Multivariate Model Driven Control Charts

[#1]

I have prepared MDMCC applying grade wise filter for different grades. After this I removed the outliers but after removing the outliers automatically few more observations appeared above the red line which were earlier below.

[#2]

Hi @Chitranshu,

The behaviour you're seeing is normal and can be seen with any control charts.

Since you're removing/excluding some points based on your previous control chart, it will change the dataset on which the new chart is created : the mean will be recalculated as well as the control limits, so points that were previously within the control limits before could be outside of the newly calculated and tighter control limits.
For the calculation of multivariate control limits, you can look at the equation here : Statistical Details for Individual Observations (jmp.com). You can see that the control limits are for example influenced by the number of variables p and the number of observations m, so if you remove some points, the number of observations won't be the same anymore, hence the changes you have seen.

I don't know the context of your project, but you can maybe have a look at STIPS course about quality methods, and discussions about the use of control charts : Statistical Thinking for Industrial Problem Solving | JMP You can also have a look at this free course by JMP : JMP Statistical Process Control Course 

The purpose of control chart is to study process outcomes over time, not to remove outliers.

It's a graphical tool to differentiate if the variability comes from random/"normal" variation (process in control, within the range of control limits), or if it comes from extra causes of variation that have to be identified and resolved.

Other members more experienced than me could give you more info, guidance and references about process control methodologies/control chart if needed.

I hope this answer will help you,

[#3]

Victor's explanation is correct.  Why are you removing the "outliers"?  These may be the most interesting and informative points in your data set.  The objective for using control charts is to partition and assign different components/sources of variation.  It is not to be inside the control limits.

[#4]

If we are preparing a model for monitoring purpose & if some parameters have behaved abnormally [ bad operation or faulty instrument at that particular time] in our historical data set and we have not removed those outliers then our current data will not flag if there is any such abnormality.

[#5 · Accepted solution]

Sorry I'm confused what you are doing?  I will admit, I use control charts as a diagnostic tool, not to monitor (which is reactionary). You can use run charts with spec limits if you want to monitor and react. Control charts as they were intended are a diagnostic to help separate and assign components of variation. Points out of control on a range chart indicate special/assignable cause variation associated with the within subgroup component of variation.  It may also be indicative of an inconsistent or unstable process. It is worthwhile to investigate what was going on at those points in time, but what action you would take to alleviate such events depends greatly on whether those are common or special cause events (see Deming). Why are "parameters behaving abnormally"?  Control limits are not an assessment of good or bad.  Points beyond control limits on an X-bar chart indicative the sources of variation associated with the between subgroup component are assignable and estimable (and may be dominant).  Again, nothing to do with good or bad.