that the response center does not shift dramatically within them, and compare their vertical thickness. If the center is curved, a broad band can mix responses from several different center levels and make the cloud appear wider than its local spread.

Worked Example: Comparing Spread Across Two Groups

A hypothetical laboratory records a response in milligrams for samples processed by two procedures. Suppose the values are displayed in a strip plot with one vertical stack for each procedure:

Procedure	Response values (mg)	Mean (mg)	Range (mg)
A	48, 49, 50, 50, 51, 52	50	4
B	40, 45, 50, 50, 55, 60	50	20

For Procedure A, the range is \(52-48=4\) mg. For Procedure B, the range is \(60-40=20\) mg. Both group means are 50 mg, so the difference in vertical thickness is not explained by a difference in these groups’ centers. The plot and ranges together suggest greater observed spread for Procedure B.

This comparison concerns the distributions in the observed samples. It does not establish that the procedure caused the difference, and a visual comparison alone does not quantify how certain we should be about a population-level difference.

A Practical Visual Check

Use a consistent sequence when reading a scatterplot or group display. First identify the response scale and the predictor values or group labels. Then look at the middle of the response cloud across the horizontal axis. Finally, compare the vertical dispersion around that middle at different locations. A smoother can help reveal a changing center, but it does not, by itself, summarize how thick the cloud is around that center.

Choose comparable locations.
For a quantitative predictor, compare narrow bands of similar width; for a categorical predictor, compare the separate groups.

Locate the center.
Check whether the middle of the response values is roughly level, rising, falling, or curved. A moving center can affect the apparent total height.

Compare vertical thickness.
Look for a consistent fan-out, tightening, or roughly stable width. Use ranges or IQRs as descriptive checks when useful.

Qualify the conclusion.
Note sparse bands and limited observations, and describe only the spread pattern visible in the observed data.

Not every plot has a clean fan shape. Spread may change gradually, abruptly, or not at all, and random scatter can make the visual pattern uneven. Similar vertical thickness across the range is a reasonable description when no systematic widening or narrowing is apparent; it does not mean every band has exactly the same range.

Common Mistake

Pitfall: “The whole scatterplot is tall, so the response spread must increase as the predictor increases.” This reasoning confuses the plot’s total vertical extent with its local vertical spread. The correction is to compare responses at similar predictor values, or within groups, while separately noting whether their centers differ.

Another useful caution is not to interpret a small difference in band widths as a definite pattern when the plot has few observations or uneven coverage. A band with many points may look fuller simply because it contains more observations. The evidence for changing spread is stronger when the change in vertical thickness is visible across a sequence of comparable regions, rather than depending on the point count or one part of the display.

Key takeaway: Diagnose changing vertical spread by comparing the thickness of the response cloud locally, at similar predictor values or within groups. Track the center separately: it may move even when spread is stable, and a wide overall plot may reflect shifting centers rather than widening spread.

Check Your Understanding

Answer these questions using the distinction between a changing center and changing vertical spread.

A scatterplot has a rising center and a band that appears equally thick from left to right. How would you describe the center and the spread separately?
At three successive predictor bands, the response ranges are 5, 9, and 16 units. What visual pattern might these summaries support, and what should you inspect before drawing a conclusion?
Two groups have nearly equal means, but one group’s vertical distribution is much wider. What does this suggest about the observed responses?
Why might a broad predictor band make a curved relationship look more vertically dispersed than it is at any particular predictor value?
What is wrong with using the total vertical height of a scatterplot as the sole measure of changing spread?