Why Look at the Predictor Edges?
In a scatterplot, the predictor values run along the horizontal axis. The observations at the smallest and largest predictor values therefore sit at the edges of the observed horizontal range. These points deserve a deliberate look: an observation far from the rest of the predictor values can extend the plot and may draw the eye, even when its response value is not especially unusual.
An edge observation is not automatically an error or a problem. It may be a valid measurement that broadens the range represented in the data. The useful question is more specific: is there other data near this edge, or is one observation separated from the rest? A point with nearby predictor values has support from observations in that part of the horizontal range. An isolated point has much less nearby coverage, so the visual pattern there depends heavily on that one observation.
This is a visual assessment, not a rule for deleting observations or a numerical test of influence. Check the point in context, including its units and how it was recorded. Keep the response axis in view as well: an extreme predictor value and an unusual response value describe different features of a point.
A Simple Way to Compare Edge Gaps
For a data set with at least two distinct predictor values, sort the observed predictor values from smallest to largest. Call them \(x_{(1)}, x_{(2)}, \ldots, x_{(n)}\), where the parentheses indicate their sorted order. The left-edge gap is the distance from the smallest value to the next value; the right-edge gap is the distance from the largest value to the value just before it. Compare each gap with the full predictor range.
A convenient descriptive comparison is the edge gap divided by the full range. For example, a right-edge gap that is a large fraction of the range suggests that the largest predictor value is relatively separated from the other observations. There is no universal cutoff at which an observation becomes isolated. The data set size, the measurement scale, the substantive context, and the rest of the plotted values all matter.
This comparison is useful because it expresses the gap relative to the horizontal spread of the data. If the predictor is converted from one unit to another by multiplying every value by the same positive constant, both the gap and the range are multiplied by that constant. Their ratio stays the same. The ratio is therefore not affected by a simple change of units, though it does not replace looking at the scatterplot or understanding the context.
Worked Example: A Separated Maximum
Worked Example: A Separated Maximum
Suppose a hypothetical greenhouse records weekly light exposure, in hours, and plant growth, in centimetres, for nine plants. The predictor values are 18, 19, 20, 21, 22, 23, 24, 25, and 36 hours. Their respective growth values are 40, 42, 43, 45, 47, 49, 50, 52, and 47 cm. Assess the largest predictor value.
The observed predictor range is \(36-18=18\) hours. The right-edge gap is \(36-25=11\) hours, so the relative right-edge gap is
The other eight predictor values are spaced one hour apart, but the final gap is 11 hours. On the horizontal scale, the observation at 36 hours is separated from its nearest neighbour. It is an isolated maximum in this data set.
The 36-hour observation has a response of 47 cm, which is within the range of the other recorded growth values. This illustrates why “extreme predictor value” does not mean “unusual response value.” Still, because the point is alone at the far right, it may noticeably change the visual impression at that edge: its response is lower than the responses at 24 and 25 hours, where growth was 50 and 52 cm. The plot does not by itself tell us whether the point is erroneous or whether such a pattern would persist with more observations. We should check the measurement and record that the visual pattern at the high end is based on one separated observation.
Worked Example: An Extreme Value With Nearby Support
Worked Example: An Extreme Value With Nearby Support
In a hypothetical data set, a researcher records daily practice time, in hours, and a performance score for 11 participants. The practice-time values are 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, and 12 hours. Is the observation at 12 hours isolated at the predictor edge?
The full predictor range is \(12-2=10\) hours. The right-edge gap is \(12-11=1\) hour, giving
There is an observation at 11 hours, only one hour below the maximum, and the values continue in one-hour steps through the range. The observation at 12 hours is at the edge, but it is not separated from the other predictor values by a large gap in this set. It is better described as an extreme but supported predictor value.
That distinction matters when describing the plot. The largest value marks the high end of the observed range, but it does not stand alone in the same way as the 36-hour value in the previous example. Nearby observations let us see what responses occur close to 12 hours, rather than relying on a single point to represent that part of the horizontal scale.
Worked Example: An Unusual Gap That Needs Context
Worked Example: An Unusual Gap That Needs Context
Suppose a hypothetical survey records weekly exercise time, in hours, and a health-score measurement for 13 people. The exercise-time values are 0, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, and 30 hours. Examine the smallest predictor value.
The full range is \(30-0=30\) hours. The left-edge gap is \(8-0=8\) hours, and its relative size is
The next value after zero is separated by a larger gap than the two-hour steps among most of the remaining values. This makes the low-edge observation worth noticing. However, the calculation alone does not establish that zero is an invalid value or that it should be removed. Zero may be a legitimate report of no weekly exercise, and a value can be valid even when it is separated from nearby observations.
A careful assessment would check what “weekly exercise time” means, whether zero was an allowed and correctly recorded response, and whether the units were applied consistently. In the plot, describe the gap and the lone observation without implying that the data point is wrong. If the recording checks are satisfactory, retain it when describing the observed data.
What an Edge Observation Can and Cannot Tell You
A separated predictor value can have a large visual presence because it extends the horizontal range and has no close predictor neighbours. If its response differs from the nearby pattern, a viewer may perceive a change in direction or form. That visual effect is a reason to inspect the observation and describe the limited coverage around it; it is not proof that the observation is erroneous or that it determines a particular statistical result.
The comparison of edge gaps is most useful as a prompt for careful viewing. Look at the actual horizontal distances, the response values of the edge points, and the units. Ask whether the edge observation is part of a smooth progression of predictor values or is set apart by a noticeable gap. Also distinguish a gap that looks large relative to the plotted range from a gap that is large in a meaningful real-world sense.
An edge point can be supported on one side but not the other: at the minimum, nearby values can only lie to its right; at the maximum, they can only lie to its left. The key is not to demand observations beyond the observed range, but to see whether the edge itself is represented by more than one isolated predictor value.
Common Mistakes
- Calling every minimum or maximum an outlier. An edge value is extreme by position, but it may have other predictor values close by. Describe the spacing before labeling it isolated.
- Treating an isolated point as a mistake. Separation is a feature of the plotted data, not evidence of a recording error. Check the value and its context before drawing that conclusion.
- Confusing predictor and response extremes. A point can be far to the right but have an ordinary response. State whether you are describing its horizontal position, vertical position, or both.
- Using a ratio as a rigid decision rule. A relative edge gap helps compare spacing with the full range, but no single fraction determines whether a point is problematic. Use it alongside the plot and the units.
- Assuming a striking point proves a pattern. A separated observation may attract attention, but one point alone does not establish what the relationship is like across that part of the predictor range.
Check Your Understanding
For each question, focus on the predictor spacing as well as the point’s position in the scatterplot.
- Sorted predictor values are 1, 2, 3, 4, 5, and 15. Find the range and the right-edge gap. What does their ratio suggest?
- Can the largest predictor value be extreme without being isolated? Explain how nearby observations would affect your answer.
- A point is far to the right but has a response value similar to several other observations. Which feature is extreme, and which feature is not necessarily unusual?
- Why does a large edge gap not, by itself, show that the observation is a data-entry error?
- What happens to the relative edge gap if every predictor value is converted from hours to minutes? Explain briefly.