A
alokb
When dealing with non normal data should we try to convert data using transforamtion or Use Capability Analysis (Nonnormal) available in Minitab .
Which is a better way of doing it.
Alok
Which is a better way of doing it.
Alok
P
pabloquintana
First thing is trying to understand why your data is non-normal. You can try to plot it in a time line to see for patterns, etc.
Then MINITAB offers two options:
1. Convert the data (Box Cox available in Normal Capability dialog) and analyze it. Be warned that your limits and specs will change, but you'll get the info.
2. Identify the distribution first using the Individual Distribution Identification and then use the Non-normal capability analysis.
MINITAB recommends the first option.
Good luck.
Then MINITAB offers two options:
1. Convert the data (Box Cox available in Normal Capability dialog) and analyze it. Be warned that your limits and specs will change, but you'll get the info.
2. Identify the distribution first using the Individual Distribution Identification and then use the Non-normal capability analysis.
MINITAB recommends the first option.
Good luck.
pabloquintana is correct about the two options. I wanted to comment on the first paragraph.
Some processes may indeed be non-normal because the process is out-of-control. However, some processes are inherently non-normal. Some good examples of this are characteristics with a boundary condition (physical limit) such as zero. You cannot have a parallelism or flatness less than zero. Therefore, the closer you approach zero, the more non-normal the process will typically appear, while flatness distributions situated at a distance from zero may appear quite normal. A process that uses a physical stop to control a dimension will typically have a non-normal distribution because you cannot get a dimension past the stop, but you can get any dimension short of the stop. Screw machines were a classic example of this situation.
In summary, check the obvious causes such as out-of-control, but do not think that a non-normal process is always caused by that. It may be the "normal" condition. If you can provide more information on the nature of the process and the characteristic in quetions, we may be able to comment on the expected distribution.
Some processes may indeed be non-normal because the process is out-of-control. However, some processes are inherently non-normal. Some good examples of this are characteristics with a boundary condition (physical limit) such as zero. You cannot have a parallelism or flatness less than zero. Therefore, the closer you approach zero, the more non-normal the process will typically appear, while flatness distributions situated at a distance from zero may appear quite normal. A process that uses a physical stop to control a dimension will typically have a non-normal distribution because you cannot get a dimension past the stop, but you can get any dimension short of the stop. Screw machines were a classic example of this situation.
In summary, check the obvious causes such as out-of-control, but do not think that a non-normal process is always caused by that. It may be the "normal" condition. If you can provide more information on the nature of the process and the characteristic in quetions, we may be able to comment on the expected distribution.
L
Licht
Hi.
Where could I find examples of non-normal processes ?
tks
Where could I find examples of non-normal processes ?
tks
Last edited by a moderator:
R
Renzo2112
Hi everyone,
This is my first suggestion on this forum, I hope it helps everyone. Please follow the next steps:
1. The capability analysis study for non-normal data should be plotted in a histogram.
2. This first analysis will help you to check your data if fits well in the specs.
3. In Minitab go to Stat>Quality Tools>Capability Analysis>Normal, click on Box-Cox button and click on Box-Cox power transformation option. For default Minitab will use and calculate the optimal lambda.
4. Finally on the screen will appear a graphical summary with Cp, Cpk and all the required information needed.
Good luck
This is my first suggestion on this forum, I hope it helps everyone. Please follow the next steps:
1. The capability analysis study for non-normal data should be plotted in a histogram.
2. This first analysis will help you to check your data if fits well in the specs.
3. In Minitab go to Stat>Quality Tools>Capability Analysis>Normal, click on Box-Cox button and click on Box-Cox power transformation option. For default Minitab will use and calculate the optimal lambda.
4. Finally on the screen will appear a graphical summary with Cp, Cpk and all the required information needed.
Good luck
Power (e.g., Box-Cox) transforming data for capability analysis (and control charting) is a contentious issue. The purpose of a capability analysis is to provide the customer with evidence that the process/production we run for him/her is running reliably and within specifications. The problem with normality transforms (like Box-Cox) is that they tend to camouflage outlying data points (those that otherwise inflate our capability indices) by "hiding" them into the bulk of the other data points. Therefore the acquired capability index will automatically be improved by a mathematical trick, rather than by any real reflection of the intrinsic process. The transform yields an ad-hoc capability index; it was not calculated objectively, instead it was optimized in a way that only benefits us as the producers. I believe that the only time you can honestly use a power transform of your data is if you have been able to model your process to one such particular transform. Unfortunately I believe that in real life this is very difficult to achieve. Bottom line: By power transforming your data before performing a capability analysis you will indeed improve your capability indices. But if your customer is statistically savvy, he/she may question your honesty.
The SPC guru Donald Wheeler has blogged about issues like this (try google).
Miner (above) has a strong point explaining that some processes are inherently non-normal. Actually it could be claimed that most processes are inherently non-normal, we are only often lucky to be able to use the normal distribution as an approximation because the standard deviation of our data happens to be much smaller than the average.
The SPC guru Donald Wheeler has blogged about issues like this (try google).
Miner (above) has a strong point explaining that some processes are inherently non-normal. Actually it could be claimed that most processes are inherently non-normal, we are only often lucky to be able to use the normal distribution as an approximation because the standard deviation of our data happens to be much smaller than the average.
F
FAndrade
Why no weibull?
Please clarify your question.
F
FAndrade
After my normality test fail (p<0,05), I use capability test for non normal distribution, using weibull model.
Similar threads
