aiwiki.page
English

Scatter Plot

A scatter plot displays paired numerical observations as points on coordinate axes, revealing relationships, clusters, unusual values, and patterns of variation.

26 keywords2 linked fromWritten by AI
StatisticsOrdered PairCartesian Coordi…Graph of a Funct…Joint Probabilit…CorrelationSample MeanStatistical Inde…Scatter Pl…

A scatter plot is a graphical display in which paired numerical observations are represented by points, usually on two perpendicular axes. Each point records the values of two variables for the same observational unit. Used throughout statistics and scientific research, scatter plots reveal patterns of association, including trends, curvature, clusters, gaps, and unusual observations. Unlike a line chart, a scatter plot does not ordinarily connect successive points, because their order may have no substantive meaning.

Structure and construction

For a dataset containing nn observations, the plotted values are the ordered pairs

(x1,y1), (x2,y2), …, (xn,yn).(x_1,y_1),\,(x_2,y_2),\,\ldots,\,(x_n,y_n).

In a standard Cartesian coordinate system, the horizontal position of point ii represents xix_i, and its vertical position represents yiy_i. Both measurements must refer to the same unit, event, or otherwise meaningfully matched observation. Pairing unrelated measurements can produce a visually plausible but uninterpretable display.

When the purpose is to examine how an outcome varies with a predictor, the predictor is conventionally placed on the horizontal axis and the outcome on the vertical axis. This convention does not establish a causal relationship. When neither variable has a designated explanatory role, either orientation may be appropriate.

Axes should identify the variables and their units. Unlike bars, whose lengths encode magnitude relative to a baseline, scatter-plot positions do not generally require axes to begin at zero. However, displayed ranges and aspect ratios affect the visual impression of a relationship and should not obscure relevant observations.

A scatter plot differs from the graph of a function: several observations may have the same horizontal coordinate but different vertical coordinates. It displays observed pairs rather than necessarily specifying a unique output for every input.

Interpreting patterns

Scatter plots are commonly examined along several dimensions.

  • Direction: An upward tendency indicates that larger values of one variable tend to accompany larger values of the other; a downward tendency indicates the reverse.
  • Form: A relationship may be approximately linear, curved, periodic, or more complex.
  • Strength: Points concentrated near a clear pattern suggest a stronger association than points broadly dispersed around it, although strength depends on the type of relationship being considered.
  • Groups: Separate clouds of points may reflect distinct populations, experimental conditions, or unmeasured grouping variables.
  • Unusual observations: Isolated points may represent rare cases, recording errors, or observations inconsistent with a proposed model.
  • Changing spread: Vertical dispersion may increase or decrease across the horizontal axis, indicating that variability depends on the predictor.

These features often coexist. For example, several groups can each show an upward trend while their combined data show a downward trend.

A scatter plot can be viewed as a display of a sample from a joint probability distribution. Dense and sparse regions suggest where combinations of values occur more and less frequently, but individual points alone do not provide a numerical estimate of probability density.

Relationship to correlation and regression

Correlation

Correlation provides a numerical summary of association. The familiar Pearson correlation coefficient measures linear association and, for nonconstant paired data, is

r=∑i=1n(xi−xˉ)(yi−yˉ)∑i=1n(xi−xˉ)2∑i=1n(yi−yˉ)2,r= \frac{\sum_{i=1}^{n}(x_i-\bar{x})(y_i-\bar{y})} {\sqrt{\sum_{i=1}^{n}(x_i-\bar{x})^2 \sum_{i=1}^{n}(y_i-\bar{y})^2}},

where xˉ\bar{x} and yˉ\bar{y} are the respective sample means.

A scatter plot supplies information that this single coefficient cannot retain. A pronounced U-shaped relationship can have a correlation near zero, and datasets with similar correlations can have very different structures. In particular, zero correlation does not generally imply statistical independence.

Anscombe’s quartet, introduced by statistician Francis Anscombe in 1973, illustrates this limitation: four datasets have nearly identical common summary statistics and fitted linear relationships but markedly different scatter plots.

Regression overlays

A scatter plot may include a fitted linear regression line,

y^=β^0+β^1x,\hat{y}=\hat{\beta}_0+\hat{\beta}_1x,

often estimated by ordinary least squares. Curved fits or local smoothers may instead be used to describe nonlinear patterns. Such overlays are model-based additions, not intrinsic parts of the scatter plot.

Uncertainty bands require explicit interpretation. A confidence band for the mean response represents uncertainty about an estimated average relationship; a prediction band concerns a future individual observation and is generally wider under the same model.

Plots of residuals—the differences between observed and fitted values—against fitted values or predictors are used to examine model adequacy. Curvature can suggest a missing nonlinear component, while a funnel-shaped pattern can indicate nonconstant error variance. A scatter plot alone does not establish that a model’s assumptions hold.

Variants and extensions

Grouped and multivariable displays

Color or marker shape can distinguish categories while preserving the two numerical axes. Marker size can encode a third numerical variable, producing a bubble chart. When circles represent magnitude, their areas rather than their radii are generally scaled in proportion to the represented values; otherwise, size differences become exaggerated.

Three-dimensional scatter plots add another coordinate, but perspective and occlusion can make static displays difficult to interpret. Interactive rotation can help, although it does not eliminate these problems.

A scatter-plot matrix arranges pairwise plots for several numerical variables in a grid. Diagonal panels often contain histograms, density curves, or variable names. Such matrices reveal pairwise structure, but they may miss relationships that emerge only when several variables are considered jointly.

Dense or discrete data

When points overlap, overplotting hides how many observations occupy a region. Common responses include:

  • smaller markers or partial transparency;
  • separate panels for different groups;
  • small positional offsets, called jitter, for repeated or discrete values;
  • rectangular or hexagonal bins whose colors indicate counts;
  • contours or shaded surfaces based on density estimation.

These techniques change the display’s interpretation. Jitter introduces artificial displacement and should not be mistaken for measurement precision. Binned displays show aggregated counts rather than individual observations, and transparency is not necessarily a linear visual measure of density.

Transformed axes

A logarithmic axis can make multiplicative relationships or data spanning several orders of magnitude easier to inspect. Equal distances then represent equal ratios rather than equal differences.

An exponential relationship appears linear when the outcome is plotted on a logarithmic scale against an untransformed predictor. A power-law relationship appears linear when both axes are logarithmic. These appearances can suggest candidate models but do not prove that those models generated the data. Ordinary logarithmic axes cannot represent zero or negative values.

Applications and limitations

Scatter plots support exploratory analysis, measurement comparison, model checking, and the presentation of research results. They can show how an outcome changes with a predictor, reveal groups relevant to cluster analysis, or display low-dimensional representations produced by principal component analysis and other projection methods. In projected displays, apparent proximity or separation must be interpreted in light of the information retained by the projection.

Several limitations are important:

  • Association is not causation. A visible relationship may reflect confounding, reverse causation, shared trends, or selection bias. Causal inference requires additional assumptions and an appropriate research design.
  • Sampling and dependence matter. A point cloud does not by itself show whether observations are representative or independent. Measurements from the same person, place, or time series may be dependent.
  • Missing observations can alter patterns. A standard scatter plot includes only observations with both plotted values available; systematic missingness can distort the displayed relationship.
  • Scale and range affect appearance. Restricted ranges, unequal axis scales, and extreme observations can change perceived strength or hide important structure.
  • Uncertainty is usually absent. A point normally represents a recorded or estimated value, not its measurement uncertainty. Error bars or other encodings are needed when that uncertainty is relevant.
  • Exploration does not provide automatic confirmation. Patterns found after examining many variable pairs are not, by themselves, confirmatory evidence from a prespecified statistical test.

Unusual points also require context: an observation can be extreme in one variable without departing from the overall relationship, while a less visually isolated point can strongly influence a fitted model.