AI Glossary / Statistics and probability / Intermediate

Range restrictions

Definition

Range restrictions occur when the observations available for analysis cover only a limited part of the variation present in the population of interest. The restriction can result from selection, eligibility criteria, measurement limits or deliberate sampling. It is especially important when interpreting correlations: restricting variation in one variable can change the observed association, often reducing it under familiar linear-model conditions. The effect is not universally a reduction, and its direction depends on the selection mechanism and underlying relationships.

What it means

A relationship observed among highly selected participants may differ from the relationship in all eligible people. For example, a study confined to individuals with similar scores contains less information about how outcomes vary across the full score range. The restricted sample is not necessarily wrong for its own population. The problem arises when its association is interpreted as though it described a broader population without examining the selection process and the purpose of the inference.

How it works

Identify the intended population and compare its available variation with the analyzed sample. Determine whether selection acts directly on a variable being studied or indirectly through another related variable. Examine plots and selection rules, not only the minimum and maximum observed values. Where a correction is contemplated, state the required assumptions and information about the unrestricted population. Consider collecting broader data or narrowing the claim. A numerical correction cannot recover missing relationships reliably when its assumptions are implausible.

Three practical examples

Example 1: An employer studies the relationship between an entry score and later performance using only hired applicants. Example 2: A manufacturer evaluates a process using only pieces that already passed an initial quality screen. Example 3: A learning study includes only advanced participants, limiting the variation in prior knowledge. These examples may involve more than simple direct restriction, because selection can depend on several variables and those variables can be related to the outcome.

Think of it this way

Think of judging a landscape through a narrow window. The visible part may look almost flat even though the full landscape rises substantially from one end to the other. The window does not alter the terrain, but it changes the evidence available about its shape. If the window is positioned according to another feature, the resulting view can be distorted in more complicated ways than merely appearing flatter.

Worked example

Construct ten observations by taking x values -2, -1, 0, 1 and 2, each twice, and setting y = x - 1 and y = x + 1 at each x. The mean of both variables is zero. Using population-style moments for this finite illustration, Var(x) = 2, Var(y) = 3 and Cov(x, y) = 2, so correlation is 2/square root of 6 = square root of 2/3, approximately 0.8165. Now retain only x = -1, 0 and 1, leaving six observations. Var(x) becomes 2/3, Var(y) becomes 5/3 and covariance becomes 2/3. Correlation is therefore square root of 2/5, approximately 0.6325. The slope pattern y = x plus the same balanced noise remains, but the restricted x variation reduces the correlation.

Illustrative case study

Fictional case study: Spruce Training evaluates an aptitude measure using only employees who cleared a high entry threshold. A modest observed correlation is initially interpreted as proof that the measure has little relationship with performance anywhere. An analyst documents the restricted sample and separates the question about current employees from the question about all applicants. The team seeks broader evidence before making the latter claim. This fictional case does not establish the validity of any assessment or justify an automatic correction.

Common misconceptions and mistakes

Common mistakes include treating a restricted-sample correlation as universal, assuming every restriction lowers correlation, and applying a correction formula without examining its assumptions. Selection on multiple variables can induce or reverse associations. Measurement error, nonlinear relationships and changes in outcome variability can complicate interpretation. A short observed range may also arise by chance in a small sample rather than from a deliberate restriction. Extrapolating a fitted relationship beyond the observed range requires additional justification.

Business use

Business use: Range restrictions matter in recruitment analytics, quality screening, customer segmentation and evaluations of specialized populations. Document who was eligible, who was selected and which values were excluded. Align conclusions with the population actually observed. When broader decisions are required, assess whether additional data can cover missing regions. Report uncertainty and selection limitations alongside association measures, and avoid treating statistical adjustment as a replacement for understanding how the sample was assembled.

Check your understanding

Why does the worked example's correlation fall despite an unchanged slope pattern?

Answer: The variation in x shrinks relative to the unchanged noise variation.

Would the same numerical correction necessarily work for applicants selected using several criteria?

Answer: No. That selection mechanism can differ substantially from direct restriction on a single variable.

Related terms

Selection bias concerns systematic differences induced by selection

Learn next

Recreate the ten-point example and compare its scatterplot with the restricted six-point plot. Calculate both correlation and slope to see why they are different summaries. Then study selection based on an outcome or a third variable. Before applying a range-restriction correction, identify the exact model assumptions and the information required about the target population.

Sources

Sources: University of Vermont, Statistical Methods glossary, Range restrictions: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html ; Penn State STAT 501, regression and correlation course materials: https://online.stat.psu.edu/stat501/ . The glossary supplies the restricted-variation context; the numerical construction independently demonstrates one attenuation mechanism. The training case and cautions about the scope of its inference are original explanations.

Previous: Range query (database) Next: Rank