Changing Spread Across a Scatterplot

Why Vertical Spread Matters

A scatterplot can show more than whether a response tends to rise or fall as a predictor changes. The vertical spread of the points can also change. For example, a hypothetical portfolio's monthly cost might be tightly clustered at low exposure levels but much more variable at high exposure levels. That pattern is different from a change in the average cost, even though both can appear in the same plot.

This tutorial focuses on recognizing that difference by eye. A fan-shaped pattern has vertical spread that grows across predictor values; a narrowing pattern has spread that shrinks. These are descriptions of the observed scatter, not conclusions about cause or proof that a particular statistical model is appropriate.

Definition: For a response \(Y\) and predictor \(X\), the average relationship describes how the conditional mean \(E[Y\mid X=x]\) changes with \(x\). The conditional vertical spread describes how the values of \(Y\) vary around that mean at a given \(x\); when the conditional variance exists, it is \(\operatorname{Var}(Y\mid X=x)\). A fan or narrowing pattern refers to this spread changing across \(x\), rather than only to the conditional mean changing.

In a real scatterplot there are usually few, if any, observations at exactly the same predictor value. Consequently, conditional spread is assessed visually by comparing the vertical thickness of nearby sections of the plot, or by using several predictor intervals as a rough summary. The intervals are a viewing aid: they do not make observations within a wide interval identical in predictor value.

How to Read a Changing-Spread Pattern

First look at the overall direction and shape of the point cloud, including any smooth curve used to help describe its average pattern. Then inspect how far points typically lie above and below that average pattern at low, middle, and high predictor values. A pattern of increasingly wide vertical deviations suggests increasing spread; increasingly narrow deviations suggest decreasing spread.

Do not assess spread only by measuring the distance from the horizontal axis. If the average response rises with \(X\), the points will also rise, even when their vertical scatter around that rising average is about the same throughout. The relevant comparison is how far the points sit from the local average pattern, not how high the points sit on the page.

For a small set of observations in a predictor interval, the range and sample standard deviation can help make a visual comparison explicit. If the response values in one interval are \(y_1,\ldots,y_n\), their sample mean and sample standard deviation are

\bar y=\frac{1}{n}\sum_{j=1}^{n} y_j, \qquad s=\sqrt{\frac{\sum_{j=1}^{n}(y_j-\bar y)^2}{n-1}}.

These are descriptive summaries of the values in that interval. They can support an inspection, but a coarse interval may combine points with different average responses. It is therefore important to consider both the location of the local center and the vertical spread around it.

Worked Example: An Increasing Fan Around a Rising Average

A hypothetical analyst plots a response against an exposure measure and divides the observations into three ordered exposure bands. The response values are 8, 10, and 12 in the low band; 17, 20, and 23 in the middle band; and 26, 30, and 34 in the high band. Determine whether the average response, vertical spread, or both change across bands. Use sample standard deviations to compare spread.

Method. Calculate the mean and sample standard deviation within each band, then compare the centers and spreads separately. This avoids concluding that spread changes merely because the response values are higher at higher exposure.

Low band. The mean is \((8+10+12)/3=30/3=10\). The deviations from 10 are \(-2,0,2\), so their squared sum is \(4+0+4=8\). Thus \(s=\sqrt{8/(3-1)}=\sqrt{4}=2\). The range is \(12-8=4\).

Middle band. The mean is \((17+20+23)/3=60/3=20\). The deviations are \(-3,0,3\), with squared sum \(9+0+9=18\). Thus \(s=\sqrt{18/2}=\sqrt{9}=3\). The range is \(23-17=6\).

High band. The mean is \((26+30+34)/3=90/3=30\). The deviations are \(-4,0,4\), with squared sum \(16+0+16=32\). Thus \(s=\sqrt{32/2}=\sqrt{16}=4\). The range is \(34-26=8\).

The band means rise from 10 to 20 to 30, and the sample standard deviations rise from 2 to 3 to 4. Both the average response and the vertical spread increase. The wider vertical scatter is consistent with a fan-shaped pattern; this small descriptive summary alone does not establish its cause or confirm a particular model.

Check. Each band's mean is the middle of its three equally spaced response values, and the sample standard deviations equal the respective distances from the mean to either outer value. The common mistake avoided here is treating higher response values by themselves as evidence of greater spread.

Why the Average and Spread Must Be Separated

The distinction can be expressed with a simple decomposition. Let \(m(x)=E[Y\mid X=x]\), assuming the conditional mean exists. For observations at a given \(x\), write the response as its local mean plus a deviation:

Y=m(x)+\varepsilon,\qquad E[\varepsilon\mid X=x]=0.

Here \(\varepsilon=Y-m(x)\) is the deviation from the conditional mean. Subtracting the local mean gives \(Y-m(x)=\varepsilon\), so

\operatorname{Var}(Y\mid X=x) =E\left[(Y-m(x))^2\mid X=x\right] =E[\varepsilon^2\mid X=x].

This explains why a change in the mean does not, by itself, imply a change in conditional spread: the mean determines the center \(m(x)\), while spread concerns the squared deviations from that center. If those deviations are about the same size across predictor values, the average relationship may change while the vertical spread remains approximately constant.

There is also a useful reason not to compare raw response variation across broad predictor groups. For finite second moments, the total variance can be decomposed as

\operatorname{Var}(Y) =E[\operatorname{Var}(Y\mid X)] +\operatorname{Var}(E[Y\mid X]).

The first term represents average variation around the conditional means; the second represents variation in those means across predictor values. Thus a broad group can have a wide range of response values because its average response changes across the group, even if the local scatter around that average is steady. This is why the local center and deviations from it should be considered together.

Worked Example: A Rising Average Without a Wider Fan

A hypothetical maintenance study records a response in three ordered predictor bands. The values are 4, 5, and 6 in the low band; 10, 11, and 12 in the middle band; and 16, 17, and 18 in the high band. A reviewer says, “The response values are much higher in the high band, so the scatter must be wider there.” Assess the claim.

Method. Compare each band's mean and sample standard deviation. The reviewer has compared vertical locations, so the correction is to center each band before measuring its spread.

Low band. The mean is \((4+5+6)/3=15/3=5\). The deviations are \(-1,0,1\), whose squares sum to \(1+0+1=2\). The sample standard deviation is \(s=\sqrt{2/2}=1\), and the range is \(6-4=2\).

Middle band. The mean is \((10+11+12)/3=33/3=11\). The deviations are \(-1,0,1\), again giving a squared sum of 2, sample standard deviation \(s=\sqrt{2/2}=1\), and range \(12-10=2\).

High band. The mean is \((16+17+18)/3=51/3=17\). The deviations are \(-1,0,1\), so the squared sum is 2, sample standard deviation is \(s=1\), and range is \(18-16=2\).

The average response rises from 5 to 11 to 17, but the sample standard deviation and range are constant across the three bands. The values occupy higher positions as the predictor increases, yet there is no increasing vertical spread in these summaries. The reviewer has mistaken a change in the average relationship for a fan-shaped pattern.

Check. In every band, the responses are exactly one unit below, at, and one unit above their mean. The identical deviations verify the constant spread in this constructed example.

Using a Scatterplot and a Smooth as Guides

A scatterplot is the primary display for this inspection. A smooth curve can help make a changing average relationship easier to see; it does not directly estimate spread or test whether spread changes. In R, a basic plot and a lowess smooth can be drawn as follows:

\texttt{plot(x, y)} \qquad \texttt{lines(lowess(x, y), col = "red")}

After viewing the point cloud and smooth, inspect vertical deviations around the smooth in several parts of the predictor range. If deviations appear to fan out, the pattern may indicate increasing variability. If they tighten, it may indicate decreasing variability. When the mean pattern is curved, judge distances from that curved local pattern rather than from one straight line or a horizontal reference.

A smooth is not a set of known conditional means. Its shape and the apparent thickness around it depend on the observations and on the smoothing choices. Treat it as a visual reference, not as proof that the mean has been correctly specified. When there are few observations in a region, the apparent width can also be unstable.

Worked Example: Recognizing a Narrowing Pattern

A hypothetical service team plots response time against daily workload and summarizes three workload bands. Response times are 24, 30, and 36 in the low band; 36, 40, and 44 in the middle band; and 47, 50, and 53 in the high band. Determine whether the average and vertical spread appear to change, showing the calculations.

Method. Compute band means and sample standard deviations. Looking at values alone could obscure the fact that their centers also move; centering within each band separates that movement from the spread.

Low band. The mean is \((24+30+36)/3=90/3=30\). Deviations are \(-6,0,6\); the squared sum is \(36+0+36=72\). Therefore \(s=\sqrt{72/2}=\sqrt{36}=6\). The range is \(36-24=12\).

Middle band. The mean is \((36+40+44)/3=120/3=40\). Deviations are \(-4,0,4\); the squared sum is \(16+0+16=32\). Therefore \(s=\sqrt{32/2}=\sqrt{16}=4\). The range is \(44-36=8\).

High band. The mean is \((47+50+53)/3=150/3=50\). Deviations are \(-3,0,3\); the squared sum is \(9+0+9=18\). Therefore \(s=\sqrt{18/2}=\sqrt{9}=3\). The range is \(53-47=6\).

The band means increase from 30 to 40 to 50, while the sample standard deviations decrease from 6 to 4 to 3. These summaries are consistent with a rising average response and a narrowing vertical spread. A narrowing pattern is the reverse of an increasing fan, but it deserves the same care: check that the comparison is around the local center and consider how many points support each part of the plot.

Check. The deviations are symmetric around each listed mean, and the computed sample standard deviations equal 6, 4, and 3. The common mistake avoided is to report only that the response rises, overlooking that the scatter around its rising center becomes narrower.

Common Traps and Practical Limits

Confusing location with spread. A higher cloud of points is not necessarily a wider cloud. Compare vertical distances from the local average pattern.
Comparing broad groups without considering their centers. If the average response changes substantially within a group, its raw range can reflect both changing average and local scatter.
Calling a few points a stable fan. A small number of observations can make a section look wide or narrow by chance. Describe the visual evidence cautiously.
Treating a smooth curve as a formal test. A smooth helps reveal average shape; it does not establish that conditional variance changes or identify why it does.
Assuming a scatterplot proves a cause. A plot can reveal a pattern worth investigating, but does not establish that the predictor caused the change in spread.

Recognizing changing spread is useful because it alerts an analyst that a single visual description of the average relationship may not capture all important features of the data. It can guide further investigation and careful model assessment, but this tutorial's visual diagnosis does not prescribe a correction or a particular model.

Key takeaway: Inspect the vertical deviations around the local average pattern at low, middle, and high predictor values. A change in the center is a change in the average relationship; a change in the typical size of the deviations is a change in spread. They can occur separately or together.

Check Your Understanding

For each problem, compare the average response and the vertical spread separately. Sample standard deviations use the denominator \(n-1\).

A hypothetical sensor has response values 5, 8, and 11 in the low predictor band; 14, 16, and 18 in the middle band; and 27, 30, and 33 in the high band. Calculate the mean and sample standard deviation in each band. What happens to the center and spread? (Answer: Means are 8, 16, and 30; sample standard deviations are 3, 2, and 3. The center rises, while spread narrows and then widens.)
Three ordered groups have response values 2, 6, and 10; 8, 10, and 12; and 15, 16, and 17. Calculate each sample standard deviation and describe the pattern. (Answer: The sample standard deviations are 4, 2, and 1; the spread narrows across the groups.)
A scatterplot's point cloud rises from left to right. Its points appear about equally far above and below a smooth curve at low, middle, and high predictor values. What can be said about the average relationship and spread? (Answer: The average response rises, while the visual evidence suggests approximately constant spread around the smooth.)
A reviewer compares the raw response ranges in two broad predictor bands and says the band with the higher values must have greater variability. State the key check needed before accepting that claim. (Answer: Compare each band's vertical deviations around its local average pattern; higher location alone does not show greater spread.)
In a hypothetical plot, the average response decreases across predictor values, while typical deviations around the local average appear larger at the high-predictor end. Identify the two patterns without asserting a cause. (Answer: The average response decreases, and the vertical spread appears to increase toward higher predictor values.)

What You'll Learn