Modern DOE methods allow you to create an experiment that fits your system. Learn how to overcome the limitations of traditional DOE by using the Custom Design platform. After creating the design, use different tools to explore and compare different designs to make sure you are creating the best option to support your goals.

Q&A / Discussion

Big thanks to @Bill_Worley and @Victor_G for help on the Q&A session for this session.

Q: What design is best if you only have continuous factors?

A: Many designs can work with continuous factors but custom DOE will give the most flexibility.

Q: What is difference between a custom design with only main effects and definitive screening design?

A: A main effects design will not provide estimates for higher order effects; a DSD will provide an estimate on two-factor interactions and polynomial terms if they are active and important (and if not all main effects are active).

Q: What does estimability "Necessary" vs "If Possible" mean?

A: Effects as "Necessary" will force the design generation to be able to estimate these terms. Effects "If Possible" will create Bayesian designs, where the effects can be estimated if you have enough degree of freedoms left, and the design generation will try to accommodate the points to be able to estimate these effects.

Q: Does a custom design allow replicates?

A: You can specify the number of replicated treatments in a custom DOE. If you want to replicate an entire design (a replicate for each treatment), you can create the design and then augment DOE to replicate all treatments.

Q: Can I run a DSD with categorical and mixture factors?

A: DSD allows for continuous factors and two-level categorical factors.

Q: For mixture factors, which is best: classical mixture design or space filling design?

A: It depends, as the methodologies are very different. It is basically a model-based vs. model-agnostic approaches comparison, and they have their benefits and drawbacks. More here.

Q: How do you evaluate if the responses generated are significant?

A: There are many model diagnostic tools to help with this. The Analysis of Variance and Parameter Estimate reports are a good starting point.

Q: Will changing a factor to Hard or Very Hard restrict the setting range?

A: No, setting a factor to Hard or Very Hard will group treatments with similar settings together (instead of completely randomized) and ensure the design is balanced between these factors that are restricted.

Q: Are there situations where D-optimal is better than other optimality criteria?

A: Using the Design Explorer platform can help find a design that has a good compromise between different optimality criterion (such as I for prediction variance and D for screening). There is also the option to use A-optimal design, as it enables putting different emphasis/precision on parameters estimates (for example, to have better precision on main effects and less on interaction effects in the context of screening designs). Bradley Jones did a wonderful talk on A-optimal designs.

Q: How do you know what optimality criterion was used to create a design?

A: The DOE table will list the optimality criteria that was used.

Q: If you don't follow the order given by DOE, what does that mean for the results? For example, if you are limited by time when stopping production to run the experiment, meaning you have to go in as logical an order as possible. Would that skew the results?

A: If possible, you should stick to the random order provided. If there limitations known ahead of time, you can create a blocking factor around a known constraint.