DSD example with possible mixturesI have a prospective customer that is interested in JMP because of the paper attached here. They were particularly interested in the applications of DSD and I had a few questions from the paper they sent. I summarized a few of the key figures and had the following question:
- This DSD appears to be for a mixture design and from what I can tell they are using the Cholesterol Molar Composition as a slack variable since it doesn't show up in the DSD Design that they generated. Is this an appropriate use of a DSD?
Any other observations or comments would be welcome! Not going to get into the detailed process of this experiment during the customer call but would appreciate having some better understanding to guide them in potential approaches or other platforms of interest. Thanks @Jeff_Upton for the input so far!
Re: DSD example with possible mixturesPlease see the work below that uses MESD and has a similar problem. The screening, however, has a narrow scope.
Development of Tapentadol Hydrochloride Loaded Proniosomal Gel using Main Effects Screening Design and in silico Verification using Parameter Sensitivity Analysis | Indian Journal of Pharmaceutical Education and Research (ijper.org)
Re: DSD example with possible mixturesThey also put a very nice promotion for JMP in the Supplementary Information so I am also fine with giving them a pass!
Re: DSD example with possible mixturesYeah, I agree, @SamGardner . They deserve some credit for using DOE and it is great if it inspires others in this field to use DOE. But they way they use DOE is definitely not an example to follow.
Re: DSD example with possible mixturesThere are correlations between main effects, quadratics, and two FI (not all, not big, but correlated). I've only skimmed the papers, but I'm not sure how it can be called a DSD given these conditions. When I created the design in JMP, only the categorical main effects were correlated. So there appears to be a better (DSD) design. I'm not sure the molar composition factors are mixture factors, but I'm not a chemist. If they are, it would be helpful to know the bounding constraint.
Re: DSD example with possible mixturesLooking at the paper, they actually used a Taguchi L18 array for the follow-on design and inaccurately called that a Fraction Factorial, I think. The L18 design is a three level design in all but one factor, and if they left that column out of the design then all of the factors would allow for estimation of quadratic effects. The L18 design is less efficient than the DSD, and it has partial confounding of most of the effects, compared to a DSD that has no confounding of main effects with higher order interactions.
But, this was an academic project, so I will cut them some slack and just be happy that they used DOE at all.
I attached the L18 design if you want to take a look at and compare to a DSD. Design comparison output of the full 8 factor L18 design and an 8 factor DSD is shown below also.
Re: DSD example with possible mixturesThank you, @Phil , for the insight as always! It does seem like they are going back in time with each step, starting with DSD and ending with OFAT.
Re: DSD example with possible mixtures Hi @Joseph_Reese ,
I haven't read the paper but from the figures and tables above I think this is a strange approach, to say the least. I would not recommend following this approach.
I am fairly sure that you are correct that Cholesterol is being used as the slack variable. That might be okay but it seems like Cholesterol and C12-200 are both big parts of the formulation and are both varying over large ranges. The slack variable approach might not be a useful approach here. But that is not the strangest thing about what they have proposed...
The really weird thing is in that flow diagram / decision tree. Why would you go from DSD to fractional factorial? A DSD is screening design (it is in the name). A DSD is primarily for screening factors. It is not for screening for "hits".
I supposed you could use a DSD to screen for hits - it will do a reasonable job of covering the factor space. But I don't know why the next step would be a fractional factorial.
And then they suggest that you move the factor space until you only have 1 significant factor. At which point you use OFAT. Very strange!
This all goes against the accepted paradigm of sequential DOE. The authors seem to think they have come up with a better version but it looks awful to me. With a bit of luck you could still find useful formulations but I can only see that it makes is harder.
Again, I haven't read the paper, so they might have some insight as to why this radical approach is uniquely useful. But my initial reaction is avoid.
