Help with DOE analysis - Mixing Study

[#1]

Hello all,

I am fairly new to DOE and would appreciate some assistance with the analysis of my data, especially with the correct approach and order to the analysis.

I have created a simple mixing design. On the face of it, this is a very simple experiment, which involves mixing three different components to create a composite, conducting some experiments and measuring several outputs. The design is repeated in triplicate (8 runs per replicate), which should allow up to the possible single three-way interaction to be determined. I also manually added in some additional controls (25-39 entries) to the proposed design. There are 5 measured outputs. Outputs 1 and 2 are separate outputs. Output 3 was measured over a wide frequency range, but for the purposes of this design, three separate measurements were made - one at low, one at medium, and one at high frequency.

[#2]

There is no 'one size fits all' pathway to analyzing any experimental design. Much depends on your specific experimental goals and objectives. The type of design. The responses. How you treat noise. And on and on. A good place to start for you might be viewing many of the Mastering JMP on demand webinars to give you some general ideas. All are 'free'. There are several devoted to mixture design scenarios. Here's a good entry point webinar:

Mixture Designs 

[#3]

Hello Matthew,

Output 1 has a  good model.

Output 2 has a moderately good model, probably due to low measurements.

Output 3 benefits of stacking the 3 measurements (obtained at Low, Medium, High frequency). The model includes the initial 3 mixture components + 2nd order interactions in addition to an aditional categorical variable Label (i.e. frequency level) and its interaction with mixture components. It has a clear outlier (row 19 = 209) which should be excluded. Then the analysis benefits of Box Cox transformation to get a better model.

I hope this helps.

Emmanuel

[#4]

Additional comments: you would benefit of JMP Pro features for Outputs 2 & 3.

Indeed Output 2 measurement is likely limited by LoD (limit of detection) leading to inacurate model. JMP Pro has a functionality to manage model considering this LoD.

Similarly Output 3 has been stacked and 2nd order interactions terms (Frequency_Level x Mixture_Components) are significants indicating a different behavior dependings on Frequency level. Using FDE (Functional data explorer) permits modelizing the behavior accross all frequency levels rather than treating this variable as a discrete one with 3 modalities (Low, Medium, High) and get the overall picture.

Emmanuel