Do Attribute Control Charts really help?

[Chennaiite]

I am learning to apply Control Charts to understand process signals. Of late, I tried to use Control chart for some of our attribute data but find it very insensitive to identify signals that require attention. I am attaching a sample here with hypothetical data though. I have constructed both Xmr chart and C Chart. XmR apparently does a better job. This is not the only instance. I tried it on 3 other data and all of them have ended up with wide limits. Please help me. Is there anything I can change in my approach or perspective?

 

[Steve Prevette]

Several things:

1. In general, if you have measurement data, it is much better to use Xmr charting rather than collapsing the data to percent that are acceptable, or percent that exceed a certain threshold.

2. A c-chart is only used when you expect something to behave as a Poisson Process. Poisson events must be independent from each other and have a constant probability of occurrence in any given time interval.

3. Your example data is obviously NOT Poisson, given the vast difference between the XmR and C chart control limits. This may be a VERY IMPORTANT conclusion if you had expected your data to be Poisson. Dr. Deming relates a p-chart story where he plotted the percent defective over time, and the data were too tightly grouped around the average. Come to find out, there was falsification of the data occurring, with the inspector stopping counting defective product once a 10% defect rate was reached.

 

[Chennaiite]

Thanks a lot.

1. In general, if you have measurement data, it is much better to use Xmr charting rather than collapsing the data to percent that are acceptable, or percent that exceed a certain threshold.

I understand that. When I measure in percent, sometimes the data point is influenced by the denominator. Especially when denominator is low, even a small change in numerator effects huge difference in percentage. However, under some circumstances measuring in percentage makes the measurement meaningful and comparable. I think XmR is the best option even for percentages. But sometimes even XmR results in wide limits.

2. A c-chart is only used when you expect something to behave as a Poisson Process. Poisson events must be independent from each other and have a constant probability of occurrence in any given time interval.

Can I use XmR chart wherever I am not sure about the distribution type?

 

[Steve Prevette]

Thanks a lot.

Can I use XmR chart wherever I am not sure about the distribution type?

Yes. When in doubt, use XmR. In fact, author Dr. Don Wheeler would say always use XmR.

 

[Bev D]

I would add two points to Steve's input.
the data you show is very 'chunky' and this will lead to larger control limits even with an XMR chart. (I realize your attachment is a simulation, but the point still holds)

If you have a large varitiation in the sample size (denominator) the limits will be very different for each data point based on the samples size. (this is how robust the proper control chart is to all kinds of variation). If teh variation is large enough, teh XMR will not be the best chart.

Unfortunately the choice of control is not a process that can be done without thought, understanding of the process and understanding of the how the different charts and rational subgrouping works.

I once explained it this way: a chain saw is a great tool, so is a nail file. but I would never use a chain saw to file my nails or a nail file to cut down a tree.

 

[DRAMMAN]

When I worked at a company who was a global household name in the quality world our stats and SPC training was weeks and weeks of great training. The Attribute SPC training consisted of a paragraph that basically said attribute data is of little value and to figure out how to get variable data.

 

[Chennaiite]

I know a QC Instructor for JUSE. He even recommends what he calls 'barbaric control limit' i.e. a control limit based on the trend, that will logically separate signals. And of course sometimes the choice of control chart to improve the process itself is a common sense decision.

 

[Steve Prevette]

I should point out that at many points in the "real world", all you have are attributes data. For example, how many injuries do you have per month (C-chart). Or how many injuries per 200,000 hours worked do you have (u-chart). Or what percent of injuries are back injuries (p-chart). Alll three of these could be very important trending charts for the company safety program.

 

[Bev D]

I agree that it always better to monitor continuous data when possible. Not only is there better resolution, but if you have a highly capable process, you can detect changes before too many - if any - defects are produced. Other than some poisson events (like particle counts, microbial contamination, etc.) most categorical data represents defects. So your process is already 'not capable'. But it may be acceptable - at least in comparison to other problems. Customer complaints, service events, visual defects... in these cases a control chart will at least keep the organization from reacting to random noise and from missign true excursions - or improvements.

 

[DRAMMAN]

As Bev stated, the biggest value of using attribute control charts for tracking items like defect counts and defect percentages if that the charts almost always show the proces is in control so it helps keep the oganization from over reacting to one bad measurement period. This helps keep the organization focused on improving the system as opposed to wild goose chases in response to one negative event.