First of all, the responses you got were fine about the math, but everyone misses the real point here.
You should not be calculating Cpk at all on this data.
Any use of a capability metric on a process that is not known to be stable is of no use. None at all. This is the first rule of all capability assessments.
Cp and Cpk are only re…[Read more]
No capability metrics should ever be calculated on a process that is not stable. So you first need enough data to demonstrate stability, generally 20-25 points is the rule for a control chart. This can be from a single batch or from a running process.
Cp and Cpk should never be used unless the population is shown to be normally distributed. I…[Read more]
A great question. I was asked the same by a student in the past month.
The answer is to use a square root transformation. It turns out that the time-between-failures typically has a poisson distribution if it is stable. In this case, the square root is the correct transformation. You can check this with simulated time data using a poisson ran…[Read more]
If the output being evaluated is a pass/fail value (binomial) then the logistic tools are appropriate.
For logistic regression, you would use inputs (predictors) that are continuous terms. Your question discussed using % of the factors. This would be OK as long as it is percent of a continuous variable. If they are discrete terms and you are…[Read more]
You are facing a common issue. There were some good suggestions provided, but in my experience you need to ask why it does not fit a common distribution.
Control charting is a key tool before a capability analysis. If it is not stable, then you should not calculate a capability from it, this is because the instability cause may change and in…[Read more]
This can be answered without all the technical knowledge.
Time measurements (usually) have a minimum of zero. There is a natural limit that can not be passed, no negative times allowed. In this case as time values approach zero, they must be skewed away from zero. In general this data has a lognormal distribution. OK.
Now has the rec…[Read more]
Just removing steps because the group think or conventional wisdom says it is Non-Value added is risky. Many a book learned Lean leader quickly drops a process step that the team voted as non-value, then claims a savings in labor, finds out in a few days or weeks that the step was important when it was instituted years ago, and no one…[Read more]
Others gave you some answers, but I want to talk about the standard deviation question.
Defective data has a binomial distribution, which has an approximate standard deviation of sqrt((p*q)/n). This uses the normal approximation of the binomial, which is adequate if n*p and n*q are both greater than 5. Your data would meet this c…[Read more]
This generally happens when you have more terms in the model than you have observations. If you are testing a lot of different factors or included all the higher level interactions, this would be the cause.
drop the terms with the smallest MS value, one at a time until you start getting an error term.
I understand what you are saying, but you may have a different history that I have seen. I mentored under Dave D. a few years back and I saw him shorting a number of topics to make it easier on that round of students and MBBs he was to be training.
I interviewed some 3M folks for employment recently, since they decided to cut nearly all six s…[Read more]
I work for a consulting firm in this area and the article about 3M was surprising at first read, at least for the Lean Six Sigma we provide and instruct. Our discussions went in these areas.
3M believes it so it was true in their case. Why is it not pervasively true? We came to a couple of ideas. Six Sigma started as a business system to impr…[Read more]
Rick Haynes replied to the topic Relation of Control Limits with Type – 1/2 Errors? in the forum General 14 years, 9 months ago
I think this question is not answerable.
Type 1 and 2 errors are associated with hypothesis testing. Control charts (all of them) are not hypothesis tests. I am not sure if discussing the relationship may cause the control charts to be inappropriately used.
Control charts are used to indicate when a possible process change event occurs. An ou…[Read more]
I would not agree with some of the guidance provided on this topic. If you are starting a quality program like Lean Six Sigma or what ever, Doing it yourself leads to some real future issues.
I would recommend any organization start with interviewing the leading consultants and providers of lean six sigma or other method. Look at their m…[Read more]
I believe that following some of the prior advice may lead to a loss of insight about your process.
Generally, just making a lower control limit 0 when it calculates as an impossible (-) value is a quick fix. Any condition that creates this derives from a distribution of data that is by its nature skewed. Since it is skewed to the (+) side, t…[Read more]
Greg, that is a question that has been seldom asked from most six sigma efforts since the DMAIC cam about.
If a student asked me, I would then ask why do you ask?
Generally the use of opportunities and DPMO is intended to provide a baseline where multiple products/process quality performance can be combined into a rolled up scorecard metric. S…[Read more]
The answer is based on your use of the measure. If the goal is to relate the average to the cost of the process or to resources needed, you would use the average of the actual numbers. If you only wanted to represent the typical value, you might use the reverse transformed average from the transformed data imr chart. In many cases the re…[Read more]
Being a Statistician, I see this question a lot. Especially since Minitab and other software packages will perform the chi-square test with counts less than 5.
A layman’s answer can be developed by considering that every cell or combination has random variation in the true count value. (easy assumption) If a cell has a count of two, what pe…[Read more]