OptiGein Design Docs
Home › Factorial designs

Factorial designs

Factorial designs answer the first question of any experiment: out of everything you could change, which factors actually move the result, and by how much. They run each factor at a small number of settings and vary them together, so a handful of runs tells you about many factors at once. This page covers the six designs in the Factorial family and the ideas you need to pick between them.

You reach all six from the New Design screen, listed under Factorial in the catalogue on the left. Select one and the panel on the right explains what it does and how many runs it costs before you build anything. If you are not sure which family you need at all, start with Which design do I need?

Resolution, and why the two-level grid is coloured

When you cannot afford to run every combination of factors, you run a fraction of them. The price of that saving is aliasing: some effects can no longer be told apart, because they were measured on the same set of runs. Resolution is the shorthand for how bad the aliasing is. A higher Roman numeral means cleaner effects. The three you will meet are these.

ResolutionWhat stays clearWhat is tangledUse it for
IIIMain effects.A main effect is aliased with two-factor interactions, so a big reading could be the factor or an interaction you cannot separate. A quick ruggedness or sorting check when you are willing to assume interactions are small.
IVMain effects are clear of every two-factor interaction. Two-factor interactions are aliased with each other, so you learn that an interaction exists but not which pair caused it. Screening a long list down to the few factors that matter.
VMain effects and every two-factor interaction are all clear of one another.Interactions only overlap three-factor and higher terms, which are usually negligible.Characterising a short list of factors, interactions included.

OptiGein Design paints this straight onto the two-level grid so you can see the trade at a glance. Green cells reach Resolution V or better, yellow cells are Resolution IV, and red cells are Resolution III. Full factorials sit in the leftmost column with no colour fill, because they alias nothing at all.

Regular Two-Level

This is the workhorse of the family. Every factor, from 2 up to 21 of them, runs at a low and a high setting. You choose the size on a grid whose rows are the number of factors and whose columns are the fraction, from the full factorial on the left through progressively smaller fractions to the right. Each cell shows the run count and its resolution colour, so you are choosing runs and aliasing together in one click.

The Regular Two-Level grid, colour coded by resolution
The two-level grid. Green is Resolution V or better, yellow is IV, red is III; the full factorials on the left have no aliasing.

Below the grid are four controls. Replicates repeats the whole design so you can estimate pure error. Blocks (1, 2, 4 or 8) splits the runs into groups when you cannot do them all under identical conditions, for example across two days. Center points per block adds runs at the midpoint of every factor. Show Generators reveals the defining relation (written I = ...) for a fractional design, which is the exact statement of what is aliased with what.

Read the aliases before you commit. A green full or fractional two-level design is often the right first step before a response surface study like the tablet example used elsewhere in this guide. Confirm which of your factors matter here, cheaply, then spend the larger run budget only on those.

Multilevel Categoric

Use this when a factor is a category rather than a dial, or when two settings are not enough. It takes 1 to 12 factors, each with its own list of levels (a supplier, a grade, three temperatures), entered as a comma-separated list. The design then crosses every level of every factor, so each combination becomes exactly one run and the effect of each level can be read straight from the averages in the run sheet.

The Multilevel Categoric configuration page
Multilevel Categoric crosses every level of every factor. Set the number of factors, replicates and blocks, then name the levels.

The run count is the product of the level counts, so it grows quickly: three factors at three levels each is already 27 runs before any replication. Replicates and Blocks work as they do for the two-level design. Because this is a full crossing, there is no aliasing to worry about, but the size is the limit rather than the resolution.

Plackett-Burman

When you have many two-level factors and you only want to know which ones matter, Plackett-Burman is the cheapest honest answer. It offers 12, 20 or 24 run sizes and handles up to 23 factors, which is why the number of factors you may enter is capped at one below the run size you pick. Every design is Resolution III.

The Plackett-Burman configuration page
Plackett-Burman. Choose the run size, then the number of factors; you can add replicates and center points.

The catch is baked into that Resolution III: main effects are partially aliased with two-factor interactions. You can conclude that a factor looks important, but if an interaction is really driving the response, Plackett-Burman will not tell you. Treat it as a sieve, not a final answer, and follow a promising result with a higher-resolution design on the survivors.

Min-Run Characterize (Resolution V)

A full Resolution V two-level design for many factors can run into hundreds of experiments. This design gives you the same clean reading of every main effect and every two-factor interaction for 6 to 20 factors, at close to the fewest runs the model allows. The software builds it by D-optimal coordinate exchange rather than from a textbook fraction, so it is not tied to a power-of-two size.

The Min-Run Characterize configuration page
Min-Run Characterize. The preview reports the run count for the full two-factor interaction model at Resolution V.

The run count equals the number of terms in the two-factor interaction model, which is 1 + k + k(k-1)/2 for k factors. That makes the plan saturated: at the minimum it has no spare runs to estimate error. This is why the Center points control matters here. Adding a few gives you replication for a pure-error estimate and a curvature check. Choose this design when you need interactions but a full Resolution V fraction is too expensive.

Min-Run Screen (Resolution IV)

This is the leanest screening design in the family, for 5 to 50 factors at roughly two runs per factor. It is built as a fold-over pair: a compact main-effects half plus its mirror image, which guarantees Resolution IV. Main effects come out clear of every two-factor interaction, while the two-factor interactions overlap one another.

The Min-Run Screen configuration page
Min-Run Screen. The Run option chooses between the bare minimum and minimum runs plus two.

The Run option offers minimum runs, or minimum runs plus 2. The plus-2 choice is the default because it lifts the fit away from saturation and gives a steadier estimate. You can conclude which factors have real main effects and that some interactions are present, but not which interactions. The software itself suggests you prefer an orthogonal Regular Two-Level fraction whenever you can afford one, and only reach for this design when you cannot.

Taguchi OA

The Taguchi orthogonal arrays are a fixed catalogue of balanced tables: L4, L8, L9, L12, L16, L18, L25 and L27. You pick the designation and the array decides the number of factors and their levels for you; the columns range from two-level (for example L8 holds seven two-level factors) to three and five levels. You only name the factors and their levels. Every column is balanced against every other, so main effects are estimated with equal precision.

The Taguchi OA configuration page with the array selector
Taguchi OA. The designation you pick fixes the number of columns and their levels; you fill in names.
A Taguchi array leaves no room for interactions. These arrays are saturated: every column carries a main effect, so nothing is left over to estimate a two-factor interaction. Any interaction that is really present will fold into the main-effect columns and quietly bias them. Read the alias notes on the design before you commit runs, and again when you read the results. If interactions are part of the question, choose a two-level design at Resolution IV or V instead.

Center points, replication and run order

Three settings show up across several of these designs and are worth understanding once.

Center points are runs with every numeric factor set to its midpoint. They do two jobs. Repeated, they give you a pure-error estimate that does not depend on the model being correct. Compared against the average of the factorial runs, they reveal curvature: if the centre sits well off the line between the corners, the response is bending. A two-level design can detect that curvature but cannot map it, so a strong signal is your cue to move up to a response surface design.

Replicates repeat runs so that random measurement scatter can be separated from real effects. More replication sharpens every estimate and is the honest way to earn degrees of freedom for the error term, at the cost of more runs.

Run order is randomised for a reason. Every design here fills in a randomised Run column and leaves the Randomize run order box ticked by default. Perform the experiments in that order. It stops anything that drifts over time, a warming room or an ageing reagent, from lining up with a factor and masquerading as its effect. Untick it only when the run order is genuinely fixed for you.