How many combinations are in a full design with five two-level factors?

2^5 = 32, before replication or added runs.

Can the response be continuous?

Yes. Two-level refers to factor settings, not response values.

Do center points fully fit a quadratic model?

No. They can detect curvature without identifying a general quadratic response surface.

AI Glossary / Statistics and probability / Intermediate

2-level designs

Definition

Two-level designs are experimental designs in which each studied factor is assigned two specified settings, often called low and high. A full factorial design tests every combination and requires 2^k combinations for k factors. Fractional designs use a carefully selected subset. Two-level designs can estimate effects efficiently, but the chosen design determines which effects can be distinguished and what assumptions are needed.

What it means

A factor is an input deliberately varied, such as processing temperature, a software setting, or the presence of a feature. Its two levels can be numerical values or categories. Low and high describe the selected settings, not necessarily the full range of possibilities.
Varying several factors together makes it possible to investigate interactions: the effect of one factor may depend on another. This differs from changing one factor at a time, which can miss those dependencies. Two levels do not mean two factors, two observations, or a binary response.

How it works

Choose a response measure and factors relevant to the experimental objective. Select meaningful, feasible levels and define how each run will be conducted. Avoid levels so close that important effects are difficult to detect or so far apart that the comparison no longer serves the intended question.
For a full factorial, enumerate all combinations. A fractional design can reduce the number of runs, but examine its alias structure before collecting data: some effects are inseparable without additional assumptions or runs.
Randomize run order where appropriate, use blocking for known nuisance variation, and plan replication to assess variability. Repeated measurements of one experimental unit are not automatically equivalent to independent replicate runs.
Analyze main effects and interactions in the context of the design. For quantitative factors, two endpoint levels alone cannot identify separate quadratic curvature along each factor. Center points can help detect overall curvature, but more design information is needed to fit a general quadratic response surface. Confirm useful findings through appropriate follow-up experiments.

Three practical examples

A manufacturing team compares two temperatures and two mixing speeds while checking whether their effects interact.
A software team varies two settings for several components to screen their effects on a defined performance measure.
A research team uses a fractional two-level design to explore many candidate factors before conducting a more detailed follow-up study.

Think of it this way

Think of testing the corners of a box whose dimensions are the factors. With two factors there are four corners; with three there are eight. The corners reveal how selected settings work together, but they do not fully describe every curve or optimum inside the box.

Worked example

Consider an invented full factorial experiment with factors A and B, each at low and high levels. The measured response is 10 at A-low/B-low, 14 at A-high/B-low, 12 at A-low/B-high, and 20 at A-high/B-high.
The average response at A-high is (14 + 20) / 2 = 17, and at A-low it is (10 + 12) / 2 = 11. A's main effect is therefore 6. B's main effect is (12 + 20) / 2 − (10 + 14) / 2 = 16 − 12 = 4.
A's simple effect is 14 − 10 = 4 when B is low, but 20 − 12 = 8 when B is high. The difference of these simple effects is 4, demonstrating interaction. Under the conventional two-level factorial effect definition, the AB interaction effect is half that difference, namely 2.
With levels coded −1 and +1, the fitted corner-value model is y = 14 + 3A + 2B + AB. Its coefficients are half the corresponding factorial effects. This four-run example has no replication, so fitting all four coefficients leaves no residual degrees of freedom for estimating random error. The numbers illustrate effects, not statistical significance.

Illustrative case study

In this fictional case, Birch Analytics evaluates two data-processing settings at two levels each. A preliminary one-factor comparison suggests that one setting has little benefit, but the full four-combination experiment reveals that its effect depends on the other setting.
The team repeats runs under a planned protocol and records variation rather than relying on a single result per combination. It also checks whether the response behaves differently at intermediate numerical settings before making a broader recommendation.
The final report distinguishes the measured factor effects, their uncertainty, and the range over which the experiment supports conclusions. It does not extrapolate beyond the tested region without evidence.

Common misconceptions and mistakes

Two-level designs are not automatically full factorial designs. A fractional design can confound effects, so an apparently large main effect may partly represent an aliased interaction.
Another mistake is interpreting regression coefficients under −1/+1 coding as the full low-to-high effects. They differ by a factor of two. Interaction conventions also require care, as the worked example shows.
Finally, endpoint comparisons cannot establish that a response is linear throughout the interval. Center points can reveal evidence of curvature but do not by themselves identify every quadratic term. Randomization and replication remain important even when the design matrix looks balanced.

Business use

These designs are useful for screening influential factors and learning about interactions with a controlled number of runs. They can support process improvement and structured software or model experimentation.
Budget for valid execution, replication, and follow-up rather than counting only nominal combinations. Check that the response measures the business objective and that nuisance changes are not mistaken for factor effects. Use the design's actual resolution and alias structure when interpreting results.

Check your understanding

What is A's main effect in the example?

Answer: Six response units.

Why is there an interaction?

Answer: A's effect changes from four to eight depending on B.

Why can the saturated four-run model not estimate residual error?

Answer: All four observations are used to fit four coefficients.

Related terms

Factorial design
Main effect
Aliasing (factorial experiments)
Replication (statistics)
fractional factorial
interaction
aliasing
replication
randomization
center point

Learn next

Learn next how fractional factorial resolution controls aliasing. Study replication and blocking, then compare two-level screening designs with response-surface designs for investigating curvature.

Sources

NIST/SEMATECH e-Handbook, Two-Level Full Factorial Designs: https://www.itl.nist.gov/div898/handbook/pri/section3/pri3331.htm NIST, Fractional Factorial Designs: https://www.itl.nist.gov/div898/handbook/pri/section3/pri334.htm NIST, Response Surface Designs: https://www.itl.nist.gov/div898/handbook/pri/section3/pri336.htm

Previous: 10-fold cross-validation Next: 3D body scanning