Outlier vs influential point in regression

By Jude Wallis · Published

In regression, an outlier is a point with a large residual: it sits far above or below the line. An influential point is one whose removal noticeably changes the slope, intercept, or correlation. An outlier can leave the slope unchanged, and an influential point can have a small residual.

AP Statistics: Unit 5 (topics 5.1 Graphical Representations Between Two Quantitative Variables, 5.4 Residuals, 5.5 Least-Squares Regression). This page belongs to Unit 5, Regression Analysis, of the Fall 2026 AP Statistics course: unusual features of a scatterplot are part of topic 5.1, residuals are topic 5.4, and the least-squares line is topic 5.5. Influential point and high-leverage point are standard statistics vocabulary for these ideas.

Outlier vs influential point: the short answer

The two labels are assigned by different tests. An outlier in a regression setting is judged by its residual, the vertical distance from the point to the line. If that distance is large compared with the other residuals, the point is an outlier. An influential point is judged by refitting: remove the point, compute the least-squares line again, and compare. If the slope, the intercept, or the correlation rr changes noticeably, the point is influential.

Because the tests differ, the labels can come apart. A point with a large residual near the middle of the xx-values is an outlier that barely tilts the line. A point far to the right or left of the others can pull the line toward itself so strongly that it ends up with a small residual, which makes it influential without looking like an outlier. The two worked examples show both cases, using the same six starting points: an outlier that leaves the slope unchanged while shifting the intercept and lowering rr, and a point that changes the slope sharply yet ends up with a small residual.

What makes a point an outlier

A residual is the observed response minus the predicted response:

residual=y−y^\text{residual} = y - \hat{y}

where y^\hat{y} ("y-hat") is the value the regression line predicts. An outlier is a point that does not follow the general pattern of the data, and in regression that shows up as a residual much larger in size than the rest. On a scatterplot it is a point well above or below the line. On a residual plot it is a point far from the horizontal line at zero.

This is a different test from the one used for a single quantitative variable, where an outlier is flagged by the 1.5×IQR1.5 \times \text{IQR} rule or by lying more than 2 standard deviations from the mean. A point can have an ordinary xx-value and an ordinary yy-value and still be a regression outlier, because the combination falls far from the line.

What makes a point influential

An influential point is defined by its effect. Fit the line with the point and without it. If the two fits are close, the point has little influence. If they differ noticeably, the point is influential. State which quantity changed: a point can change the slope, shift the intercept, raise or lower rr, or any combination of these.

Points with xx-values far from xˉ\bar{x} ("x-bar"), the mean of the xx-values, have the most potential to be influential. Such points are called high-leverage points. The least-squares line always passes through (xˉ,yˉ)(\bar{x}, \bar{y}), so a point far from xˉ\bar{x} acts like a weight at the end of a lever and can rotate the line. High leverage is only potential, though. If the far-out point lies along the same pattern as the rest, removing it changes the line very little. You can try this by dragging a point in the influential point interactive.

The differences side by side

FeatureOutlierInfluential point
Defined byA large residualA noticeable change in the fit when the point is removed
How to checkLook at the residuals or the residual plotRefit the line without the point and compare
Typical positionFar above or below the line, at any xxOften at an extreme xx, far from xˉ\bar{x}
Size of its own residualLarge, by definitionCan be small, because the line is pulled toward it
Effect on the slopeSmall when its xx is near xˉ\bar{x}Can be large
Effect on rrUsually weakens itCan weaken it or strengthen it
Can be bothYesYes

A point far from the others in xx that also falls off the pattern is usually both. A point that is off the pattern but near xˉ\bar{x} is an outlier that changes the intercept and rr and leaves the slope alone. A point that is far out in xx but on the pattern is not an outlier and barely changes the line.

One line, three added points

Six points, (2,6)(2, 6), (4,7)(4, 7), (6,11)(6, 11), (8,14)(8, 14), (10,16)(10, 16), (12,21)(12, 21), have the least-squares line y^=2+1.5x\hat{y} = 2 + 1.5x with r≈0.988r \approx 0.988. Add one more point in three different ways and refit.

Added pointResidual from the original lineNew lineNew rr
(7,30)(7, 30)17.517.5y^=4.5+1.5x\hat{y} = 4.5 + 1.5x0.6090.609
(24,14)(24, 14)−24-24y^=8.95+0.399x\hat{y} = 8.95 + 0.399x0.5570.557
(24,38)(24, 38)00y^=2+1.5x\hat{y} = 2 + 1.5x0.9970.997

The first point is a clear outlier at x=7x = 7, which is exactly xˉ\bar{x}. The slope does not change at all; the line shifts up by 2.5 and rr falls. The second point is far to the right and off the pattern, and the slope drops from 1.5 to about 0.4. The third point is just as far to the right but lies exactly on the original line, so the line is unchanged and rr rises slightly. The second and third points have the same xx-value and the same leverage. Only one of them changes the line.

The classic mix-up and how to avoid it

The first mistake is treating the two words as synonyms and writing "this outlier is influential" without checking. A large residual does not establish influence over the slope, as the point (7,30)(7, 30) shows.

The second mistake is hunting for influential points on the residual plot alone. After (24,14)(24, 14) is added and the line is refit, that point's residual is about −4.5-4.5, while the ordinary point (12,21)(12, 21) has a residual of about 7.37.3. The point that did the damage has a smaller residual than a point that did nothing wrong, because the line moved toward it. Look at the scatterplot for points far from the rest in the xx direction, and refit without them.

The third mistake is deleting a point because it earned one of these labels. Neither label says the point is an error. Check how the value was recorded, and if it cannot be explained, report the fit with and without it. More cases to practice on are in the influential points practice set.

An outlier that leaves the slope alone

Six points, (2,6)(2, 6), (4,7)(4, 7), (6,11)(6, 11), (8,14)(8, 14), (10,16)(10, 16), (12,21)(12, 21), have least-squares line y^=2+1.5x\hat{y} = 2 + 1.5x. A seventh point (7,30)(7, 30) is added. Find its residual from the original line, then find the new least-squares line.

  1. Confirm the original line. xˉ=426=7\bar{x} = \frac{42}{6} = 7 and yˉ=756=12.5\bar{y} = \frac{75}{6} = 12.5. The xx deviations are −5,−3,−1,1,3,5-5, -3, -1, 1, 3, 5, so ∑(x−xˉ)2=25+9+1+1+9+25=70\sum (x - \bar{x})^2 = 25 + 9 + 1 + 1 + 9 + 25 = 70.

  2. The yy deviations are −6.5,−5.5,−1.5,1.5,3.5,8.5-6.5, -5.5, -1.5, 1.5, 3.5, 8.5, so ∑(x−xˉ)(y−yˉ)=32.5+16.5+1.5+1.5+10.5+42.5=105\sum (x - \bar{x})(y - \bar{y}) = 32.5 + 16.5 + 1.5 + 1.5 + 10.5 + 42.5 = 105. The slope is 10570=1.5\frac{105}{70} = 1.5 and the intercept is 12.5−1.5(7)=212.5 - 1.5(7) = 2.

  3. Residual of the new point from the original line: the prediction at x=7x = 7 is 2+1.5(7)=12.52 + 1.5(7) = 12.5, so the residual is 30−12.5=17.530 - 12.5 = 17.5. The six original residuals are 1,−1,0,0,−1,11, -1, 0, 0, -1, 1, so 17.5 is far larger than any of them. The point is an outlier.

  4. Refit with seven points. The new xˉ\bar{x} is 497=7\frac{49}{7} = 7, unchanged, so the new point's xx deviation is 0. It adds 0 to ∑(x−xˉ)2\sum (x - \bar{x})^2, which stays at 70, and 0 to ∑(x−xˉ)(y−yˉ)\sum (x - \bar{x})(y - \bar{y}), which stays at 105.

  5. The new slope is 10570=1.5\frac{105}{70} = 1.5, the same as before. The new yˉ\bar{y} is 75+307=15\frac{75 + 30}{7} = 15, so the new intercept is 15−1.5(7)=4.515 - 1.5(7) = 4.5.

  6. The correlation falls from about 0.988 to about 0.609, because the new point adds a large amount of scatter around the line.

The point (7,30)(7, 30) has a residual of 17.5 from the original line, so it is an outlier. The refit line is y^=4.5+1.5x\hat{y} = 4.5 + 1.5x: the slope is still 1.5, the intercept rises by 2.5, and rr drops to about 0.609. Because the point sits at xˉ\bar{x}, it has no leverage to tilt the line. It is influential for the intercept and the correlation, and it has no effect on the slope.

An influential point with a small residual

Start again from the same six points with line y^=2+1.5x\hat{y} = 2 + 1.5x. This time the seventh point is (24,14)(24, 14). Find the new least-squares line and the new point's residual from it.

  1. New means: xˉ=42+247=667≈9.429\bar{x} = \frac{42 + 24}{7} = \frac{66}{7} \approx 9.429 and yˉ=75+147=897≈12.714\bar{y} = \frac{75 + 14}{7} = \frac{89}{7} \approx 12.714.

  2. Use the shortcut sums. ∑x2=4+16+36+64+100+144+576=940\sum x^2 = 4 + 16 + 36 + 64 + 100 + 144 + 576 = 940, so ∑(x−xˉ)2=940−6627=940−622.286=317.714\sum (x - \bar{x})^2 = 940 - \frac{66^2}{7} = 940 - 622.286 = 317.714.

  3. ∑xy=12+28+66+112+160+252+336=966\sum xy = 12 + 28 + 66 + 112 + 160 + 252 + 336 = 966, so ∑(x−xˉ)(y−yˉ)=966−66×897=966−839.143=126.857\sum (x - \bar{x})(y - \bar{y}) = 966 - \frac{66 \times 89}{7} = 966 - 839.143 = 126.857.

  4. New slope: 126.857317.714≈0.399\frac{126.857}{317.714} \approx 0.399. New intercept: 12.714−0.399(9.429)≈8.9512.714 - 0.399(9.429) \approx 8.95. The new line is y^=8.95+0.399x\hat{y} = 8.95 + 0.399x.

  5. Compare slopes: the original was 1.5 and the new one is about 0.4. One added point cut the slope to roughly a quarter of its value, so the point is influential. The correlation also falls, from about 0.988 to about 0.557.

  6. Residual of the new point from the new line: the prediction at x=24x = 24 is 8.95+0.399(24)≈18.538.95 + 0.399(24) \approx 18.53, so the residual is 14−18.53≈−4.514 - 18.53 \approx -4.5.

  7. Compare with another point. At x=12x = 12 the new line predicts 8.95+0.399(12)≈13.748.95 + 0.399(12) \approx 13.74, so the point (12,21)(12, 21) has residual 21−13.74≈7.321 - 13.74 \approx 7.3, larger in size than the residual of the point that was added.

With (24,14)(24, 14) included, the line becomes y^=8.95+0.399x\hat{y} = 8.95 + 0.399x and rr falls to about 0.557. The slope dropped from 1.5 to about 0.4, so the point is influential. Its own residual from the new line is only about −4.5-4.5, smaller in size than the 7.3 at (12,21)(12, 21). An influential point can hide on a residual plot because the line has been pulled toward it.

Frequently asked questions

Can a point be both an outlier and an influential point?

Yes. A point that is far from the other xx-values and also far off the pattern typically changes the line a great deal and can still have a sizable residual. The two labels describe different properties, so a point may have both, one, or neither.

Is every high-leverage point influential?

No. High leverage means the xx-value is far from the mean of xx, which gives the point the potential to move the line. If the point falls along the same pattern as the rest, removing it changes the line very little. On this page, adding (24,38)(24, 38) leaves the line y^=2+1.5x\hat{y} = 2 + 1.5x exactly as it was.

Should I remove an outlier or an influential point from the data?

Only with a documented reason, such as a recording error. Being unusual is not a reason. If the point cannot be explained, the standard practice is to report the regression both with and without it so a reader can see how much the conclusion depends on that one observation.