Did you know that four points can appear to tell a very tidy story, yet a fifth point can turn the apparent relationship the other way? This is one reason a scatter plot deserves more than a quick glance at its line.
A scatter plot displays paired numerical observations, with one measurement on each axis. An outlier is a point unusually far from the overall pattern in a relevant direction. When one point strongly changes a calculated trend or correlation, the right response is to investigate its source and report the sensitivity of the result—not automatically to erase the point.
First read the axes, not the shape
Before calling a group of dots ‘positive’ or ‘negative’, establish what each pair actually measures. An x-axis labelled ‘hours of revision’ and a y-axis labelled ‘practice-quiz percentage’ are not the same as revision hours and real examination outcomes. Check the units, the data collection period, whether both values come from the same person or trial, and whether all points were produced under comparable conditions.
A rising cloud can indicate an association: larger values of x tend to appear with larger values of y. A falling cloud can indicate the reverse. Neither observation, by itself, establishes that x caused y. Prior knowledge, selection effects, measurement mistakes and other factors may contribute to the pattern.
Worked example: five entirely invented pupil records
Imagine a fictional class practice exercise. The x-value is a pupil’s reported preparation time in hours; the y-value is their score on one 0–100 point practice quiz. The data are invented solely to show mathematics; they are not measurements of learners and do not predict grades.
- Pupil A: (1 hour, 60 points).
- Pupil B: (2 hours, 64 points).
- Pupil C: (3 hours, 68 points).
- Pupil D: (4 hours, 72 points).
- Pupil E: (5 hours, 35 points).
The first four points lie on the line score = 56 + 4 × hours. For example, at 3 hours the expression gives 56 + 4 × 3 = 68 points. Across those four records, an extra hour is associated with an increase of 4 points. Because they were constructed that way, their Pearson correlation is exactly +1. This is an illustrative arrangement, not evidence that real revision produces fixed gains.
Now include the fifth record. Extending the line to 5 hours would give 56 + 4 × 5 = 76 points, but the displayed fifth score is 35 points, a difference of −41 points from that simple extension. This unusually low point dramatically alters a small sample’s summary.
For readers ready for an optional correlation check, the five x-values have mean 3 and the five y-values have mean 59.8. The sums of squared deviations are 10 for x and 848.8 for y; the sum of products of deviations is −42. Pearson’s r = −42 ÷ √(10 × 848.8) ≈ −0.46. That is now a moderately negative sample correlation, even though the first four points lined up perfectly upward. The calculation can be checked by substituting the five pairs into the formula.
What changed? Not the first four observations. The fifth point, and the decision to include it, changed the statistic. This makes the central practical question unavoidable: what does the fifth point represent?
A diagnostic: error, genuine exception or different question?
There are several possible explanations, and they require different actions. Perhaps 35 was mistyped instead of 85. Perhaps the score was correct but the pupil completed a different version of the quiz. Perhaps the pupil genuinely reported five hours and scored 35 on the same assessment. In each case the plot looks similar initially, but the evidential meaning differs.
- Transcription error: compare the plotted coordinate against the original record; correct a demonstrable mistake and keep an audit note.
- Unit or scale mismatch: check whether everyone reported hours rather than minutes and whether quiz scores use the same maximum.
- Different measurement conditions: label, separate or account for a different quiz, assistance condition or test date rather than quietly combining incompatible results.
- Genuine unusual observation: keep it visible and discuss its influence. A rare but valid result may be especially informative.
- Insufficient explanation: say the reason remains unknown; do not invent a story about motivation, ability or honesty.
The key is to distinguish a measurement problem from a surprising but valid case. The latter cannot be excluded just because it disrupts a preferred conclusion.
What does an influential point mean?
An outlier is unusual in some sense, but not every unusual point strongly changes a fitted line. An influential observation substantially affects a model’s estimated parameters or conclusion. A point far to the right on the x-axis can have especially strong leverage over an ordinary least-squares line. By contrast, a point unusually high or low in y near the centre may have a different effect. The position and the modelling method matter.
An accessible first check is to describe the display with all recorded points, then compare its appearance and simple summaries with a clearly labelled sensitivity calculation that temporarily leaves the suspected point out. This is a comparison, not a licence to permanently discard an inconvenient result. Record both versions and the reason for the check.
Guided practice: do not confuse the point with the explanation
Draw a coordinate grid from 0 to 6 hours horizontally and 0 to 100 points vertically, then place the five invented pairs. First describe only what can be seen: four points rise together, the fifth is substantially lower. Next describe the numerical check: the first four follow a positive line, while all five yield r approximately −0.46. Finally write a question to investigate: Was the fifth pupil’s score recorded on the same quiz under the same scoring rules?
An answer such as ‘Pupil E was lazy’ is not supported. Neither ‘the data prove revision is harmful’ nor ‘the fifth pupil must be excluded’ is justified. A responsible description would be: ‘This five-record illustration is highly sensitive to one unusual point; the recorded conditions must be checked before interpreting the association.’
Independent practice: a second set of data
Without using the worked example, consider a new fictional school-lab exercise. Four trials record temperature in °C and dissolving time in seconds: (20, 80), (25, 70), (30, 60), (35, 50). A fifth record is (40, 92). All observations refer to the same stated substance and mass, but you have not been told whether the fifth trial used the same stirring method.
Task one: describe the first four points. They show dissolving time falling by 10 seconds for each 5 °C increase, within the illustrative observations. Task two: say what is unusual about the fifth point. Its time of 92 seconds reverses the earlier downward pattern. Task three: design a fair check. Verify timing, temperature and stirring, and repeat with the same controlled procedure if appropriate. Task four: explain why the fifth measurement should not simply disappear from the notebook.
A strong response labels the observation as a possible exception or comparability problem, checks the underlying evidence and distinguishes the raw records from any later analysis. There is no need to invent an experimental mechanism without supporting data.
How to report this honestly in a school project
- State the measured quantities, units, sample size and sampling conditions.
- Show every valid raw observation, including the unusual point, with identification labels if privacy permits.
- Describe the overall direction without claiming a cause.
- Check suspicious values against original records and document any confirmed correction.
- If a sensitivity comparison is made, label which points were included in each calculation.
- Report the limitations: small sample, self-reported preparation time, different quiz conditions, or any unknowns.
- Propose a better next measurement rather than claiming an unexplained point proves misconduct.
This reporting habit is useful well beyond quizzes. School fieldwork, consumer comparisons, scientific data and news charts can all contain a value that changes the story. The reader should be able to tell what was measured, what was inferred, and which observations drive the inference.
A parent-friendly progress check
For younger learners, skip Pearson’s r. Place four sticky notes along an ascending diagonal, then move a fifth note lower. Ask: ‘What pattern do most notes show? What does the fifth note change? What information would you need before throwing it away?’ For older learners, compare calculations with and without the unusual pair, explaining that removing a point for analysis is different from deleting a record.
A teacher can assess whether the learner gives at least two defensible checks, retains the actual evidence and refrains from assigning motives to participants. Track improvement across several unfamiliar datasets rather than counting a single correct description as mastery. No single chart exercise guarantees examination scores or scientific expertise.
Misconceptions that make a small sample misleading
- ‘If r is negative, every point slopes down.’ A correlation summarises the whole set and may mask several contrasting subsets.
- ‘An outlier must be a mistake.’ It can be real; verification comes before correction.
- ‘Removing a difficult point makes analysis more accurate.’ Removing valid evidence without a defensible method can bias the answer.
- ‘A positive trend proves study causes higher scores.’ Association does not establish causation.
- ‘One plotted trend applies to every pupil.’ The sample, its conditions and its uncertainties limit generalisation.
Frequently asked questions
Do I need to calculate correlation to notice an outlier?
No. A clear graph and careful description can identify a suspicious or unusually distant point. Correlation is useful for a more advanced quantitative check; it does not replace reading the raw observations.
Should a teacher exclude the point when marking a project?
Not automatically. The learner should explain how the data were verified and why any exclusion is appropriate. A known invalid reading calls for different treatment from a valid unusual observation.
Is a trend line the same as the line passing through every dot?
Usually not. A fitted line summarises a selected model and generally does not touch every observation. It is a description under assumptions, not proof of the cause or an exact prediction.
How many points are enough?
There is no universal magic number. Required sample size depends on the question, variation, design and precision needed. This example has only five constructed pairs and is intentionally too small to support real claims about pupil performance.
Primary references and useful companion lessons
Read NIST’s treatment of scatter plots and outliers and Penn State’s introduction to correlation and influential observations for further context. Nearby eduKate lessons on the importance of the denominator and consistent fieldwork counting rules develop the same habit of checking evidence before comparing results. The Education Hub and the Clementi progressive mathematics tutorial reference give a path from diagnosis to independent checking.
A graph can surprise you, and that is part of the joy of learning. The responsible next move is to ask what the surprising point means, verify its origin and show readers how much the conclusion depends on it. Keep the evidence visible, even when it complicates the story.