Parameter Estimates being zerod

[#1]

I am trying to analyze one of my experiments that was created in and I am having some trouble understanding the parameter estimates. 

Depending on which interactions I choose to analyze ( i chose only up to 2 way interactions since this is a fractional factorial design) the parameter estimates either say unstable, zerod and in some other cases empty. I am hoping to et some clarification on why this comes up and what it means. 

I have a included a screenshot below. 

I am currently using JMP 17 on Windows 

[#2]

I'll start, but hopefully someone with a better statistical background can say exactly what is meant by "Unstable."  I frequently see output like this from models, especially with a number of nominal variables and with interactions.  I believe the problem is that many of the combinations are largely redundant (and totally redundant in the cases of the zeroed out predictor).  Intuitively, the combinations are rare enough that the other variables in the model make those combinations unnecessary - they really don't add anything to the predictive power of the model (also see by the very large p values).  I have never understood precisely what is meant by "Unstable" in a mathematical sense, however.

[#3 · Accepted solution]

Hi @MedianPuppy4982,

I suggest you read the JMP Help section linked to your problem : Models with Linear Dependencies among Model Terms (jmp.com)

Your question is a very frequent one in the forum. You can check previously answered posts :

LostDFs for 2-Way Anova 

I am receiving Lost DFs, Biased and Zeroed parameter estimate messages, what do they mean? 

Basically, you should check the "Singularity" panel, you will see how linked/correlated might be some of the terms in your model.

• If some terms are completely correlated (confounded), then JMP is not able to differentiate them and use them independantly in the model, resulting in a "Zeroed" result.
• If some terms are correlated between each others, then their estimates are biased as they are not independant. This phenomenon is called multicollinearity, resulting in biased estimates and inflated confidence intervals.
• Finally, you may also have introduced more terms than the number of unique observations/experiments available, so this might create linear correlation between some terms that make them either impossible to estimate independently ("Zeroed" or empty), or difficult to estimate independantly ("Biased").

I hope this answer will help you,

EDIT: For "Unstable" terms, look at the estimates and standard deviations. There is a lack of data to precisely estimate those terms (and so many terms), resulting in a very large variance. Try refine your model or test other (more simple) models.

[#4]

@Victor_G Thank you so much for your response! 

I actually wanted to follow up about the singularity panel. I was looking through other post and some have responses that say singularity details well not be shown unless you add models that are not supported by the data. I also saw that it shows up on the Fit Lease Square Reports but all of the factors are categorially so we had to run a nominal logistic. I just wanted to double check that this would imply I would not see the singularity panel.

is there another way to check this correlation without it ?

[#5]

It's not correlation, but confounding.  You don't have enough DF's for the model you input.

Correlated variables in the model would result in inflated VIF's

[#6]

@statman 

Thank you for your response! 

I was under the impression with only three observations I would have only 2 degrees of freedom. Just to try it out i increased the DFs to 3 and the singularity panel still did not come up neither did any measurement of VIFs.

[#7]

I'm sorry, but I don't understand your scenario.  DF's are a function of the number of independent comparisons available from the data set.  It is a function of how the data was collected.  You can't just make up DF's.  The DF's available and how you assign them to the model will impact singularity.

[#8 · Accepted solution]

Here is a DOE with 10 degrees of freedom:

I could replicate the design, adding 10 more runs to double the DFs to 20. However, I can only estimate the slope and intercept.

If I try to estimate the quadratic term for X1, I get the singularity because the data does not support estimating this effect:

You need a combination of sufficient DFs and the correct levels in the design matrix to support the model matrix for estimation. Using DOE assures you that you can estimate all the terms in the specified model. Ad hoc data collection does not offer such a guarantee.

[#9]

@Mark_Bailey Thank you so much for the clarification, this was really helpful