Distribution Shape & Outliers — Identify distribution shape and outliers - Biostatistics

SECTION 1

Historical Context & Motivation

The systematic study of how data values cluster, spread, and deviate from typical patterns has roots stretching back to the earliest days of probability theory. Before researchers could test hypotheses or build regression models, they first needed a language for describing what a batch of numbers actually looked like — whether the values massed symmetrically around a center, trailed off in one direction, or contained anomalous observations that seemed to belong to an entirely different population. In the biomedical sciences, this descriptive step is not merely preliminary: the shape of a distribution determines which inferential tools are valid, while the presence of outliers can inflate error rates, bias parameter estimates, and even reverse the direction of an apparent effect.

1733 • De Moivre's Bell Curve

Abraham de Moivre derived the normal curve as an approximation to the binomial distribution, providing the first mathematical description of the symmetric, bell-shaped distribution that would become the default assumption in much of statistics.

1893 • Pearson's Skewness & Kurtosis

Karl Pearson formalized numerical measures of distribution shape — skewness (asymmetry) and kurtosis (tail heaviness) — enabling researchers to quantify departures from normality rather than relying on visual impression alone.

1970s • Tukey's Exploratory Data Analysis

John Tukey pioneered exploratory data analysis (EDA) and invented the box-and-whisker plot, offering a robust visual tool for identifying distribution shape and flagging outliers using the interquartile range.

2000s • Modern Biostatistical Guidelines

Reporting guidelines such as CONSORT and STROBE began emphasizing the importance of assessing distributional assumptions before applying parametric tests, making distribution-shape assessment a standard part of the biostatistical workflow.

The central question this lesson addresses is deceptively simple: given a set of biomedical measurements — blood-glucose levels, hospital length-of-stay data, gene expression counts — how do we characterize the shape of the distribution, and how do we decide whether extreme observations are genuine outliers that warrant special treatment? Mastering this skill is prerequisite to virtually every inferential procedure you will encounter in biostatistics.

SECTION 2

Core Principles & Definitions

Distribution shape and outlier identification rest on a small number of foundational concepts. Understanding these ideas provides the vocabulary and conceptual framework for everything that follows, from choosing the right summary statistics to selecting an appropriate hypothesis test.

Symmetry vs. Skewness

A distribution is symmetric when the left and right halves mirror each other. Skewness quantifies the direction and degree of asymmetry: positive (right) skew indicates a long right tail; negative (left) skew indicates a long left tail.

Kurtosis & Tail Weight

Kurtosis measures the heaviness of the tails relative to a normal distribution. Leptokurtic distributions have heavier tails (more extreme values), while platykurtic distributions have lighter tails. In biostatistics, heavy tails signal a higher probability of extreme observations.

Modality

The number of prominent peaks in a distribution is its modality. Unimodal distributions have one peak; bimodal distributions have two, often suggesting a mixture of subpopulations (e.g., responders vs. non-responders to a drug).

Outliers

An outlier is an observation that lies far from the bulk of the data. It may result from measurement error, data-entry mistakes, or represent a genuinely rare biological phenomenon. The decision to retain, transform, or exclude an outlier must be guided by domain knowledge, not convenience.

The Normal Benchmark

Many parametric procedures assume data follow a normal (Gaussian) distribution — symmetric, mesokurtic, unimodal. Assessing shape is largely about measuring how far the empirical distribution departs from this benchmark.

✦

KEY TAKEAWAY

Think of distribution shape as the fingerprint of your data. Just as a forensic analyst examines ridge patterns before making an identification, a biostatistician examines the symmetry, modality, and tail behavior of a distribution before selecting a statistical test. Applying a parametric procedure to heavily skewed data is like using a tool designed for one fingerprint pattern on a completely different one — the match is poor, and your conclusions become unreliable.

SECTION 3

Visual Explanation — Distribution Shapes

The most immediate way to assess distribution shape is through graphical displays. Histograms, density plots, and Q-Q plots each reveal different aspects of the data's geometry. The following diagram illustrates the three canonical distribution shapes — left-skewed, symmetric, and right-skewed — alongside markers indicating where the mean and median fall relative to one another.

In a left-skewed distribution the long tail extends to the left, pulling the mean below the median. In a symmetric distribution the mean and median coincide. In a right-skewed distribution the long tail extends to the right, pulling the mean above the median.

Notice the relationship between the mean and median in each panel. Because the mean is sensitive to extreme values while the median is not, a large gap between the two statistics is itself a diagnostic indicator of skewness. In biomedical data, right-skewed distributions are extremely common — hospital length-of-stay, serum biomarker concentrations, and health-care costs all tend to have a mass of typical values near the left with a long tail of rare, high values stretching to the right. Left-skewed distributions, while less frequent, appear in contexts such as age-at-onset data for diseases that primarily affect older adults.

SECTION 4

Mathematical Framework

While visual inspection is the first step, quantifying distribution shape and identifying outliers formally requires a small set of summary statistics and decision rules. We begin with the moment-based measures of skewness and kurtosis, then introduce the interquartile-range (IQR) rule that underpins Tukey's boxplot method for flagging outliers.

SAMPLE SKEWNESS (FISHER'S MEASURE)

g₁ = (n / ((n − 1)(n − 2))) × Σᵢ((xᵢ − x̄) / s)³

where n is the sample size, x̄ is the sample mean, and s is the sample standard deviation. A value of g₁ = 0 indicates perfect symmetry; g₁ > 0 indicates right skew; g₁ < 0 indicates left skew.

EXCESS KURTOSIS

g₂ = ((n(n + 1)) / ((n − 1)(n − 2)(n − 3))) × Σᵢ((xᵢ − x̄) / s)⁴ − (3(n − 1)²) / ((n − 2)(n − 3))

Excess kurtosis subtracts 3 so that the normal distribution has g₂ = 0. Positive values (leptokurtic) mean heavier tails and a sharper peak; negative values (platykurtic) mean lighter tails and a flatter peak.

TUKEY'S IQR OUTLIER FENCES

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

where Q₁ is the 25th percentile, Q₃ is the 75th percentile, and IQR = Q₃ − Q₁. Any observation below the lower fence or above the upper fence is classified as a potential outlier. The multiplier 1.5 corresponds roughly to ±2.7σ for a normal distribution.

Z-SCORE METHOD

zᵢ = (xᵢ − x̄) / s

Observations with |zᵢ| > 3 are commonly flagged as outliers under the assumption of approximate normality. However, because the mean and standard deviation are themselves sensitive to outliers, this method can mask extreme values in small samples — a phenomenon known as masking.

⚠

IQR vs. Z-Score: Which to Use?

The IQR method is robust because it relies on percentiles rather than means and standard deviations. It is preferred for skewed data or when the distribution's normality is unknown. The z-score method, while intuitive, assumes approximate normality and can underperform in small or heavily skewed samples.

SECTION 5

Detailed Breakdown — Outlier Detection with Box Plots

The box-and-whisker plot, introduced by John Tukey in 1977, remains the single most effective visual tool for simultaneously displaying distribution shape and flagging outliers. The box spans the interquartile range from Q₁ to Q₃, a horizontal line marks the median, and the whiskers extend to the most extreme data points that fall within the fences. Any point beyond the whiskers is plotted individually and considered a potential outlier. The following diagram shows a box plot for a hypothetical dataset of serum C-reactive protein (CRP) concentrations measured in a clinical cohort.

A box plot of serum CRP concentrations. The box spans Q₁ (3.5) to Q₃ (9.0), giving IQR = 5.5. The median (6.0) lies below the box's midpoint, hinting at right skew. Whiskers extend to the farthest observations within the fences. Two points at 20.0 and 22.0 exceed the upper fence and are plotted as individual outliers.

Several additional features of this box plot are worth noting. First, the median is positioned closer to Q₁ than to Q₃, which visually confirms the right skew already suggested by the skewness coefficient. Second, the upper whisker is considerably longer than the lower whisker, reinforcing the presence of a stretched right tail. Third, the two points at 20.0 and 22.0 mg/L are plotted as individual circles beyond the upper fence, identifying them as potential outliers that warrant clinical investigation — perhaps these patients had acute infections or underlying inflammatory conditions that elevated CRP well beyond the cohort norm.

💡

Bimodality Clue in a Box Plot

A box plot cannot directly reveal bimodality because it collapses the distribution to five summary numbers. If you suspect a mixture of subpopulations (e.g., different genotypes), supplement the box plot with a histogram or kernel density estimate. Bimodality may explain apparent outliers — what looks extreme in a pooled distribution may be perfectly typical within one subgroup.

SECTION 6

Worked Example — Assessing Shape & Outliers in Clinical Data

Consider the following dataset of fasting blood-glucose levels (mg/dL) measured in 15 adult participants enrolled in a diabetes-screening study: 78, 82, 85, 88, 89, 91, 93, 95, 97, 100, 103, 108, 115, 142, 210. We wish to determine the distribution shape, compute skewness, and identify any outliers using Tukey's IQR rule.

Fasting Blood-Glucose Analysis

1
Step 1 — Compute the Five-Number Summary

With 15 observations sorted in ascending order, identify: Minimum = 78, Q₁ = median of observations 1–7 = 88, Median (Q₂) = 8th value = 95, Q₃ = median of observations 9–15 = 108, Maximum = 210.

Five-number summary: {78, 88, 95, 108, 210}
2
Step 2 — Calculate the IQR and Fences

IQR = Q₃ − Q₁ = 108 − 88 = 20. Lower Fence = Q₁ − 1.5 × IQR = 88 − 30 = 58. Upper Fence = Q₃ + 1.5 × IQR = 108 + 30 = 138.

IQR = 20; Lower Fence = 58; Upper Fence = 138
3
Step 3 — Identify Outliers

Any observation below 58 or above 138 is flagged. The values 142 and 210 exceed the upper fence. No values fall below the lower fence. Thus, 142 and 210 are classified as potential outliers.

Outliers: 142 mg/dL, 210 mg/dL
4
Step 4 — Assess Distribution Shape

Compare mean and median. Mean = (sum of all values) / 15 = 1476 / 15 = 98.4 mg/dL. The mean (98.4) exceeds the median (95), indicating the right tail is pulling the mean upward. Additionally, the distance from Q₃ to the maximum (210 − 108 = 102) far exceeds the distance from the minimum to Q₁ (88 − 78 = 10), confirming pronounced right (positive) skew.

Distribution is right-skewed (positively skewed)
5
Step 5 — Clinical Interpretation

The two outlier values (142 and 210 mg/dL) likely represent individuals with impaired fasting glucose or overt diabetes, respectively. Rather than removing these values, a biostatistician might report results both with and without outliers, or use a non-parametric test that is less sensitive to extreme values. A log-transformation could also be considered to stabilize the variance and reduce skew before applying parametric methods.

Recommendation: retain outliers and use robust/non-parametric methods, or apply log-transformation.

SECTION 7

Strengths, Limitations & Comparisons of Shape-Assessment Methods

No single technique for assessing distribution shape or identifying outliers is universally superior. Each method has tradeoffs in terms of robustness, sensitivity, and the assumptions it requires. The table below compares the most commonly used approaches in biostatistical practice.

Method	Strengths	Limitations
Histogram	Shows full shape, modality, and gaps; intuitive for all audiences	Sensitive to bin width choice; appearance can change dramatically with different binning
Box Plot (Tukey)	Robust outlier flagging via IQR; compact comparison across groups; invariant to bin choice	Cannot reveal bimodality or other fine-grained distributional features
Q-Q Plot	Directly compares empirical quantiles to a theoretical distribution; sensitive to tail deviations	Requires familiarity with interpretation; less intuitive for non-statisticians
Skewness / Kurtosis	Provides a single numeric summary; easily incorporated into automated screening pipelines	Sensitive to outliers themselves; sample skewness can be misleading in small samples (n < 30)
Z-Score Rule	Straightforward calculation; relates directly to standard deviations from the mean	Assumes approximate normality; prone to masking effect in small or contaminated samples

Comparison of common methods for assessing distribution shape and detecting outliers

✦

BEST PRACTICE

In practice, biostatisticians rarely rely on a single tool. A recommended workflow combines a visual method (histogram or density plot to assess overall shape), a robust summary display (box plot for outlier detection), and a numerical check (skewness and kurtosis statistics to quantify the departure from normality). This multi-pronged approach is analogous to a diagnostic workup in medicine — no single test is definitive, but converging evidence from several tests builds confidence in the diagnosis.

SECTION 8

Connections to Formal Normality Tests & Robust Statistics

The descriptive techniques covered so far — visual inspection, skewness, kurtosis, and IQR-based outlier rules — provide the groundwork for more formal approaches. Once you progress to inferential biostatistics, you will encounter formal normality tests that convert shape information into a hypothesis-testing framework, as well as robust statistical methods that accommodate non-normal distributions and outliers by design rather than by post-hoc correction.

Descriptive Approach (This Lesson)	Advanced Extension
Visual inspection via histogram / Q-Q plot	Shapiro-Wilk test, Kolmogorov-Smirnov test, Anderson-Darling test — formal hypothesis tests for normality
Sample skewness and kurtosis coefficients	D'Agostino-Pearson omnibus test — combines skewness and kurtosis into a single chi-squared statistic
IQR-based outlier flagging (Tukey fences)	Grubbs' test, Dixon's Q-test, Rosner's test — formal tests for single or multiple outliers
Reporting mean ± SD for symmetric data	M-estimators, trimmed means, Winsorized means — robust location estimators less influenced by outliers
Choosing parametric vs. nonparametric tests by shape	Generalized linear models (GLMs) — directly model non-normal response distributions (e.g., Poisson, gamma)

How descriptive shape assessment connects to advanced inferential and robust methods

As you advance in biostatistics, you will find that distribution shape assessment never truly goes away — it simply becomes embedded in more sophisticated frameworks. Generalized linear models, for instance, allow you to specify a non-normal error distribution (Poisson for counts, gamma for right-skewed continuous outcomes) rather than forcing a normality assumption. Bayesian methods incorporate prior distributions whose shapes must also be evaluated. In every case, the intuition you develop here — reading histograms, interpreting skewness, recognizing when outliers distort summary statistics — will remain the foundation upon which more advanced analyses rest.

SECTION 9

Practice Problems

Question 1: CONCEPTUAL

A researcher reports that the mean systolic blood pressure in her sample is 135 mmHg, while the median is 128 mmHg. Without seeing the data, what can you infer about the shape of the distribution? Explain why the relationship between the mean and median supports your conclusion.

Show answer

The distribution is right-skewed (positively skewed). When the mean exceeds the median, it indicates that high values in the right tail are pulling the arithmetic mean upward while leaving the median — which depends only on the rank ordering of observations — relatively unaffected. In clinical data, right skew is common for blood-pressure measurements because a subset of hypertensive patients can produce very high readings that stretch the upper tail. The 7 mmHg gap between mean and median is a moderate indicator of skew and suggests that the median (or a non-parametric test) may be a more representative summary of central tendency for this sample.

Question 2: BASIC CALCULATION

A dataset of 12 BMI values (kg/m²) has Q₁ = 22.0, Q₃ = 29.5. An individual in the sample has a BMI of 41.2. Using Tukey's IQR rule, determine whether this value qualifies as an outlier.

Show answer

First, compute the IQR: Q₃ − Q₁ = 29.5 − 22.0 = 7.5. Next, calculate the upper fence: Q₃ + 1.5 × IQR = 29.5 + 1.5 × 7.5 = 29.5 + 11.25 = 40.75. Since 41.2 > 40.75, this BMI value exceeds the upper fence and is classified as a potential outlier. The margin is narrow (41.2 − 40.75 = 0.45 kg/m²), so clinical judgment is especially important here: a BMI of 41.2 places the individual in the Class III obesity range, which is clinically meaningful, and the observation should likely be retained as a genuine extreme rather than a data error.

Question 3: INTERMEDIATE

You are analyzing hospital length-of-stay (LOS) data for 200 surgical patients. The histogram shows a strong peak near 3 days with a long right tail extending to 45 days. The sample skewness is g₁ = 2.8 and excess kurtosis is g₂ = 11.4. (a) Characterize the distribution shape. (b) Explain why you should not report mean ± SD as the primary summary. (c) Suggest a transformation and explain its rationale.

Show answer

(a) The distribution is strongly right-skewed (g₁ = 2.8 >> 0) and leptokurtic (g₂ = 11.4 >> 0), indicating both pronounced asymmetry and very heavy tails. (b) In a right-skewed, heavy-tailed distribution, the mean is inflated by extreme values and the standard deviation is dominated by tail observations. Reporting mean ± SD would overstate the 'typical' LOS and create a symmetric interval that may include negative days on the lower end, which is impossible. The median and interquartile range (or the median with the range) provide a more accurate and interpretable summary of central tendency and spread. (c) A natural-log transformation (ln(LOS)) is commonly used for right-skewed, strictly positive data. The log transform compresses the right tail and stretches the left, often producing an approximately normal distribution on the log scale. This facilitates the use of parametric methods; results can then be back-transformed to geometric means and multiplicative confidence intervals for clinical interpretation.

Question 4: APPLIED

In a pharmacokinetic study, you measure peak plasma drug concentrations (Cₘₐₓ) in 50 volunteers. The box plot shows two points above the upper whisker. The study protocol states that volunteers were fasted, but chart review reveals that one of the two outlier subjects consumed a high-fat meal before dosing. Describe how you would handle each outlier and justify your approach.

Show answer

The two outliers require different handling because they have different explanatory contexts. The subject who consumed a high-fat meal violated the study protocol; the elevated Cₘₐₓ is likely a direct consequence of altered drug absorption (high-fat meals can increase bioavailability for many lipophilic drugs). This observation should be flagged as a protocol deviation. It is appropriate to exclude this data point from the primary analysis (per pre-specified exclusion criteria) but to include it in a sensitivity analysis to demonstrate its impact. The second outlier, with no identifiable protocol deviation or data-entry error, should be retained in the primary analysis because it may represent genuine pharmacokinetic variability (e.g., a CYP450 polymorphism that slows metabolism). Removing it without justification would bias the results and underestimate population variability. A transparent report would present results with and without this observation and discuss the potential biological mechanism.

Question 5: CRITICAL THINKING

A colleague argues that formally testing for normality (e.g., Shapiro-Wilk test) makes visual shape assessment via histograms and Q-Q plots unnecessary. Construct a counterargument, addressing at least two scenarios in which reliance on a formal test alone would be misleading.

Show answer

Formal normality tests suffer from two well-known limitations that make them insufficient as stand-alone tools. First, in large samples (n > 200), even trivially small departures from normality — departures that have no practical impact on the validity of a t-test or ANOVA — will produce statistically significant p-values, leading the researcher to reject normality unnecessarily and switch to less powerful nonparametric methods. This is the 'over-sensitivity' problem: the test is answering a question (is the distribution exactly normal?) that is stricter than the question the analyst actually needs answered (is the distribution approximately normal enough for parametric methods to perform adequately?). Second, in small samples (n < 20), formal tests have low power and frequently fail to reject normality even when the data are substantially non-normal — the 'under-sensitivity' problem. A histogram or Q-Q plot would immediately reveal bimodality, heavy tails, or strong skewness that a p-value of 0.12 from a Shapiro-Wilk test might mask. Visual tools also reveal the specific nature of the departure (which tail is heavy, whether there are gaps or clusters), information that a single p-value cannot convey. The best practice, therefore, is to use graphical methods to understand how the distribution departs from normality and to use formal tests, if at all, as a supplementary quantitative check — never as a replacement for looking at the data.

SUMMARY

Lesson Summary

Summary

Characterizing distribution shape involves assessing three properties: symmetry vs. skewness (whether the tails are balanced or one extends farther), kurtosis (the heaviness of the tails relative to the normal distribution), and modality (the number of distinct peaks). The relationship between the mean and median provides a quick diagnostic: when the mean exceeds the median, the distribution is right-skewed; when the mean falls below the median, it is left-skewed; and when they coincide, the distribution is approximately symmetric.

Outliers are observations that lie far from the bulk of the data and can be detected visually through box plots or numerically via the IQR fence rule (Q₁ − 1.5 × IQR, Q₃ + 1.5 × IQR) or the z-score method (|z| > 3). The IQR approach is preferred for skewed data because it is robust to the very outliers it seeks to detect. In biostatistics, the decision to retain, transform, or exclude an outlier must always be guided by domain knowledge, transparent documentation, and sensitivity analysis — never by the desire to achieve a convenient statistical result.