Build a descriptive-statistics table around your variables
Match each variable to a useful descriptive summary, show valid and missing denominators, and build a clear table from a checked fictional dataset.
About 11 min read · Allow extra time for the practical checks
Statistical software can produce a page of means, standard deviations, quartiles, minima, maxima and percentages in seconds. That output is not yet a useful descriptive-statistics table. The table still needs decisions about what each variable means, which observations contributed and which summary makes the distribution understandable.
A category coded as 1, 2 and 3 does not become a continuous measure because the software can calculate its mean. A percentage is incomplete when the reader cannot tell whether its denominator is the full sample or only valid responses. A single overall sample size can also conceal that different variables have different amounts of missing data.
The practical task is to build the table around the variables, not around the order of a software menu.
Begin with a variable-to-summary specification
Before formatting the table, make a short specification from the cleaned dataset and codebook. Record the display label, meaning, unit, variable type, valid values, missing codes and intended summary.
If your next decision is how to show a pattern visually rather than summarise values in rows, use the separate guide to choose a chart for your data question.
This step connects the table to the study’s measurement decisions. If the meaning or scoring rule is still unsettled, return to the operational-definitions guide. Descriptive statistics cannot repair a variable that represents the wrong construct or mixes incompatible codes.
The UCLA Statistical Consulting introduction to R illustrates the basic distinction: continuous variables may be summarized with measures such as the mean, median and variance, whereas categorical variables are commonly represented by frequency tables and proportions. Kent State University’s frequency-table guidance also warns that numeric category codes do not make means and standard deviations meaningful.
On a narrow screen, scroll the table sideways to read every column.
| Variable situation | Useful starting summary | Decision to document | Common mistake |
|---|---|---|---|
| Nominal categories | Count and percentage for each relevant category | Whether percentages use all cases or valid responses | Reporting the mean of arbitrary category codes |
| Ordered categories | Counts and percentages; a median or quartiles only when their interpretation is justified | Order, labels and treatment of unused categories | Hiding the distribution behind one average |
| Binary indicator | Counts for both states and a proportion for the stated event | Which state is the event and what forms the denominator | Reporting 72.7% without saying 8 of 11 |
| Continuous, reasonably symmetric | Mean and standard deviation, usually with valid n and units | Evidence that the distribution and extreme values make this pair informative | Using the pair automatically because it is a default |
| Continuous, skewed or long-tailed | Median and quartiles or another justified robust summary | Quartile convention, units and whether a range adds useful context | Choosing the summary that makes the result look preferable |
| Derived scale or score | The summary appropriate to the validated score and analysis plan | Version, scoring rule, possible range and missing-item rule | Inventing a total from items not designed to be combined |
This is a decision map, not a universal style sheet. A discipline, reporting guideline, supervisor or target journal may require a particular form. The important point is that the choice should be traceable.
Inspect the distribution before selecting centre and spread
For a continuous variable, do not decide between mean with standard deviation and median with quartiles by habit alone. Inspect the distribution, units, impossible values and extreme observations. Check whether an extreme value is a valid observation, a coding error or a value needing separate investigation.
The NIST/SEMATECH discussion of measures of scale explains that spread measures emphasize different parts of a distribution. Standard deviation gives substantial weight to tail values, while the interquartile range focuses on the middle portion. For data with long tails, a rank-based measure can be more stable.
That does not mean every visibly skewed variable must be reported in one prescribed way. It means the summary should answer a defensible descriptive question. If a mean matters substantively—for example, because total resource use divided by all eligible cases is the quantity of interest—you may report it with appropriate context. Do not switch summaries after seeing which one supports a preferred narrative.
Keep units attached to the variable label or table note. A value of 27.5 is uninterpretable until the reader knows whether it represents minutes, days, points or currency.
Make denominators visible
A descriptive table should let the reader reconstruct what contributed to each row. At minimum, distinguish:
- the intended or overall sample size;
- the valid observations for the variable;
- the observations missing for that variable;
- the denominator used for each percentage.
The official STROBE checklist asks observational-study reports to indicate missing data for each variable of interest. The principle is useful beyond observational health research: variable-specific missingness can change what a descriptive value represents.
Do not turn missing into zero unless zero is genuinely the recorded value. A blank commute-time response does not establish a zero-minute commute. Likewise, a questionnaire item marked “not applicable” may need to remain distinct from an unanswered item. Cleaning and missing-data analysis require their own justified decisions; the descriptive table should report those decisions accurately rather than conceal them.
For categorical percentages, choose and label the denominator. If 12 people are in the sample but one does not answer an item, “8 yes” can be:
- 66.7% of all 12 cases; or
- 72.7% of the 11 valid responses.
Both calculations are arithmetically correct. They answer different descriptive questions. Reporting “8 (72.7%)” without a denominator leaves the reader to guess.
Worked fictional example
The organisation, people and data below are wholly fictional. They exist only to demonstrate the calculations and table decisions. They do not represent a real sample or support any population claim.
A fictional service team has 12 records from a small internal planning exercise. The file contains three variables:
- study arrangement: on-site, hybrid or remote;
- weekly commute minutes: a numeric duration;
- completed planning step: yes or no.
The variable specification comes before the final table.
On a narrow screen, scroll the table sideways to read every column.
| Variable | Type and unit | Valid observations | Missing observations | Chosen summary |
|---|---|---|---|---|
| Study arrangement | Nominal category | 12 | 0 | Count and percentage for each category |
| Weekly commute minutes | Continuous duration in minutes | 10 | 2 | Median, first and third quartiles, and range after distribution inspection |
| Completed planning step | Binary indicator | 11 | 1 | Yes and no counts with percentages of valid responses |
Check the category counts and percentages
Study arrangement contains 5 on-site, 4 hybrid and 3 remote records. The counts total 12.
- On-site:
5 / 12 × 100 = 41.7% - Hybrid:
4 / 12 × 100 = 33.3% - Remote:
3 / 12 × 100 = 25.0%
Rounding to one decimal gives 100.0% here. In another table, separately rounded percentages may total 99.9% or 100.1%; a note can explain rounding rather than forcing a category to absorb the difference.
Check the continuous summary
The ten valid commute values, sorted in minutes, are:
0, 15, 15, 15, 25, 30, 60, 60, 60, 120
Two of the 12 records are missing. The median is halfway between the fifth and sixth values:
The calculation is (25 + 30) / 2 = 27.5.
Using the stated median-of-halves convention, the first quartile is 15 and the third quartile is 60. The interquartile-range width is therefore 60 - 15 = 45 minutes. The observed range is 0–120 minutes.
The arithmetic mean is 40 minutes because the values total 400. It is higher than the median and affected by the 120-minute observation. The 120-minute value is not automatically an error; it should be checked against the record and definition. For this teaching table, the median and quartiles describe the observed centre and middle half more transparently.
Statistical packages use more than one quantile algorithm. An actual report should use the analysis plan or software convention consistently and document it when exact quartiles matter. The example states its convention so the displayed values can be reproduced.
Check the binary denominator
For completed planning step, 8 records are yes, 3 are no and 1 is missing. The valid denominator is 11:
- Yes:
8 / 11 × 100 = 72.7% - No:
3 / 11 × 100 = 27.3%
The percentages describe the valid responses, not all 12 records. The missing count remains visible beside them.
Assemble the reader-facing table
On a narrow screen, scroll the table sideways to read every column.
| Variable | Valid n | Missing n | Descriptive summary |
|---|---|---|---|
| Study arrangement | 12 | 0 | On-site 5 (41.7%); hybrid 4 (33.3%); remote 3 (25.0%) |
| Weekly commute, minutes | 10 | 2 | Median 27.5; Q1–Q3 15–60; range 0–120 |
| Completed planning step | 11 | 1 | Yes 8 (72.7%); no 3 (27.3%) |
Note. The dataset is fictional. Percentages use valid observations within each variable. Quartiles for commute minutes use the stated median-of-halves convention.
The APA Style sample tables demonstrate why familiar, standard formats help: category tables show counts and percentages with sample context, while descriptive rows identify variable-specific n and relevant summaries. You do not need to imitate one sample mechanically. Use the style required for your work and retain the information a reader needs.
Keep description separate from inference
The fictional table describes these 12 records. It does not show that one arrangement causes a different commute, that the completion proportion generalizes to another team or that any groups differ beyond ordinary sample variation.
Do not add a p value simply because a table contains groups or categories. Inferential analysis begins with a research question, design, sampling process, assumptions and planned comparison. It cannot be selected from the visual appearance of descriptive values alone.
Similarly, a descriptive table should not become a hiding place for selective results. Decide which variables belong from the research question, protocol or transparent reporting plan—not from whether their summaries look favourable.
If the table supports a later analysis, keep the roles distinct. The descriptive table establishes the observed sample and variables. A later model or test addresses a specified inferential question under its own assumptions. When denominators change across several analyses, maintain a separate analysis-specific denominator record; that is a broader task than this variable summary.
Audit the table before submission
On a narrow screen, scroll the table sideways to read every column.
| Audit question | Defect it catches | Repair |
|---|---|---|
| Does every row name the actual variable and unit? | Labels such as “Score 2” or a duration without units | Use a concise display label and define abbreviations or scoring in a note |
| Does the summary match the variable type? | Means for nominal codes or a single percentage for a multi-category variable | Return to the codebook and variable-to-summary specification |
| Are valid and missing counts reconcilable? | One overall N hiding variable-specific loss | Display valid n and missing n or provide an equally explicit note |
| Is every percentage denominator clear? | Percentages of valid responses presented as percentages of the full sample | State the denominator rule and show counts beside percentages |
| Can the statistics be reproduced? | Undocumented filtering, recoding, weighting or quartile conventions | Retain the script or calculation record and document material choices |
| Does the prose stay descriptive? | Causal, representative or inferential claims unsupported by the table | Describe the observed records and reserve wider claims for justified analysis |
Recalculate totals and percentages from the analysis output, not from a copied table cell. Check that displayed rounding is consistent and that category labels match the codebook. Where confidentiality requires suppressing small cells or combining categories, follow the applicable ethical, institutional and disclosure-control rules rather than improvising after publication.
Once the summaries are checked, audit the display’s reader-facing context. The guide to table and figure captions separates informative titles, labels, decoding notes and body-text callouts without choosing statistics or changing the analysis.
A practical 12-step workflow
- Confirm the research question and reporting requirements.
- Freeze or version the cleaned analysis file.
- Read the codebook and identify the unit represented by each row.
- List the variables that genuinely belong in the table.
- Record each variable’s meaning, type, unit, codes and missing-value rule.
- Inspect categorical distributions and continuous-variable shapes.
- Choose a summary pair or category display for each variable.
- Compute valid n, missing n and every percentage denominator.
- Generate the table from reproducible output where possible.
- Check calculations, rounding, labels, notes and confidentiality rules.
- Remove statistics that do not help readers understand the variables.
- Audit the surrounding prose for unsupported inference or causation.
Save the specification and calculation output with the project records. A table is easier to revise when each displayed value can be traced to a variable, rule and computation.
A clear descriptive table does not need every statistic the software can produce. It needs the right summary for each variable, a visible denominator and enough context for the reader to understand what the numbers describe. If you need help checking summaries, assumptions or reproducible output from your own data, statistics and data-analysis support can assist while you retain responsibility for the study, analysis and reporting.
