Control Chart Builder and skewed variables

[#1]

Hi All. Thanks for reading my question and your time;

I am very inexperienced about the Quality and Process tools. I want to use the Control Chart Builder for the compositional data measured in our lab for a Chinese product. This type of data is skewed some times for some elements, like CaO.

Can I still use the control charts if the data is not normally distributed?

By the way, I have the min and max spec limits for the elements.

Thanks.

[#2]

If you are inexperienced at basic statistical process control methods, before blindly using any software application or charting method, I strongly suggest some general education and skill building. A good place to start is the JMP Statistical Thinking for Industrial Problem Solving free online course. There is much more to appropriate use of SPC methods than distributional issues. Here is the URL for that web site:

https://www.jmp.com/en_us/online-statistics-course.html

If your work is strictly limited to SPC methods, studies and techniques then at a minimum I recommend completing the first three modules in the sequence of modules.

[#3]

Thanks for your suggestions. I understand and I enrolled a few weeks ago. I also have watched 2-3 hours of webinar videos with JMP showing how these tools are used. But the type of question I had was not covered in the videos. I dig deeper into the course materials then.

[#4]

Regarding your question about the assumption of normality for control charts, I suggest you read Wheelers book "Understanding Statistical Process Control".  There is NO normal assumption for control charts as originally derived by Shewhart.  Rational subgrouping and sampling strategies are imperative to properly use control charts. It sounds like you are wanting to use control charts for incoming materials.  This not what they are intended for, but that doesn't mean you can't adapt the charts for your intended purpose.  Since you will have limited or no rational subgrouping (The more you understand how the supplied materials are manufactured, the better you will be able to develop rational subgroups) strategies, you might be limited to looking at the time series using Individual Moving Range charts.  This implies you KNOW the time series the incoming material was manufactured in.

I have attached a good article that covers the basics.

[#5]

Thanks for your help. I will also read the article you sent, thanks.

The data are grouped as below and in the attached file to make sure if you have any more suggestions. I think I am not worried about the moving average here.

[#6]

What questions are you trying to answer?  Always start with questions & hypotheses before deciding the appropriate data collection strategy and subsequent analysis needed.

[#7]

Some of the XRF elements and some of the other columns in the previous attached image are very important to us. I am going to see the materials we are buying has a consistent quality; if the variations are so high in the abovementioned columns, then we need to contact the supplier. Before that, we surely replicate the testing to make sure again. We can also capture the variations visually and show them to suppliers if needed. Does this justify the use of the graphical control charts in JMP for my case?

The variations may still be within the spec limits, but we don't like to work with a highly variable product. It causes other issues sometimes which are out of the scope of this discussion.

Thanks.

[#8 · Accepted solution]

If you are trying to determine if the incoming materials are consistent, an Individual, Moving Range chart will work as long as you know the production order of the materials (not just the ship date).  Here are some additional thoughts:

1. You can assess your companies measurement systems for both consistency and precision using control charts.  Design a measurement study to incorporate components of variation of the measurement system (e.g., precision - repeatability of the equipment, reproducibility across multiple technicians and stability).  Since you don't have control over the manufacturing, you will likely use randomization to get samples over a wide enough inference space (this is not ideal) to determine your measurement systems adequacy.

2. Using simple Variability charts will go a long way to visually communicating the variability of the incoming materials.

3. Realize consistency of the incoming product has multiple components.  Consistent within "batch", consistent "between batch", etc.

4. Most spec. limits are NOT rigorously or scientifically determined.  They usually have no relationship to the natural variation of the materials they are applied to.  They are not useful in determining consistency, nor are they useful in determining causality.  About the only use is to determine if you want/need to spend resources investigating the process.  Control charts are designed to do two things:

• give you insight into which set of x's has greatest leverage in the process (where to your investigation should focus)
• what is the nature of that variability (is it special or common, so in essence how should your investigation proceed)

[#9]

Thanks again for your time and insightful answer. I accepted your final answer as the solution. I feel so dumb now because I figured there are so many things, terms, etc. according to your answers which I should learn in this area. To me, it seems that Quality and Process tools are even harder than DOE and predictive modeling!! Or it seems it's a very deep section covered in JMP software.

Just a last question, what do you mean exactly by production order?

I think in my case, the shipment order/date is the production order.

Thanks again.

[#10]

No worries, I've been at it for over 30 years and still learning.  I also started with a bias towards DOE, but learned directed sampling can be extremely effective and efficient.

By production order, the order the product was manufactured in.  It may or may not be reflected in the ship date or even the date code.  Much depends on how the product is manufactured (e.g., continuous, batch or one-piece flow). The Moving Range chart is dependent on the time series (or some rational series).  If you randomize the samples and put those on a MR chart you get completely different answers, so the time ordering is imperative.  You are trying to see if the process is consistent.  How is that done....compare something to itself over time.  This is what all range charts including the MR chart do.  There will always be variation, so the control limits are there so you don't over react to "common cause" or "normal" variation.