# Cpk roughly double Ppk How to fix this?

#### Christopher Lear

##### Registered
Hello Bev and Minor,

1. Each subgroup is a measurement of each lane laterally across knives.
2. We take measurements at three different intervals throughout the day. This is to catch is a knife moves for a various amount of reasons.

If I measure one lane the variability would be very small unless something goes wrong with a knife then we would discard those parts. Also, if I measure one lane over and over would I really be capturing the process of a multiple knife run?

#### Miner

##### Forum Moderator
Well, you are actually mixing different process streams. With each strip being a different process stream. Normally, you should chart different process streams separately, UNLESS you can show that they are statistically equivalent. In this case, they happen to be equivalent (see ANOM below), so you can subgroup them as you have done. This then brings us back to my initial judgment, which was that your process is unstable. It appears that your knives are moving or wearing over time. The next step is to determine whether this is due to deliberate movement by operators, unintentional loosening and moving due to inadequate clamping/fixturing, or wear over time (or anything else that may be possible for this process).

#### Bev D

##### Heretical Statistician
Super Moderator
So your subgroup is made up of 1 measurement from each of the 5 knives. And each ‘subgroup’ is measured at 3 different times each day?

This is not a rational subgroup as each knife is independent (and not randomly different) than the other knives. This non random variation falsely increases the standard deviation within each subgroup. Then you have variation from subgroup to subgroup that is quite large compared to teh within subgroup variation most likely due to adjustments - this too is creates a non-random variation between subgroups.

The idea of leveraging the ‘bell curve’ relies on randomness of the variation of the data points in relationship to each other. This is clearly not present here. Teh non-random variation between knives and between subgroups overinflated teh Sd. Do you know when the adjustments are made? To which Knife? Can you annotate this in your data?

You have confused SPC with Process capability (although they are intertwined by teh subgrouping scheme you chose), they are not the same.). Both are affected by not rationally subgrouping.

#### Bev D

##### Heretical Statistician
Super Moderator
Well, you are actually mixing different process streams. With each strip being a different process stream. Normally, you should chart different process streams separately, UNLESS you can show that they are statistically equivalent. In this case, they happen to be equivalent (see ANOM below), so you can subgroup them as you have done. This then brings us back to my initial judgment, which was that your process is unstable. It appears that your knives are moving or wearing over time. The next step is to determine whether this is due to deliberate movement by operators, unintentional loosening and moving due to inadequate clamping/fixturing, or wear over time (or anything else that may be possible for this process).

View attachment 29982
Not so sure about this. Homogeneity isn’t about the averages (ANOM) it’s about the individual values in relationship to each other within the subgroup. I don’t see that in the multi vari? There is a clear pattern of knife 1 being ‘higher’ than knife 5 in almost every subgroup.

#### Christopher Lear

##### Registered
Ok, so if I understand correctly... I should measure say lane one 25 times and have only 1 subgroup for this analysis?

#### Christopher Lear

##### Registered
My current subgroups were made up of 25 lanes collected 5 times.

#### Bev D

##### Heretical Statistician
Super Moderator
Oooh. That’s different than what Miner graphed.

#### Bev D

##### Heretical Statistician
Super Moderator
Ok, so if I understand correctly... I should measure say lane one 25 times and have only 1 subgroup for this analysis?
Uh not necessarily. We need to look at the data to determine the best approach. This cannot be done blindly to the process and its actual performance.

#### Miner

##### Forum Moderator
Yeah, that is totally different from what I understood. So the subgroup size is 25?