Scatter Chart Maker
Two numeric columns produce a scatter plot; add a third and each point is sized by it, giving a bubble chart. Scatter plots are the standard way to look at the relationship between two continuous variables, and they are also the chart type most often over-interpreted.
How to use it
- Paste X and Y columns, optionally followed by a Size column.
- Look at the overall pattern before looking at individual points.
- Check for overplotting where points are dense.
- Choose a theme and export a high-resolution PNG.
Position is the strongest encoding available
A scatter plot places each observation by position on two common scales, which is the most accurately read visual encoding there is. That is why it remains the default for exploring a relationship: the reader is doing the thing human vision is best at.
It also shows every observation rather than a summary, which is its main advantage over a correlation coefficient or a fitted line. Anscombe quartet is the standard demonstration — four datasets with identical means, variances, correlations, and regression lines, and four completely different scatter plots. One is linear, one curved, one linear with a single outlier dragging the fit, and one a vertical stack with one distant point. Looking at the data catches all of that; the summary statistics catch none of it.
Both axes are auto-scaled to the data. Unlike a bar chart, a scatter plot has no requirement to include zero, because position rather than length carries the meaning and the reader is comparing points to each other rather than to a baseline.
Overplotting and what to do about it
With enough points, markers overlap and the densest regions saturate into a solid mass. At that stage the plot shows the outline of the data and says nothing about where within that outline the observations concentrate — which is usually the interesting part.
It gets worse with rounded or discrete data, where many observations share exactly the same coordinates and stack invisibly. A hundred points at the same location look identical to one point.
Several fixes are standard. Transparency lets density show through as darkness, and works up to a few thousand points. Smaller markers delay saturation. Jittering — adding a small random offset — separates coincident points where the data is discrete, at the cost of showing positions that are slightly wrong. Beyond roughly ten thousand points, binning into a heatmap or hexbin plot is more honest than any of these, because it shows density directly rather than approximating it.
Bubble size is area, not radius
When a Size column is present, values are scaled to marker diameters between 8 and 48 pixels. The important detail is that human perception reads the area of a circle, not its radius, and area grows with the square of the radius.
A circle with twice the radius has four times the area, so if size is mapped linearly to radius, a doubled value looks quadrupled. Correct implementations map value to area and derive the radius from its square root, which is what produces an honest comparison.
Even done correctly, size is a weak encoding. People compare areas poorly, and the practical resolution is coarse — you can reliably distinguish small, medium, and large, and not much more. Size is best used for a secondary variable providing context, with the two important variables on the axes where position does the work.
The 8 to 48 pixel range also compresses the extremes. A dataset spanning three orders of magnitude will have its smallest and largest values clipped toward the ends of the range, so the visual ratio understates the real one. Where the size variable is highly skewed, a log transform before plotting usually represents it better.
What a pattern does and does not establish
A visible relationship in a scatter plot establishes association, and association has several possible explanations of which causation is only one.
The variables may both be driven by a third factor. Ice cream sales and drowning deaths correlate because both rise with temperature. The causation may run the other way from the assumed direction, which is common in observational business data — a feature that engaged users adopt is not necessarily a feature that creates engagement. And with enough variables, apparent relationships arise by chance: testing twenty pairs at a 5 percent threshold produces about one spurious result on average.
Two structural traps are worth watching for. Selection effects can create or reverse a relationship — if the data only includes cases that passed some filter, the relationship within the survivors can be the opposite of the relationship in the population. And aggregated points can reverse a within-group relationship entirely, which is Simpson paradox: every subgroup trends one way and the pooled data trends the other. If your points are group averages, the pattern you are seeing may not hold for any individual.
The scatter plot is still the right tool for finding these questions. It is just not the tool that answers them.
At a glance
| Input format | X and Y columns, optional Size column |
|---|---|
| Marker sizing | 8 to 48 pixels, scaled by value |
| Axis scaling | Auto to data range, zero not required |
| Practical point limit | A few thousand before overplotting dominates |
| Export | PNG at 2× display resolution |
Frequently asked questions
What columns are required?
X and Y. A third Size column is optional and turns the plot into a bubble chart.
Why do my points merge into a solid mass?
Overplotting. Use transparency or smaller markers, and above roughly ten thousand points prefer a binned heatmap.
How is bubble size calculated?
Values map to marker diameter between 8 and 48 pixels. Perception reads area rather than radius, so highly skewed size data is better log-transformed first.
Does the axis need to include zero?
No. Position rather than length carries the meaning, and readers compare points to each other rather than to a baseline.
Does a visible trend prove causation?
No. A third variable, reversed direction, selection effects, or Simpson paradox can all produce the same picture.
Read more
Distribution and correlation — Summary statistics hide things that a plot of every observation does not. That is the entire argument for these two chart types.