Data Distribution

1 Data Distribution

When we collect data — whether it is exam scores, household incomes, the heights of plants, or the number of daily website visits — those values do not simply exist as a flat list of numbers. They form a pattern. Some values appear frequently, others rarely. Some datasets cluster tightly around a central point, while others are scattered broadly across a wide range. This pattern of how data values are spread across their possible range is called a data distribution.

Understanding a distribution means asking: where do most of the values sit? How far do they spread? Is the pattern balanced, or does it lean more to one side? These questions are not merely aesthetic — they are fundamental to choosing the right statistical tools, drawing valid conclusions, and communicating findings accurately. Before calculating any average or measure of spread, a statistician's first task is almost always to examine the shape of the distribution.

Visualising a distribution is the most natural starting point. Three common tools achieve this:

• Histograms group data into intervals (bins) and display the frequency of values within each bin as a bar. The overall silhouette of the bars reveals the shape of the distribution at a glance.
• Frequency tables list each value or group of values alongside how often it occurs. While less visual, they are precise and useful for smaller datasets or exact counts.
• Density plots smooth the histogram into a continuous curve, making it easier to perceive the overall shape without the distraction of bin choices.

A critical insight is that two datasets can share the same mean yet have completely different distributions. Consider two classes of students, both with a mean exam score of 65 out of 100. In Class A, nearly every student scored between 60 and 70 — a tight cluster. In Class B, half the students scored below 40 and half scored above 85 — a wide spread with almost nobody near the mean. Reporting only the mean hides this crucial difference. The shape and spread of the distribution provide the essential context that averages alone cannot.

With that foundation in place, we can explore the specific types of distributions that arise most frequently in data analysis.

A symmetrical distribution is one where the left and right halves are mirror images of each other when folded at the centre. The most famous example is the normal distribution (also called the bell curve), which appears in many natural phenomena: human heights, measurement errors, standardised test scores. In a perfectly symmetrical distribution, three measures of central tendency coincide at a single point:

• The mean (the arithmetic average) sits at the centre.
• The median (the middle value when data is ordered) sits at the centre.
• The mode (the most frequently occurring value) sits at the centre.

Because all three measures agree, any of them can fairly represent a "typical" value. Symmetry also tells us that the chances of a randomly selected data point falling a given distance above the centre are equal to the chances of it falling the same distance below — the distribution is balanced. Recognising symmetry is practically important: many powerful statistical techniques (such as parametric tests like the t-test) assume or work best with roughly symmetrical data, so confirming or ruling out symmetry is an early diagnostic step.

Real-world data, however, is often not perfectly symmetrical. It may have a longer "tail" — a stretch of values — extending further to one side than the other. This property is called skewness. Skewness measures the degree and direction of asymmetry in a distribution. There are three possibilities:

• Positive skew (right-skewed): the tail extends to the right.
• Negative skew (left-skewed): the tail extends to the left.
• Zero skew: the distribution is symmetric.

A key practical consequence of skewness is its effect on the relationship between the mean and the median. Because the mean is computed using every value, it is sensitive to extreme values at the end of a tail and is therefore pulled in the direction of the tail. The median, which depends only on the rank ordering of values, resists this pull. This means that when skewness is present, the mean and median diverge, and that divergence is itself a signal of skewness. Identifying skewness therefore directly informs which measure of central tendency — mean or median — is more honest and representative for a given dataset.

In a positively skewed (right-skewed) distribution, the bulk of data values cluster at the lower end of the scale, with a long tail stretching out to the right toward higher values. The classic real-world example is household income. Most households earn a moderate income, but a small number of extremely high earners pull the tail far to the right. Other examples include:

• The number of times a webpage is shared (most posts are shared rarely; a few go viral).
• Property prices in a city (many affordable homes, a few ultra-premium properties).
• Insurance claim amounts (most claims are small; rare catastrophic events produce enormous claims).

In a right-skewed distribution the relationship between the three measures of centre follows a consistent pattern:

Because the mean is inflated by the high-value tail, it overstates what a "typical" observation looks like. If a government report states that the mean household income is £45,000 but the median is £28,000, the median is the more honest figure for describing the experience of a typical household. Analysts working with right-skewed data often apply transformations (such as taking the logarithm of values) to reduce the skew before further statistical analysis.

A negatively skewed (left-skewed) distribution is the mirror image: most data values cluster at the higher end of the scale, with a long tail stretching to the left toward lower values. Real-world examples include:

• Age at retirement: the vast majority of people retire in a relatively narrow window around 60–70 years old, but some retire much earlier (illness, early pension), creating a tail to the left.
• Exam scores on an easy test: most students score very highly, but a small number perform unusually poorly.
• Time until machine failure when machines are regularly maintained: most last a long time; a few fail unusually early.

The ordering of measures in a left-skewed distribution is the reverse of the right-skewed case:

Again, the median provides a more representative description of the typical data point, since the mean is deflated by the outlying low values in the tail. The general principle holds in both directions of skew: the median is more robust to skewness than the mean.

Beyond shape and skewness, the spread of a distribution is equally important for a full characterisation. Two distributions can be the same shape — both symmetric bell curves, for instance — and yet one might span a narrow range of values while the other spans a wide range. The spread describes how much variability exists in the data:

• A narrow distribution (low spread) means most data points are bunched close to the centre. Measurements are consistent and predictable. For example, the heights of students in a single school year group vary relatively little.
• A wide distribution (high spread) means data points are scattered far from the centre. Values are highly variable. For example, annual salaries across an entire economy span an enormous range.

The primary numerical measures used to quantify spread are:

• Range: the difference between the maximum and minimum values. Simple to calculate but highly sensitive to outliers.
• Variance: the average of the squared differences from the mean. Gives a mathematical measure of overall spread, though the squaring makes the units less intuitive.
• Standard deviation: the square root of the variance, expressed in the same units as the original data. The most widely used measure of spread, it describes a typical distance from the mean.
• Interquartile range (IQR): the range of the middle 50% of data (from the 25th to the 75th percentile). Robust to outliers, making it preferable when data is skewed.

Shape and spread must be described together for a distribution to be fully understood. Saying "the mean is 50" is incomplete. Saying "the distribution is approximately symmetric with a mean of 50 and a standard deviation of 5" is far more informative — it tells us that most values fall in a predictable range around 50, that no tail is distorting the picture, and that we can anticipate roughly how extreme any given observation might be.

Bringing all of this together, here is a practical step-by-step approach to recognising and describing a distribution:

• Step 1 — Visualise. Plot a histogram or density plot. This immediately reveals the overall silhouette: roughly bell-shaped, skewed, flat (uniform), bimodal (two peaks), or something else entirely. For smaller datasets, a sorted dot plot or stem-and-leaf plot can serve the same purpose.
• Step 2 — Assess symmetry or skew. Does the histogram look roughly balanced around a central peak, or does one tail extend significantly further than the other? If a tail stretches to the right, suspect positive skew; if to the left, suspect negative skew.
• Step 3 — Compare the mean and median. Calculate both. If the mean is notably higher than the median, the data is likely right-skewed. If the mean is notably lower than the median, the data is likely left-skewed. If they are close, the distribution is probably roughly symmetric. This is a quick numerical diagnostic that complements the visual inspection.
• Step 4 — Describe the spread. Calculate the standard deviation (or IQR for skewed data) and the range. Note whether the spread is narrow or wide relative to the scale of the data.
• Step 5 — Communicate clearly. Summarise the distribution in plain language. For example: "The distribution of claim amounts is strongly right-skewed: most claims fall below £500, but a small number of very large claims stretch the distribution to over £50,000. The median claim of £210 is a more representative typical value than the mean of £1,340, which is inflated by these outliers." This kind of description is accessible to both technical colleagues and non-specialist stakeholders.

To illustrate these steps concretely, consider the following small dataset of daily call durations (in minutes) recorded at a customer service centre:

2, 3, 3, 4, 4, 4, 5, 5, 6, 7, 8, 12, 18, 25

A histogram of these values would show a cluster of short calls on the left and a long tail of a few very long calls stretching to the right — a clear right-skewed shape. Computing the measures:

The mean (7.57 minutes) is substantially higher than the median (5.0 minutes) and mode (4 minutes), confirming the right skew observed in the histogram. The three outlying long calls (12, 18, and 25 minutes) have dragged the mean upward, making it an unrepresentative summary of a typical call. A manager reporting that the "average" call lasts 7.57 minutes would give a misleading picture; reporting the median of 5 minutes would be far more informative about what a typical agent experiences. The large standard deviation (≈ 6.2 minutes) relative to the median (5 minutes) also signals the high variability introduced by those extreme values.

Developing the habit of examining distributions — rather than jumping straight to calculating means — is one of the most valuable skills in statistics and data analysis. It prevents misleading conclusions, guides the choice of appropriate methods, and ensures that summaries genuinely reflect the data rather than obscure it.