Looking for best DOE design

[#1]

Hello all,

I am struggling to find the best DOE design for finding best conditions to my protocol.

I looked into documentation, assist to many webinar but canot manage to find the best one. So I need your help on that.

I have 2 categorical responses (4 levels), 1 continuous factor and 1 categorical (2 levels) factor.
My continuous factor is concentration of a reagent. I need to include negative control for that, meaning that I want to see the result when concentration of this factor is 0. I work on biological samples and this negative control is not an option.

I tried Custom Design but didn't get range of value for my continuous factor.

So I tried Space Filling Design but, despite giving me a large amount of tests, this design doesn't use the value 0 for my continuous factor.

Many thanks for any insight, help that you may have.

Marie

[#2]

Hi @mgerusdurand ,

You need to add higher order terms in the Model outline for your continuous factor (X1 here):

This should give you a design that uses more values for X1.

Some explanation that might help...

The default model is a main effects model and the optimal placement of runs to estimate the main effect of a continuous factor is at the low and high extremes of the range. Placing runs at the low and high extremes is the best for estimating a simple linear main effect of a continuous factor. Custom Design is doing what it should do: giving you the optimal design for estimation of the model that you have specified. When you specify higher order terms in the model, this forces Custom Design to place runs in the middle of the range - you need points in the middle to estimate a curved line.

Another point to consider is that it is more challenging to gain insight from categorical responses than for continuous responses. There is less information in a categorical response than a continuous response, so you typically need more runs in your experiment to learn something useful about the system. If you can, it is better to use a continuous response.

I hope this helps,

Phi

[#3]

Thanks Phil.

The responses are discrete numeric, to get quantitative ones (continuous) this will involved other processes for which I would need to include other factrs. I may consider it if proposed model still not optimal.

I will try to add higher order first. 

Thanks for your help

Marie

[#4]

Marie,

I do not understand your situation, so I have some questions/comments:

1. I do not understand what you mean by "best" DOE.  First, how do you define best? It is unlikely anyone can do this, so my advice is to create multiple designs and compare and contrast them (precision, resolution, inference space, efficiency, etc.).  Predict what information/knowledge is possible from each experiment and weigh his against the resources required.

2. I completely agree with Phil that using categorical response variables can be very inefficient.  You will lack discrimination especially with less than 5 categories.  This usually results in needing large sample sizes to evaluate. It would be helpful to know what the responses are as we may provide thoughts on how to modify them, but understand this may be proprietary.

3. I am confused by some of your statements.  They seem to be contradictory:

"I need to include negative control for that..." and "negative control is not an option" ?

and then "this design doesn't use the value 0 for my continuous factor."

So do you want 0 as a level in the experiment or not for the continuous factor?

[#5]

Hello @statman ,

I may not have been talking in DOE langage, sorry I am novice.

1. Best DOE for me will allow to cover all range of values. Having only extreme limts didn't allowed me to get the right information based of my knowledge of the biological process I am assessing. I want to find the DOE that will help me to get the most relevant informations.

2. Our responses (reportable parameters) depend on our targets and are often categorical. When working in an IVD environmet you don't have the choice of the output you need to assess and unfortunately in this case it is categorical 4 levels. If I want to use quantitative data this involved other processes and then it is an all new story. 

3. I need a negative control to check that the response I observed is due to the presence of the reagent (which concentration is my continuous factor). So I need some test where my continuous factor is 0 so that I can assess this reagent effect on my response.

Hope this is a better explanation,

Marie

[#6]

You might be able to simplify the design of your experiment if you leave the negative control to the end and add it manually after Custom Design makes the data table. The design is about estimating model parameters. Controls are about assurance of the runs. Separating them will make it easier. You might want to exclude the control row from the regression analysis but do not hide it from the plots to help your assessment.

Your case might also benefit from using a Discrete Numeric factor instead of a Continuous factor. You can specify the levels you need this way. @Phil_Kay explained how the design and the model are related. Leave the higher-order terms in the model that JMP adds automatically if you choose a Discrete Numeric factor.

[#7]

Hello @Mark_Bailey ,

Thanks for the input. Indeed adding my control at then end could help. I would just need to be sure that I can add it to the model after because I need to ensure that this one will take into consideration if I got same results from a test than from the control this can be discarded. Not sure if I explain it well. 

I willa lso try the discrete numeric to see how it works.

Thanks again for your help,

Marie