Why Look for Curvature?
A scatterplot may show a clear relationship without following a straight path. For example, a response might rise quickly at first and then level off, or it might rise and later fall. A straight-line summary can describe an overall direction, but it cannot show these changes in form. Recognizing curvature helps you describe what the plotted data actually suggest before deciding whether a straight-line summary is adequate.
Curvature is not a claim that every point lies on a perfect curve. Real observations vary around any broad pattern. The question is whether the typical response appears to change with the predictor at a rate that itself changes. Look first at the raw points: do they show a sustained bend, flattening, or reversal across a meaningful range of the predictor? A smoother can help make that broad pattern easier to see, but it does not replace the observations or establish a mathematical model.
Look for Changes in the Rate of Change
Direction describes whether the response tends to increase or decrease as the predictor increases. Form describes the shape of that pattern. A curved relationship can have one direction throughout—for instance, it can increase while becoming less steep—or it can change direction, increasing at first and then decreasing.
A useful way to make the visual description more precise is to compare changes in response across successive predictor values. When predictor values are equally spaced, the differences between successive responses are adjacent changes. If those changes are similar, the sequence is consistent with a straight-line pattern. If they steadily grow, shrink, or change sign, they can indicate bending or a turning pattern. These comparisons are descriptive, not a formal test.
When predictor values are not equally spaced, raw response differences are not directly comparable: a change over a wide predictor interval may be larger simply because that interval is wider. Instead, compare adjacent slopes, calculated as the response change divided by the predictor change. Either way, the raw points remain important; a short run of irregular values does not automatically establish a curved relationship.
Worked Example: Increasing but Bending Upward
Suppose a hypothetical scatterplot shows the following response values at equally spaced predictor values. The points are described as the typical pattern in this constructed example, not as results from a real study.
| Predictor \(x\) | Response \(y\) | Change from preceding response |
|---|---|---|
| 0 | 2 | — |
| 1 | 3 | 1 |
| 2 | 6 | 3 |
| 3 | 11 | 5 |
| 4 | 18 | 7 |
The adjacent response changes are \(3-2=1\), \(6-3=3\), \(11-6=5\), and \(18-11=7\). The predictor advances by 1 each time, so the corresponding adjacent slopes are \(1/1=1\), \(3/1=3\), \(5/1=5\), and \(7/1=7\) response units per predictor unit. The relationship is increasing, and the increases become larger across the range. That is evidence of an upward bend: the pattern is getting steeper, rather than following a line with a roughly constant slope.
The plotted points in this example happen to follow \(y=x^2+2\): for \(x=3\), for instance, \(3^2+2=11\). That equation explains these constructed values exactly, but visual evidence from a real scatterplot would not, by itself, justify claiming that its relationship follows this particular equation.
Plateaus and Turning Points
A plateau is a region where the typical response changes little as the predictor continues to increase. It may come after a steep increase or decrease. A plateau across part of the plot is not the same as a flat relationship across the whole plot: the earlier portion may have a strong slope even when the later portion is nearly level. Describe where the flattening occurs rather than calling the entire relationship “flat.”
A turning point is a location where the direction of the broad pattern reverses. In an inverted-U pattern, the response tends to rise, reaches a high region, and then falls. A U-shaped pattern falls before rising. The highest or lowest observed point is not automatically the turning point of the underlying pattern; individual observations can sit above or below the broader trend. Look for a sustained change in direction in the cloud of points, preferably also visible in a smoother.
Worked Example: Increasing and Then Leveling Off
Consider these hypothetical values, again equally spaced in the predictor.
| Predictor \(x\) | Response \(y\) | Change from preceding response |
|---|---|---|
| 0 | 1 | — |
| 1 | 4 | 3 |
| 2 | 7 | 3 |
| 3 | 9 | 2 |
| 4 | 10 | 1 |
| 5 | 10.4 | 0.4 |
The response changes are \(4-1=3\), \(7-4=3\), \(9-7=2\), \(10-9=1\), and \(10.4-10=0.4\). Because each predictor step is 1, those are also the adjacent slopes in response units per predictor unit. The responses increase throughout, but the slopes shrink from 3 to 0.4. The pattern therefore suggests an increase that becomes progressively flatter and approaches a plateau near the high-\(x\) end.
It would be misleading to call this relationship flat: the response rose by \(10.4-1=9.4\) units from \(x=0\) to \(x=5\). It would also be too strong to say that the response has reached a permanent maximum. The plotted values show flattening over the displayed range; they do not reveal what happens beyond it.
Worked Example: An Inverted-U Turning Pattern
A third hypothetical set of equally spaced points has the following values.
| Predictor \(x\) | Response \(y\) | Change from preceding response |
|---|---|---|
| 0 | 2 | — |
| 1 | 6 | 4 |
| 2 | 9 | 3 |
| 3 | 10 | 1 |
| 4 | 9 | −1 |
| 5 | 6 | −3 |
| 6 | 2 | −4 |
The adjacent changes are \(6-2=4\), \(9-6=3\), \(10-9=1\), \(9-10=-1\), \(6-9=-3\), and \(2-6=-4\). Since each predictor step is 1, the adjacent slopes have those same numerical values. They are positive before the high point, zero nowhere between these sampled values, and negative afterward. The observed maximum is 10 response units at \(x=3\). Together, these values support describing the pattern as an inverted U, with a high region near \(x=3\).
The adjacent slopes change sign between the steps from \(x=2\) to \(x=3\) and from \(x=3\) to \(x=4\). That places the change in direction near the peak, but does not identify an exact turning point between observations. With real scattered data, describe the location approximately and check whether the broader cloud supports the reversal.
A Short Justification Using Differences
For equally spaced predictor values, adjacent differences provide a simple way to distinguish a straight-line sequence from one that bends. Suppose \(x\) increases by a fixed step \(h\), and the responses follow a straight line \(y=a+bx\). At successive predictor values \(x\) and \(x+h\), the response change is
Because \(b\) and \(h\) are constant, every adjacent response change is the same. The difference between consecutive adjacent changes—the second difference—is therefore zero. If second differences are consistently positive or negative in a clean sequence, the rate of change is itself changing, which is consistent with bending.
For instance, in the first worked example, the adjacent changes are \(1,3,5,7\), so the second differences are \(3-1=2\), \(5-3=2\), and \(7-5=2\). They are consistently positive. This calculation supports the visual description of a bend that gets steeper. In a real scatterplot, second differences will often vary because observations scatter around the broad pattern. Their purpose here is to clarify the idea of changing slope, not to provide a stand-alone decision rule.
Using Raw Points and a Smoother Together
A smoother can help reveal the broad form of a scatterplot when individual points make it hard to see. If the raw point cloud and smoother both show a sustained bend, flattening, or reversal, the evidence for describing curvature is stronger than if the shape appears only in the smoother. A smoother is still a visual summary: its shape can depend on the data and on how it is drawn. Do not describe a brief wiggle in a smoother as a meaningful turning pattern unless the raw observations support that interpretation.
Before describing curvature, make sure the plot lets you see the relationship clearly. An axis scale can make a modest bend appear dramatic or make real changes hard to see, so read the axis labels and units and judge the pattern over the displayed range. Curvature should describe how the typical response changes across the predictor—not a change in the vertical spread of points, which is a separate feature of a scatterplot.
Useful descriptions name both direction and form. For example: “The response generally increases, but the increase becomes less steep at higher predictor values,” or “The pattern rises to a high region and then declines.” These statements describe the visual evidence without claiming a cause, an exact equation, or behavior outside the observed range.
Common Mistakes
- Calling any irregularity curvature. Wrong reasoning: “The points go up, down, and up, so the relationship must turn several times.” Correction: individual observations vary. Look for a sustained pattern in the cloud, not every local fluctuation.
- Calling a plateau a flat relationship. Wrong reasoning: “The right side is nearly level, so there is no relationship.” Correction: describe the whole range. The response may rise substantially before it levels off.
- Treating the smoother as the data. Wrong reasoning: “The curve bends, so the observations must follow that exact curve.” Correction: check the raw points and treat the smoother as a visual aid, not proof of a specific model.
- Claiming an exact turning point from a plot. Wrong reasoning: “The highest plotted point gives the exact peak of the relationship.” Correction: a point may be unusually high or low. Describe a high or low region only when the broader pattern supports it.
- Extrapolating beyond the plotted range. Wrong reasoning: “The response is flattening, so it will stay at that level.” Correction: the plot only shows the observed range; it cannot establish the pattern beyond it.
Check Your Understanding
Answer in terms of the broad pattern and the evidence you would use.
- As a predictor increases in equal steps, the response changes are \(2, 2, 2, 2\). What form is this sequence consistent with, and why?
- A plot rises steeply at first and then becomes nearly level. Give a precise description that does not incorrectly call the entire relationship flat.
- In an inverted-U pattern, what feature of the adjacent slopes signals a turning pattern?
- Why should you be cautious if a bend appears in a smoother but is not supported by the raw points?
- When predictor values are unequally spaced, why are adjacent response differences alone not enough to compare rates of change?