Recognizing Skewness Across Plots

Reading Skew Across Plots

Skew describes asymmetry in a distribution. In exploratory analysis, it is useful to check that interpretation across more than one display: a histogram shows the overall shape, while a boxplot summarizes the center, quartiles, and tails. These plots emphasize different features, so they can support a diagnosis without always telling an identical story.

The direction of skew is named for the longer tail, not for where most observations are concentrated. A distribution with a long tail toward larger values is right-skewed (or positively skewed); one with a long tail toward smaller values is left-skewed (or negatively skewed). The main task is to compare the tail directions in the histogram with the asymmetry visible in the boxplot.

Definition: A distribution is right-skewed when its upper-value tail extends farther or tapers more gradually than its lower-value tail. It is left-skewed when its lower-value tail extends farther or tapers more gradually than its upper-value tail. These are descriptions of shape, not guarantees about a particular statistic or a formal test.

What Each Plot Contributes

A histogram groups observations into bins and displays how their frequencies vary across the value scale. To assess skew, look for the direction in which the bars continue through increasingly sparse values. For example, a large concentration of small values with a few progressively larger values suggests a right tail. A concentration of large values with a thinning trail toward smaller values suggests a left tail.

A histogram can reveal features a boxplot cannot, including multiple peaks, gaps, and concentrations at particular values. Its appearance also depends on bin width and bin boundaries. A skew diagnosis should not rest on a single unusual bar or on a histogram whose bins conceal substantial detail.

A boxplot gives a compact comparison of the median and quartiles, together with whiskers and, under a common convention, separately plotted observations beyond the fences. Let $Q_1$, $Q_2$, and $Q_3$ denote the first quartile, median, and third quartile, and let $IQR=Q_3-Q_1$. Compare the two half-spreads, $Q_2-Q_1$ and $Q_3-Q_2$. A larger upper half-spread is a clue to right skew; a larger lower half-spread is a clue to left skew.

For a modified boxplot using the 1.5 IQR rule, the lower and upper fences are $Q_1-1.5IQR$ and $Q_3+1.5IQR$. The whiskers extend to the most extreme observed values that remain within those fences; observations beyond them are plotted separately. Whisker lengths are measured from the quartile to the whisker endpoint, not from the median. A longer upper whisker can support a right-skew interpretation, but it is not decisive by itself: the quartile spacing and the histogram should also be considered.

Key comparison: Look for agreement among (1) the direction of the histogram’s longer tail, (2) the relative lengths of $Q_2-Q_1$ and $Q_3-Q_2$, and (3) the whisker lengths. Treat these as complementary clues, not interchangeable measurements.

Worked Example: Diagnosing a Right Tail

A hypothetical portfolio’s annual claim amounts are summarized by $Q_1=\$12$, $Q_2=\$15$, and $Q_3=\$20$, in thousands of dollars. A histogram has its highest bars at lower amounts and then tapers across several bins toward larger amounts. The smallest observation is 4, the largest observation at or below the upper fence is 29, and the largest observation is 36. Calculate the fences, identify the boxplot whisker endpoints, and assess the skew.

Method. Calculate the IQR and fences first; then use the stated observations to locate the whiskers. Finally compare the shape clues. This avoids assuming that every extreme observation is a whisker endpoint.

Quartile spacing. The IQR is $20-12=8$. The lower half-spread is $15-12=3$, and the upper half-spread is $20-15=5$. The upper half of the box is wider by $5-3=2$, which is a clue to right skew.

Fences. The lower fence is $12-1.5(8)=12-12=0$. The upper fence is $20+1.5(8)=20+12=32$. The minimum, 4, is above the lower fence, so the lower whisker ends at 4. The value 29 is within the upper fence, while 36 is above 32. Therefore the upper whisker ends at 29, and 36 is plotted separately.

Conclusion and check. The histogram’s tail points toward larger values, and the upper half of the box is wider. The whisker lengths are $29-20=9$ above $Q_3$ and $12-4=8$ below $Q_1$, also slightly longer above. Together these clues support a right-skew diagnosis. The whisker is correctly at 29 because it is the largest stated observation within the fence; 36 is beyond it.

Using Quartile Spacing Carefully

One way to summarize the box’s asymmetry is the quartile-based Bowley coefficient:

B=\frac{(Q_3-Q_2)-(Q_2-Q_1)}{Q_3-Q_1} =\frac{Q_3+Q_1-2Q_2}{IQR}, \qquad IQR>0.

The numerator subtracts the lower half-spread from the upper half-spread. Thus $B>0$ indicates that the upper half of the box is wider, and $B<0$ indicates that the lower half is wider. Because each half-spread is nonnegative and their sum is $IQR$, $B$ lies between $-1$ and $1$. If $IQR=0$, the coefficient is undefined; quartile spacing then cannot provide this measure of asymmetry.

This is a descriptive indicator based on three quantiles. It does not establish a distributional model, and it does not capture every feature of a distribution. In particular, it cannot reveal multiple modes, and its sign need not settle a case in which the histogram and whiskers give conflicting impressions.

Worked Example: Using the Quartile-Based Indicator

A hypothetical service-time distribution has $Q_1=20$, $Q_2=25$, and $Q_3=27$, measured in minutes. Its histogram has a thin trail toward shorter times. Calculate the IQR and $B$, find the fences, and explain what the boxplot can and cannot establish.

Method. Compare the two half-spreads and calculate $B$, then apply the 1.5 IQR rule. This separates evidence about the central half of the data from evidence about the most extreme observations.

The IQR is $27-20=7$. The lower half-spread is $25-20=5$, while the upper half-spread is $27-25=2$. Therefore,

B=\frac{2-5}{7}=-\frac{3}{7}\approx -0.429.

The negative value is consistent with left skew: the lower half of the box is wider. The lower fence is $20-1.5(7)=20-10.5=9.5$, and the upper fence is $27+1.5(7)=27+10.5=37.5$. These fences identify which observations would be plotted beyond the whiskers if any fall outside them. The supplied information does not give the sample minimum or maximum, so the exact whisker endpoints and any separately plotted observations cannot be determined.

The histogram’s trail toward shorter times and the wider lower half of the box agree, supporting a left-skew description. The coefficient summarizes quartile asymmetry, but neither it nor the fences tells us whether the histogram has one peak or several.

Why a Long Tail Can Pull the Mean

The direction of a tail is not defined by the mean, but an algebraic fact helps explain why a few extreme values can accompany a difference between the mean and median. Suppose there are $n$ observations with mean $\bar{x}$, and a new observation $x$ is added. The new mean is

\bar{x}_{\text{new}} =\frac{n\bar{x}+x}{n+1}.

Subtract the original mean from both sides:

\bar{x}_{\text{new}}-\bar{x} =\frac{n\bar{x}+x-(n+1)\bar{x}}{n+1} =\frac{x-\bar{x}}{n+1}.

If $x$ is above the original mean, then $x-\bar{x}>0$, so adding it raises the mean. If it is below the mean, adding it lowers the mean. A few unusually large observations can therefore pull the mean toward a right tail; unusually small observations can pull it toward a left tail. This reasoning explains sensitivity to extreme values. It does not prove skew from a mean-median comparison, because the full distribution may have other features affecting those summaries.

Worked Example: Resolving Conflicting Plot Clues

For a hypothetical claim-processing-time sample, the quartiles are $Q_1=8$, $Q_2=11$, and $Q_3=17$ days. A histogram appears to taper toward larger times. The minimum is 3, the maximum is 30, and there are no observations outside the 1.5 IQR fences. Compare the clues and state a measured conclusion.

Method. Calculate the quartile asymmetry and whisker lengths, then compare them with the histogram rather than forcing a single interpretation.

The IQR is $17-8=9$. The lower half-spread is $11-8=3$, and the upper half-spread is $17-11=6$. Thus $B=(6-3)/9=3/9=0.333$, indicating a wider upper half of the box. The lower fence is $8-1.5(9)=8-13.5=-5.5$, and the upper fence is $17+1.5(9)=17+13.5=30.5$. Since 3 and 30 are within the fences, the lower and upper whiskers end at 3 and 30, respectively.

The lower whisker length is $8-3=5$; the upper whisker length is $30-17=13$. Both the quartile spacing and whiskers suggest greater extension toward larger values, in agreement with the histogram. The appropriate conclusion is that the displays consistently support right skew. The calculation checks out: both observed endpoints lie within their fences, so neither is plotted separately under the stated convention.

Common Traps in Skew Interpretation

Naming the side with most observations. The direction is named for the longer tail, not the side where the bars are tallest. A cluster of small values with a sparse trail to large values is right-skewed.
Reversing the boxplot clue. If $Q_3-Q_2$ is larger than $Q_2-Q_1$, the upper half of the box is wider; that is evidence toward right skew, not left skew.
Calling a flagged point the whisker endpoint. Under the 1.5 IQR convention, a whisker ends at the most extreme observation still within its fence. A value outside the fence is plotted separately.
Treating one clue as a verdict. Unequal quartile spacing describes the central half; whiskers depend on the extremes; and histogram shape depends in part on binning. Check whether the clues agree and qualify the conclusion when they do not.
Assuming every skewed distribution has a particular mean-median ordering. The mean can be influenced by extreme observations, but a mean-median comparison alone does not define the direction of skew or reveal the whole shape.

Key takeaway: Diagnose skew by comparing tail direction in the histogram with quartile spacing and whisker lengths in the boxplot. State which clues support the conclusion, and do not claim more than the displays show.

Check Your Understanding

For each item, use the stated plot or summary information and distinguish a direct calculation from a visual clue.

A histogram of hypothetical repair costs has its tallest bars at low costs and a gradually thinning trail toward high costs. What direction of skew does the histogram suggest? (Answer: Right skew.)
A boxplot has $Q_1=14$, $Q_2=17$, and $Q_3=26$. Calculate the IQR and $B$, and state the quartile-spacing clue. (Answer: $IQR=12$, $B=(9-3)/12=0.500$; the wider upper half suggests right skew.)
A reported summary gives $Q_1=30$, $Q_2=24$, and $Q_3=20$. Can these be ordered quartiles for one distribution? (Answer: No. Quartiles must satisfy $Q_1\le Q_2\le Q_3$, but the stated values are in descending order.)
A hypothetical sample has $Q_1=5$, $Q_3=15$, and observed upper-end values 24 and 33. Calculate the upper fence and identify the upper whisker endpoint and any separately plotted value, assuming 24 is the only other upper-end observation. (Answer: $IQR=10$, upper fence $=15+1.5(10)=30$; the whisker ends at 24 and 33 is plotted separately.)
A histogram appears to have a longer right tail, but a boxplot has $Q_1=10$, $Q_2=18$, and $Q_3=20$. What does the quartile spacing suggest, and how should the disagreement be reported? (Answer: The lower half-spread is 8 and the upper half-spread is 2, suggesting left skew by quartile spacing. Report that the plots give conflicting clues and inspect the histogram’s bins and the data before making a stronger diagnosis.)

What You'll Learn