Choosing a Design

 Review A Design and Analysis Process

 Several design choices are usually available from a set of requirements. The trade-off is between the better estimating capability of a design with many runs against the economy of a small design. Final design choices are best made by comparing several different final designs using the average variation of the coefficients (avc) and the minimum detectable effect (mde). In many experiments it will be found (when the data is analyzed) that not all factors or interactions are significant. If they are dropped from the analysis, the avc and mde will improve (decrease).

 After a design is tentatively selected, a few runs may be added. Performance indices may also be checked after a Final design is assembled.

 Design runs may sometimes be removed without introducing confounding.

 For nonlinear situations having more than a few factors, a Central Composite or a Box-Behnken design will have fewer runs than a fractional factorial design.

 There are often a few fractional factorial designs from which to choose: John 3/4 designs usually have more estimating power than the corresponding size Plackett-Burman. For mixed resolution requirements, the mixed-resolution design and any available fixed-resolution designs are made available.

 When almost nothing is known about a process, a screening design may be appropriate. Usually, a screening design is done as a minimal design (Resolution III), which would be incapable of estimating any interactions if all the factors are truly significant. The results from a screening design should never be accepted without estimating the effects of the interactions between the significant parameters. A resolution IV design may be a reasonable compromise for a screening design. A Resolution IV design has main factors not confounded with 2-factor interactions. In the event that some factors are not found to be significant, it is often true that some interactions may be estimated. However, there is no assurance that the desired interaction will be estimable.