Why the Locations of Points Matter
A scatterplot can suggest a clear pattern, but that impression depends on where the observations actually lie. A smooth-looking trend may be supported by many points across the predictor range, or it may be based on dense clusters separated by a wide gap. Those situations do not provide the same visual evidence about the relationship between the predictor and response.
The earlier material on mapping predictor density with rugs introduced marks along an axis to show where predictor values occur. Here, the key question is how that distribution changes what we can responsibly read from a plot. Dense regions provide many nearby observations to inspect. Sparse regions provide fewer. A gap provides no observations at all to show what happens within it. And beyond the observed predictor range, the plot contains no direct evidence about the pattern.
These are descriptive terms, not formal guarantees. “Dense” and “sparse” depend partly on the scale and width of the regions being compared. For example, a dozen points within a narrow interval may indicate substantial concentration, while the same dozen spread over a much wider interval may appear sparse. Always judge point density in relation to the predictor scale and the rest of the plot.
Density and Local Visual Support
A trend is visually better supported in a region when observations are present across that region and their positions consistently show the pattern being described. A large number of nearby points can make a local direction or curvature easier to see. In a sparse region, a few points may have a strong influence on the apparent shape. In a gap, the plot cannot show whether the relationship continues smoothly, changes direction, or behaves in some other way.
A rug plot can help reveal these differences along the predictor axis. Closely spaced rug marks indicate that predictor values are concentrated in that part of the range; widely separated marks indicate sparse coverage; a long empty stretch signals a gap. The scatterplot still matters: rugs show where predictor values occur, not how response values vary there. To assess the relationship, inspect the point positions in both dimensions.
A numerical count can help make a visual description precise, provided the interval boundaries are stated clearly. For a chosen interval, the count is simply the number of plotted observations whose predictor values fall within it. Counts depend on the interval width and placement, so they summarize coverage rather than establish a universal threshold for “enough” data.
Worked Example: Locating a Gap in Predictor Coverage
Suppose a hypothetical scatterplot shows predictor values, in hours, of 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 18, 19, and 20. The response is a laboratory reading. Describe the observed range and the coverage.
The smallest predictor value is 2 hours and the largest is 20 hours, so the observed predictor range is \([2,20]\) hours. The values from 2 through 12 occur below the higher cluster at 18, 19, and 20. There are no observations strictly between 12 and 18 hours.
For instance, the interval \((12,18)\) has width \(18-12=6\) hours and contains 0 observations. By contrast, the interval \([2,12]\) has width \(12-2=10\) hours and contains 11 observations. This comparison shows that the first interval is a gap and that observed values are concentrated in the lower portion of the range. It does not tell us what response values would have occurred between 12 and 18 hours.
A rug would show the same separation along the horizontal axis. Any description of a trend through the gap should be cautious: the plotted data do not reveal the path of the relationship there.
Gaps Can Hide Changes in Form
When observations occur in two clusters separated by a gap, the pattern within each cluster may be visible while the pattern connecting them is not. The points on either side might be consistent with a steady increase, but they could also be consistent with curvature or a change in direction inside the unobserved interval. The endpoints alone do not determine what happens between them.
This is one reason to distinguish describing what the plot shows from drawing a continuous path across it. A line or smooth curve can visually bridge a gap, but the connecting segment is not supported by observations in the empty region. It may be a convenient summary, not evidence that the response follows that exact path there.
Worked Example: Reading Two Clusters Without Filling the Gap
In a hypothetical plot of practice time \(x\), in hours, and a task score \(y\), the observed points are clustered at \(x=1,2,3,4\) and \(x=10,11,12,13\). Scores generally rise within each cluster. Describe what is and is not visible.
The observed predictor range is \([1,13]\) hours. The interval \((4,10)\) has width \(10-4=6\) hours and contains no observations. Within the lower cluster, the plotted scores generally increase as practice time increases; the same visual description applies within the higher cluster.
Those local descriptions do not establish how scores vary from 4 to 10 hours. A steadily increasing pattern across the full range is one possible continuation, but a flatter, curved, or changing pattern in the gap is also compatible with the locations of the observed points. The appropriate conclusion is that the two local clusters show increasing patterns, while the intervening relationship is not directly displayed.
The fact that the two clusters are separated does not mean that no relationship exists between the variables. It means the plot offers limited direct information about the form of that relationship in the empty interval.
Why Extrapolation Is Especially Uncertain
A pattern inside the observed range can be judged against actual plotted points. Outside that range, there are no observations to check whether the apparent pattern continues. Extending a line or curve beyond the data therefore assumes that its form remains appropriate beyond the range where it was observed. The scatterplot alone cannot verify that assumption.
This limitation is not merely a warning that a prediction might be imprecise. The continuation itself is not determined by the observed pattern. To see why, let \(b\) be the largest observed predictor value, and suppose one candidate curve is \(f(x)\). Another candidate can agree with it for every \(x\leq b\) but differ beyond \(b\):
For every observed predictor value \(x\leq b\), the added term is absent, so \(g(x)=f(x)\). The two curves therefore match all observations within the observed range, yet they give different values for \(x>b\). Since the plotted data cannot distinguish these continuations, the appearance of the observed trend does not by itself establish what happens beyond \(b\). This is a logical limitation of the information shown, not a special property of any one plotting method.
Worked Example: Classifying an In-Range and an Out-of-Range Question
A hypothetical plot relates a machine setting \(x\), measured in units, to a recorded response. Predictor values range from 10 to 30 units, with many points between 14 and 24, a few near 11, and none below 10 or above 30. A reader asks what the plot suggests at 22 units and at 36 units.
At 22 units, the question is within the observed range \([10,30]\), and there are many nearby observations. It is an interpolation question with relatively good local coverage. The reader can inspect the points around 22 to describe the visible local pattern, while still avoiding a claim that the plot proves an exact response at that setting.
At 36 units, the question is outside the observed range because \(36>30\). It is an extrapolation question. There are no plotted observations between 30 and 36, or beyond 36, to show whether the pattern continues. Extending the visible trend to 36 would require assuming a continuation not checked by the plot.
The sparse observations near 11 also deserve a different description from the dense region near 22: the local pattern near 11 is based on fewer nearby points. It is still within the observed range, but its visual support is thinner. Being in range does not automatically mean every part of the range is equally well represented.
A Practical Reading Sequence
Before describing a trend as applying across a predictor range, locate the observations first. The following sequence keeps claims tied to what the plot displays.
Read the smallest and largest plotted predictor values. Mark questions beyond those limits as extrapolation.
Look for clusters, sparse stretches, and gaps. Use a rug if available, while remembering that it summarizes predictor locations only.
State whether the visible relationship rises, falls, curves, or appears otherwise, and identify the region to which the description applies.
Do not treat a smooth connector through an empty interval, or an extension beyond the range, as a pattern demonstrated by the observations.
Common Mistakes and Why This Matters
- Treating a smooth curve as evidence everywhere it is drawn. A curve can make a gap look filled, but no observations in the gap verify its shape. Describe the observed clusters and identify the interval without data.
- Calling every in-range value equally well supported. A predictor value can lie between the minimum and maximum yet still fall in a sparse region. Mention the local coverage, not just the overall range.
- Assuming a trend must continue beyond the largest observed value. The trend describes the plotted range. Beyond that range, alternative continuations can agree with every observed point.
- Interpreting an empty region as proof of no relationship. A gap says that predictor values were not observed there; it does not reveal the response pattern that would have appeared there.
- Comparing counts without considering interval width. Ten observations in a wide interval and ten in a narrow interval represent different concentrations. State interval boundaries and widths when using counts to describe density.
Reading density and gaps carefully matters because the range of a visual claim should match the range and distribution of the data. A plot may support a useful description in a dense region, a qualified description in a sparse region, and no direct description in a gap or beyond the observed limits. Being explicit about these distinctions makes a visual interpretation more accurate without pretending the plot answers questions its data cannot address.
Check Your Understanding
For each question, connect your answer to the locations of observations on the predictor axis.
- A plot has predictor values from 3 to 15, with a clear gap from 7 to 11. Which predictor values are outside the observed range, and what does the gap tell you?
- Why can a trend near a dense cluster be easier to assess visually than a trend in a sparse region?
- A smooth curve crosses a predictor interval containing no points. What can and cannot be concluded from the curve in that interval?
- A question asks about a predictor value just below the minimum observed value. Is this interpolation or extrapolation? Explain.
- Two possible curves match throughout the observed range but differ beyond its maximum. What does this show about claims based only on visual extrapolation?