Slides 📊

Open the Chapter 2 slides in a new tab

The original PowerPoint slides remain available as a legacy download:

Download Chapter 2 slides (PPTX)

2.4. Exploring Quantitative Distributions: Modality, Skewness & Outliers

Histograms reveal patterns that help us interpret the overall shape of a data distribution. In this lesson, we will develop a systematic approach to reading histograms by focusing on three key characteristics: modality, skewness, and outliers.

Road Map 🧭

  • Identify patterns in histograms by examining their shape, center, and spread.

  • Distinguish between unimodal, bimodal, and multimodal distributions.

  • Recognize symmetric versus skewed distributions.

  • Learn to spot potential outliers that may require further investigation.

2.4.1. Looking for Patterns in Histograms

Histogram of the furnace dataset

Fig. 2.23 Histogram of the furnace data set (see Section 2.3)

Describe this figure in detail

Histogram titled “Residential furnace energy consumption”, drawn on a density scale with light blue bars. The horizontal axis is BTU, from about 2.5 to 19, with ticks at 5, 10 and 15; the vertical axis is Density, from 0 to about 0.14. From left to right the 11 bar heights are about 0.007, 0.015, 0.065, 0.11, 0.137, 0.131, 0.087, 0.065, 0.015, 0.015 and 0.007 on the density axis. The tallest bar spans about 8.4 to 10 BTU.

The red curve, the kernel density estimate, peaks at about 0.134 near 9.6 BTU. The blue curve, a normal density with the sample mean and standard deviation, peaks at about 0.138 near 10 BTU. Both curves run from about 3 to 18 BTU. The red curve lies above the blue one between about 5.5 and 9.7 BTU. It drops below the blue curve between about 9.7 and 13 BTU, with a slight shoulder near 12.3 BTU. The two curves nearly coincide from about 13 to 14.5 BTU, and the red curve lies just above the blue one beyond about 14.5 BTU.

Once we generate a histogram, we begin our data-reading process by noting the basic properties such as the center and spread of the data.

  • Center: Where is the bulk of the data located? Is there one central location or multiple?

  • Spread: How widely dispersed are the values around these central locations?

According to Fig. 2.23, the center of the furnace data set is located around 10, and it spans a range between 0 and 20, approximately.

After covering the basics, we move on to classify the shape of the data distribution based on a few key characteristics:

  • Modality

  • Skewness

  • Presence of any potential outliers

Let us discuss each of these characteristics in detail.

2.4.2. Modality: Counting the Peaks

Modality tells us whether our data is concentrated around a single location or multiple locations.

A mode is a local peak in a histogram or density curve. Counting them helps us understand whether the data represents one homogeneous population or multiple distinct subgroups.

Unimodal distributions have a single peak, suggesting that the data is coming from a single population centered around a unique location. Most textbook examples show unimodal distributions in the classic “bell curve” shape just like Fig. 2.23, but a unimodal distribution can take many forms—it only needs to have one primary peak.

Bimodal distributions show two peaks in their histograms. The existence of two centers might indicate that the data was sampled from two different populations with distinct characteristics.

Multimodal distributions have two or more peaks in their histograms, potentially indicating that the data was collected from several different populations.

Caution 🛑

Sometimes, apparent multimodality in a histogram can be an artifact of using too many bins rather than a genuine feature of the data. If the bin count is too large, you’ll tend to see several modes because individual data points are highlighted rather than the overall summary. This is why choosing an appropriate bin size is crucial when creating histograms.

Example 💡: Old Faithful Eruption Data

Let’s examine the eruption data of the Old Faithful geyser in Yellowstone National Park. This built-in R dataset contains measurements on both the waiting time between eruptions and the duration of eruptions themselves. Let us plot the histograms of both variables.

library(ggplot2)
data(faithful) # Load Old Faithful dataset
nbins <- round(sqrt(nrow(faithful))) + 2

# Plot eruption times
ggplot(faithful, aes(eruptions)) +
geom_histogram(aes(y = after_stat(density)), bins = nbins, fill = "dodgerblue", colour = "black", linewidth = 1) +
geom_density(colour = "red", linewidth = 1.2) +
labs(title = "Old Faithful eruption lengths – bimodal distribution",
      x = "Eruption time (minutes)", y = "Frequency")

# Plot waiting times using your own code
Histogram of length of eruption of the Old Faithful geyser

Fig. 2.24 Length of eruption

Describe this figure in detail

Histogram titled “Histogram of Old Faithful Length of Eruption”, with bright blue bars. The horizontal axis is “Eruption time in (minutes)”, from about 1.4 to 5.2, with ticks at 2, 3, 4 and 5. The vertical axis is labeled Frequency, but its values, from 0 to about 0.65 with ticks at 0.0, 0.2, 0.4 and 0.6, are densities.

The bars form two clusters separated by a stretch of very short bars. The left cluster, from about 1.6 to 2.5 minutes, has heights of about 0.45, 0.555, 0.335 and 0.185; its tallest bar spans about 1.9 to 2.1 minutes. A single very short bar, about 0.015, sits just to its left at about 1.4 to 1.6 minutes. Between about 2.5 and 3.4 minutes the four bars are all below about 0.07. The right cluster, from about 3.4 to 5.1 minutes, has heights of about 0.17, 0.12, 0.385, 0.52, 0.57, 0.655, 0.32 and 0.12; its tallest bar, the tallest overall, spans about 4.5 to 4.7 minutes.

The red curve, the kernel density estimate, has two peaks: about 0.34 near 2 minutes and about 0.48 near 4.4 minutes, with a low point of about 0.065 near 3 minutes. The blue curve, a normal curve, is a single broad hump that peaks at about 0.35 near 3.5 minutes, where the bars are short.

The histogram of lengths of eruption shows two distinct peaks—one around 2 minutes and another around 4.5 minutes. This indicates that eruptions tend to fall into two categories: shorter eruptions and longer eruptions, with relatively few eruptions of intermediate length.

Histogram of waiting duration of the Old Faithful geyser

Fig. 2.25 Waiting duration

Describe this figure in detail

Histogram titled “Histogram of Old Faithful Waiting Times between Eruptions”, with bright blue bars. The horizontal axis is Waiting Time (minutes), from about 40 to 100, with ticks at 40, 60, 80 and 100. The vertical axis is labeled Frequency, but its values, from 0 to about 0.051 with ticks every 0.01, are densities.

A single very short bar, about 0.001, sits at the far left at about 41 to 45 minutes. The bars from about 45 to 64.6 minutes form a lower, wider cluster with heights between about 0.012 and 0.023; its tallest bar, about 0.023, spans about 51 to 54.6 minutes. The shortest bar between the clusters, about 0.007, spans about 64.6 to 67.9 minutes. The heights then rise through about 0.013, 0.016 and 0.032 to the tallest bar overall, about 0.051 at about 77.8 to 81.2 minutes, followed by about 0.040. After that they fall to about 0.016 and 0.018, then to about 0.0045 and 0.001 at the right end, near 97.7 minutes.

The red curve, the kernel density estimate, has a low peak of about 0.0195 near 53.5 minutes, a low point of about 0.011 near 66 minutes, and a higher peak of about 0.0365 near 80 minutes. The blue curve, a normal curve, is a single broad hump that peaks at about 0.029 near 71 minutes.

The histogram of waiting times also reveals two modes—shorter waiting periods and longer ones, with a dip in the middle region around 60 minutes.

The bimodality in both variables reflects the physical processes within the geyser. Short eruptions tend to be followed by shorter waiting times, while long eruptions precede longer waiting intervals.

This example illustrates an important principle: when you encounter multimodality, it often signals the presence of distinct subgroups that should be analyzed separately rather than averaged together.

2.4.3. Symmetry and Skewness

We call a data symmetric if its distribution is reasonably balanced around a central value. When the data is not symmetric, we call it skewed. A data can be right (positively) skewed or left (negatively) skewed depending on the direction of the longer tail:

  • If the tail stretches toward higher values, the distribution is right (or positively) skewed.

  • If the tail stretches toward lower values, the distribution is left (or negatively) skewed.

Symmetric, left skewed, and right skewed Histograms

Fig. 2.26 Histograms with different degrees and directions of skewness

Describe this figure in detail

Three histograms side by side, each with purple bars, Counts on the vertical axis and x on the horizontal axis. The titles printed above them are “Left (negatively) skewed”, “Symmetric” and “Right (positively) skewed”.

Left (negatively) skewed: x runs from 0 to 500 and counts from 0 to about 710. The tallest bar, about 710, spans about 400 to 420. To its right the bars drop quickly, to about 520, 325, 85 and 13, ending at 500. To its left the heights decline gradually through about 640, 520, 420, 375, 285, 213, 136 and 107, then to about 50 at 200 to 240 and under 25 from 140 to 200. Bars barely above the axis sit at about 100 to 140, and a single tiny bar sits at 0 to 20, with nothing between it and those bars.

Symmetric: x runs from about -10 to 105 and counts from 0 to about 1220. The tallest bar, about 1220, spans about 49 to 53. The heights fall off at a similar rate on both sides: the neighboring bars are about 1140 on the left and 1085 on the right, and four bars out from the peak they are about 630 on the left and 560 on the right. Both ends drop below about 30, near -5 and near 100.

Right (positively) skewed: x runs from 0 to about 480 and counts from 0 to about 1330. The tallest bar, about 1330, is the first, from 0 to about 18. The heights then fall quickly, to about 945, 655, 450, 340, 228 and 173 by about 120. After that, a long thin tail of bars under about 105 stretches from about 120 to about 365, with gaps near its end.

2.4.4. Identifying Potential Outliers

When examining a histogram, we also look for potential outliers—observations that deviate significantly from the overall pattern of the data. When encountered, they should not immediately be discarded but investigated thoroughly. This is important because:

  1. They may represent errors in data collection or entry.

  2. They could be valid but unusual observations that provide valuable insights.

  3. They can significantly influence statistical measures like the mean.

  4. They may indicate violations of assumptions in statistical inference procedures.

For now, we identify potential outliers visually. In Chapter 3.4, this eyeballed judgment is replaced by a formal criterion—the 1.5 × IQR rule—and the points it flags are called explicit points.

When R creates a histogram, it automatically determines an appropriate range for the axes to ensure all data points are included in the plot. If there are values far from the majority of the data, the plotting range will be extended to include them, creating visible gaps in the histogram where no observations fall.

Histogram with potential outliers

Fig. 2.27 Histogram with potential outliers marked in red circles

Describe this figure in detail

Histogram drawn on a density scale with purple bars. The horizontal axis is x, from about 30 to 103, with ticks at 30, 50, 70 and 90; the vertical axis is density, from 0 to about 0.08. Almost all the data form a single mound of bars from about 60 to 94. The tallest bar, about 0.082, spans about 77 to 79.5. The bars on either side of it are about 0.069 and 0.071, and the heights fall below about 0.004 under about 67 on the left and above about 92 on the right.

Three bars far from the mound are so short they are barely visible on the axis, and each is marked by a translucent red circle. They are at about 33 to 35.5, about 47.5 to 50, and about 99 to 101.5. Between these bars and the mound the axis is empty, which stretches the horizontal axis well beyond the main body of the data.

The red curve, the kernel density estimate, and the blue curve, the normal curve, both follow the mound. The red curve peaks at about 0.080 and the blue curve at about 0.074, both near 78. Across the empty stretches both curves lie along zero.

2.4.5. Bringing It All Together

Key Takeaways 📝

  1. Histograms reveal the shape, center, spread, and potential outliers in a quantitative data.

  2. Potential outliers require careful investigation—they may represent errors or important observations.

  3. Extreme values influence the sample mean more strongly than the sample median. In many skewed distributions, the mean lies farther toward the longer tail than the median does. This is a common pattern, not a universal rule. The ordering of mean and median does not determine skewness alone.

2.4.6. Exercises

These exercises develop your skills in describing the shape of distributions by identifying modality, skewness, and potential outliers.

Exercise 1: Identifying Modality

For each histogram description below, identify the modality (unimodal, bimodal, or multimodal) and suggest a possible explanation for the pattern observed.

  1. A histogram of response times for a web server shows a single peak around 50 ms with values ranging from 20 ms to 150 ms, tapering off gradually on both sides.

  2. A histogram of employee commute times shows two distinct peaks—one around 15 minutes and another around 45 minutes—with relatively few values in between.

  3. A histogram of customer ages at a family restaurant shows three peaks: one around age 8, another around age 35, and a third around age 65.

  4. A histogram of exam scores shows a single peak around 75 with a fairly symmetric spread from 50 to 100.

  5. A quality engineer measures the diameter of ball bearings from two production lines (mixed together). The histogram shows two peaks: one at 10.0 mm and another at 10.5 mm. What does this suggest, and what should the engineer do?

Solution

Part (a): Web Server Response Times

Unimodal — Single peak around 50 ms.

Explanation: The response times likely come from a single, consistent server process. The tapering on both sides suggests natural variation around a typical response time, which is expected behavior for a well-functioning system.

Part (b): Employee Commute Times

Bimodal — Two distinct peaks at 15 and 45 minutes.

Explanation: This likely reflects two distinct groups of employees:

  • Those who live close to work (peak at 15 min)

  • Those who commute from farther away, possibly suburbs (peak at 45 min)

This could also reflect different transportation modes (driving vs. public transit).

Part (c): Customer Ages at Restaurant

Multimodal (trimodal) — Three peaks at ages 8, 35, and 65.

Explanation: Family restaurants attract distinct demographic groups:

  • Children (peak ~8 years)

  • Parents (peak ~35 years)

  • Grandparents (peak ~65 years)

The multimodality reflects the family-oriented customer base.

Part (d): Exam Scores

Unimodal — Single peak around 75.

Explanation: Students likely come from a single population with similar preparation levels. The symmetric spread suggests the exam difficulty was well-calibrated for the class.

Part (e): Ball Bearing Diameters

Bimodal — Two peaks at 10.0 mm and 10.5 mm.

This strongly suggests the two production lines are producing bearings with different mean diameters. This is a quality control concern.

The engineer should:

  1. Separate the data by production line and create individual histograms

  2. Investigate why the lines are producing different diameters

  3. Determine which (if either) meets specifications

  4. Calibrate the production lines to produce consistent output

Analyzing the mixed data together would be misleading—the overall mean would be between the two peaks, representing neither line accurately.


Exercise 2: Identifying Skewness

The histograms below show distributions for different engineering and business scenarios. For each, identify whether the distribution is symmetric, right-skewed (positively skewed), or left-skewed (negatively skewed). Explain your reasoning.

Four histograms showing different skewness patterns

Fig. 2.28 Four distributions with different skewness patterns

Describe this figure in detail

Four histograms in a 2-by-2 grid, each drawn on a density scale with a red curve, the kernel density estimate. There is no normal curve.

A: Bug Resolution Time (hours), blue bars, top left. The horizontal axis is Hours, from 0 to 20; density runs from 0 to 0.15. The tallest bar, about 0.145, is the first, from about 0.6 to 1.9 hours. From there the heights fall: about 0.126, 0.122, 0.098, 0.074, 0.051 and 0.055 out to about 9.6 hours, then about 0.020, 0.027, 0.031 and 0.016 out to about 14.7 hours, and finally four bars under 0.01 continuing to about 19.8 hours. The red curve starts at about 0.085 at 1 hour, peaks at about 0.127 near 2.6 hours, and declines to about 0.004 at 19 hours, with a slight shoulder near 12 hours.

B: Standardized Test Scores, orange bars, top right. The horizontal axis is Score, from about 36 to 100; density runs from 0 to about 0.095. The tallest bar, about 0.094, is the last, from about 96 to 100. Moving left, the heights drop to about 0.053, 0.028 and 0.034, then to bars under 0.01 between about 58 and 83.5. There are no bars between about 41 and 58, and one very short bar, about 0.001, sits alone at about 36 to 41. The red curve stays near zero from 40 to about 60, rises steeply after 80, peaks at about 0.077 near 97, and turns down to about 0.056 at 100.

C: Product Lifespan (months), green bars, bottom left. The horizontal axis is Months, from about -3 to 76; density runs from 0 to about 0.032. The bars from about 3 to 29 months are all between about 0.024 and 0.0315; the tallest, about 0.0315, spans about 8 to 13 months, and the next tallest, about 0.0296, spans about 18 to 24 months. To the left there is only one bar, about 0.0105, from about -3 to 3 months. To the right the heights fall through about 0.018, 0.0105, 0.0057, 0.0038, 0.0028 and 0.0028, ending at about 60 months. Then two bins are empty, and a single bar of about 0.001 sits at about 71 to 76 months. The red curve starts at about 0.0125 at 0 months, has two low bumps (about 0.028 near 10 months and about 0.0265 near 22 months), and falls to near zero by about 65 months.

D: Precision Measurements, purple bars, bottom right. The horizontal axis is Value, from about 38 to 64; density runs from 0 to about 0.092. The tallest bar, about 0.091, spans about 48.5 to 50.2, with bars of about 0.082 and 0.085 on either side. Further left the heights are about 0.044, 0.050 and 0.0205, then one bin at about 40 to 41.7 is empty, and a short bar of about 0.006 sits at about 38.3 to 40. Further right they are about 0.050, 0.0645, 0.038, 0.0235, 0.0175, 0.0115 and 0.003, ending at about 63.8. The red curve peaks at about 0.085 near 49.5, with slight shoulders near 44 and near 54.

  1. Histogram A: Software bug resolution time (hours)

  2. Histogram B: Standardized test scores

  3. Histogram C: Product lifespan until failure (months)

  4. Histogram D: Precision measurements from a calibrated instrument

  5. For each skewed distribution above, explain why that type of skewness makes sense given the context.

Solution

Part (a): Bug Resolution Time — Right-skewed (positively skewed)

The tail extends toward higher values (longer resolution times).

Part (b): Test Scores — Left-skewed (negatively skewed)

The tail extends toward lower values (lower scores), with most scores clustered at the high end.

Part (c): Product Lifespan — Right-skewed (positively skewed)

The tail extends toward higher values (longer lifespans).

Part (d): Precision Measurements — Symmetric

The distribution is balanced around the center with similar spread on both sides.

Part (e): Why the Skewness Makes Sense

Bug resolution time (right-skewed): Most bugs are resolved quickly (simple fixes), but some complex bugs take much longer. There’s a natural lower bound (can’t be negative) but no strict upper bound, creating a right tail.

Test scores (left-skewed): If the test is relatively easy or students are well-prepared, most scores cluster near the top. The upper bound (100% or maximum score) creates a ceiling effect, while a few students score much lower, creating a left tail.

Product lifespan (right-skewed): Most products fail within a typical range, but some last much longer than expected. There’s a lower bound (products can’t fail before time 0) and no upper bound on how long something might last, creating a right tail.

Precision measurements (symmetric): A well-calibrated instrument produces measurements that vary randomly around the true value, with equal probability of being slightly above or below. This creates the classic symmetric, bell-shaped pattern.


Exercise 3: Identifying Potential Outliers

A manufacturing engineer collects data on the tensile strength (in MPa) of 50 steel samples. The histogram and data summary are shown below.

Histogram of tensile strength with potential outliers

Fig. 2.29 Tensile strength distribution with potential outliers

Describe this figure in detail

Histogram titled “Tensile Strength of Steel Samples”, drawn on a density scale with gray bars. The horizontal axis is Tensile Strength (MPa), from about 380 to 550, with ticks at 400, 450, 500 and 550; the vertical axis is Density, from 0 to about 0.0225. The bars are about 13.5 MPa wide.

From left to right: a short bar of about 0.0015 at about 382 to 395 MPa; an empty bin; bars of about 0.003 and 0.0015 at about 409 to 436 MPa; another empty bin at about 436 to 449 MPa; then the main group from about 449 to 543 MPa, with heights of about 0.0015, 0.012, 0.009, 0.0224, 0.015, 0.006 and 0.003. The tallest bar spans about 489 to 503 MPa.

The red curve, the kernel density estimate, has a small bump of about 0.002 near 418 MPa, dips to about 0.0008 near 440 MPa, and rises to a peak of about 0.0172 near 500 MPa. The blue curve, the normal curve, has a single peak of about 0.0135 near 489 MPa. The red curve lies above the blue one below about 427 MPa. The blue curve lies above the red one between about 427 and 487 MPa, below it between about 487 and 526 MPa, and above it again beyond about 526 MPa.

Summary Statistics:

  • Mean: 485 MPa

  • Median: 492 MPa

  • Standard Deviation: 28 MPa

  • Minimum: 387 MPa

  • Maximum: 538 MPa

  • Q1: 472 MPa

  • Q3: 505 MPa

  1. Based on the histogram, are there any potential outliers? If so, approximately where are they located?

  2. The mean (485 MPa) is less than the median (492 MPa). What does this suggest about the shape of the distribution?

  3. Using the 1.5 × IQR rule (previewed here; taught fully in Chapter 3.4), calculate the lower and upper fences. Are there any explicit points—values beyond the fences—according to this rule?

  4. One of the low values (387 MPa) came from a sample that was accidentally heated to a higher temperature during processing. Should this value be removed from the analysis? Explain your reasoning.

  5. Another low value (412 MPa) has no known processing issues. How should the engineer treat this value?

Solution :class-container: sd-border-success

Part (a): Potential Outliers from Histogram

Yes, there appear to be potential outliers on the left side (low end) of the distribution. The histogram shows a gap followed by a few isolated observations around 380-420 MPa, separated from the main body of data centered around 490 MPa.

Part (b): Mean vs. Median

The mean (485 MPa) being below the median (492 MPa) is consistent with left-skewness, but this comparison alone does not establish the distribution’s shape.

Here, the histogram also shows a left tail with a few unusually low values. These observations influence the mean more strongly than the median, helping explain why the mean is lower.

Comparing the mean and median is a useful rule of thumb. Examine the histogram and the overall pattern before concluding which direction a distribution is skewed.

Part (c): 1.5 × IQR Rule

First, calculate the IQR:

IQR = Q3 − Q1 = 505 − 472 = 33 MPa

Calculate the fences:

  • Lower fence = Q1 − 1.5 × IQR = 472 − 1.5(33) = 472 − 49.5 = 422.5 MPa

  • Upper fence = Q3 + 1.5 × IQR = 505 + 1.5(33) = 505 + 49.5 = 554.5 MPa

Checking the extremes:

  • Minimum (387 MPa) < 422.5 → Explicit point (below lower fence)

  • Maximum (538 MPa) < 554.5 → Not an explicit point (within upper fence)

According to the 1.5 × IQR rule, at least one explicit point exists on the low end (387 MPa, and possibly others below 422.5 MPa).

Part (d): Value with Known Processing Error

Yes, this value should likely be removed (or analyzed separately).

Reasoning:

  • The value has a known assignable cause (improper heat treatment)

  • It does not represent the normal manufacturing process

  • Including it would bias the analysis of the standard process

  • It should be documented and investigated, but excluded from the main analysis

This is an example of an outlier caused by a measurement or process error.

Part (e): Value with No Known Issues

This value should NOT be automatically removed.

Reasoning:

  • No assignable cause has been identified

  • It may represent natural variation in the manufacturing process

  • It could indicate a real issue that needs investigation

  • Removing data without justification can bias results

The engineer should:

  1. Investigate whether there were any unusual circumstances

  2. Consider whether this represents a real subpopulation of weaker samples

  3. Report results both with and without this observation if doing sensitivity analysis

  4. Never remove data simply because it’s inconvenient


Exercise 4: Comprehensive Shape Description

For each histogram below, provide a complete description of the distribution including:

  • Center: Approximate location of the typical value

  • Spread: Approximate range of the data

  • Shape: Modality and skewness

  • Outliers: Any potential outliers

Three histograms for comprehensive description

Fig. 2.30 Three distributions requiring complete description

Describe this figure in detail

Three histograms side by side, each drawn on a density scale with a red curve, the kernel density estimate.

A: CPU Utilization (%), red-orange bars, left. The horizontal axis is Utilization (%), from about 45 to 105, with ticks at 60, 80 and 100; density runs from 0 to about 0.052. Bars run from about 57 to 101. Their heights rise from about 0.0024 in the first bar through about 0.0086, 0.0098, 0.023, 0.031 and 0.0355 to the tallest, about 0.051, at about 90 to 95.5, and then drop to about 0.022 in the last bar, which ends at about 101. The red curve runs along zero from 45 to about 55, rises to a peak of about 0.044 near 92, and falls steeply to about 0.001 at 105.

B: Time Between Arrivals (min), blue bars, middle. The horizontal axis is Time (minutes), from about -0.6 to 14, with ticks at 0, 5 and 10; density runs from 0 to about 0.235. The first bar, about 0.188, spans about -0.6 to 0.6 minutes, and the tallest, about 0.233, spans about 0.6 to 1.8 minutes. To the right the heights fall to about 0.16, 0.072, 0.083, 0.044, 0.011 and 0.0055 by about 9 minutes, then rise to about 0.0275 at about 9 to 10.3 minutes. Two empty bins, from about 10.3 to 12.7 minutes, separate a final short bar of about 0.0055 at about 12.7 to 13.9 minutes. The red curve starts at about 0.155 at 0, peaks at about 0.233 near 1.1 minutes, levels off at about 0.073 near 4.6 minutes, has a small bump of about 0.019 near 9.4 minutes, and is near zero at about 12 minutes.

C: Battery Capacity (mAh), green bars, right. The horizontal axis is Capacity (mAh), from about 2850 to 3150, with ticks at 2900, 3000 and 3100; density runs from 0 to about 0.0095. The tallest bar, about 0.0094, spans about 2975 to 3000 mAh, next to a bar of about 0.0087 on its left. Further left the heights fall to about 0.0032, 0.0016, 0.0005 and 0.0005, ending at about 2850 mAh. To the right of the tallest bar the next three bars are about 0.0046, 0.0049 and 0.0043, followed by about 0.0022. Then one bin, at about 3100 to 3123 mAh, is empty, and a final bar of about 0.0005 sits at about 3123 to 3148 mAh. The red curve peaks at about 0.0087 near 2980 mAh and has a shoulder of about 0.0048 near 3036 mAh.

  1. Distribution A: CPU utilization (%) during a stress test

  2. Distribution B: Time between customer arrivals at a service desk (minutes)

  3. Distribution C: Battery capacity measurements (mAh) for smartphone batteries

Solution

Distribution A: CPU Utilization

  • Center: Approximately 85-90%

  • Spread: Ranges from about 50% to 100%

  • Shape: Unimodal, left-skewed (negatively skewed). The distribution is bunched up against the upper limit (100%) with a tail extending toward lower values.

  • Outliers: No clear outliers; the lower values appear to be part of the natural left tail.

Interpretation: During a stress test, CPU utilization is expected to be high most of the time. The 100% ceiling creates a natural boundary, and occasional dips in utilization create the left tail. This is typical behavior for stress test data.

Distribution B: Time Between Arrivals

  • Center: Approximately 2-3 minutes (but note: for right-skewed data, the median is a better measure than the visual “center”)

  • Spread: Ranges from near 0 to about 15 minutes

  • Shape: Unimodal, right-skewed (positively skewed). Most inter-arrival times are short, with fewer long gaps.

  • Outliers: No clear outliers; the long times are part of the expected right tail.

Interpretation: This pattern is typical of arrival processes. Most customers arrive in quick succession, but occasionally there are longer gaps. The distribution resembles an exponential distribution, which is commonly used to model waiting times.

Distribution C: Battery Capacity

  • Center: Approximately 3000 mAh

  • Spread: Ranges from about 2850 to 3150 mAh

  • Shape: Unimodal, approximately symmetric. The distribution is roughly bell-shaped with balanced tails.

  • Outliers: No obvious outliers visible.

Interpretation: This suggests a well-controlled manufacturing process. Battery capacities cluster tightly around the target value with natural, symmetric variation. This distribution appears approximately normal, which is desirable for quality control purposes.


Exercise 5: Modality Investigation

A data analyst notices that a histogram of employee salaries at a tech company appears bimodal. Before concluding that there are two distinct salary groups, they should investigate further.

Bimodal salary distribution

Fig. 2.31 Employee salary distribution showing apparent bimodality

Describe this figure in detail

Histogram titled “Employee Salary Distribution”, drawn on a density scale with gray bars and a red curve, the kernel density estimate. The horizontal axis is Salary ($1000s), from about 38 to 164, with ticks at 40, 80, 120 and 160; the vertical axis is Density, from 0 to about 0.0155.

At the far left, a short bar of about 0.0008 spans about 38 to 46.5, followed by one empty bin. From about 55 the heights rise through about 0.0034 and 0.0051 to about 0.0153, the tallest bar overall, at about 72 to 80. They then fall through about 0.0127, 0.0085 and 0.0051 to about 0.0025 at about 105 to 113.5. From about 113.5 to 164 the bars are tall again: three of about 0.011, then about 0.0136 at about 139 to 147, then about 0.011 and 0.0076.

The red curve has two humps. It peaks at about 0.0105 near 81, drops to a low point of about 0.0063 near 106, rises to a second peak of about 0.0119 near 138, and ends at about 0.0047 at 164.

  1. List three possible explanations for why the salary distribution might be bimodal.

  2. The analyst creates faceted histograms by job category (Engineering vs. Sales). What would they expect to see if job category explains the bimodality?

  3. If the bimodality disappears when the data is separated by job category, what conclusion can be drawn about analyzing the combined data?

  4. If the bimodality persists even within each job category, what other variables might the analyst investigate?

  5. Why would it be misleading to report the “average salary” for this company without acknowledging the bimodal distribution?

Solution

Part (a): Possible Explanations for Bimodality

Three possible explanations:

  1. Job category/role: Different job families (e.g., engineering vs. sales vs. administrative) may have different salary structures

  2. Experience level: Entry-level vs. senior employees may form two distinct groups

  3. Location: Employees in different geographic regions (e.g., San Francisco vs. remote/other cities) may have different salary bands due to cost-of-living adjustments

Other valid explanations: full-time vs. part-time, different departments, acquired company with legacy salary structure, hourly vs. salaried employees.

Part (b): Expected Result if Job Category Explains Bimodality

If job category explains the bimodality, the faceted histograms would show:

  • Engineering: A single peak (unimodal) around the higher salary value (~$140k)

  • Sales: A single peak (unimodal) around the lower salary value (~$80k)

Each category would have its own unimodal distribution, and combining them creates the apparent bimodality.

Salary distribution faceted by job category

Fig. 2.32 Faceted histograms reveal the source of bimodality

Describe this figure in detail

Two histograms side by side under the title “Salary Distribution by Job Category”, in panels labeled “Engineering” (left) and “Sales” (right). They share the horizontal axis Salary ($1000s), from about 40 to 180 with ticks at 40, 80, 120 and 160, and use bins about 13 wide. Each panel has its own Density axis and a red curve, the kernel density estimate.

Engineering: density runs from 0 to 0.025. Bars occupy about 98 to 176, with heights of about 0.0019, 0.0144, 0.021, 0.0239, 0.0134 and 0.0019; the tallest spans about 137 to 150. The red curve runs along zero from about 46 to 98, rises with a slight shoulder near 118 to a peak of about 0.0237 near 140, and ends at about 0.008 at 164.

Sales: density runs from 0 to about 0.031. Bars occupy about 46 to 111, with heights of about 0.0038, 0.0127, 0.0306, 0.0229 and 0.0063; the tallest spans about 72 to 85. The red curve peaks at about 0.029 near 76, has a shoulder of about 0.0265 near 85, reaches zero by about 105, and runs along zero to 164.

The Engineering bars start at about 98, before the Sales bars end at about 111, so the two groups overlap only between about 98 and 111.

Part (c): Conclusion About Combined Analysis

If bimodality disappears when separated by job category, then:

  • The combined data represents a mixture of two distinct populations

  • Analyzing the combined data obscures important differences between groups

  • Summary statistics (mean, median, standard deviation) for the combined data are misleading

  • Separate analyses should be conducted for each job category

  • Any comparisons or decisions should account for job category

Part (d): Other Variables to Investigate

If bimodality persists within job categories:

  • Years of experience or tenure

  • Education level (Bachelor’s vs. Master’s/PhD)

  • Performance ratings or job level/title within category

  • Geographic location or office

  • Gender (to check for pay equity issues)

  • Full-time vs. part-time status

  • Acquisition history (legacy employees vs. new hires)

Part (e): Why Average Salary Would Be Misleading

Reporting only the average salary would be misleading because:

  1. No one earns the average: The mean falls in the “valley” between the two peaks where few employees actually earn that amount

  2. Hides important structure: It obscures the fact that there are two distinct salary groups

  3. Poor representation: Neither the high-salary group nor the low-salary group is well-represented by the average

  4. Incorrect expectations: Job candidates might expect salaries near the mean when they should expect salaries near one of the peaks depending on their role

A more honest summary would report the median (or mean) salary separately for each major job category, or at minimum acknowledge the bimodal nature of the distribution.


Exercise 6: Shape and Statistical Measures

Understanding distribution shape helps us choose appropriate summary statistics. For each scenario, answer the questions about the relationship between shape and statistics.

  1. A distribution is strongly right-skewed. Which is larger: the mean or the median? Why?

  2. For a symmetric, unimodal distribution, what is the approximate relationship between the mean and median?

  3. A quality engineer has data on product weights that is left-skewed with a few unusually light items. Which measure of center (mean or median) better represents the “typical” product weight? Why?

  4. Two datasets have the same mean and standard deviation, but one is symmetric and one is right-skewed. Would you expect them to look similar? Explain.

  5. A dataset has these summary statistics: Mean = 50, Median = 50, Mode = 50. What can you conclude about the shape of the distribution?

  6. Another dataset has: Mean = 45, Median = 50, Mode = 55. What shape does this suggest?

Solution

Part (a): Right-Skewed Distribution

The mean is larger than the median.

Why: In a right-skewed distribution, the tail extends toward high values. These extreme high values pull the mean upward, while the median (the middle value) is resistant to extremes. The mean “chases” the tail.

Part (b): Symmetric Unimodal Distribution

The mean and median are approximately equal.

In a perfectly symmetric distribution, the mean equals the median exactly. In real data that is approximately symmetric, they will be very close to each other.

Part (c): Left-Skewed Product Weights

The median better represents the typical weight.

Why: The few unusually light items pull the mean downward, making it lower than what most products actually weigh. The median is resistant to these extreme values and better captures the center of the main body of data.

Rule of thumb: For skewed distributions, the median is usually the better measure of “typical.”

Part (d): Same Mean and SD, Different Shapes

No, they would NOT look similar.

The mean and standard deviation alone don’t capture the shape of a distribution. Two distributions can have identical means and standard deviations but very different appearances:

  • The symmetric one would be balanced around the mean

  • The right-skewed one would have most values below the mean with a long right tail

This illustrates why visualizing data with histograms is essential—summary statistics alone can be misleading.

Part (e): Mean = Median = Mode = 50

This suggests the distribution is symmetric and unimodal.

When all three measures of center coincide, the distribution is balanced around a single central value. This is characteristic of symmetric distributions like the normal distribution.

Note: This doesn’t guarantee perfect symmetry, but it’s a strong indication.

Part (f): Mean = 45, Median = 50, Mode = 55

This suggests a left-skewed (negatively skewed) distribution.

The ordering Mean < Median < Mode indicates:

  • The mean is pulled left by low values in the left tail

  • The median is in the middle

  • The mode (peak) is on the right side

This pattern is characteristic of left-skewed distributions where the tail extends toward lower values.


2.4.7. Additional Practice Problems

True/False Questions (1 point each)

  1. A bimodal distribution always indicates that data was collected from two different populations.

    Ⓣ or Ⓕ

  2. In a right-skewed distribution, the mean is typically greater than the median.

    Ⓣ or Ⓕ

  3. Any observation that falls outside the 1.5 × IQR fences should automatically be removed from the dataset.

    Ⓣ or Ⓕ

  4. A histogram that appears multimodal might actually be unimodal if too many bins were used.

    Ⓣ or Ⓕ

  5. The presence of outliers always indicates errors in data collection.

    Ⓣ or Ⓕ

  6. For a perfectly symmetric distribution, the mean equals the median.

    Ⓣ or Ⓕ

Multiple Choice Questions (2 points each)

  1. A histogram of home prices in a city shows a long tail extending toward high values, with most homes clustered at lower prices. This distribution is:

    Ⓐ Left-skewed (negatively skewed)

    Ⓑ Right-skewed (positively skewed)

    Ⓒ Symmetric

    Ⓓ Bimodal

  2. An engineer analyzes failure times for electronic components and finds the distribution is right-skewed. Which measure of center would BEST represent the typical failure time?

    Ⓐ Mean

    Ⓑ Median

    Ⓒ Mode

    Ⓓ Standard deviation

  3. A dataset has Mean = 72, Median = 68, and a single mode. What is the most likely shape of this distribution?

    Ⓐ Symmetric

    Ⓑ Left-skewed

    Ⓒ Right-skewed

    Ⓓ Bimodal

  4. Which of the following is the BEST reason to investigate (but not automatically remove) an outlier?

    Ⓐ It makes the standard deviation larger

    Ⓑ It may represent a valid but unusual observation that provides insight

    Ⓒ It changes the median

    Ⓓ It makes the histogram look asymmetric

Answers to Practice Problems

True/False Answers:

  1. False — Bimodality can have many causes: two populations, artifact of bin selection, natural phenomena with two states, etc. It doesn’t always mean two populations.

  2. True — In right-skewed distributions, the tail of high values pulls the mean upward while the median remains resistant. Thus mean > median.

  3. False — Points beyond the fences are explicit points (Chapter 3.4) and should be investigated, not automatically removed. They may represent valid observations, important anomalies, or genuine features of the data; only those that fail investigation are treated as real outliers. Removal requires justification (e.g., known data entry error).

  4. True — Using too many bins can create artificial peaks due to sampling variability, making a unimodal distribution appear multimodal. This is why appropriate bin selection is important.

  5. False — Outliers can represent errors, but they can also be valid unusual observations, important discoveries, or natural extreme values. Their presence doesn’t automatically indicate errors.

  6. True — By definition, a perfectly symmetric distribution has its mean and median at the same location (the center of symmetry).

Multiple Choice Answers:

  1. Ⓑ — A long tail toward high values with most data clustered at lower values describes right-skewed (positively skewed). This is typical for home prices, incomes, and other economic data.

  2. Ⓑ — For skewed distributions, the median is the better measure of center because it is resistant to extreme values. The mean would be pulled toward the long right tail and overestimate the typical failure time.

  3. Ⓒ — When Mean > Median, the distribution is right-skewed. The high values in the right tail pull the mean upward while the median remains closer to the bulk of the data.

  4. Ⓑ — Outliers should be investigated because they may represent valid but unusual observations that provide insight. They might reveal equipment problems, special cases, or important phenomena that deserve attention rather than deletion.