Presenting Data Using Tables and Graphs

Overview

Tables and graphs are analytical tools, not decoration. A well-designed table supports precise lookup and comparison. A well-designed graph makes patterns, distributions, trends and uncertainty visible. Poor presentation can conceal missing data, exaggerate small differences or imply precision that the study does not possess.

The choice should follow the communication objective:

• use text for one or two central results;
• use a table when readers need exact values across several variables or groups; and
• use a graph when the shape, pattern, relationship or trend is more important than exact lookup.

A result should not be repeated in full in text, a table and a graph. Each format should contribute something distinct.

Why this matters

Readers often see the table or figure before reading the Methods section. Presentation can therefore shape interpretation strongly.

Misleading practices include:

• percentages without denominators;
• error bars that are not defined;
• bar charts that hide all individual numerical values;
• axes truncated to exaggerate differences;
• connecting independent groups with a line;
• three-dimensional bars that distort height;
• pie charts with many similar slices;
• reporting only p-values; and
• showing subgroup estimates without confidence intervals.

Experimental research on graphical perception shows that people compare position along a common scale more accurately than angles, areas or volumes [1]. Good displays therefore favour straightforward position and length encodings over decorative complexity.

Choose the format by purpose

Table 1. Selecting text, table or graph

Note: A graph must preserve the study’s pairing, clustering and time structure.

Principles for clear tables

Write an informative title

A table title should identify:

• what is shown;
• population or analysis set;
• groups or time point; and
• study context where needed.

Avoid titles such as Table 2. Results.

A better title is:

Table 2. Week-12 blood-pressure outcomes among randomised participants

Make rows and columns self-explanatory

Use labels rather than internal variable names. State units in row or column headings.

Prefer:

Systolic blood pressure, mean (SD), mmHg

rather than:

SBP

Define abbreviations in a note.

Show denominators

For categorical variables, report n/N (%) when the denominator may differ.

For example:

Blood-pressure control: 74/114 (64.9%)

is clearer than:

74 (64.9%)

when outcomes are missing.

Match summaries to variables

Use:

• number and percentage for categorical variables;
• mean and standard deviation when the mean is a meaningful description;
• median and interquartile range when distribution, scale or outliers make them more informative;
• effect estimates with 95% confidence intervals for comparisons.

Do not report mean ± standard error as though it described participant variability. The standard deviation describes spread among observations; the standard error describes sampling uncertainty of an estimate.

Align numbers and use sensible precision

Numbers should be aligned by decimal point where possible.

Use precision justified by measurement and analysis:

• blood pressure to the nearest whole mmHg when measured that way;
• percentages commonly to one decimal place;
• effect estimates with enough precision for interpretation;
• exact p-values to three decimal places, except $p < 0.001$.

Excessive decimals imply false precision.

Include explanatory notes

Notes should define:

• abbreviations;
• denominators;
• missing data;
• analysis populations;
• reference categories;
• direction of contrasts;
• statistical methods; and
• whether values are observed or model-based.

Baseline tables in randomised trials

A baseline table describes the randomised groups. It should not be used to test whether randomisation “worked” through baseline significance tests.

Chance differences are expected after random allocation. Baseline p-values do not determine whether adjustment is required. Adjustment variables should be prespecified using design and prognostic knowledge.

Current CONSORT guidance emphasises transparent participant flow, group summaries, analysed numbers, effect estimates and uncertainty [2].

Simulated baseline table

simulated data

Table 2. Baseline characteristics by allocated group

Characteristic	Telemonitoring (n = 120)	Usual care (n = 120)
Age, mean (SD), years	59.8 (10.7)	60.4 (10.2)
Female sex, n/N (%)	64/120 (53.3%)	62/120 (51.7%)
Baseline systolic blood pressure, mean (SD), mmHg	151.8 (12.6)	151.1 (11.9)
Diabetes mellitus, n/N (%)	46/120 (38.3%)	43/120 (35.8%)
Number of antihypertensive medicines, median (IQR)	2 (2–3)	2 (1–3)

Note: IQR = interquartile range; SD = standard deviation. No hypothesis tests are presented for baseline differences.

(Percentage calculations use the displayed counts and denominators; calculations for the means, standard deviations, medians and interquartile ranges are not shown here.)

Principles for clear graphs

Use a common scale

Position along a common axis supports accurate comparison [1]. Bar lengths and point positions are generally easier to compare than pie-slice angles, bubble areas or three-dimensional volumes.

Label axes and units

Every numerical axis needs:

• a descriptive variable name;
• unit;
• readable tick labels; and
• a scale appropriate to the data.

Do not rely on a caption to supply essential units.

Use honest axis limits

For bar charts, the numerical axis should ordinarily begin at zero because bar length represents magnitude.

For point or line graphs, a non-zero origin may be reasonable when it reveals clinically relevant variation. The scale must be clearly labelled and not chosen to exaggerate a trivial difference.

Show distributions where possible

For small or moderate samples, display individual observations using:

• dot plots;
• jittered plots;
• paired-value plots; or
• box or violin displays combined with points.

Bar-and-error plots can conceal skewness, multimodality, unequal sample size and outliers. Biomedical researchers have specifically recommended moving beyond bar and line graphs that hide continuous data distributions [3].

Define error bars

Error bars may represent:

• standard deviation;
• standard error;
• confidence interval;
• interquartile range; or
• another interval.

The caption must state which.

For treatment effects, a 95% confidence interval is usually more informative than a standard error because it directly displays uncertainty on the effect scale.

Preserve pairing and repetition

For paired data, connect measurements from the same participant or display paired differences.

For repeated measurements, show time in order and represent within-person correlation in the analysis. Group mean trajectories should include numbers observed at each time and clearly defined intervals.

Do not connect independent categories with lines merely to make a graph appear dynamic.

Use colour carefully

Colour can distinguish groups, but meaning should not depend on colour alone.

Use:

• direct labels;
• different line types or point shapes;
• sufficient contrast;
• palettes that remain interpretable with common colour-vision deficiencies; and
• grayscale-compatible design for printing.

Avoid red–green-only distinctions.

Reduce non-data ink

Remove:

• three-dimensional effects;
• decorative backgrounds;
• unnecessary gridlines;
• heavy borders;
• gradients;
• pictograms scaled by area; and
• legends that could be replaced by direct labels.

Minimal design should not remove essential context.

Selecting a graph by data type

Table 3. Common graph choices

Worked example

Presenting missing data and participant flow

Missing data should be visible rather than hidden in variable-specific denominators.

For each major outcome, report:

• expected number;
• observed number;
• missing number and percentage;
• reasons where available; and
• analysis population.

In a trial, a participant-flow diagram should show:

• assessed for eligibility;
• randomised;
• allocated;
• received intervention where relevant;
• lost to follow-up;
• discontinued intervention;
• included in each analysis; and
• reasons for exclusions.

Withdrawal from treatment and missing outcome data are different and should not be combined.

Presenting regression and adjusted analyses

A regression table should state:

• outcome;
• effect measure;
• predictor or contrast;
• reference category;
• crude or adjusted status;
• adjustment variables;
• estimate;
• 95% confidence interval;
• p-value where relevant; and
• analysis sample.

Do not present a software coefficient without translating its scale.

Examples include:

• mean difference in mmHg;
• risk ratio rather than an unexplained log coefficient;
• odds ratio explicitly labelled as an odds ratio;
• hazard ratio interpreted as a relative instantaneous event rate under the model, not a risk ratio.

A forest plot can display several effect estimates and confidence intervals on a common scale. Subgroup displays should include interaction assessment and avoid implying differences merely because one subgroup is statistically significant and another is not.

Observed summaries and model-based estimates MMed extension

A table may contain both raw group summaries and adjusted model estimates, but they should be clearly distinguished.

Observed summaries describe participants with available measurements. Model-based estimates may represent:

• adjustment for baseline covariates;
• all randomised participants through multiple imputation;
• standardisation to a target population; or
• a transformation back to the clinical scale.

The caption or note should state:

• which values are observed;
• which are estimated by a model;
• the model and adjustment set;
• the population represented; and
• how missing data were handled.

Mixing these without explanation can make denominators and estimands ambiguous.

Accessibility and responsive presentation

A table or graph should remain understandable:

• without colour;
• with screen magnification;
• when printed;
• on a narrow screen; and
• through a text alternative where appropriate.

For tables:

• use genuine column and row headers;
• avoid merged cells;
• keep headings concise;
• allow horizontal scrolling within the table container rather than the entire page; and
• do not present tables as images.

For figures:

• provide alternative text describing the educational or analytical meaning;
• use a caption that states what readers should notice;
• label lines or points directly where possible;
• maintain sufficient contrast; and
• do not rely on hover interaction for essential information.

Reproducibility

Tables and graphs should be generated from the analysis dataset using reproducible code.

The code should control:

• labels;
• ordering;
• units;
• rounding;
• denominators;
• confidence intervals;
• captions;
• footnotes; and
• export dimensions.

Manual copying creates opportunities for disagreement between text, tables and figures.

Suggested reporting

Statistical reporting guidance recommends naming methods, reporting estimates with uncertainty and providing sufficient information for readers to verify the analysis [4]. Design-specific guidance such as CONSORT and STROBE should also be followed [2], [5].

Common mistakes

Knowledge check

When is a table preferable to a graph?

Answer and reasoning

A table is preferable when readers need exact values across several variables or groups. A graph is better when the main message is a pattern, trend, distribution or relationship.

Why is a bar chart of mean ± standard error often inadequate for a numerical outcome?

Answer and reasoning

It hides individual observations, distribution shape, overlap, sample size and outliers. The standard error also describes uncertainty in the mean, not variability among participants.

What should an error bar caption state?

Answer and reasoning

It should identify whether the bar represents a standard deviation, standard error, confidence interval, interquartile range or another interval. Readers should not have to guess.

Why should baseline p-values generally be omitted from a randomised-trial table?

Answer and reasoning

Random allocation produces chance differences by design. Baseline significance tests do not determine whether randomisation succeeded or which covariates should be adjusted for. Adjustment should be prespecified.

How should a table show an outcome observed for 114 of 120 participants?

Answer and reasoning

Report both numerator and denominator, such as 74/114 (64.9%), and state the overall expected group size and missing outcome count in the table or note.

Key take-home messages

References

1. Cleveland WS, McGill R. Graphical perception: theory, experimentation, and application to the development of graphical methods. J Am Stat Assoc. 1984;79(387):531–554. https://doi.org/10.1080/01621459.1984.10478080
2. Hopewell S, Chan AW, Collins GS, Hróbjartsson A, Moher D, Schulz KF, et al. CONSORT 2025 statement: updated guideline for reporting randomised trials. BMJ. 2025;389:e081123. https://doi.org/10.1136/bmj-2024-081123
3. Weissgerber TL, Milic NM, Winham SJ, Garovic VD. Beyond bar and line graphs: time for a new data presentation paradigm. PLoS Biol. 2015;13(4):e1002128. https://doi.org/10.1371/journal.pbio.1002128
4. Lang TA, Altman DG. Basic statistical reporting for articles published in biomedical journals: the SAMPL guidelines. Int J Nurs Stud. 2015;52(1):5–9. https://doi.org/10.1016/j.ijnurstu.2014.09.006
5. von Elm E, Altman DG, Egger M, Pocock SJ, Gøtzsche PC, Vandenbroucke JP, et al. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. PLoS Med. 2007;4(10):e296. https://doi.org/10.1371/journal.pmed.0040296