Descriptive statistics summarise what is in a dataset: its categories, typical values, spread, distribution, unusual observations and patterns over time or across groups. The appropriate summary depends on the variable type, the shape of the data, missingness and the question being asked. A mean is useful for some numerical distributions, but can be distorted by extreme values; a median can better represent a typical value in a skewed distribution; a percentage needs a clear denominator; and a chart should be chosen to show the data structure rather than to decorate a report. Strong interpretation combines a table or display with a sentence about what it shows, for whom, over what period and with what limitation (NIST/SEMATECH, 2025; USC Libraries, 2026).

Why descriptive statistics come first

Descriptive analysis is not a preliminary chore to be rushed through on the way to a p value. It is the stage where a researcher learns what the data actually contain. It reveals coding errors, missingness, outliers, group sizes, skewed distributions and patterns that affect every later decision. A hypothesis test or model cannot make a poorly understood dataset credible.

Begin with the research question and variable definition.

  • If the question is “How did monthly passenger journeys change?”, a time plot and percentage change may be more useful than a single average.
  • If the question compares two groups, show the number of records in each group and compare distributions before making a claim.
  • If a survey item has ordered categories, a frequency table and proportions may communicate more than treating response labels as precise distances.

The Statistics Project Ideas resource provides the upstream feasibility check for selecting a question and data route before this descriptive stage.

Data formExamplesUseful first summariesInterpretation focus
Nominal categoriesTravel mode, region, product type.Counts, proportions, bar chart.Which categories are most/least frequent and what denominator is used?
Ordered categoriesSatisfaction band, agreement scale, income band.Frequency table, cumulative percentage, median category where appropriate.Distribution across ordered responses; do not assume equal spacing without justification.
Numerical measurementJourney time, revenue, temperature, test score.Mean/median, spread, distribution plot and range.Typical value, variation, skewness and unusual observations.
Time-based seriesMonthly cases, daily demand, annual rate.Time plot, change from period to period, rolling/seasonal context where relevant.Trend, seasonality, breaks and missing periods.
Area-level indicatorLocal rate, deprivation score, emissions level.Map/table/ordered comparison, distribution and outliers.Geographic variation; avoid individual-level conclusions.

The five questions before selecting a summary

A useful decision routine asks: What is the variable? What is the question? What does the distribution look like? How complete are the data? Who or what do the records represent? These questions reduce mechanical use of the mean, percentage or chart type.

1. What is the variable type?

Categories need counts and proportions. Numerical measures may support centre and spread statistics. Time series need an ordered time display. Variables with labels such as “very dissatisfied” to “very satisfied” are ordered but the numerical distance between categories may not be equal; report them with that caution in mind.

2. What is the analytical purpose?

A single dataset may support several summaries, but the question determines which one is primary. If the purpose is to show variation, a range alone may be insufficient; a distribution plot can show clustering and outliers. If the purpose is to compare groups, use the same units, time period and scale for both groups.

3. What does the distribution look like?

Inspect a table or graph before selecting mean or median. A right-skewed distribution, such as a small number of very high sales values, can pull the mean above the value most records are near. The median is often a more stable representation of a middle observation in such a distribution. This does not make the mean wrong; it means both the distribution and the summary must be explained.

4. Are values missing or anomalous?

A missing value is not automatically zero, and an unusual value is not automatically an error. Investigate the source, coding and context. State the number of usable records and how missing values were handled. USC Libraries specifically highlights reporting missing-data handling, data-cleaning choices and descriptive statistics as part of transparent quantitative reporting (USC Libraries, 2026).

5. What do records represent?

A local-authority rate, a respondent and a month are different units. A summary of areas cannot describe individual residents. A summary of survey respondents cannot automatically describe non-respondents. Interpretation needs to remain at the level supported by the data.

Decision questionIf the answer is “yes”Descriptive response
Are values numerical and roughly symmetric without dominant extremes?Mean may communicate a useful average.Report mean with standard deviation and a distribution display.
Are values skewed or influenced by extremes?Median may better show a central observation.Report median with interquartile range and display.
Is the variable categorical?A numerical average may be meaningless.Use counts, proportions and a bar chart.
Is the variable ordered?Order matters, but distances may not be equal.Use ordered table/plot, median category or cumulative percentages.
Does the question concern change over time?Order is essential.Use a time plot and identify periods, gaps and possible seasonality.
Are groups being compared?Group sizes and distributions matter.Show group n, comparable summary and a clear shared scale.

Centre and spread: use them together

A centre statistic answers “where are values typically located?” A spread statistic answers “how much do they vary?” Reporting one without the other can be misleading. Two groups can have the same average but very different consistency or inequality. A single “average” does not show the range of experience or observation.

MeasureWhat it describesBest used whenCaveat
MeanArithmetic average.Numerical data where values and extremes are meaningful.Sensitive to extreme values and skew.
MedianMiddle value after ordering.Skewed numerical data or ordered values.Does not show the size of values away from the middle.
ModeMost frequent value/category.Categories or discrete numerical values.May be absent or multiple; rarely enough alone.
RangeDifference between maximum and minimum.A quick indication of total span.Depends heavily on two values.
Interquartile rangeSpread of the middle half of ordered values.Skewed distributions and median-based summaries.Does not describe extremes.
Standard deviationTypical spread around the mean.Numerical data where mean/variation interpretation is meaningful.Can be hard to interpret without distribution context.

Take a fictional example of weekly study time. If most respondents report 5–12 hours but two record 45 hours, the mean may be pulled upward. Reporting the median and interquartile range, alongside a histogram or dot plot, tells the reader whether the high values are recording errors, genuine but unusual cases or a sign of a long right tail. The decision should be documented, not hidden.

Tables and charts should answer a question

Tables provide exact values; charts show overall pattern. Both need a title, clear labels, units, source and note where needed. A chart cannot stand alone: the text should tell the reader what to notice. NIST’s handbook is built around problem-oriented statistical analysis and exploratory data analysis because effective interpretation relies on seeing data structure, not only calculating a summary (NIST/SEMATECH, 2025).

PurposeUsually suitable displayWhat to checkAvoid
Compare category frequenciesBar chart or ordered bar chart.Categories are mutually understood; y-axis starts at zero where lengths imply quantity.Pie chart with many small categories or unclear total.
Show a numerical distributionDot plot, histogram or box plot.Bins/scale are reasonable; sample size is stated.Hiding the distribution behind a mean alone.
Compare numerical groupsSide-by-side box plots or comparable distribution plots.Same scale and group n visible.Separate axes that create false visual differences.
Show change over timeLine chart.Time spacing, missing periods and units are clear.Joining unrelated categories or treating time order casually.
Show two numerical measuresScatterplot.Axes, units, clusters and unusual cases examined.Adding a trend line and declaring causation.

Describing distributions in words

A useful interpretation has four elements: the subject, pattern, evidence and boundary. For example: “In the available retailer dataset, monthly online orders were higher in November and December than in most other months; the line chart shows a marked seasonal peak, but the data do not identify which promotional or market factors produced it.” The wording anchors the statement in the dataset and resists a causal leap.

Avoid vague statements such as “the data were significant” when describing a distribution. Significance is not a synonym for importance, and descriptive analysis may not involve a significance test at all. Describe direction, magnitude, spread and context instead.

Fictional planning application

The following scenario is fictional. Amara is analysing a public dataset of monthly bicycle counts at one urban monitoring point. Her research question asks how counts varied from January to December 2025. At first, she calculates an average monthly count and writes that cycling “was typical at 8,400 journeys a month”.

Her tutor asks her to examine the time pattern. Amara creates a line chart, checks whether all months have the same number of observation days and notes a gap in April due to equipment maintenance. The chart reveals lower counts in winter, a gradual rise through spring and a peak in late summer. She reports monthly counts, identifies the incomplete month and gives a simple range rather than treating the annual mean as the main story.

Amara’s final interpretation is: “At this monitoring point, recorded bicycle journeys varied across 2025, with higher counts during late summer. April is incomplete because of maintenance, and the series describes one location rather than cycling across the city.” The descriptive work gives her a stronger answer than a single average could.

Practical analysis sequence

First, confirm the unit of analysis, variable definitions, coding and data period. Second, make a basic frequency or summary table and count missing values. Third, choose a display that matches the data type and question. Fourth, inspect the distribution or time pattern before selecting mean, median, proportion or variation measure.

Fifth, compare summaries across groups only when measurements, denominators and periods are comparable. Sixth, write an interpretation using the subject, pattern, evidence and boundary structure. If an inferential analysis is then needed, its design and assumptions should be justified; Hypothesis Testing explains that next stage rather than replacing descriptive analysis with a p value.

Critical limitations and safeguards

Descriptive statistics can conceal as well as reveal. An average can hide inequality, a percentage can hide a small denominator, a chart scale can exaggerate change and a total can confuse count with rate. Always state the denominator, unit and timeframe. Use rates only where the population at risk or exposure is defined appropriately.

Do not delete unusual values merely because they affect a summary. Investigate whether they are valid, record the decision and, where useful, show an analysis with and without a justified sensitivity check. For sensitive data, avoid tables or charts that could identify individuals through small cells. Follow institutional and source-specific rules on disclosure control and secure handling.

FAQs

Q1. Should I always use the mean?
No. The mean is useful for many numerical variables, but the median can be more informative in skewed data or where extreme values have strong influence. Inspect the distribution first.

Q2. What is the difference between a count and a percentage?
A count is the number of records in a category. A percentage expresses that count relative to a stated total. Both are needed when group sizes or denominators matter.

Q3. Can a bar chart show numerical measurements?
Bar charts are generally for categories or distinct values. Continuous numerical data are usually better represented with a histogram, dot plot, box plot or time plot, depending on the question.

Q4. Does a descriptive pattern prove why something happened?
No. A pattern can motivate a further question, but descriptive data alone usually do not establish a causal explanation.

References

American Statistical Association (2016) Guidelines for Assessment and Instruction in Statistics Education: College Report 2016 (Accessed: 25 August 2026).

NIST/SEMATECH (2025) e-Handbook of Statistical Methods (Accessed: 25 August 2026).

UK Statistics Authority (2022) Code of Practice for Statistics (Accessed: 25 August 2026).

USC Libraries (2026) Quantitative Methods (Accessed: 25 August 2026).