Why Begin With a Scatterplot?
When a data set contains paired measurements, each observational unit contributes a value of a predictor and a corresponding value of a response. A scatterplot displays those pairs as points. Before calculating a regression estimate or relying on a straight-line model, inspect the plot: a summary based on a straight line can be misleading if the pattern curves, has an unusual observation, or has little clear direction.
In an actuarial setting, a hypothetical study might pair a policy feature with a claim-related measure for each policy. The plot can reveal whether larger predictor values tend to accompany larger or smaller responses, whether that tendency is approximately straight, and whether some observations differ from the overall pattern. A plot describes association; by itself, it does not establish that changing the predictor causes a change in the response.
The predictor is conventionally placed on the horizontal axis and the response on the vertical axis. Label both axes with the variable and its units. Choose scales that display the observed values without hiding important variation or exaggerating small differences. A point is unusual relative to a scatterplot when it lies apart from the main cloud or has a predictor value far from the others. Such a point deserves investigation, but unusual does not automatically mean erroneous.
A Systematic Reading of the Pattern
Read a scatterplot in a consistent order. First check the axes, units, and range of observed values. Then describe direction: an upward pattern is positive, a downward pattern is negative, and a cloud with no clear tendency has no evident direction. Next assess strength by how tightly the points follow the pattern. Finally examine form and unusual observations. These are distinct features: a pattern can be strongly curved, weakly linear, or mostly straight with one point that does not fit.
- Positive direction: larger \(x\) values tend to occur with larger \(y\) values.
- Negative direction: larger \(x\) values tend to occur with smaller \(y\) values.
- Strength: points close to a coherent pattern indicate a stronger association than widely dispersed points.
- Form: a roughly straight band is different from a bend, curve, or other systematic shape.
- Unusual observations: inspect points separated from the rest or located at an extreme of the predictor range.
“Strong” and “weak” are visual descriptions, not precise numerical cutoffs. A scatterplot can support a judgment that a relationship is approximately linear, but the plot does not prove that a linear model is appropriate for every purpose. Also, the apparent strength can depend on the scales chosen for display, so check the numerical axis labels rather than judging only the picture’s shape.
Worked Example: A Nearly Straight Upward Pattern
Hypothetical scenario. An analyst records the number of completed training modules, \(x\), and a quality-assurance score, \(y\), for five employees. The scores are on a 30-point scale. Assess direction, strength, form, and unusual observations before considering a straight-line description.
| Modules, \(x\) | Score, \(y\) |
|---|---|
| 1 | 12 |
| 2 | 14 |
| 3 | 16 |
| 4 | 18 |
| 5 | 20 |
Method. Plot the five coordinate pairs: \((1,12)\), \((2,14)\), \((3,16)\), \((4,18)\), and \((5,20)\). Compare the response changes as the predictor increases by one module each time.
From \(x=1\) to \(x=2\), the response changes by \(14-12=2\). From \(x=2\) to \(x=3\), it changes by \(16-14=2\). The next changes are \(18-16=2\) and \(20-18=2\). Thus the observed response rises by 2 points for each one-module increase throughout these five observations.
Assessment. The direction is positive because the scores rise as the number of modules rises. The points form an exact straight pattern in these observed data, so the association is strong and there is no visible curvature. None of the five points separates from that pattern. This makes a straight-line description plausible over the observed range of 1 to 5 modules; it does not establish causation or justify predictions beyond that range.
Common mistake avoided. Do not call this pattern downward: both the listed response values and their successive changes increase. Also, an upward association alone does not show that training caused higher scores; other differences among employees could be relevant.
Using Changes to Notice Curvature
When predictor values are equally spaced, successive response changes provide a useful descriptive check for curvature. If the response changes are roughly constant, the pattern is consistent with a straight line. If those changes systematically grow or shrink, a curve may describe the observed pattern better. This is a visual diagnostic, not a substitute for inspecting the plotted points: irregular noise can also make individual changes vary.
The reason is direct. For a straight pattern described by \(y=a+bx\), increasing \(x\) by a fixed amount \(h\) changes the corresponding \(y\) value by:
The change is the same at every starting value of \(x\), because \(b\) and \(h\) are fixed. Conversely, equal response changes for equal predictor steps are consistent with a straight pattern. They do not prove that a straight-line model is suitable in a broader population; they describe the observed points.
Worked Example: A Positive Pattern That Curves
Hypothetical scenario. A facilities team records equipment age in years, \(x\), and annual maintenance cost in hundreds of dollars, \(y\), for six machines. Decide whether the scatterplot supports a straight-line description.
| Age, \(x\) (years) | Cost, \(y\) (hundreds of dollars) |
|---|---|
| 1 | 2 |
| 2 | 3 |
| 3 | 6 |
| 4 | 11 |
| 5 | 18 |
| 6 | 27 |
Method. Plot the coordinate pairs and calculate the successive response changes. The age increases by exactly one year at each step, so these changes can be compared directly.
The changes are \(3-2=1\), \(6-3=3\), \(11-6=5\), \(18-11=7\), and \(27-18=9\) hundreds of dollars. The response increases each time, so the direction is positive. But the increases are not approximately constant: they rise from 1 to 3 to 5 to 7 to 9. The plotted points therefore bend upward rather than following a roughly straight band.
Assessment. A straight-line description is questionable because the cost increases more rapidly at the larger equipment ages in this small data set. The pattern has a clear direction, but direction does not determine form. The six points do not establish that costs will continue to accelerate for older equipment; the observed ages only range from 1 to 6 years.
Common mistake avoided. Calling the data “linear” just because both variables increase confuses positive direction with straight-line form. The changing increments are the warning: equal one-year steps correspond to successively larger cost changes.
Unusual Points and the Limits of Visual Judgment
An unusual point can be separated vertically from a pattern, far away horizontally from most observations, or both. A point with an extreme predictor value is especially important to notice because it occupies a region where there may be little other data for comparison. At this stage, describe where the point lies and how it relates to the rest of the plot. Do not assume that it should be discarded or assign it a formal influence label based only on appearance.
Check an unusual observation against its source record, units, and measurement process. If the data are correct, the observation may represent a genuine case that the analysis should retain. If there is a documented recording error, correction or exclusion may be appropriate, with the reason made explicit. In either case, compare the overall pattern with and without the point as a sensitivity check, not as a license to remove inconvenient data.
Worked Example: An Observation Apart From the Main Pattern
Hypothetical scenario. For a small set of service contracts, an analyst plots a risk index \(x\) against annual service visits \(y\). The following data are recorded. Assess the main pattern and the final observation.
| Risk index, \(x\) | Visits, \(y\) |
|---|---|
| 1 | 4 |
| 2 | 6 |
| 3 | 8 |
| 4 | 10 |
| 5 | 12 |
| 10 | 8 |
Method. First examine the five observations with risk indices from 1 through 5. Their successive visit changes are \(6-4=2\), \(8-6=2\), \(10-8=2\), and \(12-10=2\). They form a tight upward straight pattern. Then locate the remaining point, \((10,8)\), relative to that pattern.
The final point is different in both coordinates: its risk index, 10, is 5 units above the next-largest index, 5, while its visit count, 8, is 4 below the preceding observation’s count, 12. If the five-point pattern continued by increasing visits by 2 for each one-unit increase in risk index, an index of 10 would correspond descriptively to \(12+5(2)=22\) visits. The observed value is 8, which is \(22-8=14\) visits below that continuation. This comparison flags the point; it does not claim that 22 is a validated prediction.
Assessment. The first five points have a clear positive, nearly straight pattern. The last observation is unusual relative to it and lies far from the predictor values of the other five. Check its record and context. If it is valid, report that the overall plot includes an unusual point and that the visual pattern depends on considering it; do not silently omit it.
Common mistake avoided. Do not delete the final point merely because it weakens the main pattern. Nor should a continuation from the first five observations be treated as certain at \(x=10\): that value is outside their range, so the comparison is extrapolative.
Deciding Whether a Straight-Line Candidate Is Reasonable
A straight-line candidate is more plausible when the scatterplot shows one general direction, the points cluster around a roughly straight band, and no clear bend or separate group dominates the picture. A straight line may still be an imperfect description; the point is to identify whether its basic form is defensible before using a linear model. If the plot shows a curve, a cluster, or an unusual observation, describe that feature rather than forcing a single straight-line summary.
A scatterplot is also limited by its sample and context. A strong visual association does not imply causation, and a weak-looking pattern does not prove that no relationship exists. The axes may omit important variables, observations may have been selected in a nonrepresentative way, and the plot cannot by itself establish the assumptions needed for statistical inference. It is an essential first check, not a complete analysis.
Check Your Understanding
For each hypothetical data set, assess the requested scatterplot features before deciding whether a straight-line description is plausible.
- A service analyst records training hours \(x=1,2,3,4\) and completion counts \(y=9,7,5,3\). State the direction and form, and give the successive response changes. (Answer: Negative, exactly straight in the listed observations; changes are \(-2,-2,-2\).)
- A hypothetical sensor has paired readings \((1,2),(2,5),(3,10),(4,17)\). Find the successive response changes for equal predictor steps and assess the form. (Answer: Changes are \(3,5,7\); the pattern is positive and curves upward, so a straight-line description is questionable.)
- A hypothetical portfolio review gives pairs \((2,5),(3,7),(4,9),(5,11),(12,4)\). Describe the first four points and identify the issue with the fifth. (Answer: The first four form a positive straight pattern with response changes \(2,2,2\); \((12,4)\) is unusual and far in predictor value, so it should be checked, not automatically removed.)
- In a hypothetical study, equal increases of 2 units in \(x\) are paired with response changes of \(1,1,1,1\). Explain what this suggests about form, and state one limitation of the conclusion. (Answer: The constant changes are consistent with a straight pattern; they do not prove population linearity or establish causation.)
- A scatterplot appears to show a strong upward pattern, but the axes are unlabelled and the points cover only a narrow range of \(x\). State two reasons not to rely on it for a broad prediction. (Answer: The variables and units cannot be interpreted without labels, and prediction beyond the narrow observed \(x\)-range would be extrapolation.)