Shape & Outliers — Shape: Skewness, Symmetry, and Outliers - College Statistics

SECTION 1

Historical Context & Motivation

The notion that data can be described by the shape of its frequency distribution is so deeply embedded in modern statistics that it is easy to forget it had to be invented. Before the nineteenth century, scientists collected measurements but rarely thought systematically about how those measurements clustered, spread, or trailed off asymmetrically. The development of concepts like skewness, symmetry, and outliers arose from attempts to reconcile idealized mathematical models—particularly the normal curve—with the messy reality of empirical data. Understanding this history illuminates why descriptive statistics treats distributional shape as foundational: without knowing a distribution's shape, we cannot choose appropriate summary measures, construct valid confidence intervals, or run hypothesis tests whose assumptions actually hold.

1733 • De Moivre's Normal Curve

Abraham de Moivre derived the bell-shaped curve as an approximation to the binomial distribution, establishing the symmetric normal distribution as a theoretical benchmark against which other shapes would eventually be compared.

1893 • Karl Pearson's Moment Coefficients

Karl Pearson introduced the system of frequency curves and formalized moment-based measures of skewness and kurtosis, providing the first rigorous quantification of distributional asymmetry and tail behavior.

1977 • Tukey's Exploratory Data Analysis

John Tukey published his landmark work on exploratory data analysis, popularizing the box plot and the 1.5 × IQR rule for identifying outliers, shifting the focus from parametric assumptions to data-driven pattern recognition.

1986 • Grubbs and Robust Methods

Building on Frank Grubbs's earlier work, the 1980s saw a surge in robust statistics—methods explicitly designed to handle outliers and non-symmetric distributions without distorting inferential conclusions.

These historical developments converge on a central question that motivates this lesson: given a dataset, how do we characterize its distributional shape, and what do we do when individual observations fall far from the bulk of the data? The answers require a precise vocabulary—left-skewed, right-skewed, symmetric, unimodal, bimodal—and quantitative tools ranging from graphical displays to moment coefficients and formal outlier criteria.

SECTION 2

Core Principles & Definitions

Before computing any formula, it is essential to build a clear conceptual framework for what distributional shape actually means. When we collect a sample and construct a histogram or density plot, the resulting curve can be described along three independent dimensions: its number of peaks (modality), its degree of mirror-image balance around the center (symmetry), and the direction and extent to which one tail stretches further than the other (skewness). Each dimension carries practical implications for which summary statistics are most informative and which inferential procedures are valid.

Symmetry

A distribution is symmetric if the left half of its density curve is a mirror image of the right half about the center. In a perfectly symmetric distribution, the mean, median, and mode coincide.

Skewness

Skewness quantifies asymmetry. A right-skewed (positively skewed) distribution has a longer right tail, pulling the mean above the median. A left-skewed (negatively skewed) distribution has a longer left tail, pulling the mean below the median.

Modality

A distribution may be unimodal (one peak), bimodal (two peaks), or multimodal. Multiple modes often suggest the data come from distinct subpopulations.

Outliers

An outlier is an observation that lies an abnormal distance from other values in the sample. Outliers can arise from measurement error, data entry mistakes, or genuine extreme phenomena, and they disproportionately affect the mean and standard deviation.

Tail Behavior

Beyond skewness, the kurtosis of a distribution describes how heavy or light its tails are relative to a normal distribution. Heavy tails increase the probability of extreme values and are closely related to outlier frequency.

✦

KEY TAKEAWAY

Think of a distribution's shape as the personality of your dataset. Symmetry is like a balanced seesaw—equal weight on both sides. Skewness is what happens when a heavy child sits on one end: the seesaw tilts, and the center of gravity (the mean) shifts toward the heavier side. Outliers are the unexpected guest who sits at the very tip of the seesaw, exerting outsized leverage on the balance point. Before running any analysis, you must read the personality of your data—because using the wrong summary statistic for a skewed distribution is like prescribing medicine without first diagnosing the patient.

SECTION 3

Visual Explanation — Shapes of Distributions

The most immediate way to assess distributional shape is to look at it. The diagram below presents three canonical distribution shapes side by side: a left-skewed distribution, a symmetric distribution, and a right-skewed distribution. Pay close attention to the relative positions of the mean, median, and mode in each case, as these relationships form the diagnostic signature that lets you classify a distribution's skewness from numerical summaries alone.

Three canonical distribution shapes. In a left-skewed distribution, the long tail extends to the left and the mean is pulled below the median. In a symmetric distribution, all three central tendency measures coincide. In a right-skewed distribution, the long tail extends to the right and the mean is pulled above the median.

The key diagnostic to internalize is the ordering of the three measures of central tendency. When a distribution is perfectly symmetric, the mean, median, and mode all coincide at the center. When the distribution is right-skewed, extreme values in the right tail inflate the arithmetic mean, dragging it to the right of the median; the mode, which simply marks the highest frequency, remains at the peak. The mirror-image logic applies to left-skewed distributions. This ordering—mean < median < mode for left-skewed and mode < median < mean for right-skewed—is the single most useful heuristic for classifying skewness when you have only numerical summaries.

SECTION 4

Mathematical Framework

While graphical inspection provides an intuitive sense of shape, formal analysis requires numerical coefficients. The most widely used measure of skewness is Pearson's moment coefficient of skewness, which is based on the third standardized moment of the distribution. We also present Pearson's second coefficient of skewness (the median-based formula) and the 1.5 × IQR rule for outlier detection. These formulas give you precise, reproducible characterizations of shape and extreme values.

PEARSON'S MOMENT COEFFICIENT OF SKEWNESS

g₁ = (1/n) Σᵢ₌₁ⁿ [(xᵢ − x̄) / s]³

where n = sample size, xᵢ = individual observation, x̄ = sample mean, and s = sample standard deviation. A positive value indicates right skewness; a negative value indicates left skewness; zero indicates perfect symmetry.

PEARSON'S SECOND COEFFICIENT (MEDIAN-BASED)

Skₚ = 3(x̄ − median) / s

This simplified formula leverages the relationship between the mean and median. It typically ranges between −3 and +3. Values outside ±1 generally indicate substantial skewness, while values near zero suggest approximate symmetry.

1.5 × IQR OUTLIER FENCES

Lower Fence = Q₁ − 1.5 × IQR | Upper Fence = Q₃ + 1.5 × IQR

where Q₁ = first quartile (25th percentile), Q₃ = third quartile (75th percentile), and IQR = Q₃ − Q₁. Any observation falling below the lower fence or above the upper fence is flagged as a potential outlier. Observations beyond 3 × IQR from the quartiles are sometimes called extreme outliers.

Z-SCORE OUTLIER CRITERION

zᵢ = (xᵢ − x̄) / s → |zᵢ| > 3 flags a potential outlier

This criterion standardizes each observation relative to the mean. Under a normal distribution, approximately 99.7% of observations fall within ±3 standard deviations, so observations exceeding this threshold are rare and warrant investigation. Note that this method is sensitive to the very outliers it seeks to detect, since both the mean and standard deviation are influenced by extreme values.

ℹ

Why Cubing Matters

In the moment coefficient of skewness, each deviation is cubed. Cubing preserves the sign of the deviation—negative deviations remain negative, and positive deviations remain positive—while amplifying the contribution of observations far from the mean. If the distribution has a long right tail, the large positive cubed deviations dominate the sum, yielding a positive skewness coefficient. This is why skewness is a more informative measure than simply comparing the mean and median: it weights extreme deviations cubically.

SECTION 5

Outlier Detection — Methods & Visual Diagnostics

Identifying outliers is one of the most consequential steps in exploratory data analysis. An outlier may represent a data entry error that should be corrected, a legitimate but rare observation that reveals important variation, or a signal that the assumed model is inadequate. The challenge is that no single rule definitively separates outliers from non-outliers; rather, multiple complementary criteria should be applied and evaluated in context. The diagram below illustrates a box plot alongside a dot plot, showing how the 1.5 × IQR fences partition observations into the central bulk and potential outliers.

A box plot of a sample with Q₁ = 23, Q₃ = 43, and IQR = 20. The upper fence is at Q₃ + 1.5 × 20 = 73. The observation at 78 (red dot) exceeds this fence and is flagged as an outlier. Notice that the whiskers extend only to the most extreme non-outlier observation, not to the fence itself.

Method	Criterion	Strengths	Weaknesses
1.5 × IQR Rule	x < Q₁ − 1.5·IQR or x > Q₃ + 1.5·IQR	Resistant to outliers themselves; based on quartiles, which are robust	Does not account for sample size; may flag too many points in large samples
Z-Score Method	\|z\| > 3 (or sometimes > 2)	Intuitive; directly tied to normal distribution probabilities	Mean and s are themselves distorted by outliers (masking effect)
Modified Z-Score (MAD)	M = 0.6745(xᵢ − median) / MAD; \|M\| > 3.5	Uses median and MAD, which are highly robust to outliers	Less widely taught; requires understanding of MAD
Grubbs' Test	G = max\|xᵢ − x̄\| / s compared to critical value	Formal hypothesis test with controlled Type I error	Assumes normality; tests only one outlier at a time

Comparison of outlier detection methods

SECTION 6

Worked Example — Analyzing Shape and Outliers

Consider the following dataset representing the annual salaries (in thousands of dollars) of 12 employees at a small company: 35, 38, 40, 42, 44, 45, 47, 48, 50, 52, 55, 120. We will determine the distribution's skewness using Pearson's second coefficient, identify outliers using the 1.5 × IQR rule, and interpret the results.

Salary Distribution Analysis

1
Step 1 — Compute Summary Statistics

Order the data (already sorted). With n = 12, the mean is x̄ = (35 + 38 + 40 + 42 + 44 + 45 + 47 + 48 + 50 + 52 + 55 + 120) / 12 = 616 / 12 ≈ 51.33. The median is the average of the 6th and 7th values: (45 + 47) / 2 = 46. The standard deviation is s ≈ 21.80 (computed from the usual formula).

x̄ ≈ 51.33, median = 46, s ≈ 21.80
2
Step 2 — Compute Pearson's Second Coefficient of Skewness

Apply the formula Skₚ = 3(x̄ − median) / s = 3(51.33 − 46) / 21.80 = 3(5.33) / 21.80 = 15.99 / 21.80 ≈ 0.733.

Skₚ ≈ 0.733 → moderately right-skewed
3
Step 3 — Find Q₁, Q₃, and IQR

For n = 12, the lower quartile Q₁ is the median of the lower half (positions 1–6): Q₁ = (40 + 42) / 2 = 41. The upper quartile Q₃ is the median of the upper half (positions 7–12): Q₃ = (50 + 52) / 2 = 51. Thus IQR = Q₃ − Q₁ = 51 − 41 = 10.

Q₁ = 41, Q₃ = 51, IQR = 10
4
Step 4 — Compute Outlier Fences

Lower fence = Q₁ − 1.5 × IQR = 41 − 15 = 26. Upper fence = Q₃ + 1.5 × IQR = 51 + 15 = 66. Check each observation: all values from 35 to 55 fall within [26, 66]. The value 120 falls well above the upper fence of 66.

120 is an outlier (120 > 66)
5
Step 5 — Interpret the Results

The positive skewness coefficient confirms that the distribution is right-skewed, consistent with the long right tail created by the salary of 120 (likely a senior executive). Because outliers inflate the mean relative to the median, the median ($46K) is a more representative measure of center than the mean ($51.33K) for this dataset. If reporting a single summary, the median paired with the IQR would better communicate the typical employee's salary.

Report median = $46K and IQR = $10K as preferred summaries for this skewed distribution.

SECTION 7

Practical Implications — Choosing Summary Statistics

The shape of a distribution directly determines which summary statistics are appropriate and which inferential methods are valid. A common error in applied work is to report the mean and standard deviation for data that are substantially skewed, giving stakeholders a misleading picture of the 'typical' observation. The table below summarizes the recommended summary measures and graphical displays for different distributional shapes.

Distribution Shape	Best Measure of Center	Best Measure of Spread	Preferred Graph
Symmetric, no outliers	Mean	Standard deviation	Histogram, normal Q–Q plot
Symmetric with outliers	Median (or trimmed mean)	IQR (or MAD)	Box plot, dot plot
Right-skewed	Median	IQR	Box plot, right-skewed histogram
Left-skewed	Median	IQR	Box plot, left-skewed histogram
Bimodal	Report both modes; median can mislead	Range or separate IQRs by group	Histogram (reveals two peaks)

Matching summary statistics to distribution shape

✦

KEY TAKEAWAY

In engineering, before selecting a material for a bridge, you would never use 'average strength' if the material has wildly inconsistent performance at the tails—you'd focus on the weakest specimens. Similarly, the mean of a skewed distribution misrepresents the center because it is dragged by tail values. The median is the 'load-bearing' statistic: it tells you where the bulk of the data actually sits, regardless of how extreme the tails are. Always let the shape of the data dictate your choice of summary.

SECTION 8

Connection to Inferential Statistics & Advanced Theory

Understanding skewness and outliers is not merely a descriptive exercise—it has profound consequences for inferential statistics. Many classical procedures, including t-tests, ANOVA, and linear regression, assume that residuals are approximately normally distributed. Severe skewness or the presence of influential outliers can inflate Type I error rates, reduce statistical power, or bias parameter estimates. The table below contrasts the descriptive perspective you have learned in this lesson with the inferential perspective you will encounter in later coursework.

Aspect	Descriptive Level (This Lesson)	Inferential Level (Future Courses)
Skewness	Measured via moment coefficients and visual inspection of histograms; guides choice of summary statistics	Checked via Q–Q plots and formal tests (Shapiro–Wilk, D'Agostino) to validate normality assumptions; data transformations (log, sqrt) applied to reduce skew
Outliers	Identified by 1.5 × IQR fences or z-scores; decision to keep, remove, or investigate based on context	Assessed via Cook's distance, leverage, and DFFITS in regression; robust estimation (M-estimators, bootstrapping) used to limit outlier influence
Symmetry	Determines whether mean or median better represents center	Underpins the Central Limit Theorem's rate of convergence; highly skewed populations require larger samples for x̄ to be approximately normal
Tail Behavior	Described by kurtosis; heavier tails increase outlier frequency	Modeled by t-distributions (heavier tails than normal); extreme value theory used in finance and engineering for tail risk

Descriptive vs. inferential perspectives on shape and outliers

As you advance through your statistics curriculum, the skills developed here—interpreting histograms, computing skewness, and flagging outliers—will become prerequisite diagnostic steps that you perform before every model you fit. The Central Limit Theorem guarantees that sample means become approximately normal for large samples, but for small to moderate samples drawn from skewed populations, this approximation can be poor. Recognizing distributional shape early allows you to choose appropriate methods—nonparametric tests, data transformations, or robust estimators—rather than blindly applying procedures whose assumptions are violated.

SECTION 9

Practice Problems

Question 1: CONCEPTUAL

A researcher reports that for a particular dataset, the mean is 72 and the median is 68. Without seeing the data, what can you conclude about the likely direction of skewness? Explain your reasoning using the relationship between the mean, median, and the tail that produces the skew.

Show answer

Because the mean (72) exceeds the median (68), the distribution is likely right-skewed (positively skewed). In a right-skewed distribution, the long right tail contains extreme high values that pull the arithmetic mean upward, away from the median. The median, which depends only on the rank ordering of the data, is resistant to these extreme values and remains closer to the bulk of the distribution. The magnitude of the gap (4 units) suggests moderate rather than extreme skewness, but a histogram or formal skewness coefficient would be needed for a definitive assessment.

Question 2: BASIC CALCULATION

Given a dataset with Q₁ = 15, Q₃ = 35, and IQR = 20, compute the lower and upper outlier fences using the 1.5 × IQR rule. Would an observation of x = −18 be classified as an outlier? What about x = 62?

Show answer

Lower fence = Q₁ − 1.5 × IQR = 15 − 1.5(20) = 15 − 30 = −15. Upper fence = Q₃ + 1.5 × IQR = 35 + 1.5(20) = 35 + 30 = 65. The observation x = −18 falls below the lower fence of −15, so it is classified as an outlier. The observation x = 62 falls within the fences (−15 ≤ 62 ≤ 65), so it is not classified as an outlier despite being well above Q₃. Note that the fences define a range of ±1.5 IQRs beyond the quartiles, and only observations outside this range are flagged.

Question 3: INTERMEDIATE

A sample of 10 exam scores is: 55, 60, 62, 65, 68, 70, 72, 75, 78, 95. Compute Pearson's second coefficient of skewness (Skₚ = 3(x̄ − median) / s) and classify the distribution. Then identify any outliers using the 1.5 × IQR method.

Show answer

First, compute summary statistics: x̄ = (55+60+62+65+68+70+72+75+78+95)/10 = 700/10 = 70. The median is the average of the 5th and 6th values: (68 + 70)/2 = 69. The sample standard deviation is s ≈ 11.05 (using the n−1 formula). Pearson's second coefficient: Skₚ = 3(70 − 69)/11.05 = 3/11.05 ≈ 0.27, indicating slight right skewness. For outliers: Q₁ = (60 + 62)/2 = 61, Q₃ = (75 + 78)/2 = 76.5, IQR = 15.5. Lower fence = 61 − 23.25 = 37.75. Upper fence = 76.5 + 23.25 = 99.75. All observations fall within [37.75, 99.75], so there are no outliers by this criterion, though the score of 95 is notably high.

Question 4: APPLIED

A hospital records the waiting times (in minutes) in its emergency department over one week: 5, 8, 10, 12, 12, 14, 15, 18, 20, 22, 25, 30, 45, 90, 180. A hospital administrator proposes reporting the mean waiting time to the public. Evaluate this proposal by analyzing the distribution's shape and outlier status. What measure of center and spread would you recommend instead, and why?

Show answer

The dataset is strongly right-skewed: most patients wait 5–30 minutes, but the values 90 and 180 represent extremely long waits. The mean is approximately 33.7 minutes, which is higher than 80% of the observations—it would mislead patients into expecting a much longer wait than most actually experience. Computing quartiles: Q₁ ≈ 12, Q₃ ≈ 30, IQR = 18, upper fence = 30 + 27 = 57. Both 90 and 180 are outliers. The median of 18 minutes paired with the IQR of 18 minutes provides a much more representative summary of the typical patient experience. The administrator should report the median and IQR, supplemented by noting that occasionally patients experience waits exceeding 90 minutes (identifying these as outlier cases requiring further operational investigation).

Question 5: CRITICAL THINKING

A dataset has a sample skewness of g₁ = 0.05, yet a box plot reveals two observations flagged as outliers on the upper end. Explain how a distribution can have near-zero skewness while still containing outliers. Under what conditions might these two diagnostics give seemingly contradictory signals? Discuss the limitations of relying on any single measure of distributional shape.

Show answer

A near-zero skewness coefficient means the distribution is approximately symmetric on average—the positive and negative cubed deviations approximately cancel. However, this does not preclude the existence of individual extreme values. If the distribution has outliers on both tails (or a compensating pattern of mild asymmetry elsewhere), the third moment can be near zero. In this specific scenario, two upper outliers could be counterbalanced by a slight leftward lean in the bulk of the data, or by a single extreme lower value that is not flagged by the IQR rule due to a wider left spread. This illustrates a fundamental limitation: skewness is a global summary (a single number describing the entire distribution), while the IQR outlier rule is a local diagnostic (examining individual points relative to the quartiles). Neither alone captures the full story. Best practice is to combine numerical summaries with graphical displays—histograms, Q–Q plots, and box plots—to develop a holistic picture of distributional shape.

SUMMARY

Lesson Summary

Summary

The shape of a distribution—characterized by its symmetry, skewness, and modality—is the first thing a statistician should assess before choosing summary statistics or inferential procedures. In a symmetric distribution, the mean, median, and mode coincide, and the mean paired with the standard deviation fully describes the center and spread. In a right-skewed distribution the mean is pulled above the median, and in a left-skewed distribution the mean is pulled below it—making the median and IQR the preferred summaries for asymmetric data.

Outliers are observations that lie an abnormal distance from the bulk of the data, identified through tools such as the 1.5 × IQR rule, z-scores, or Grubbs' test. Because the mean and standard deviation are sensitive to extreme values while the median and IQR are resistant, recognizing outliers is essential before reporting any summary. Skewness is quantified by Pearson's moment coefficient (the third standardized moment) or Pearson's second coefficient (a simpler median-based formula). Mastering these concepts prepares you for the normality checks, residual diagnostics, and robust methods that underpin the entire inferential statistics curriculum.