How many experiments does a DOE need? Fewer than you fear, but only if you match the design to the job. The honest answer has two layers. First, every design has a fixed minimum — a formula that turns your factor count into a run count. Second, that minimum is only useful if the study has enough statistical power to detect the effect you actually care about. This guide gives you both: the DOE run counts for every common design, and the sample size DOE maths that tells you whether those runs can resolve a real effect. When you are done, plug your numbers into a free design of experiments calculator and watch the run count update as you configure factors and levels.
Run counts by design (comparison table)
The design you pick fixes the minimum number of runs. A two-level full factorial tests every combination, so it needs 2k runs for k factors; screening and fractional designs deliberately test only a slice; response-surface designs add axial and center runs to fit curvature. The table below is the fastest way to see the trade.
Run count is the total number of distinct experiments (before replication) a design requires. The whole art of DOE is buying the most information for the fewest runs.
| Factors (k) | Full factorial 2k | Half-fraction 2k−1 | Central composite | Box-Behnken | Definitive screening |
|---|---|---|---|---|---|
| 2 | 4 | — | 13 | — | — |
| 3 | 8 | 4 | 20 | 15 | — |
| 4 | 16 | 8 | 31 | 27 | 9 |
| 5 | 32 | 16 | 32 | 46 | 11 |
Screening designs sit outside this per-factor table because their whole point is to hold runs nearly flat as factors grow. A Plackett-Burman design comes in multiples of four: 8 runs screen up to 7 factors, 12 runs up to 11, and 20 runs up to 19. That is why Plackett-Burman is the go-to when you have a long list of candidate medium components or process parameters — the run count barely moves as the factor list grows.
The single most important thing to notice in Table 1 is how fast the full factorial column climbs. Each new factor doubles the runs. That exponential growth — and why fractioning tames it — is the subject of the next section, and it is worth seeing as a chart.
Figure 1. Full factorial runs (2k) explode with factor count, while a screening design keeps run count nearly flat — the core reason you screen before you optimize.
What drives run count
Three levers set how many experiments a DOE needs: the number of factors, the complexity of the model you want to fit, and how much you replicate. Understand these three and you can predict any design's run count without a lookup table.
Number of factors is the dominant lever. For a full factorial it is exponential (2k); for screening designs it is roughly linear. This is why the first strategic move in almost every study is to reduce the factor list by screening before you commit to an expensive full design — a point developed in our guide to choosing a DOE design.
Model complexity is the second lever. To estimate only main effects, two levels per factor suffice. To estimate two-factor interactions you need enough distinct combinations to keep them clear of the main effects (higher resolution). To estimate curvature — a quadratic optimum — you need at least three levels, which is why response surface methodology designs like central composite and Box-Behnken cost more runs than a two-level factorial. A useful floor: you need at least as many runs as terms in your model, and comfortably about twice as many for a trustworthy fit.
Replication is the third lever, and the only one you tune after choosing the design. Running the whole design twice doubles the runs but does not change what the model can estimate — it sharpens the estimates you already have and buys statistical power, which is the next section.
Power, effect size & α/β
DOE power is the probability that your experiment detects a real effect of a given size as statistically significant. It is the answer to the question the run count alone cannot answer: "if temperature really does matter, will my design actually catch it?" Power is 1 − β, where β is the false-negative (missed-effect) rate; the convention is to design for 80% power (β = 0.20), sometimes 90%.
Four quantities are locked together, and fixing any three fixes the fourth:
- α — the significance level, the false-positive rate you tolerate (usually 0.05).
- Power (1 − β) — the chance of detecting a true effect (target 0.80).
- Effect size δ — the smallest change in the response you consider worth detecting.
- Noise σ — the run-to-run standard deviation of the response.
For a two-level factorial where effects are estimated as the difference between two averages, the total number of runs N needed is approximately:
N ≈ 4σ2 (zα/2 + zβ)2 / δ2
Read it as a set of levers. Halving the effect you want to detect (δ) quadruples the runs. Halving the noise (σ) quarters them — which is why tightening assay precision often beats adding runs. Demanding higher power or a stricter α nudges the z-values up and adds runs more gently. This is the whole of sample size DOE reasoning in one equation.
Worked example: sizing a titer factorial for power
You are screening factors for mAb titer. Historical runs give a run-to-run noise of σ = 0.10 g/L. The smallest titer effect worth acting on is δ = 0.15 g/L. You want α = 0.05 (two-sided, zα/2 = 1.96) and 80% power (zβ = 0.84).
N ≈ 4 × (0.10)² × (1.96 + 0.84)² / (0.15)²
= 4 × 0.01 × (2.80)² / 0.0225
= 0.04 × 7.84 / 0.0225
= 0.3136 / 0.0225 ≈ 13.9 runs
Round up to 16 runs. A 24 full factorial (16 runs) clears this bar; an 8-run half-fraction would not — it would be underpowered for a 0.15 g/L effect at this noise, and a real effect could easily slip past. If runs were cheaper you might instead replicate an 8-run design twice (16 runs total) to reach the same power while keeping fewer distinct conditions.
See the run count before you commit
Configure your factors, levels, and design in a free DOE generator and read the exact run count instantly — then compare designs side by side to fit your budget.
Replication & center points
Replication buys power; center points buy an error estimate and a curvature check — they are not interchangeable. This is the most common run-budgeting mistake: adding center points hoping to strengthen the main effects, when center points do almost nothing for main-effect power.
Replication means repeating design points (ideally the full factorial corners). Because the standard error of an effect scales as 2σ/√N, doubling the runs by replication shrinks the error by a factor of √2 and directly raises power. Replicate when a first analysis leaves an important effect borderline, or when you know upfront that biological noise is high — a theme in our getting-started guide for biologists.
Center points are runs with every factor at its mid-level. Three to five of them do two valuable jobs: they give a model-independent estimate of pure error (essential for the lack-of-fit test), and they detect curvature — if the center response departs from the average of the corners, a straight-line model is inadequate and you need a response-surface design. But because a center point sits at the middle of every factor, it carries no information about how the response changes when a factor moves from low to high. It cannot raise main-effect power.
| Extra runs | Estimates pure error | Detects curvature | Raises main-effect power |
|---|---|---|---|
| Replicate corner runs | Yes | No (still 2 levels) | Yes — the main reason to do it |
| Add center points | Yes | Yes | No — negligible |
Balancing runs vs information
The goal is never the fewest runs or the most information — it is the most information per run you can afford. In bioprocessing a single bioreactor run can cost a week and hundreds of dollars in media, so this balance is not academic.
Three habits keep the balance right. First, screen before you optimize: spend a small unreplicated screening design to kill the factors that do not matter, then invest your runs in the survivors. Studying inert factors is the largest source of wasted experiments, so cutting a factor list from ten to three before the expensive stage saves far more runs than any clever fractioning. Our DOE for bioprocess optimization walkthrough follows exactly this screen-then-optimize path with a worked example.
Second, size for the effect, not for the model. Once a design can fit your model, check its power against the smallest effect worth detecting using the formula above. A design that fits the model but misses a real 0.15 g/L effect has wasted every run. Third, stage your budget: reserve runs for a confirmation experiment at the predicted optimum. A DOE is not finished until you have run the settings it recommends and verified the response — and if you then read the output correctly, you avoid chasing noise, as covered in how to read your DOE results.
Put concretely: for an 8-factor problem, do not reach for a 256-run full factorial. Screen the eight in ~13 runs (definitive screening design), confirm the ~3 that matter, then spend ~15 runs on a Box-Behnken to map the optimum — about 28 runs total to go from eight unknowns to a confirmed setpoint. That is the balance DOE run counts and power analysis are meant to find.
Build a right-sized design in the browser
The free DOE software builds factorial, screening, and response-surface designs, shows the run count for each, and randomizes the run order — no install, no coding.
Frequently Asked Questions
How many experiments does a DOE need?
It depends on your design and factor count. A two-level full factorial needs 2^k runs (8 for 3 factors, 16 for 4, 32 for 5). Screening designs need far fewer: a 12-run Plackett-Burman screens up to 11 factors, and a definitive screening design needs about 2k+1 runs. Response-surface designs need more: a 3-factor central composite is about 20 runs and a Box-Behnken is 15. As a rule of thumb, budget at least twice as many runs as terms you want to estimate, plus 3-5 center points.
What is statistical power in a DOE?
DOE power is the probability that your experiment detects a real effect of a given size as statistically significant. Power = 1 - beta, where beta is the false-negative rate. Aim for 80% power (beta = 0.20). Power rises with more runs (or replicates), a larger true effect, and lower process noise, and falls as you demand a stricter significance level. A design can be big enough to build the model yet too small to detect the effect you care about, so check power, not just run count.
How do I calculate the sample size for a DOE?
For a two-level factorial, the number of runs N to detect an effect of size delta at significance alpha and power 1-beta is approximately N = 4 sigma^2 (z_alpha/2 + z_beta)^2 / delta^2, where sigma is the run-to-run standard deviation. Estimate sigma from historical data or replicate runs. This is the sample-size-doe calculation: it tells you whether your planned run count can actually resolve the effect you are chasing, or whether you need replication.
Do center points add statistical power?
Not for main effects. Center points (runs at the mid-level of every factor) let you estimate pure error and test for curvature, but because they sit at the center they contribute almost nothing to the estimate of a factor's main effect. To increase power for main effects you must replicate the factorial (corner) runs or add more distinct design points. Typically 3-5 center points is enough for the error and curvature checks.
Is it better to run more factors or more replicates?
Early on, favor more factors over more replicates. Screening many factors in a single unreplicated design finds the critical few far more efficiently than replicating a small design, because the biggest source of wasted runs is studying factors that turn out not to matter. Once screening has isolated 2-4 important factors, add replication (and center points) to sharpen the estimates and gain power on the factors you have confirmed matter.
Related Tools
- DOE Experiment Generator — Builds factorial, screening, and response-surface designs and shows the run count for each, free in the browser.
- Media & Feed Estimator — Cost the conditions across a planned DOE before you commit the runs.
- Seed Train Planner — Plan the inoculum trains behind a bank of DOE bioreactor runs.
References
- Montgomery, D.C. (2017). Design and Analysis of Experiments, 9th ed. Wiley. ISBN 978-1119113478.
- Mandenius, C.F. & Brundin, A. (2008). Bioprocess optimization using design-of-experiments methodology. Biotechnology Progress, 24(6), 1191–1203. DOI: 10.1002/btpr.67
- NIST/SEMATECH (2012). e-Handbook of Statistical Methods, Section 5.3: Choosing an experimental design. itl.nist.gov