Мазмұнға өту

Статистика · Бастапқы

Use scatter, line and paired plots for the right relationships

Choose among scatter, line and paired displays to preserve the relationship in the data.

10 мин оқу · 2026 ж. 26 қаз.

Бұл бет әлі аударылмаған, сондықтан ағылшын нұсқасын оқып отырсыз.

Негізгі тұжырымдар

  • Show numeric association.
  • Use meaningful horizontal ordering.
  • Preserve actual pairing.
Мазмұны

Choose among scatter, line and paired displays to preserve the relationship in the data.

Scatter, line and paired plots can contain almost the same marks while answering different scientific questions. A scatter plot asks how two measurements vary together. A line plot makes an ordering, often time, visible. A paired plot shows corresponding values from the same unit under two conditions. Choosing among them depends on the design and the claim, not on a software menu.

This tutorial explains how to construct and interpret each display, preserve matching, distinguish association from agreement and write a figure legend that readers can evaluate. The controller values discussed below are synthetic teaching data, not observations from a tested robot.

First decide what one point means

Imagine a spreadsheet with columns for participant identifier, date, study time and test score. A point could represent a participant's total study time and final score. Alternatively, it could represent one daily observation from a participant. These plots answer different questions, and the second contains repeated observations that are not independent people.

Write the unit in your planning notes: “one point equals one participant at the final visit”. Then specify the horizontal and vertical measurements and their units. Do this before fitting a line. A graph cannot establish the independence assumptions needed by a statistical model.

For a social survey, distinguish personal income from household income. For a scientific experiment, distinguish a culture mean from one cell. For an engineering benchmark, distinguish one independently generated scenario from another random run in the same scenario. The count of marks and the independent sample size may differ.

Read a scatter plot before calculating a correlation

A scatter plot places paired numeric measurements on two axes. Inspect direction, shape, spread, subgroups, gaps and unusual points. Overplotting occurs when several observations occupy the same or nearby positions, hiding how many values are present. Transparency, smaller marks, carefully disclosed jitter or bins can help. NIST's scatter plot guidance describes its use for investigating associations and their shapes. S32

A positive linear association means larger values of one variable tend to accompany larger values of the other along an approximately straight pattern. Heteroscedasticity means the spread of outcomes changes across the predictor range. A fan shaped cloud might suggest this. A curve might indicate that a straight line is a poor description. A single influential point can strongly change a fitted slope without representing the dominant pattern.

Do not connect observations merely in spreadsheet order. If the rows happen to be sorted by score, connecting them can manufacture a striking trajectory that no participant actually followed. Use separate shapes or panels for meaningful groups, but explain whether those groups were defined before examining the result.

Association is not agreement

Suppose, hypothetically, a new sensor always reads exactly 5 units above a reference instrument. The measurements can have perfect positive correlation while disagreeing systematically. A fitted regression line does not answer whether the instruments can be used interchangeably.

An equality line is the reference line y = x. Values below it have a smaller vertical than horizontal measurement. It differs from a regression line estimated from the data. For instrument comparisons, inspect the equality line and consider a plot of differences against measurement averages. Bland and Altman's original work explains why correlation is insufficient for assessing measurement agreement. S403

An agreement analysis also needs a scientifically acceptable error range. A difference of 5 units may be trivial for one application and unacceptable for another. Repeated measurements, proportional bias and changing variability require appropriate extensions; do not apply a simple agreement formula automatically to a complex dataset.

In the controller illustration, A is on the horizontal axis and B on the vertical axis. Every point below the equality line indicates lower B error for that synthetic case. This is a performance comparison, not a claim that the controllers are interchangeable measurement instruments.

Use line plots for meaningful sequences

A line implies that movement from one horizontal position to the next is meaningful. Time, distance, dose and ordered simulation settings can justify it. Unrelated subject categories usually do not. If measurements occur at 0, 2 and 10 hours, their horizontal distances should reflect those times unless the axis explicitly uses a different representation.

Separate observations from fitted values. A straight segment joining two measured points is interpolation, not evidence that all intermediate values were measured. A smooth curve may come from a model or smoother. State which. Avoid extrapolating a line far beyond the observed range without a substantive model and clear uncertainty.

For a hypothetical materials test, suppose load is increased at defined intervals and displacement is recorded. A line can show the sequence of measurements. If different specimens contribute at each load because testing is destructive, the figure must not imply that the same specimen was followed. A summary line can still be useful if the legend identifies independent groups at each setting.

For a social science panel study, show how many participants remain at each wave. A rising average can reflect genuine improvement, changing sample composition, or both. If weaker performers are disproportionately missing later, connecting the means does not by itself demonstrate improvement within individuals.

Preserve genuine pairing

A pair consists of measurements linked by the design, such as the same participant before and after an intervention, two sensors reading the same specimen, or two controllers evaluated in the same layout. The connection must be established from identifiers or a documented matching procedure. Two rows occupying the same spreadsheet position are not sufficient.

A paired plot places conditions on the horizontal axis and connects values from the same unit. The slope and direction of each segment reveal change. With many units, an additional plot of within unit differences may be easier to read. Keep the direction of subtraction consistent throughout the paper.

Three synthetic examples show an A versus B scatter plot with an equality line, a time series with observed points and a paired condition plot.

For the twelve supplied synthetic controller cases, define difference as A minus B. The differences are 4, 3, 4, 4, 3, 4, 3, 2, 5, 5, 2 and 4 mm. Their sum is 43 mm; dividing by twelve gives 3.58 mm after rounding. Every difference is positive. The mean alone would not tell readers whether this consistency existed, whereas the paired plot makes it visible.

The scatter plot and paired plot therefore reveal complementary features. The scatter plot shows whether layouts difficult for A also tend to be difficult for B. The paired plot emphasises the direction and size of change within each layout. Use both only if both aspects matter to the argument.

Add summaries without concealing variation

A group line or fitted curve should be visually distinguishable from individual observations. Use the caption to state whether a shaded band is a confidence interval for an average response or a prediction interval for a new observation. Neither is the same as the spread of all observed values.

For clustered or repeated measurements, the calculation must respect the structure. Sampling individual time points as though they were independent participants can create misleadingly narrow bands. If you use a bootstrap, for example, the resampling unit must match the inferential design; describing a band simply as “bootstrapped” is incomplete.

Avoid visual overload. With hundreds of individual trajectories, consider a small number of clearly justified summaries, separate panels or a distribution of changes. Disclose any selection of “representative” trajectories and use a documented selection rule. Do not show only the trajectories with the strongest response.

Work through a complete reporting example

Here is an inadequate statement about the synthetic controller data:

The scatter plot proves that B improves navigation. The results are highly correlated, and all runs were significant.

The sentence confuses correlation with improvement, makes a causal empirical claim from constructed values and does not define a test. A stronger reporting paragraph is:

The synthetic data contain twelve matched layout cases, with positioning error recorded for each controller. Every B value lies below the equality line in the A versus B scatter plot. Within case differences, defined as A minus B, range from 2 to 5 mm. The mean paired difference is 3.58 mm, with a t based 95% confidence interval of 2.95 to 4.22 mm. These constructed values illustrate consistent reporting of a paired comparison and do not establish performance in a real navigation task.

The first sentence defines the design, the second describes the graphic, the third describes variability in change and the fourth supplies an estimate and uncertainty. The final sentence limits the claim to what the data actually represent.

A corresponding legend is:

Each point represents one synthetic layout case. Horizontal and vertical coordinates show positioning error in millimetres for controllers A and B. The diagonal is the equality line, not a fitted model. In the paired view, lines connect the two values from the same case. All twelve cases are shown. The illustrations use constructed teaching values.

This legend does not substitute for the Methods. A real paper must additionally explain how layouts were generated, whether order was randomised, how failures were treated and what constituted an independent case.

Apply the same logic across disciplines

For a hypothetical ecology project relating rainfall to seedling survival, a scatter plot can show an association, but region, soil and temperature may also matter. A fitted curve should not be described as the effect of rainfall unless the design supports that interpretation.

For a hypothetical education project with baseline and final scores, a paired plot can show within student change. Without a suitable comparison, however, improvement may reflect practice, maturation or other changes. A useful result describes the change while the discussion evaluates these alternatives.

For a hypothetical battery experiment, line plots can show voltage over time, but comparing thousands of time points from one battery with thousands from another does not provide thousands of independent batteries. Label device identities and report the number of independently tested devices.

Practice and model answer

Task: Fifteen students have baseline scores, but only twelve have final scores. You plot two means using all available scores and join them. Can the connecting line be described as the average individual improvement?

Model answer: No. The baseline and final means use different sets of students. First identify the twelve complete pairs and show their within student changes for a descriptive paired analysis, while reporting the three missing final scores and investigating missingness. An appropriate model may use additional observations under stated assumptions, but its estimate must be described accurately. The complete pair result is also vulnerable to selection if missing students differ systematically.

Task: Two instruments show a correlation of 0.99. Is that enough to claim agreement?

Model answer: No. Inspect systematic differences, how differences vary across the measurement range, and whether those differences are acceptable for the intended use. A strong correlation can coexist with a constant or proportional measurement bias.

Use Template 11 to record the unit, pairing, reference line and uncertainty definition before finalising the figure.

Sources

S32: National Institute of Standards and Technology. Scatter Plot. NIST/SEMATECH e Handbook of Statistical Methods, section 1.3.3.26. Accessed 11 November 2026.

S403: Bland JM, Altman DG. 1986. Statistical methods for assessing agreement between two methods of clinical measurement. The Lancet 1(8476), pages 307 to 310. PMID: 2868172.