Why fraction a design?
A fractional factorial design runs only a carefully selected fraction of the runs in a full factorial — written 2^(k-p), where p is the degree of fractionation. The motive is economy: a full factorial design doubles in size with every factor, so a 2⁵ study already costs 32 runs and a 2⁷ study costs 128. Most of those runs are spent estimating high-order interactions that are almost never important in biology.
Fractioning reclaims that wasted effort. A half-fraction (2^(5-1)) runs a 2⁵ study in 16 runs; a quarter-fraction (2^(5-2)) does it in 8. The design still estimates every main effect and the important low-order interactions — it simply gives up the ability to separate some effects from one another. That sacrifice has a name: aliasing, and it is the central concept you must understand before trusting a fractional result.
The sparsity-of-effects principle is what makes this work: in most systems, a small number of main effects and two-factor interactions dominate, while three-factor and higher interactions are negligible. Fractional factorials deliberately confound the effects you care about with the ones you do not.
Lay out a fraction in seconds
The free DOE generator builds 2^(k-p) fractions, reports the resolution and alias structure, and randomizes the run order for you.
Resolution III, IV, V explained
Resolution grades how severely a fractional factorial confounds its effects — the higher the resolution, the cleaner the design. It is defined by the length of the shortest "word" in the defining relation (explained below), and it tells you at a glance which effects you can trust.
| Resolution | Main effects aliased with… | Two-factor interactions aliased with… | Use when |
|---|---|---|---|
| III | two-factor interactions | main effects | Screening many factors; main effects only |
| IV | three-factor interactions (safe) | other two-factor interactions | Screening when some interactions matter |
| V | four-factor interactions (safe) | three-factor interactions (safe) | Near-optimization; clean interactions |
In a resolution III design, a main effect is aliased with a two-factor interaction — risky, because a real interaction can disguise itself as a main effect. These are cheapest and suit pure screening, where you assume interactions are negligible. A resolution IV design protects every main effect from two-factor interactions but lets two-factor interactions alias each other. A resolution V design keeps both main effects and two-factor interactions clear, confounding them only with three-factor and higher interactions — effectively as good as a full factorial for practical purposes.
Aliasing & confounding (the catch)
Aliasing (also called confounding) is the price every fractional factorial pays: two or more effects share a single column of the design, so their estimates are summed and cannot be told apart. Whatever you read off that column is the sum of all its aliases — if a value is large, you cannot be certain which aliased effect produced it.
Where do the aliases come from? When you fractionate, you build the extra factor from an interaction of the existing ones — for example, in a 2^(4-1) design you set D = ABC. That generator implies a defining relation, found by multiplying both sides by D: I = ABCD (since D×D = I, the identity). Every alias pair is obtained by multiplying an effect by the defining relation.
Reading this map: each main effect (A, B, C, D) is aliased only with a three-factor interaction, which the sparsity principle says is almost certainly negligible — so the main effects are trustworthy. The two-factor interactions, however, come in confounded pairs (AB+CD, AC+BD, AD+BC). If AB looks large you cannot tell whether it is really A×B or C×D without a follow-up run. That is the defining feature of a resolution IV design.
Choosing your fraction (2^(k-p))
Choosing a fraction means trading runs against resolution. The more you fractionate (larger p), the fewer runs — and the lower the resolution. The right choice depends on whether you are screening or optimizing and how many factors you have.
| Factors k | Design | Runs | Resolution | vs full 2^k |
|---|---|---|---|---|
| 4 | 2^(4-1) | 8 | IV | 16 → 8 |
| 5 | 2^(5-1) | 16 | V | 32 → 16 |
| 5 | 2^(5-2) | 8 | III | 32 → 8 |
| 6 | 2^(6-2) | 16 | IV | 64 → 16 |
| 7 | 2^(7-3) | 16 | IV | 128 → 16 |
| 7 | 2^(7-4) | 8 | III | 128 → 8 |
The practical rule: for screening, where you only need to find which factors matter, a resolution III or IV fraction is ideal — pair it with a follow-up if a confounded interaction turns out to be important. For work near an optimum, choose resolution V (or a full factorial) so interactions are clean. When five or more factors are in play, screening first with a fraction and then running a small full factorial on the survivors is almost always the most efficient path, as described in our DOE for bioprocess optimization workflow.
Order table & DOE fractional plan (randomized run order)
Once the fraction is chosen, you need an order table doe fractional plan: the design matrix laid out with an explicit, randomized run order. Building a design in standard order (the tidy Yates pattern) is convenient on paper, but running it that way is dangerous — any slow drift in the lab gets aligned with a factor and biases its estimate.
Randomization breaks that link. By shuffling the run order, drift from a warming incubator, an aging feed stock, or operator fatigue is spread evenly across factor levels, becoming noise rather than bias. A good doe generator and randomizer produces both the standard-order design (for analysis) and the randomized run sheet (for execution) at once.
| Std order | A | B | C | D=ABC | Run order |
|---|---|---|---|---|---|
| 1 | − | − | − | − | 4 |
| 2 | + | − | − | + | 1 |
| 3 | − | + | − | + | 7 |
| 4 | + | + | − | − | 2 |
| 5 | − | − | + | + | 8 |
| 6 | + | − | + | − | 3 |
| 7 | − | + | + | − | 6 |
| 8 | + | + | + | + | 5 |
Note how column D follows the generator D = ABC exactly — each D sign equals the product of the A, B, and C signs in that row. The "Run order" column is the order you actually execute the eight runs on the bench.
Worked example
Suppose you are screening four upstream factors for their effect on titer in a microbial fermentation: A = temperature, B = pH, C = feed rate, and D = agitation. A full 2⁴ factorial would need 16 runs; the 2^(4-1) half-fraction above does it in 8 — a 50% saving.
Setting up and analyzing the fraction
1. Generator and resolution. With D = ABC, the defining relation is I = ABCD — the shortest word has length 4, so this is a resolution IV design. Main effects are clear of two-factor interactions; the two-factor interactions alias in three pairs (AB+CD, AC+BD, AD+BC).
2. Run it randomized. Execute the 8 runs in the randomized run order from Table 3, not standard order, so any drift in the feed or temperature control does not bias a factor.
3. Estimate effects. Each effect is the mean response at its + level minus the mean at its − level — identical arithmetic to a full factorial design, just on 8 rows. The four main-effect columns give clean estimates of A, B, C, and D.
4. Interpret aliases. Suppose the AB column comes back large. Because AB is aliased with CD, you cannot yet tell whether the driver is temperature×pH or feed×agitation. The fix is a cheap follow-up: a few extra runs (a "fold-over" or a targeted full factorial on the two suspect pairs) de-aliases them.
Result: in 8 runs you have ranked four factors and flagged one interaction worth resolving — the information a 16-run full factorial would give, at half the bench cost, with the single caveat that one interaction needs a confirmatory run.
Build this fraction and its order table free
The DOE matrix generator returns the 2^(k-p) design, its resolution and alias structure, plus a randomized run sheet.
Generate one free
You do not need a licensed package to design a fractional factorial. A free doe matrix generator picks a sensible generator for your factor count, reports the resolution and the full alias structure, and outputs the order table doe fractional plan — the randomized run sheet you take to the bench.
Our free DOE software works as a doe generator and randomizer: choose the number of factors, set the levels, and it builds the fraction, shows what is confounded with what, and randomizes the run order — all in the browser. When you are ready to resolve interactions, it builds full factorial and response-surface designs too. For the bigger picture of how screening fits with optimization, see our guide on designing a DOE for bioprocess optimization.
Frequently Asked Questions
What is a fractional factorial design?
A fractional factorial design runs only a carefully chosen fraction of a full 2^k factorial — a 2^(k-p) design. For example, a half-fraction of a 2⁵ design (32 runs) uses just 16 runs, and a quarter-fraction uses 8. The trade-off is aliasing: some effects become confounded, so the design estimates the most important effects efficiently while sacrificing the ability to separate certain interactions.
What does aliasing or confounding mean in DOE?
Aliasing (confounding) means two or more effects share the same column in the design, so their estimates are added together and cannot be separated. In a resolution IV fractional factorial, main effects are aliased with three-factor interactions (usually safe), but two-factor interactions are aliased with each other. The alias structure follows directly from the design's defining relation.
What is design resolution (III, IV, V)?
Resolution describes how badly effects are aliased, set by the shortest word in the defining relation. Resolution III aliases main effects with two-factor interactions (cheapest, riskiest). Resolution IV keeps main effects clear of two-factor interactions but aliases two-factor interactions with each other. Resolution V keeps main effects and two-factor interactions all clear, aliasing them only with three-factor or higher interactions.
Why randomize the run order in a fractional factorial?
Randomizing the run order spreads any time-related drift — a drifting feed pump, a slowly warming room, batch-to-batch media variation — evenly across the factor levels so it does not masquerade as a real effect. The standard order is convenient for building the design, but the actual experiments should follow a randomized run order table, which a DOE generator and randomizer produces automatically.
How do I generate a fractional factorial plan for free?
A free DOE matrix generator builds the 2^(k-p) design, shows the alias structure and resolution, and outputs a randomized order table — the DOE fractional plan — ready for the bench. The BioProcess Tools DOE generator does this in the browser with no software install or licence.
Related Tools
- DOE Experiment Generator — A free doe matrix generator and randomizer for fractional, full factorial, and screening designs.
- E. coli Expression Optimizer — Identify expression factors worth screening in a fractional factorial.
- Fed-Batch Feed Strategy Calculator — Set feed-rate factor levels for a screening design.
References
- Montgomery, D.C. (2017). Design and Analysis of Experiments, 9th ed. Wiley. Chapter 8 (two-level fractional factorial designs).
- NIST/SEMATECH (2012). e-Handbook of Statistical Methods, Section 5.3.3.4: Fractional factorial designs. DOI: 10.18434/M32189
- Box, G.E.P., Hunter, J.S. & Hunter, W.G. (2005). Statistics for Experimenters: Design, Innovation, and Discovery, 2nd ed. Wiley.