03 · What You Need to Know
A Figure Communicates More Than the Values Printed on It
Visual Encoding Changes How Readers Perceive the Same Numbers
A graph translates numerical information into visual properties such as position, length, area, angle, shape, color, and spatial arrangement. Readers perceive those visual relationships before, and often more readily than, they inspect individual numbers.
That is why visualization choices matter scientifically. Research on misleading visualizations has documented how inappropriate scaling, shape, spatial orientation, omitted information, and other graphical choices can produce inaccurate impressions even when the underlying values themselves remain unchanged.
A figure should therefore be evaluated at two levels: are the plotted values correct, and does the visual representation communicate their relationships faithfully?
A Truncated Axis Can Magnify a Small Difference
Consider two groups with values of 98 and 100. A bar chart whose vertical axis begins at zero displays the two bars as nearly equal. If the axis begins at 97, the second bar may appear several times taller than the visible portion of the first.
The values 98 and 100 remain perfectly correct. What changes is the perceptual magnitude of their difference.
For bar charts, this is especially consequential because bar length encodes magnitude. Research on misleading graphs therefore commonly identifies truncated axes as a potential source of distortion.
This does not mean every axis in every graph must begin at zero. Line graphs, for example, may legitimately use restricted ranges to show variation that would otherwise be invisible. The appropriate scale depends on the graph type and purpose. The key is whether the choice clarifies the data or creates an exaggerated impression.
Two Axes Can Make Unrelated Trends Look Remarkably Similar
Dual-axis figures allow two variables with different units or ranges to appear on the same plot. By changing the range of either axis, researchers can make trajectories appear more or less aligned.
The underlying data need not change at all.
If the visual alignment encourages readers to infer a relationship stronger than the analysis supports, the figure can become misleading. Axis labels and numerical accuracy do not automatically neutralize the visual impression created by carefully chosen scaling.
Means Can Hide the Distribution Behind Them
Two datasets can have the same mean and very different distributions.
One group might contain tightly clustered observations. Another might contain two distinct subgroups. A third might contain substantial skew or several extreme values. If all three are represented by the same bar height, the figure conceals those differences.
This is one reason visualization guidance increasingly encourages showing individual observations, distributions, box plots, violin plots, or other representations where appropriate rather than relying exclusively on bars showing summary statistics.
The figure is not numerically false because the mean is correct. It may simply be insufficient for the inference readers are likely to make.
Aggregation Can Create a Pattern That Is Not Visible Within the Groups
Combining data across participants, sites, time periods, demographic groups, experimental batches, or other meaningful strata can alter apparent relationships.
An aggregate trend may differ from the patterns within the constituent groups. Conversely, separating data selectively can make a weak overall relationship look striking within a chosen subgroup.
Aggregation is not inherently misleading. Researchers routinely summarize data because figures cannot display every dimension simultaneously. But the level of aggregation should match the research question and should not conceal structure essential to interpreting the result.
Percentages Without Denominators Can Distort Practical Meaning
A figure reports that an event occurred in 50% of Group A and 25% of Group B. That sounds like a substantial difference.
But 50% could mean 1 of 2 participants, while 25% could mean 250 of 1,000. The percentages remain mathematically correct, yet readers lack important information about precision and evidentiary weight.
Denominators also matter when groups differ substantially in size. Reporting raw counts alone can create the opposite problem by making a larger group appear to have a greater rate simply because it contains more observations.
Where interpretation depends on the denominator, figures should make the relevant sample sizes or proportions sufficiently clear.
Relative Change Can Sound Much Larger Than Absolute Change
Suppose an outcome increases from 1% to 2%. That is a 100% relative increase but an absolute increase of one percentage point.
Both descriptions are mathematically correct.
Displaying only the relative increase can nevertheless produce a different impression of practical magnitude from showing the absolute risks. Which representation is appropriate depends on the question, but researchers should consider whether the chosen metric gives readers enough context to interpret the effect.
Uncertainty Is Part of the Result
A point estimate is not the entire statistical result. Sampling uncertainty, variability, measurement error, and other forms of uncertainty may materially affect interpretation.
Visualization guidance has emphasized that omitting uncertainty can itself be misleading. A graph showing two means as precise points may make their difference appear far more definite than a figure showing confidence intervals or the underlying distribution.
Even error bars can be ambiguous if readers are not told whether they represent standard deviations, standard errors, confidence intervals, or something else.
Precision of the Mean Is Not the Same as Predictability of Individual Outcomes
A particularly subtle problem arises when a figure shows narrow confidence intervals around group means. Readers may interpret those narrow intervals as evidence that individual outcomes are tightly concentrated.
They are not the same thing.
Experimental research has shown that displaying only inferential uncertainty can lead even scientifically trained readers to overestimate the predictability or magnitude of effects when individual outcomes remain highly variable.
Where individual variability is important to the substantive claim, showing the underlying observations or distribution can prevent the mean from doing more interpretive work than it deserves.
Normalization Can Clarify Data or Hide Their Original Scale
Researchers often normalize measurements to a baseline, control, maximum value, reference category, or other quantity. This can be analytically appropriate and make comparisons easier.
But normalization can also obscure meaningful differences in the original values.
For example, two groups may both be plotted as “100% of baseline” even though their baseline measurements differ substantially. A figure displaying only normalized values may conceal that difference unless readers are given enough information to understand what was normalized and why.
The issue is not that normalization is inherently misleading. The transformation and reference should be clear enough for readers to interpret the plotted quantities correctly.
A Selective Time Window Can Change the Apparent Trend
A trend can look dramatic over one interval and unremarkable over another.
Starting a time series immediately before an unusual peak, ending it before a reversal, or omitting earlier observations can change the visual narrative without changing any displayed value.
Researchers therefore need a defensible rationale for the period shown. If the figure displays only part of the available time series, readers should not be led to assume that the selected interval represents the broader trend when it does not.
Color and Area Can Exaggerate Differences
Visual magnitude is not communicated only through axes.
If circles represent quantities, doubling the diameter quadruples the area. A reader may therefore perceive a fourfold visual difference when the intended numerical difference is only twofold.
Three-dimensional bars, perspective effects, color saturation, unequal icon scaling, and other design choices can similarly distort perceived magnitude.
A figure may contain accurate labels while its visual encoding communicates something else at first glance.
Selective Data Ranges Can Hide Important Observations
A scatterplot may display only observations between specified limits, leaving extreme values outside the visible range. A histogram may use bin widths that conceal multimodality. A heatmap may use a color scale that compresses variation in one region while exaggerating another.
These choices can sometimes be justified. But researchers should ask what information disappears because of them.
This is closely related to selectively presenting images, cases, quotations, or examples. Selection can occur through graph design as well as through deciding which observations to discuss.
A Figure Can Be Accurate but the Caption Can Make It Misleading
Visual interpretation also depends on labels, legends, captions, and surrounding prose.
A figure may accurately show an association while the caption calls it an “effect.” A graph may show a subgroup while the caption describes the result as applying to the entire sample. Error bars may be present but never defined.
Technical accuracy therefore includes more than pixel placement. The figure and its explanatory text work together as a representation of the research.
Misleading Presentation Is Not Automatically Research Misconduct
Poor figure design can arise from inexperience, software defaults, disciplinary conventions, honest mistakes, or genuine disagreement about the clearest way to display data.
That should not automatically be equated with falsification.
Under the PHS framework, falsification involves manipulating research materials, equipment, or processes, or changing or omitting data or results so that the research is not accurately represented in the research record. A formal misconduct finding additionally requires the specified departure from accepted practices, culpability, and evidentiary standard.
Still, a technically accurate figure should not be treated as beyond integrity scrutiny merely because its individual numbers can be verified. The relevant question remains what the figure represents to the reader.
Watch Out
Do not audit a figure only by asking whether each plotted number is correct. Ask what visual conclusion the figure encourages and whether the complete underlying data support that conclusion. Accuracy of coordinates is not the same as accuracy of communication.
