nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

Which Run order I have to choose for my DoE?

ELH

Hello,

I want to perform a DoE in order to optimize an extraction method. My plan is to start with a screening of the main factors using Plackett Burman followed by a central composite design in order to do a response surface. Before making a Design tube I should choose a Run order (sort left to right, sort right to left, randomize...) and I do not know which run order I should choose for each of my designs and why?

thank you in advance

Accepted answers

All comments

Mark_Bailey

I am fond of Plackett-Burnham designs but I think in this case you might want to use an alternate. Sequential experimentation using classical designs usually favors the Central Composite Design (Box-Wilson) because you can run the two-level screening portion with a few center points and determine if the rest of the runs are helpful.

I think that you should consider a definitive screening design in this case.

And you should randomize the run order AND reset all the factor levels before each run. If you can't (or won't) then use a Custom Design and define which factors are hard to change.

statman

Welcome to the JMP Community.

While I agree with Mark , that you should consider sequential experimentation, I would not choose Plackett Burman or DSD. I would choose a fractional factorial so you can easily know the aliasing (and there is no partial aliasing).

Some questions/comments:

1. How many factors are you considering for screening?

2. Ensure you choose bold level setting to increase the likelihood the factors will demonstrate their effects.

3. Make a list of predicted effects (main effects and 2-factor interactions). Prioritize this list. If you have 2-factor interactions high on the list, consider higher resolution designs, if they are at the bottom, Res. II is adequate.

4. The most challenging part of experimentation is not the matrix for factors, but the strategy that you use for handling noise. If you cannot identify the noise is in the extraction process, then you would randomize to prevent the noise from being confounded with factor effects (and potentially to get an unbiased estimate of the noise). If you are able to identify the noise variables are in the extraction process (using tools such as a process map), then your options are better (RCBD, BIB, Repeats, split-plots, covariates, etc.) To quote G.E.P. Box:

"Block what you can, randomize what you cannot"

statman

Sorry, that was res. III not res. II. Here are some criteria for design selection:
 Constraints: Time, money, measurement capability, etc. How many treatments can be made?
 How many factors are to be manipulated (the number of hypotheses to be compared)? Are the factors continuous or categorical?
 How will noise be managed or partitioned?
 Are some factors harder to change than others? Are there other restrictions on randomization?
 Are higher order effects suspected/predicted (e.g., interactions, curvature)?
o What is the desired resolution? (What effects do you want to estimate/separate?)
o What order polynomial is necessary?

ELH

Thank you for your answer. This is my first experience with DoE therefore I will need further details.
As I said I want to optimize an extraction method and according to some published articles they start with a screening design using the Plackett Burman (it gives good results even it is studying just the mean effects) and after the screening they use a central composite design. In my field this is the most used strategy, a screening followed by a response surface design. And here are further details about what I want to do
- The extraction technique is pressurized liquid extraction
- I have 8 for screening and from them I am planning to choose 3 for a response surface design.
- I have 2 responses to maximize
- They are no restrictions on randomization, I can randomize my trial order if it is necessary
thank you

ELH

Jeff_Perkinson

I'll let others give you specific advice on your experiment, but I'll encourage you to take a look at two good tutorials on DOE. They won't take you very long and they'll give you some good background on the theory and practice of Design of Experiments.

The first is the Design of Experiments module of Statistical Thinking for Industrial Problem Solving. Here's an overview of the module:

&amp;amp;lt;p&amp;amp;gt;&amp;amp;nbsp;&amp;amp;lt;/p&amp;amp;gt; &amp;amp;lt;p&amp;amp;gt;&amp;amp;nbsp;&amp;amp;lt;/p&amp;amp;gt;

The second is the JMP DOE Introductory Kit. This will give you a good overview of the process of designing, conducting, and analyzing your experiment in JMP itself.

These are both really good resources to help get you started.

ELH

many thanks

statman

The tutorials are good for understanding why to use multi-factor experiments and I agree these would be a good place to start. Unfortunately, they lack a thorough discussion about how to handle noise. The question of randomization requires a complete understanding of the options available. These are difficult to document (or include in a tutorial) as they are often situation dependent.

Regarding the "path" of sequential experimentation, it is impossible to know what you will learn from the first experiment or to know what the second experiment will be. Whether you find 2, 3 or 4 or no factors significant in the first screening experiment is dependent on:

1. Having a measurement system with appropriate discrimination,

2. Setting factor levels bold, but reasonable enough to expose the factors effect,

3. The "truth" (what we are trying to uncover) those factors actually matter,

4. Whether the noise is truly representative of the future noise in the extraction process.

The "test" for factor significance is typically made against all "other" factors that change during the experiment (including noise) or as is often suggested against the noise estimated by randomized replicates. If the "other" factors have a larger affect, you will conclude the factor of interest is insignificant and you will likely abandon that factor (or at least not continue to study it). If you minimize the noise in an experiment (e.g., hold noise constant or run the experiment over a time period where the noise does not change) you could easily conclude a factor is significant only to learn later that its effect depends on the noise.

The purpose of the first experiment, in a sequential approach is to find the factors that are worthy of further investigation. All of the factors you are experimenting on should be supported by scientific hypotheses. You should consider and predict ALL possible outcomes. If a factor is not significant, what are those ramifications? If a factor is significant, what will the next set of work look like? Will it be considering non-linear effects, how that factor interacts with other factors, other higher order effects, etc.? A second experiment is only one option (for example, you might consider sampling as the experiment will not provide much insight into consistency). And whether the second experiment is a central composite design or some other design can only be predicted at this point. You won't know what to do until after you have analyzed the first experiment ant thought about the scientific implications of that knowledge. "There is no instant pudding", Deming.

ELH

Dear Sir,
Following your answer, how I can do a sequential experiment since I have 9 factors which is too much? Is not better to start with a main effect screening(PB) than a sequential experimentation using CCD?
many thanks

ELH

Dear Sir,
Following your answer, how I can do a sequential experiment since I have 9 factors which is too much? Is not better to start with a main effect screening (PB) and then sequential experimentation using CCD?
many thanks