08Turn Data into Accurate Narrative and Visuals

Numbers create an illusion of solidity. A value with two decimal places looks authoritative even when its denominator is wrong, its population is unclear, or its analysis was exploratory. The medical writer's job is not to make numerical output sound confident. It is to preserve what the analysis actually means.

Context comes before calculation. Who was studied? What was measured? How was it measured? What comparison was made? How large was the difference? How precise was it? What would count as clinically meaningful? What can be generalized?^1,2

Find the source of truth

Before writing from a table, identify its status. Is it a validated output, draft output, exploratory analysis, slide transcription, or manually created summary? Check whether the same result appears elsewhere and whether those versions agree.^3,4

For every material number, confirm:

• Analysis population
• Treatment or comparison groups
• Endpoint definition
• Time point and analysis window
• Unit and transformation
• Denominator and missing-data handling
• Estimate and measure of variability or precision
• Statistical method
• Prespecified, sensitivity, supportive, subgroup, or exploratory status
• Data cutoff or version

Do not infer these details from layout. A column header can be abbreviated or inherited from a template. Use the protocol, SAP, output notes, and statistical review as appropriate.

Build number provenance

Keep a compact provenance record for high-risk values:

The record does not need to accompany every final document. Its value is that a reviewer can reconstruct the number without searching hundreds of pages.

Number provenance is especially important when the same value is transformed. Percentages may be calculated from counts, rates may use person-time, changes may be derived from baseline and follow-up, and summaries may round. Record the transformation and avoid repeated manual recalculation in several sections.

Do not report a p-value alone

Never report a p-value alone: estimates and confidence intervals carry meaning that a p-value cannot.¹ A p-value addresses a narrow question under assumptions; it does not state the size, clinical importance, or precision of an effect. Report the estimate and an appropriate measure of uncertainty in the form required by the document.

Likewise, a nonsignificant result does not prove equivalence or absence of effect. It may reflect imprecision, limited sample size, variability, or an effect smaller than the study could reliably detect. Use wording such as “the study did not detect a statistically significant difference” when that is what the analysis supports, rather than “the treatments were the same.”²

Statistical significance and clinical importance are separate. If a minimum clinically important difference or other threshold is invoked, identify its source and applicability. Do not label an effect meaningful merely because the p-value crossed a convention.

Respect analysis status

Prespecification affects credibility. Primary and secondary endpoints, multiplicity control, analysis populations, and sensitivity analyses belong to an agreed statistical plan. Post-hoc and exploratory analyses can be valuable, but they should be labeled as such and interpreted as hypothesis-generating unless stronger justification exists.

Subgroup findings require particular care. Small numbers, multiple comparisons, interaction testing, and lack of prespecification can make an apparently striking pattern unstable. Do not write a subgroup result as a general treatment claim. State the subgroup, analysis status, estimate, precision, and relevant limitation.

The same principle applies to real-world evidence and observational analyses. Confounding, selection, misclassification, missingness, and data provenance affect inference. Adjustment can reduce some problems but cannot transform an observational design into randomization. Let the design govern the verbs.⁵

Report the population, not just the percentage

Percentages without counts can hide small denominators. Counts without denominators can hide differing group size. Rates without time can hide exposure differences. Always ask what a reader needs to interpret the value.

For adverse events, distinguish patients from events and incidence from exposure-adjusted rates. For diagnostic measures, keep sensitivity, specificity, predictive values, and prevalence context distinct. For survival analyses, report the relevant estimate, time horizon, censoring context, and uncertainty. For patient-reported outcomes, explain the instrument, direction, scale, timing, and interpretive threshold when used.

Units and decimal places carry meaning. Use precision appropriate to the measurement and analysis. Excess decimal places imply information the study may not possess; inconsistent rounding creates apparent disagreement.

Write Results in the order of the questions

Results should follow the Structure established by objectives and methods. A common sequence is participant flow and analysis populations, baseline characteristics, primary outcome, secondary outcomes, other analyses, and safety. The exact order depends on the document, but it should not be rearranged merely to place the most favorable finding first.

Use parallel Methods and Results headings where possible. This makes it easier to see whether every promised analysis was reported and whether any result appears without a described method.

Narrative text should highlight the pattern, magnitude, and interpretation needed to navigate the outputs. It should not transcribe every cell. If a table provides the full values, the text can state the main result and refer to the table. If the exact value is central to the claim, include it.

Make safety visible

Fair balance requires appropriate attention to safety and tolerability.⁶ Avoid placing detailed efficacy language beside a single generic sentence that treatment was “well tolerated.” Report the defined safety population, exposure, adverse-event categories, serious events, discontinuations, deaths where applicable, and events of special interest according to the document and source data.

Do not infer causality from temporal occurrence unless the analysis and assessment support it. Preserve the distinction among adverse events, adverse reactions, and other safety classifications used by the project. Use consistent terminology across narrative, tables, and conclusions.

Choose the right visual

Visuals are analytical communication, and the form should be chosen according to the reader's question, the structure of the data, and the risk of distortion.^7,8 Choose the form according to the question:

• Use a table when readers need exact values or comparison across several dimensions.
• Use a line graph for change or trend over an ordered continuum when the scale supports it.
• Use a bar or dot display for category comparisons when position and magnitude matter.
• Use a flow diagram for progression through study or review stages.
• Use a forest plot when effect estimates and confidence intervals across studies or subgroups are the point.
• Use a diagram when relationships, processes, or mechanisms matter more than exact values.

Avoid decorative three-dimensional effects, truncated axes that exaggerate differences, inconsistent scales, and colors that imply categories or importance without explanation. Elaborate designs can confuse and distract rather than inform.^8,9 Visual choices can mislead even when every plotted value is correct.⁷

Make the visual stand alone

A reader may encounter a table or figure without reading the surrounding paragraph. Give it an informative title, define populations and time points, explain symbols and abbreviations, identify units, describe statistical summaries, and include material footnotes.⁶ A title should say what is shown, not merely “Results.”

Keep the visual concise enough to read. Large tables may need grouping, hierarchy, or movement to supplementary material. Do not solve crowding by shrinking text until the information is technically present and practically inaccessible.

For public-facing material, visuals may reduce cognitive burden, but they also require testing. Icons, risk graphics, timelines, and diagrams can be misunderstood. Health literacy, cultural relevance, accessibility, patient involvement, and human review all affect whether a visual communicates as intended.^8,10,11 A familiar-looking symbol is not automatically universal.

Run the cross-document number check

Compare repeated values across:

• Abstract or executive summary
• Main narrative
• Tables and figures
• Conclusions
• Plain-language derivative
• Poster or slides
• Response letter or briefing points

Check endpoint names, populations, denominators, direction, units, rounding, significance statements, and time points. If two values intentionally differ, explain why.

Open the statistics and visual-reporting deep dive ↗