Let us say you have a running process. What circumstances and/or events trigger you to recalculate the control limits?

Let us say you have a running process. What circumstances and/or events trigger you to recalculate the control limits?

Only a statistically significant change in the data. Davis Ballestracci once told me - "a baseline is innocent until proven guilty". And it is only proven guilty once it trips a trend rule in the rules you use.

I do not rebaseline on the exhortation of someone stating the process has changed. Yes, we should record that circumstance and/or event, but the confirmation will be when a trending rule is tripped.

OK - Let's say you have a process parameter that you change. That would trigger a recalculation, wouldn't it?

OK - Thinking about this - You make a process change. You don't recalculate until it shows up in the data, correct?

OK - Thinking about this - You make a process change. You don't recalculate until it shows up in the data, correct?

I think you have it!:yes:

By the way, I have a paper posted somewhere on the Cove called "Life Cycle of a Trend" that walks through this idea.

Thanks, Steve!

I think you have it!:yes:

By the way, I have a paper posted somewhere on the Cove called "Life Cycle of a Trend" that walks through this idea.

That paper was written on glass - so clear!
Thanks for it.

When I saw this question yesterday, I thought it was a set-up (experimeter bias)!! I figured I would let our residents guru… oops sorry, that word slipped out… resident expert on control charts, weigh in.

Interesting: when I entered “recalculate control limits” in Google, the thread in SPC overview on the Cove was one of the top five hits. Steve’s paper on the Hanford site was a couple down from that one.

I am interested in different theoretical approaches (if there are any) to recalculating control limits.

Recalculating control limits is fairly straightforward (I believe) when an outlier is removed.

I’m interested in other times to recalculate control limits. Would it depend on the nature of the process when I introduce the control chart? If the process is stable and in control when I introduce the control chart, then the calculated control limits are probably good limits for the control. However, if the process is not in control, or has a lot of noise, then my control limits may be wide. As my process improves, the standard deviation should become smaller, yes? I may want tight control on the process. So maybe I’m interested in process improvement here.

Another approach is the “good enough” approach. If at any time I have my control limits, and they demonstrate the process limits as being within the specifications, isn’t my process “good enough”? The customer can be satisfied that they have statistical confidence good product is being shipped.

Any thoughts/ opinions/corrections?

There are different schools of thought on this topic. Many people subscribe to the approach that Steve proposed. I am a contrarian on this topic. I believe that your should rebaseline after a preselected time period (every 3 months, or every year). Here is my logic:

Most people do not use all the trending rules because it diminished the Average Run Length to a point that there will be too many false alarms. If the control chart does not trigger an out of trend signal when the process improves, the control limits could be too wide. If you recalculate the limits on a periodic basis you ensure that the limits reflect the current process variability. If the process is truly stable then the new limits will not be significantly different than the old ones.

Also, you would recalculate and rebaseline the limits if the initial limits were set on a small data set or a pilot process.

Just my thoughts!

If you recalculate the limits on a periodic basis you ensure that the limits reflect the current process variability. If the process is truly stable then the new limits will not be significantly different than the old ones.



Just want to clarify.
I presume that you mean re-calculating is different from actually changing control limits.
are you saying that we must periodicaly re-calculate control limits to check for stability but not necessarily actually changing it?

Then if we see that the recalculated control limits is significantly different from the previous one, it would trigger a Root Cause Analysis?

Separate Question:
We have a practice here where i work (manufacturing) that im having a hard time understanding the basis for doing so.
We periodically change control limits (per month) even if there are no improvements made. The re-calculated control limits are always significantly different (due to incoming material's variability, as concluded by an old MSA).
What we do is when the re-calculated control limits are not different or tighter, we use it.But if the recalculated limits are wider, we retain the previous one.
Anyone knows same practice and how to react/deal with it?

Most people do not use all the trending rules because it diminished the Average Run Length to a point that there will be too many false alarms. If the control chart does not trigger an out of trend signal when the process improves, the control limits could be too wide.

That is true - if you only perform the x chart, there is not a good rule for detecting a decrease in variation. If you are after a decrease in variation, then you could plot either a s chart or a range chart to go along with the x chart.

I am definitely against the theory of recalculating average and control limits at a certain time interval. You could end up hiding a slowly moving trend. And yes, how would you know that if the shift to the new average and control limits represents a signal worthy of action (such as root cause analysis)?

Anyone knows same practice and how to react/deal with it?

Yes. Dr. Deming would call your practice "tampering". His funnel experiment (I believe I have also posted a writeup on the funnel experiment somewhere here) is a good "answer" to show what is happening. Another thing to do is to do Dr. Deming's Red Bead Experiment, or some other data generator that is known to be a stable process, and see what happens when you follow the rules you are using. See how much misinterpretation of what is happening occurs.

That is true - if you only perform the x chart, there is not a good rule for detecting a decrease in variation. If you are after a decrease in variation, then you could plot either a s chart or a range chart to go along with the x chart.

I am definitely against the theory of recalculating average and control limits at a certain time interval. You could end up hiding a slowly moving trend. And yes, how would you know that if the shift to the new average and control limits represents a signal worthy of action (such as root cause analysis)?

You and I will disagree on this topic. Even with a R or S chart, you will not be able to pick up subtle changes in variability or average unless you use all the Western Bell rules. For example, waiting for 6 times points in a row below the center line on a S chart might never happen though the overall S might have been decreased.

I understand your aversion to recalculating control limits (and changing the limits) after a set period of time. How do you propose to find slow moving trends? If a process is trending down and it takes 18 months to get 8 in a row below the center line, then I would contend that the trend is not statistically meaningful. Call me a rogue statistician :)

Let us say you have a running process. What circumstances and/or events trigger you to recalculate the control limits?

The most correct answer to any question is "It depends."

If, for example, this is a precision turning operation and the only significant variation in the process is tool wear, the answer is never. First, the distribution is the uniform distribution - which is non-normal - so the typical control limit calculations do not apply. Secondly, the notion of compressing the control limits as a form of continuous improvement also is erroneous - it would actually increase operator intervention, which would result in overcontrol and increased variation.

For a truly normal process, it would make sense to at least review the data one per year...to check for degradation or improvement.

Bob Doering