Nowhere in the article was it suggested that EWMA Charts should replace Shewart charts...they are a valuable tool to detect certain kinds of shifts just as Shewhart Charts are. A EWMA Chart uses a single test - one point more than 3 standard deviations from center - whereas you are suggesting using four tests. It is more difficult for operators to use one test than to use the same test PLUS three others? The only complexity is in calculations that are done by computer, invisible to the user. Just as more complexity does not necessarily make something more useful, more simplicity does not either. A car is more complex than a skateboard for a reason. As long as the car is simple to operate, it does not matter that what's behind the dash is complex. Additionally, the "sensitivity" you suggest is obtained by doing more tests and increasing the false alarm rate as well. Simulation demonstrates that the rate of catching small, sustained shifts is similar between the two charts, but on an I Chart with the four Western Electric rules the real signals are hidden in a sea of false alarms that are occurring at a higher rate than the real signals. The boy who cried wolf was also good at detecting real wolves.

Well, in an effort to make things simple, we get things confused.  An EWMA chart has specific uses.  Its derivation is from a non-stationary IMA (1,1) model (integrated moving average).  Blah, blah, blah, but the key is the fact that the assumption is that the underlying model has a non-stationary mean.  This is entirely plausible for something like business data.  It is absurd to assume something like inflation resets itself year after year just because Jan 1 first rolls around and each point is independent of each other.  Rather it rises and falls somewhat independently of an arbitrary time subgrouping.  Unemployment is another example – you can go out to the BLS website and find serial data to your heart’s content! 

 

Within manufacturing, chemical process data can often benefit from using an EWMA chart rather than a IX chart, because of mixing, residence time, consumption, etc.  If you want to use a IX-chart effectively here, you can solve the issue of the sample rate (how often to sample) by looking at such things as autocorrelation function plots (ACF) and partial autocorrelation function plots (PACF) but that is beyond most engineers knowledge.  Simple lag plots, seeking for no correlation, is a simple enough tool to get close to determine how often to sample.  But now to balance the sample rate with the IX-chart effectiveness, you may increase risk of not seeing a real shift/trend and that’s a business decision. 

 

Sticking one's head in the sand and saying the IX chart is a universal tool for all data is problematic at best.  If the data is non-stationary, then it is non-stationary.  An IX chart with any set of rules will flag all the time and people will also look at you funny when you say it's always out of control.  You are looking at a significant rate of change, not for "stability".

 

I don’t use EWMA very often, but I do use it when it is appropriate for data that have non-stationary means by definition.  And I think that was the point of Joel’s article, which is “here is another tool for specific moments.”