How to identify the cause of decrease in a product quality attribute?

Question

I am new user to JMP and would like to know which factors are actually causing the decrease in a quality attribute. The quality attribute is normally affected by different variables, including composition of raw material sources (e.g. 70% A and 30% B, or 30%A, 40%B and 30% C, etc) and some process variables. What kind of analysis in JMP can I perform for this case? Thank you.

Mark_Bailey · Accepted Answer

First of all, some of the methods are only available in the JMP Pro version of the software.

Second, some of the methods depend on the modeling type of the response and predictor variables, though it is not usually a limitation.

Assuming that you have a data set with observations on all those variables, there are many approaches that you might use. JMP is very interactive. The dynamic linking feature is a simple way to explore potential factors. When you select observations in the data table, they are highlighted in every open platform. This way can be expanded on using a Rows > Data Filter. You could select response levels that are important to you and see if the selection has a meaningful distribution in the other variables.

You can use a matrix of scatter plots to explore potential two-variable relationships.

You can use regression analysis of various kinds to determine if variables are related in some way.

You can use recursive partitioning to assess the contributions of each variable.

There is a Predictor Screening platform that is designed to help with this specific task.

We offer training about this very subject. The course is called Finding Important Predictors. There is a public class scheduled soon.

statman · Accepted Answer

Welcome to the JMP User Community!  You will find JMP incredibly useful for assisting in understanding relationships both graphically and quantitatively.  Your situation, however, requires more than just data analysis and more than just some thoughts in a community forum.  You might want to explore some of the on-line tutorials offered to get a better understanding.  Probably the most generally useful platform will be the Fit Model platform.

Here are some of my thoughts:

1. Your journey to understanding causal relationships may be most effective if you follow scientific method.  That is a continuous iteration (induction-deduction).  Start with hypotheses (rooted in the Sciences) as to what variables (x's) you think affect the quality attribute.

2. Question how you measure the quality attribute (Y).  Is the measure a continuous measurement (which will be far more efficient than a categorical response)? Is there more than one Y to describe the "product"?  Has the measurement process been studied?  How confident are you in the measurement system?

3. Consider the data source (what data is available?, what data do you need to provide insight into your hypotheses?) e.g.,

* Historical or Observational: In this case we might try some sort of regression procedure to look for clues.
* Directed Sampling: In this case we sample specifically to separate and assign sources of variability as well as check for stability
* Experimentation: In this case we accelerate the learning by manipulating x's (linked to your hypotheses) and study the relationships across changing noise conditions.

JMP will help in the analysis of the data, but it is important to keep in mind:

The information revealed, the questions that can be answered, appropriate actions to take, confidence in extrapolating the results, what tools you use for analysis are ALL entirely dependent on how the data was acquired.