Modelling response with a lot of zeros in JMP (not Pro)

[#1]

Hi community,

I have performed an initial design which indicated one of my impurities could be minimized by changing the range (I was able to create a model for the impurity response for the initial design). I then augmented the design to a new range however now i have a lot of zeros for the response of the impurity and therefore a linear model no longer fits. I have gone into distribution to see whether i could use a transform but the suggested transform also did not give me a model that passes criteria. I would need a model to predict expected ranges for the impurity response within the new design space.

I only have JMP (not JMPpro)

Many thanks,

Hanna

[#2]

Please elaborate on your claim, "i have a lot of zeros for the response of the impurity and therefore a linear model no longer fits." What are the issues you encountered when modeling the data after augmenting the initial design?

[#3]

Upon modeling the data after the augmentation i get a significant lack of fit.

Using the recommended transformation based on the distribution still gives me a significant lack of fit:

[#4]

Some questions:  What are you trying to achieve?  What exactly do you mean by "I then augmented the design"?

[#5]

We are trying to validate the process into a range, where there is little variability within the design space. In that regard we would like to keep the level of the impurity low. Lets say we initially had a range of 10-30°C for temperature the impurity response was between 0-15%. We then augmented the design to 20-40 °C so we added experiments on where the range was extended. The problem now is that I would also like to predict what the response range will be for this new design space (20-40 °C) as well as potentially a tighter range (25-35 °C). However without having a model i cannot predict ranges.

The augmentation was done for 3 out of 6 parameter so as the space was already 10-30 °C most additional experiments were in the higher range e.g. in the example 40°C. So I cant even really say from the distribution what it would be for 20-40 °C.

[#6]

Just a clarification regarding the use of your word augment....usually this means filling the design space more thoroughly.  

https://www.jmp.com/support/help/en/17.1/?os=mac&source=application#page/jmp/augment-designs.shtml

You are actually moving the design space.  Nothing wrong with sequential experimentation, in fact it is preferred, but I wouldn't consider this augmentation (might be my own bias, star points may extend the space).  Since the space you are seeking to understand is outside your inference space, you will need to get data to support the model works there.

A question about the 0's.  Is it just your measurement system lacks the capability to detect the impurities?

[#7]

The augmentation (lat. augmentare = to increase) included the addition of experiments in order to explore the additional space. So that's what was achieved in the augmentation. In JMP the change of ranges is easily done in the augment design option.

[#8]

OK, sorry misread your last paragraph:

The augmentation was done for 3 out of 6 parameter so as the space was already 10-30 °C most additional experiments were in the higher range e.g. in the example 40°C. So I cant even really say from the distribution what it would be for 20-40 °C.

I actually don't understand this paragraph.  Could you elaborate on what exactly you did?

Again, to increase, as you provide the latin meaning, can mean to increase the understanding of what is going on inside the design space.

[#9]

My initial design had 6 parameters i explored the design space with 16 experiments. I then widened the range for 3 of the 6 parameters to a slightly higher range for 2 and a lower ranger for 1 of the parameter. Using 'augment design' in JMP i then did the suggested additional 8 experiments. 

I then proceeded via JMP's model function. However any model I produced after augmentation had a significant lack of fit and could therefore not be used for predictions anymore. I then thought, if I can't get a model maybe I could at least make predictions from the analyse->distribution however as the augmentation mostly added values at the high range of the design space (e.g. 40 °C), I would not have values in the distribution plots of 20 °C (as it wasn't necessary to add these as the space of 10-30 °C) was already fully explored.

My main question is still, what do I do if I cannot fit a model due to the fact that there are a lot of zero values for the impurity.