Data visualization is an act of translation: it converts numbers into a visual form that a human can process faster than a table. The risk of translation is that the translator can introduce meaning the original did not have. A chart that makes a 2% difference look dramatic, a trend that only exists because the y-axis starts at 94, a pie chart with eight slices where four of them are under 3%: all of these communicate something the data does not say. The person who made them may not have intended to mislead. The reader who acts on them is misled regardless.
The charts I publish represent one choice among many, and every choice carries an implicit argument about what the data shows. Making those choices deliberately, and refusing to make the ones that trade accuracy for persuasion, is what I mean by honest visualization.
The y-axis that starts where the data starts
For bar charts, the y-axis starts at zero. This is not a style preference: it is required by the encoding. A bar's visual length is proportional to its value, and that proportionality only holds when zero is at the baseline. A bar chart with a y-axis starting at 94 and ending at 100 makes a difference of 2 units look like it represents 30 units of the bar's height. The reader's eye measures the bars and draws the wrong conclusion.
For line charts, starting at zero is not always required. A line chart encodes change over time through the slope of the line, and the relevant question is whether the slope is perceptible at the natural data range. A vital sign trending from 97.2 to 98.8 over three hours is more clearly visible with a y-axis from 96 to 100 than one from 0 to 200, and the scale does not distort the encoding because line position encodes absolute value while bar length encodes it. The rule: include zero when the chart encodes magnitude through length or area. Truncating is acceptable for lines and scatter plots when the data range is far from zero and the within-range variation is what matters.
When truncating a line chart axis, label the break explicitly. An unlabeled truncated axis is a trap. A labeled one is an honest visual decision that the reader can account for.
Choosing the right chart for the right claim
Every chart type is appropriate for specific kinds of comparisons and inappropriate for others. Using the wrong chart type does not just look odd: it makes the wrong comparison salient and the right one invisible.
| Claim type | Right chart | Common wrong choice |
|---|---|---|
| Part-to-whole composition | Stacked bar, treemap | Pie chart with more than 4 slices |
| Ranking and comparison | Horizontal bar (sorted) | 3D bar, exploded pie |
| Change over time | Line chart | Bar chart with many time periods |
| Distribution of values | Histogram, box plot, violin | Mean-only bar chart hiding variance |
| Relationship between two variables | Scatter plot | Dual-axis line chart |
| Geographic distribution | Choropleth (rate), bubble map (count) | Choropleth for raw counts (area bias) |
The pie chart entry deserves elaboration. Humans are poor at estimating angles. A pie chart with four slices where the values are 48%, 28%, 15%, and 9% is readable because the dominant slice is visually obvious. The same chart with eight slices, where five of them range between 4% and 11%, is not readable: the reader cannot reliably order the slices by size without reading the labels, which means the visual is doing nothing the table wasn't already doing. A sorted horizontal bar chart makes the ranking immediately visible and allows comparison of similar values that a pie chart obscures.
Distortions I refuse to make
Dual-axis charts with mismatched scales. A chart with revenue on the left axis and customer count on the right axis can be made to show any visual relationship between the two series by adjusting the scales. Want them to move together? Scale the right axis until they do. Want them to diverge? Scale it the other way. The apparent relationship is a product of the axis choices, not the data. If two series need to be compared, index them both to a common baseline (100 at a reference period) or plot them separately with clear annotations linking the panels.
3D charts. Three-dimensional bar charts and pie charts introduce perspective distortion: bars at the back of a 3D chart look shorter than bars of equal height at the front. They add visual complexity without adding information. They exist to make simple data look more impressive, which is a different thing from making it more informative.
Cherry-picked time ranges. A trend line that starts at the trough and ends at the peak looks like growth. The same trend line extended to include the year before the trough and the year after the peak might look flat. The time range should be the one that is analytically relevant to the claim, not the one that makes the claim look strongest.
Misleading aggregations that hide variance. A bar chart showing mean revenue by segment with no error bars or distribution context is a summary that hides the shape of the data. A segment with mean revenue of $1,200 and a standard deviation of $80 is a different business situation from one with mean revenue of $1,200 and a standard deviation of $800. The mean-only chart looks identical for both. Box plots, violin plots, or at minimum confidence interval bars show what the mean alone conceals.
Annotation as honesty
The most underused tool in honest visualization is the annotation: a text label on the chart that explains what the reader is seeing. A line chart of monthly revenue with an annotation at the point where a pricing change took effect gives the reader the context they need to interpret the slope before and after that point. Without the annotation, the reader invents their own explanation, which may be wrong.
Annotations are also where limitations live. A chart derived from a sample should say it is derived from a sample. A metric that excludes a category of data should say what is excluded and why. A trend based on six months of data should not be presented the same way as one based on five years. The limitation is part of the claim the chart makes, and the claim is incomplete without it.
The goal of a chart is not to be beautiful or impressive. It is to give the reader an accurate understanding of what the data says, faster than a table would. Every choice that sacrifices accuracy for impressiveness fails that goal, regardless of how clean the underlying data was. Data integrity ends at the source only if the chart earns the trust the data built.