Reducing Overplotting With Jitter and Transparency

Why Points Can Hide Each Other

In a scatterplot, each observation is represented by a mark at its predictor and response values. When multiple observations have exactly the same coordinates, their marks lie on top of one another. Even when coordinates are only slightly different, large marks can overlap enough to conceal how many observations are present. The display may then make a dense location look like a single observation, or make the relationship’s pattern difficult to see.

Two common display adjustments address this problem. Transparency makes each mark partly see-through, so overlapping marks can create darker regions. Jitter adds a small, deliberate perturbation to plotted coordinates, separating marks that would otherwise coincide. Both adjustments change how marks are drawn; neither changes the recorded observations.

The goal is not to make every point look different at any cost. A useful adjustment reveals overlap while preserving the broad pattern and the scale’s meaning. If the adjustment is too strong, the display can suggest variation or relationships that are not in the recorded data.

Using Transparency to Show Dense Regions

A transparent point allows some of the background to show through. Where several marks overlap, their combined appearance is darker than that of one mark. This can reveal that a location contains more observations than nearby locations. The effect depends on the point opacity, the number of overlapping points, mark size, color, background, and rendering method. Darkness is therefore a visual cue to density, not automatically an exact count.

A simple calculation describes repeated marks with the same opacity. Assume the marks have the same color and opacity, are drawn over an opaque, uniform background, and use ordinary layered compositing. Let \(a\) be the opacity of one mark, between 0 and 1. The fraction of the background still visible after one mark is \(1-a\). After \(k\) identical marks at the same location, the background fraction is \((1-a)^k\), so the combined opacity is:

This result follows by multiplying the remaining background fraction by \(1-a\) for each additional layer. For example, after two layers the remaining fraction is \((1-a)(1-a)=(1-a)^2\). Repeating this step \(k\) times gives \((1-a)^k\); subtracting from 1 gives the fraction covered by the combined marks.

Lower opacity can help in a crowded plot, but it is not a complete solution. If marks are too small, overlapping areas may remain difficult to detect. If opacity is extremely low, isolated observations may nearly disappear. If opacity is high, a dense cluster can become an almost solid patch that hides its internal structure. The useful setting depends on the number and distribution of points and on the intended display size.

Using Jitter Without Changing the Data

Jitter is most useful when values are recorded on a limited grid, such as whole-number counts or rounded measurements, and several observations consequently share coordinates. For an observation with recorded coordinates \((x_i,y_i)\), a jittered display position can be written as:

Here \(u_i\) and \(v_i\) are small offsets used only for plotting. A common choice is to draw each independently from a uniform distribution centred at zero, with fixed bounds:

The widths \(w_x\) and \(w_y\) are expressed in the original variables’ units. This makes their size interpretable: an \(x\)-offset of 0.1 means 0.1 units of the predictor, not 10% of the plotting window. Offsets centred at zero avoid a systematic shift in one direction, but any individual displayed point still moves. For this reason, jittered positions should not be used as the recorded values for calculations or analysis.

Choosing an Adjustment That Preserves the Pattern

Jitter widths should be chosen in the variables’ units and in relation to the spacing between recorded values. If values on an axis lie on a regular grid with nearest-neighbour spacing \(d\), then a simple sufficient condition for the jitter bands around adjacent grid values not to overlap is \(w<d/2\). Each original value can move at most \(w\), so the two adjacent values can move toward one another by at most \(2w\). When \(2w<d\), their display ranges remain separated. This is a practical guideline, not a guarantee that every jittered plot will be easy to read.

There is no single width or opacity that works for every plot. Useful checks include whether the axis units make the perturbation easy to understand, whether neighbouring recorded values remain distinguishable, and whether the display still resembles the original pattern. It can also help to compare the adjusted plot with the unadjusted version, particularly when the amount of overlap is important to interpretation.

Choosing Between Transparency and Jitter

Transparency is often a sensible first choice when there are many points and the main aim is to show where points accumulate. It leaves each point centred at its recorded coordinates, although the darkness of overlapping marks is approximate evidence of density rather than a direct count. Jitter can be useful when many observations fall on a small number of exact grid locations and the individual repeated positions need to be made visible.

The two adjustments may be used together, but combining them does not remove the need for care. Jitter spreads marks away from their recorded coordinates; transparency helps overlapping marks show through one another. If jitter is too wide, the marks can obscure a genuine pattern. If opacity is too low, individual marks can vanish into the background. For genuinely continuous measurements with few exact repeats, jitter is usually unnecessary: it may suggest that the recorded coordinates are less precise than they are.

Use the original recorded coordinates whenever a question concerns the data themselves. If a point’s exact value matters, read it from the data or an unjittered display, not from a displaced mark. If the display settings use random offsets, using a fixed random seed can make the same jittered arrangement appear again when the plot is regenerated. That improves reproducibility of the picture; it does not make the jittered positions observed values.

Common Mistake: Treating Display Positions as Measurements

A common mistake is to look at a jittered point at \((4.12,7.06)\) and report that the recorded values were 4.12 and 7.06. That reasoning confuses the location of a drawn mark with the observation that it represents. The correction is to retain the original values and regard the offsets as a way of seeing how many marks overlap. Jitter may reveal repeated values, but it cannot supply extra precision or change what was measured.

A related mistake is to interpret a darker patch in a transparent scatterplot as an exact count. The correction is to describe it cautiously as a region where marks appear more concentrated, unless the plot uses a separate, explicit method for reporting counts. Mark size, opacity, and rendering affect the shade, so equal counts do not always look equally dark across different settings or plots.