A two-level factorial can only ever fit a flat plane. It measures your response at the corners of a box and draws the tightest straight-sided surface through them, and it will do that cheerfully even when the truth is a dome, a valley, or a ridge. Nothing in the corner data itself will warn you. Center points in DOE are the three or four extra runs that supply the missing warning: sit them in the middle of the design, replicate them, and you get both a clean estimate of pure error and a direct test of whether the plane is lying to you. This guide covers what center points do, how the curvature test works arithmetically, how many to run, and what the answer tells you about escalating to a response-surface design. You can add them to any layout in the free design of experiments calculator.
What are center points in DOE?
A center point is a run set at the midpoint of every factor's range, coded 0 on all factors, added to a two-level factorial or screening design and normally replicated three to five times. If glucose is studied at 3 and 7 g/L and temperature at 34 and 38 °C, the center point is 5 g/L and 36 °C. It is not one of the ±1 corners; it is the geometric middle of the design region.
The crucial property is orthogonality. Because every factor sits at 0 in a center run, the center points contribute nothing to any main-effect or interaction contrast, which are all differences between the plus and minus halves of the design. Adding center points therefore cannot shift your estimate of the glucose effect, the temperature effect, or their interaction. They are pure addition: you buy new information without disturbing anything you already had. That is unusual in experimental design, where extra runs normally rebalance everything, and it is why center points are close to free lunch.
What they buy is two things that a bare two-level full factorial design cannot give you at all:
- Pure error — the run-to-run repeatability of the process and assay, measured directly rather than inferred from a model residual.
- A curvature signal — evidence that the response bends over the region you tested, which invalidates the flat model before you act on its optimum.
Antony makes the underlying point about levels plainly in Design of Experiments for Engineers and Scientists: for quantitative process variables, two levels are generally sufficient in the early stages, but if a non-linear function is expected, it is advisable to study variables at three or more levels in order to quantify the curvature effect on the response. Center points are the cheap way to get a third level without paying for a full three-level design: one extra setting, a handful of runs, no change to the factorial structure.
Pure error and lack of fit from replicated center points
Pure error is the spread between genuine replicate runs at identical settings; it is the noise floor your process and assay impose no matter how good the model is. Replicated center points are the standard way to measure it, because they are true replicates of a single condition rather than repeats of a measurement on one sample.
That distinction matters more than it sounds. Re-assaying the same supernatant three times measures analytical precision only. Running three independent bioreactors at 5 g/L glucose and 36 °C measures everything: inoculum variation, control-loop drift, sampling, and the assay together. Only the second is pure error in the DOE sense, and only the second is a fair yardstick for judging whether a factor effect is real. Genuine replication is also what separates a real error bar from a falsely narrow one, a point developed further in how many experiments a DOE needs.
With nc center runs, the pure error sum of squares is the sum of squared deviations from the center mean, carrying nc − 1 degrees of freedom:
Once you have pure error, the residual from your fitted model splits into three named pieces rather than one anonymous lump. Minitab's documentation describes the partition exactly this way: lack of fit is what remains once curvature and pure error have been backed out of the residual error. Reading those three separately is what turns an ambiguous "the model does not fit well" into an actionable diagnosis.
| Component | Where it comes from | df | What a large value means |
|---|---|---|---|
| Pure error | Spread among replicate center runs | nc − 1 | Your irreducible process + assay noise. Not a model problem. |
| Curvature | Corner mean − center mean | 1 | The response bends. A first-order model is inadequate. |
| Lack of fit (remainder) | Residual after removing the two above | by design | Model form is wrong some other way (missing term, bad range). |
Note the degrees of freedom column. Pure error is the only one you control directly, and you control it purely by choosing nc. Two center points give one degree of freedom, which produces a critical t value above 12 and a test that will miss almost any real curvature. That single line explains most of the practical guidance in the rest of this article.
The curvature test: are the center runs off the plane?
Testing for curvature in a DOE compares the average of the factorial corner runs against the average of the center runs, and judges the difference against pure error. The logic is direct: a first-order model's prediction at the coded origin is exactly the mean of the factorial points, so if the measured center mean differs from it by more than noise, the surface is not flat.
With nf factorial runs (mean ŷf) and nc center runs (mean ŷc), the curvature sum of squares carries a single degree of freedom:
Equivalently, and often easier to reason about at the bench, the same test is a two-sample t test on the difference of the two means, with t = √F. That form makes the sensitivity question concrete: you will detect curvature when
where s is the pure-error standard deviation from the center replicates. Everything on the right is either fixed by your process (s) or chosen by you (nc), which is exactly the lever the how many center points section pulls on.
One limitation is fundamental and worth stating early, because it is the single most common misreading of a significant curvature test. The test has one degree of freedom because it detects the sum of the pure quadratic effects, not the individual ones. A 2² plus center points can prove that the surface bends, but it cannot apportion that bend between glucose and temperature. Separating them requires runs at intermediate levels of each factor individually, which is what axial points buy you.
Worked example: a CHO titer 2² with four center runs
Here is the whole calculation on a realistic mAb dataset, start to finish. A CHO fed-batch screen studies glucose setpoint (A: 3 and 7 g/L) and culture temperature (B: 34 and 38 °C) on day-14 titer, as a 2² factorial with four replicated center points at 5 g/L and 36 °C.
| Run type | A (glucose) | B (temp) | Coded | Titer (g/L) |
|---|---|---|---|---|
| Factorial | 3 g/L | 34 °C | (−1, −1) | 2.10 |
| Factorial | 7 g/L | 34 °C | (+1, −1) | 3.00 |
| Factorial | 3 g/L | 38 °C | (−1, +1) | 2.50 |
| Factorial | 7 g/L | 38 °C | (+1, +1) | 3.60 |
| Center | 5 g/L | 36 °C | (0, 0) | 3.78 |
| Center | 5 g/L | 36 °C | (0, 0) | 4.02 |
| Center | 5 g/L | 36 °C | (0, 0) | 3.86 |
| Center | 5 g/L | 36 °C | (0, 0) | 4.06 |
Step by step
1. Effects from the four corners (unchanged by the center runs).
Read on its own, this is an encouraging, tidy result: push glucose up, push temperature up, expect about 3.6 g/L at the (+1, +1) corner and more beyond it.
2. The two means.
3. Pure error from the center replicates.
4. The curvature test.
5. The verdict. Curvature is overwhelming. The measured center sits 1.13 g/L above what the plane predicts, and it is also 0.33 g/L above the best corner (3.60 g/L). The tidy first-order conclusion in step 1 — "go higher on both" — is wrong. The real surface is domed, with its peak inside the design region, and the corner-only model would have marched the process in exactly the wrong direction. Four extra runs caught it.
Note what the test cannot tell you. F = 146 says the sum of the quadratic effects is large; it does not say whether the dome is mostly in glucose, mostly in temperature, or both. Interpreting the effects table and the ANOVA around this result is covered in how to read your DOE results, and pure error is what makes those p-values trustworthy in the first place. Because pure error folds in assay repeatability as well as process repeatability, tightening the analytics feeds straight through into a sharper curvature test — if the titer readout itself is the noisy part, an ELISA analyzer that surfaces replicate CV and standard-curve fit is the place to look first.
Add center points to any design in one click
Build a factorial, screening, or response-surface layout, set the number of center-point replicates, and get a randomized run order with the center runs interleaved. Free, in the browser.
How many center points should you run?
Three to five replicated center points is the standard recommendation, and four or five is the practical sweet spot for a bioprocess DOE. Screening designs typically carry three; response-surface designs such as central composite designs usually carry three to six. Common guidance is that two or three center points per block are enough merely to detect curvature, while five to seven are wanted if you also need a solid pure-error estimate.
The reason the curve flattens so fast is the t multiplier. With nc center points you get nc − 1 degrees of freedom, and the critical t collapses from 12.71 at 1 df to 4.30 at 2 df to 3.18 at 3 df, then barely moves. Applying the detection formula from the previous section to the worked example (s = 0.13 g/L, nf = 4) gives the smallest curvature you could call significant:
Figure 2. Smallest detectable curvature vs number of center points (2² factorial, pure-error SD 0.13 g/L, α = 0.05). The jump from 2 to 4 center runs improves sensitivity roughly fivefold; past 6 the return is negligible.
| Center points nc | df for pure error | t0.975 | Smallest detectable curvature (g/L) | Verdict |
|---|---|---|---|---|
| 2 | 1 | 12.71 | 1.43 | Effectively blind. Avoid. |
| 3 | 2 | 4.30 | 0.43 | Workable minimum for screening. |
| 4 | 3 | 3.18 | 0.29 | Good default. |
| 5 | 4 | 2.78 | 0.24 | Good default. |
| 6 | 5 | 2.57 | 0.22 | Diminishing returns begin. |
| 8 | 7 | 2.36 | 0.19 | Runs better spent elsewhere. |
Two practical riders. First, spread the center runs through the campaign rather than clustering them at the start or the end, otherwise the "pure error" you measure is really the drift between the beginning and end of the study. If the design is split across days or media lots, put center points in every block, as described in blocking in design of experiments — that also lets you see block-to-block movement directly, because the same condition is measured in each.
Second, center points only exist for continuous factors. There is no midpoint between two host cell lines or two media formulations. In a mixed design, run center points at the middle of the continuous factors at each level of the categorical factor, so every categorical level carries its own replicated center. Deciding how those extra runs fit inside a fixed budget is the same trade-off discussed in how many experiments a DOE needs.
When center points tell you to escalate to CCD or Box-Behnken
A significant curvature test is the signal to stop optimizing with a first-order model and move to a response-surface design. It does not invalidate your effect estimates — the main effects and interactions are still correctly estimated — but it does invalidate the model's predicted optimum, because a plane extrapolated to a corner cannot represent a surface that bends.
The efficient escalation is augmentation rather than restart. Add axial (star) points at ±α on each factor to the runs you already have, and the factorial plus center points becomes a central composite design. NIST describes CCDs exactly this way: they start with a factorial or fractional factorial with center points and add star points to estimate curvature. Your existing corner and center runs are not wasted; they become two-thirds of the RSM design. The alternative, a Box-Behnken design over the same region, avoids the extreme corner combinations altogether, which matters when the corners are hostile to the cells.
The essential gain is resolution of the curvature. A factorial plus center points estimates the sum of quadratic effects with one degree of freedom. Adding axial points supplies runs at intermediate levels of each factor separately, so each pure quadratic term gets its own coefficient and its own p-value, and the fitted surface can locate an interior optimum. The choice between CCD and Box-Behnken, their run counts, and how to read the resulting contour plots are covered in response surface methodology DOE.
| Curvature test | Situation | Next move |
|---|---|---|
| Not significant | Effects clear, region flat | Keep the first-order model. Move along the steepest ascent and re-test. |
| Significant | Optimum bracketed inside the region | Augment to a CCD: add axial points + extra center runs to the existing design. |
| Significant | Extreme corners hostile to cells | Run a Box-Behnken design over the same ranges (no corner combinations). |
| Significant | Pure error large vs effects | Fix the measurement system and replication first. Noise, not shape, is the bottleneck. |
| Not significant, but effects tiny | Ranges too narrow | Widen the factor ranges before concluding the factors do not matter. |
Add center points free
You do not need commercial software to run any of this. Build the factorial, set the number of center-point replicates, and let the tool interleave them into a randomized run order so they are spread across the campaign rather than clumped. After the data is in, the same tool fits the model and flags whether curvature is significant, which is the branch point for whether you augment to a response surface design or act on the first-order result.
The free DOE software on this site supports center points on factorial, screening, and response-surface layouts, reports the effects with their significance, and will tell you when curvature means the linear model has run out of road. If you are not yet sure which base design to build, start with the DOE design decision guide; if you already have data in hand, go to reading DOE output.
Test your own design for curvature
Enter your factor ranges, choose the number of center points, run the experiment, then paste the responses back in for effects, ANOVA, and an automatic curvature flag.
Frequently Asked Questions
What are center points in DOE?
Center points are extra runs placed at the midpoint of every factor range, coded as 0 on all factors, added to a two-level factorial or screening design. They are usually replicated three to five times. Because they sit at the middle of the design rather than at a corner, they are orthogonal to the main effects and interactions and do not change those estimates. What they add is a direct estimate of pure experimental error and a test for whether the response is curved rather than flat across the design region.
How do center points test for curvature?
A two-level factorial can only fit a flat plane, so its prediction at the center of the design is simply the average of all the factorial corner runs. Replicated center points measure what the response actually is at that point. If the measured center mean sits significantly above or below the corner mean, the surface bulges off the plane and curvature is present. The formal test compares the difference between the two means against the pure error estimated from the center replicates, using an F test with 1 numerator degree of freedom.
How many center points should I include in a DOE?
Three to five replicated center points is the standard recommendation, and four or five is the practical sweet spot. Two center points give only one degree of freedom for pure error, which makes the curvature test almost powerless. Moving from two to four center points can cut the smallest detectable curvature by roughly a factor of five, while going beyond six buys very little extra sensitivity. Screening designs typically use three; response-surface designs such as central composite designs usually use three to six.
What is pure error in a DOE and how is it different from lack of fit?
Pure error is the variation between genuine replicate runs at identical factor settings, so it measures the irreducible repeatability of your process and assay. Lack of fit is the part of the residual that the model fails to explain beyond that irreducible noise, which signals the model form itself is wrong. Replicated center points supply the pure error estimate, and once pure error and curvature are subtracted from the total residual, what remains is lack of fit.
What should I do if the curvature test is significant?
A significant curvature test means a first-order model cannot describe your response over the tested ranges, so stop trusting its predicted optimum. Augment the factorial into a response-surface design by adding axial or star points, which turns the existing runs into a central composite design and lets you fit the individual quadratic terms, or run a Box-Behnken design over the same region. A factorial plus center points can detect that curvature exists but cannot tell you which factor is responsible.
Can I use center points with categorical factors?
Not directly. A center point requires a meaningful midpoint, which only exists for continuous factors such as temperature, pH, glucose concentration, or feed rate. There is no midpoint between two media formulations or two host cell lines. When a design mixes continuous and categorical factors, the usual approach is to place center points at the middle of the continuous factors while running them at each level of the categorical factor, so each categorical level gets its own replicated center.
Related Tools
- DOE Experiment Generator — Build factorial, screening, and response-surface designs with a configurable number of center-point replicates and a randomized run order.
- ELISA Analyzer — Check replicate CV and standard-curve fit, because assay repeatability feeds straight into your pure error estimate.
- Media & Feed Estimator — Size the glucose and feed ranges you are about to bracket with a factorial plus center points.
References
- Antony, J. (2003). Design of Experiments for Engineers and Scientists. Butterworth-Heinemann/Elsevier. ISBN 0-7506-4709-4. (§4.3.1 on studying variables at three or more levels to quantify curvature; §4.6 confidence interval for the mean response.)
- Box, G.E.P. & Wilson, K.B. (1951). On the experimental attainment of optimum conditions. Journal of the Royal Statistical Society Series B, 13(1), 1–45. doi:10.1111/j.2517-6161.1951.tb00067.x
- Box, G.E.P. & Behnken, D.W. (1960). Some new three level designs for the study of quantitative variables. Technometrics, 2(4), 455–475. doi:10.1080/00401706.1960.10489912
- Montgomery, D.C. (2017). Design and Analysis of Experiments, 9th ed. Wiley. ISBN 978-1119113478. (Ch. 6, adding center points to a 2k design; Ch. 11, response surface methods.)
- NIST/SEMATECH (2012). e-Handbook of Statistical Methods, Section 5.3.3.7: Adding centerpoints. itl.nist.gov