Creating a Design

Entering Missing data

Leave it blank; only use a “0” if a zero is representative of the true outcome.

Only numeric data is allowed in the response column.

Changes to the design

Many things on the design layout screen can be changed by right-clicking on the factor column headers and selecting Edit info…

Right-click on the row header to delete a run or change its status to ignore, highlight or verification.

It is often easiest to rebuild the design if you need to make changes to how the design was created. Click on File, New Design and click YES on the “Use previous design info?” dialog. Doing this preserves all the factor names and levels and the response names. Make the necessary changes and re-build without re-typing everything.

If only factor names and numeric factor limits are being adjusted, use the Column Info Sheet view from the Design Layout screen.

Can’t run one combination of factor levels

If there is only one combination that is difficult or impossible to perform, you might just leave that run out of the experiment and treat it as missing data. Another option is to build a candidate set containing only the possible combinations and building an optimal design. Combine factors if a particular combination of two factors cannot be part of the design. You lose the interactions but can still test for differences (effects) between the combinations. Optimal designs can also have constraints applied to them if the excluded combinations only involve numeric factors.

Messed-up coding levels

This can happen if the factor settings in the design are replaced (over-typed) with substantially different settings as used during the experiment.

Follow these steps:

1. Right-click on a messed-up factor column header

2. Choose the Recode option from the list.

3. Click Yes on the dialog.

4. Repeat for the other messed-up factors.

Unfortunately, this trick won’t work with mixtures. Copy the current data set, rebuild the design to match the true component ranges and paste the data set into the “new” design.

Planning a two-factor experiment

When an experiment only includes one or two factors, there must be replicates. When the results of all the two-factor combinations are only gathered one time each there is not enough information (degrees of freedom) to test all of the effects. Effects can be estimated (coefficients, but no p-values), but they are based upon the smallest sample size possible.

A single additional replicate of the two-factor design is enough to allow the estimate of a p-value. More design replicates will serve to further increase power of the design.

This advice applies to two-level and multilevel categoric designs that only have two factors.

Too many runs on Multilevel Categoric designs

Multilevel Categoric designs provide a design that has all possible combinations of the factor treatments. When the factor have many levels or there are a large number of factors there are too many runs to allow for an efficient experiment.

Optimal designs are the correct method for reducing the combination to only those necessary to fit a reasonable model. Optimal factorial design by default are designed for a two-factor interaction model. By leaving out the ability to estimate the three-factor and higher order interaction many runs can be saved.

In a design with categoric factors, each treatment must be used at least once. If the optimal design is still too large, then some of the treatments and/or levels will need to be dropped from the design to fit your budget.

How Many Levels Do I Need?

When working with numeric factors, the true response surface is assumed to be a continuous function across the area of interest. Continuous functions can be approximated through a Taylor polynomial. The number of levels required depends on the order of the polynomial you believe will best represent the true response surface’s shape. If you believe the shape of the response surface is a hill, valley, ridge, or saddle then a quadratic model only requires three levels (extreme low, extreme high, and in the middle). If you believe the surface is wavy, then four levels are required for a cubic function. By default, these levels are evenly spaced across the testing range, but are not required to be evenly spaced.

More levels allow the analysis to fit higher-order polynomial models. If a higher-order polynomial is not necessary to model the trend in the data, then more levels does not help fit the correct model. Optimal RSM designs include lack of fit points, which add a few more levels than required to fit the model. This provides a check to make sure the model fits throughout the area of interest.

Designed experiments are used to fit models to data; the models for several responses are combined to optimize the process. Designed experiments are not intended to be used for a brute force approach, where many possible combinations are tested and the winner is picked.

Unless the build is restricted to discrete or categoric factors, optimal designs may have many more levels per factor than are actually necessary, but only a select few combinations of those levels will be in the design.

Minimum Number of Levels to Fit a Polynomial