Skip to content
Back to search results

Box-Behnken Design — Three-Level Response Surface Methodology

Introduced by
George E. P. Box and Donald W. Behnken
Year
1960
Type
Response surface design (incomplete three-level factorial)
Family
Engineering methods
Font size

What does it do?

The Box-Behnken design (BBD) is an efficient response surface methodology design that fits a full second-order polynomial model using three levels of each factor.

Introduced by Box and Behnken in 1960, it places experimental points at the midpoints of the edges of a hypercube and at the center, avoiding the corner points where all factors are simultaneously at their extreme levels. This structure makes BBD particularly attractive when extreme-level combinations are physically impossible, costly, or unsafe to test.

The intuition behind the method

Suppose you are optimizing a chemical reaction and have three controllable factors — temperature, pressure, and catalyst concentration — each with a low, medium, and high setting. A full factorial at three levels would require 27 runs. The Box-Behnken design achieves the same second-order fitting goal with only 15 runs (plus center points) by cleverly selecting which factor-level combinations to test: it always holds at least one factor at its middle level, so no experiment simultaneously pushes all factors to their extremes. Think of it as a strategic sampling of the experimental space that uses far fewer runs while still capturing curvature and interactions.

History

George Box and Donald Behnken introduced the design in their landmark 1960 paper in Technometrics, where they proposed a family of three-level designs that could estimate all coefficients of a second-order polynomial while avoiding the extreme vertices of the design cube. Their motivation was both economic — corner runs can be expensive — and practical, since many industrial and chemical processes cannot sustain all-extreme-level operating conditions. The design became a standard in response surface methodology, widely disseminated through the textbooks of Box and Draper (1987) and Myers, Montgomery, and Anderson-Cook, and is now routinely implemented in statistical software such as JMP, Minitab, and R's rsm package.

When to use it

Use Box-Behnken design when you need to fit a full second-order (quadratic) response surface model with 3 to 7 continuous factors and the corner points of the experimental space are unsafe, infeasible, or prohibitively expensive to test — for example, when extreme settings of multiple process variables simultaneously could damage equipment or pose safety hazards. BBD is appropriate when the number of experimental runs must be kept moderate and all factors can each individually reach their high and low settings. Do NOT use BBD if you need a design that supports higher-order terms beyond quadratic, if you have categorical factors, if extrapolation to extreme vertex points is the goal, or if only two factors are studied (a central composite design is more efficient for k = 2). Also avoid BBD when a corner-point run is actually needed to meet a downstream specification requirement.

Limitations and considerations

  • Not suitable for sequential experimentation: unlike central composite designs, the BBD cannot be built from an earlier two-level factorial screening experiment.
  • Does not estimate interactions at the extreme factor combinations, since corner points are absent; if behavior at the corners is critical, a central composite or full factorial design is needed.
  • Coverage of the experimental region is not uniform — the corners of the hypercube are entirely unsampled, which can introduce prediction bias if the true optimum lies near a corner.
  • For five or more factors the number of runs can grow substantially, reducing its efficiency advantage over central composite designs.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Requires fewer experimental runs than a three-level full factorial, making it cost- and time-efficient for fitting second-order models.
  • No corner point runs — particularly valuable when simultaneously extreme settings of all factors are physically hazardous, prohibitively expensive, or process-unfeasible.
  • All factor-level combinations in the design matrix are near-center or on mid-edge positions, which are often the safest and most representative operating conditions.
  • Near-rotatability: prediction variance is approximately constant at equal distances from the design center, giving uniform estimation accuracy across the experimental region.
  • Straightforward to augment with center-point replicates that provide a pure estimate of experimental error and support a lack-of-fit test.

Applications

  • Optimization of pharmaceutical formulations, where simultaneously extreme excipient concentrations may render a tablet uncompressible or unstable.
  • Process parameter optimization in chemical engineering — reactor temperature, pressure, and feed rate — where combined extremes could exceed safety margins.
  • Food science experiments optimizing texture, color, or sensory attributes as functions of ingredient proportions and processing conditions.
  • Materials science studies fitting the relationship between sintering temperature, pressure, and time on ceramic or powder-metal part properties.
  • Agricultural field experiments optimizing fertilizer rates, irrigation levels, and planting density when logistical constraints prevent testing all extreme combinations.
  • Environmental remediation studies optimizing pH, oxidant dose, and contact time for contaminant degradation, where extreme combinations could be corrosive or hazardous.

How it works

How it works

1

The experimenter identifies k continuous quantitative factors (typically 3 to 7) and sets three coded levels for each: low (−1), center (0), and high (+1). The response variable must be measurable and continuous. A key prerequisite is that the extreme levels of each factor are individually feasible — the design avoids corner vertices but does test each factor at its extremes while other factors are held at their center.

2

The BBD matrix is constructed by combining factors in incomplete blocks of two at a time. For each pair of factors, a two-level factorial (four runs) is created while the remaining factors are held at their center level. Center point replicates (typically 3–5) are added to the full design. For three factors this yields 15 runs; for four factors, 27 runs. The resulting design is rotatable or near-rotatable and has no corner points where all factors are simultaneously at their extreme settings.

3

Experiments are performed in random order to protect against lurking variables and time trends. Each design row specifies the factor settings for one experimental run. Careful measurement of the response at each run, along with proper randomization and blocking if needed, is essential for valid model estimation.

4

A second-order polynomial regression model is fitted to the data: the response is modeled as a function of the linear effects of all factors, their squared (quadratic) terms, and all two-factor interaction terms. Ordinary least squares (or weighted least squares) estimation provides the regression coefficients. Analysis of variance and lack-of-fit tests assess model adequacy.

Formula
\hat{y} = \beta_0 + \sum_{i=1}^{k}\beta_i x_i + \sum_{i=1}^{k}\beta_{ii}x_i^2 + \sum_{i<j}\beta_{ij}x_i x_j
5

Once a satisfactory model is confirmed, the fitted response surface is explored to locate the optimum — the combination of factor settings that maximizes, minimizes, or hits a target value of the response. Methods include canonical analysis, steepest ascent/descent, and numerical desirability-function optimization when multiple responses are involved.

Common pitfalls

  • Confusing BBD with a central composite design (CCD): the CCD does include corner (factorial) points and axial (star) points beyond the factor range; BBD has neither. Choosing the wrong design relative to experimental constraints leads to wasted runs or infeasible experiments.
  • Applying BBD to fewer than three factors: for k = 2, BBD reduces to a much less efficient design than a CCD; always prefer CCD or a simple rotatable design for two-factor studies.
  • Insufficient center-point replicates: using only 1–2 center points leaves the pure-error estimate poorly defined, undermining the lack-of-fit test and the overall ANOVA. At least 3 center points are strongly recommended.
  • Ignoring the need for randomization: running the BBD in non-random order (e.g., grouping all low-temperature runs) confounds the factor effects with time-related nuisance variables, invalidating the model.

Frequently asked questions

How does Box-Behnken design differ from central composite design?

Both designs fit a second-order polynomial model for response surface analysis, but they differ in structure. The central composite design (CCD) includes a two-level factorial (corner) block plus axial (star) points that extend beyond the factor range, and can be built sequentially from a prior screening experiment. The BBD has no corner points and no points outside the factor range — it uses only three levels per factor. Choose BBD when corner runs are infeasible; choose CCD when sequential building or properties at the extreme corners are important.

How many runs does a Box-Behnken design require?

The number of runs depends on the number of factors k. For k = 3: 15 runs (12 edge midpoints + 3 center points); for k = 4: 27 runs; for k = 5: 46 runs. These counts include a standard number of center-point replicates. Compare this to a three-level full factorial, which would require 27, 81, and 243 runs for k = 3, 4, and 5 respectively — BBD offers dramatic savings.

Can I use Box-Behnken design with more than seven factors?

BBD is generally used for 3 to 7 factors. For more than seven factors the run count grows substantially and alternative approaches such as definitive screening designs or optimal computer-generated designs are usually more efficient. The original Box and Behnken (1960) paper described designs up to seven factors.

What software can run a Box-Behnken analysis?

BBD is supported in most major statistical packages: JMP (Design of Experiments module), Minitab (Response Surface section of DOE), SAS (PROC OPTEX or PROC RSREG), and R (the rsm package with the bbd() function). Python users can use the pyDOE2 library. All of these generate the design matrix, fit the second-order model, and provide ANOVA, response surface plots, and optimization output.

Is Box-Behnken design rotatable?

BBD is approximately rotatable, meaning the prediction variance is nearly constant at equal distances from the design center, but strict rotatability depends on the number of factors. For k = 3, the standard BBD is exactly rotatable with the commonly used center-point count. For other values of k, rotatability is approximate. In practice this near-rotatability is sufficient for most engineering and scientific applications.

Related methods

Sources

  1. Box, G. E. P., & Behnken, D. W. (1960). Some new three level designs for the study of quantitative variables. Technometrics, 2(4), 455–475.

    Open DOI reference
    Copied.
  2. Myers, R. H., Montgomery, D. C., & Anderson-Cook, C. M. (2016). Response Surface Methodology: Process and Product Optimization Using Designed Experiments (4th ed.). Wiley.

    ISBN: 978-1118916025
    Copied.