Can a fixed variable take several values?

Yes; the chosen levels can vary across experimental units.

Does fixed mean the response has no random error?

No. Responses can vary around the mean at each selected level.

Does choosing fixed levels prevent randomization?

No. Units can be randomly assigned to those levels, serving a different purpose from selecting the levels themselves.

AI Glossary / Statistics and probability / Intermediate

Fixed variable

Definition

A fixed variable, in the experimental-design sense used here, is an independent variable whose particular levels are specified by the researcher. For example, a study might deliberately compare training durations of 10, 20 and 30 minutes. The term concerns how the predictor's values enter the study and the scope of the comparisons. It does not mean that the variable has only one value, that outcomes are certain or that observations cannot be assigned randomly to the selected levels. Other disciplines may use fixed differently, so the statistical context matters.

What it means

The researcher chooses the levels because those settings are of interest. Differences among their responses can be represented by fixed parameters in an appropriate model. This differs from treating sampled factor levels as draws from a broader population whose variability is the main target. The distinction is especially important when interpreting a result: evidence about three named settings is not automatically evidence about every possible setting. A fixed numerical predictor can also enter a regression model, where the assumed response relationship determines what interpolation or extrapolation would require.

How it works

Define the independent variable, its levels and the outcome before collecting observations. Decide whether the levels should be represented as separate categories or through a justified numerical function. Randomize allocation where appropriate and arrange replication or blocking to address relevant sources of variation. In a fixed-effects one-way model, each selected level has a mean represented by the overall mean plus its effect, with residual uncertainty around observations. Estimate contrasts or a response relationship according to the intended question. Treating the levels as fixed does not remove assumptions about errors, independence, measurement or model adequacy. In multifactor designs, include and interpret interactions where justified.

Three practical examples

Training evaluation: compare three prespecified course durations and estimate differences among those durations.
Process development: study a chosen set of temperatures while randomizing experimental runs to reduce order-related confounding.
Interface research: compare two particular layouts as selected alternatives, rather than assuming they are randomly sampled representatives of all possible layouts.

Think of it this way

Think of placing marked stops on a test track. You decide which locations to examine, but repeated measurements at each stop can still vary. Choosing the stops does not make the measurements constant or tell you what happens everywhere between them.

Worked example

Suppose a hypothetical experiment uses fixed duration levels of 10, 20 and 30 minutes, with observed group means of seven, 11 and 12 outcome units. A categorical model can represent all three means separately. If an analyst instead fits an equally weighted straight line to these means, the mean duration is 20 and the mean response is 10. The least-squares slope is [(10 - 20)(7 - 10) + (30 - 20)(12 - 10)]/[(10 - 20) squared + (30 - 20) squared] = 50/200 = 0.25. The intercept is five. Predicted means are 7.5, 10 and 12.5, leaving residuals of -0.5, one and -0.5. The selected levels are fixed in both formulations, but the straight-line model imposes an additional shape assumption. These three means alone do not provide a within-group error estimate or justify a significance test.

Illustrative case study

Fictional case: Rowan Skills Lab selects three course lengths because those are the formats its scheduling team can offer. Participants are randomly allocated within suitable scheduling blocks, and outcomes are measured consistently. Analysts distinguish comparisons among the three actual formats from a proposed prediction for a much longer course. They explain that the latter requires further support for the response model. This invented case shows why selecting fixed levels and randomizing participants are compatible decisions, and why the scope of the conclusion should follow the design.

Common misconceptions and mistakes

Do not equate fixed with constant: a fixed factor can have several levels. Do not equate fixed levels with nonrandom assignment. A numerical label does not require a linear response relationship; temperatures or durations can be modeled categorically when appropriate. Conversely, a categorical factor is not automatically random merely because its labels are names. The fixed-versus-random distinction concerns the modeling target and selection of levels. Measurement error in a supposedly controlled predictor can still matter. Treating an observed predictor as fixed for conditional analysis does not itself establish causality.

Business use

Businesses can use fixed variables to make comparisons closely aligned with concrete operational choices. Record why each level was chosen, which population of units was studied and what comparisons were planned. Report effect estimates and uncertainty rather than relying only on an overall test. If the decision expands to unseen locations, operators or settings, reconsider whether the model and design support that generalization. A clear account of fixed levels helps prevent a limited experiment from being presented as a universal rule for every possible configuration.

Check your understanding

An experiment chooses two particular layouts and randomly assigns users between them. Are the layouts necessarily a random-effects factor?

Answer: No. Random assignment of users does not make the selected layout levels a random sample from a population of layouts.

Related terms

Fixed model Anova
Independent variable
Factor level
Fixed effects
random effects
treatment contrast
experimental design
random assignment
regression

Learn next

Learn the distinction between selecting factor levels and assigning experimental units. Then compare categorical and numerical representations of a predictor, and study how fixed, random and mixed-effects models answer different questions about the population of interest.

Sources

Howell, Statistical Methods for Psychology glossary, fixed variable: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html NIST/SEMATECH, The one-way ANOVA model and assumptions: https://itl.nist.gov/div898/handbook/prc/section4/prc432.htm

Previous: Fixed sample Next: Flap (phonetic)