Can twelve runs screen eleven factors?

Yes, in a saturated main-effects model, but that leaves no residual degrees of freedom.

Does the design estimate all interactions?

No. Interaction effects can be confounded with main effects and usually require additional experimental information.

Are low and high necessarily the minimum and maximum possible settings?

No. They should be meaningful, feasible settings chosen for the experimental question.

AI Glossary / Statistics and probability / Advanced

Plackett - Burman design

Definition

A Plackett - Burman design is a family of economical two-level experimental designs used mainly to screen many candidate factors for influential main effects. A design with N runs can accommodate up to N minus one factor columns, with the familiar run counts being multiples of four. Factor levels are commonly coded as -1 and +1. The efficiency comes with a trade-off: interactions can contaminate apparent main effects, so the design is most useful when interactions are plausibly small and follow-up experiments are planned.

What it means

Screening asks which variables deserve closer attention before investing in a more detailed study. For example, a team may have eleven possible process settings but limited capacity for experimental runs. A 12-run Plackett - Burman design can examine all eleven in a main-effects model. It cannot reveal the complete response surface or independently estimate every interaction. The factor columns are balanced and mutually orthogonal, but orthogonality between main-effect columns does not make them independent of every possible omitted interaction. That distinction is central to interpreting the results.

How it works

Choose plausible low and high settings, define a response and select a validated design matrix. Assign factors to columns and randomize the actual execution order where the operational constraints permit. Record all settings and experimental conditions, then estimate main effects using high-versus-low response contrasts or the corresponding regression. Check the design's aliasing structure and whether replication or unused columns provide a defensible error assessment. Rank promising effects cautiously and confirm them in a follow-up design capable of resolving important interactions. If factors cannot be independently changed or experiments are restricted, the design must represent those constraints rather than ignore them.

Three practical examples

Manufacturing: screen processing settings before a smaller response-surface experiment.
Laboratory methods: examine sensitivity to small changes in several operating conditions during a ruggedness study.
AI infrastructure: screen two selected levels of several configuration settings on a defined performance benchmark, while recognizing that software settings can interact strongly and that repeated benchmark runs may be needed to assess variability.

Think of it this way

Think of an initial survey that directs a more detailed investigation. It helps identify promising suspects with limited effort, but it does not establish every causal pathway. A setting that looks influential may partly be carrying the signal of combinations of other settings that the initial design cannot separate.

Worked example

Take the standard 12-run, 11-column design and create an original noiseless response y = 50 + 4A - 2B, with A and B coded -1 or +1. Each column has six observations at each level, and the columns are mutually orthogonal. The mean response at A = +1 is 54, while the mean at A = -1 is 46, giving a main-effect estimate of 8. For B, the corresponding means are 48 and 52, giving -4. All nine other main-effect contrasts are zero in this constructed additive example. The regression coefficients are half the high-minus-low effects: 4 and -2. A full two-level factorial for eleven factors would require 2^11 = 2,048 runs. However, fitting an intercept and all eleven main effects to just twelve observations uses all twelve parameters, leaving zero residual degrees of freedom. The attractive run reduction therefore does not supply an ordinary residual error estimate or demonstrate that interactions are absent.

Illustrative case study

In a fictional case, Summit Compute screens candidate batch-processing settings using a 12-run plan. A memory setting and a scheduling setting appear influential. Engineers know those settings might interact, so they avoid declaring either one universally optimal. They repeat selected conditions to assess run variation and design a follow-up experiment with the shortlisted factors. The final recommendation comes from that confirmation work, while the initial screening record explains the assumptions that made the economical first stage useful.

Common misconceptions and mistakes

A factor effect is not the same numerical quantity as its coefficient under -1/+1 coding. A saturated fit does not provide residual degrees of freedom merely because the software produces coefficients. Unused columns are not automatically pure measurement error: they can also carry interaction signals. Main-effect orthogonality does not remove interaction contamination. Two tested levels cannot establish a quadratic optimum or support unlimited extrapolation. Designs with power-of-two run counts have special regular-design relationships, so broad statements about partial aliasing should be checked against the particular matrix rather than assumed identical for every family member.

Business use

The business value is economical prioritization when experimentation is expensive. Use a clear response metric and a documented baseline so effect sizes have operational meaning. Reserve budget for confirmation instead of spending the entire experimental allowance on one saturated screen. Consider replication, blocking and execution order when results can change with time or equipment. For AI systems, a mean runtime improvement may hide a worsening error rate or tail latency, so track the outcomes needed for the decision. Screening evidence should guide the next experiment, not automatically become a permanent operating policy.

Check your understanding

A coded regression coefficient is 3 for one factor. What is its high-minus-low main effect in the additive model?

Answer: 6, because changing the code from -1 to +1 changes it by two units.

Related terms

Design of experiments
Main effect
Aliasing (factorial experiments)
Replication (statistics)
screening design
interaction
aliasing
factorial experiment
replication

Learn next

Study contrast coding and aliasing, then learn how replication and follow-up factorial designs help distinguish genuine main effects from interactions and experimental noise.

Sources

NIST/SEMATECH, Plackett–Burman designs: https://itl.nist.gov/div898/handbook/pri/section3/pri335.htm JMP, Plackett–Burman Designs: https://www.jmp.com/en/statistics-knowledge-portal/design-of-experiments/screening-designs/plackett-burman-designs

Previous: Placebo-controlled study Next: Plackett-Luce model