I've got a distribution of parts, for say their length in millimetres, roughly 5000 parts in total with the median of 3.53 (see below)
What I want to do with this distribution of parts is essentially create 2 distributions from it and run 2 evaluations on each distribution. One distribution needs to have a median of 3.52 and the other 3.62, both with similar sigma's (0.28). Parts cannot go into both groups it must be one or the other. So each distribution should look like below:
3.52 Eval distribution
3.62 Eval distribution
I currently have a method of doing this but its quite long and tedious. I was wondering is there any quick way of doing this in JMP?
What is current method? I would guess that it can be scripted to automate.
Given the variety of possible subset combinations (there are multiple ways to generate subsets A and B which mean your criteria of median and sigma), I am not sure of one generic response.
What is your goal of these subsets?
Scott, I am wondering why you want to split the distribution. Is the time-ordered data showing a system in or out of statistical control. Are the parts coming from different processes/machines? Why does one median have to be 3.52 and the other 3.62? What is the purpose of the study?
I need to split the distribution because both eval's will go through the same testing scenarios (next stage) but we want to see what effect the length has on the next set of test results. In order to do this I must give a set of part numbers for the 3.52 distribution and part numbers for the 3.62 distribution in order keep track of the eval groups.
Is there a way of using the random() function to manually create a distribution of parts with a set mean and sigma? and from that able to select the two groups of parts without overlapping parts into the two groups?
Scott, check my understanding of what you are trying to do: You have a population of almost 5000 parts. You want to create two sub-populations from the parent population having medians of 3.52 and 3.62, and having similar sigma values (0.28). Then you will evaluate the two sub-populations to determine if the difference in medians has an effect on a succeeding test. If this is what you are trying to do, then I would question the statistical validity of your approach. If this is the case, then I suggest you build a DOE with length being an independent variable. I have been using Definitive Screening Designs and have found them to be very efficient.
Another point regarding the validity of your approach: If your proces for producing the 5000 parts shows a "reasonable degree of statistical control" for length, then your approach is certainly not valid. However, if the process is not stable (and the good-looking histogram does NOT confirm whether or not your process is stable because the histogram loses time-orderliness), then you can select samples from periods of high and/or low length values and compare them for evaluation. However, I would still want to do a properly coonducted DOE.
Yea you pretty much understand what I'm trying to do. I do agree with your methods but there's a few underlying reasons why it must be done this way. Basically we cant do a DOE at this stage, we know length plays the main role into the next testing phase. Selecting samples from high parts and low parts would not have the same sigma for the length parameters that the next testing stage would see if full production quantity was tested through, so that's why we would like our sigma's similar.
I don't think you will get the answers you need with your current approach. HOWEVER, you might try Recursive Partitioning and/or Neural Networks to "mine" the data to understand the effects of length. I have found this to be a useful approach to understanding key variables from the profiler(s) developed.
Thanks Lou but i need to take a set of real data and split this into two, your method creates two random datasets with a set mean and sigma. Nearly there