Jun 2 2026
Some Hazardous Ideas | Don Wheeler | Quality Digest
On 5/28/2026, in QualityDigest, Donald J. Wheeler bashed “hazardous ideas” from a list he received from Allen Scott. Allen is a frequent contradictor of mine on LinkedIn, and I welcome his comments whenever they go beyond calling my work “garbage.” Although Wheeler does not name me as the originator or a propagator of these “hazardous ideas,” I’ll take the bait and respond about the ones I actually support.
The “hazardous ideas” are all about SPC, which, to Wheeler, boils down to XmR charts, relabeled “Process Behavior Charts.” SPC receives more attention than it should, given that it is used in practice almost exclusively for ceremonial purposes. The most dramatic quality problems of the past 25 years were tread separation on Firestone tires mounted on Ford Explorers in 2001, Toyota’s unintended acceleration in 2010, and Boeing 737 MAX crashes in 2019. None of them was related to or solved with SPC.
Google’s ngram viewer shows the relative frequency of quality-related key phrases in English-language books peaking in the 1980s and 90s and waning since. “Quality Control” is now down to its 1940 level. All the other phrases followed similar trajectories, peaking later, the latest one being “Six Sigma” in the late 2000s. Oddly, the only one with any uptick since 2010 is “SPC,” and it dominates all the others.
The overall decline in interest in the field is worrisome, given that there is no evidence that the quality of manufactured goods has been improving since the turn of the century. The JD Power and Associates Initial Quality Surveys of car brands show massive increases in customer complaints, and the life of household appliances has been getting shorter. If the quality of manufactured goods is deteriorating, we should see more publications on fixing it rather than fewer, and they should focus on approaches that address today’s problems with today’s technology.
So, to contribute further to the excessive attention to SPC, let’s run through some of the ideas Wheeler deems hazardous.
Contents
SPC is not needed with a stable process
“It’s been said that once we have reengineered our process and gotten things working smoothly we no longer need to use SPC to monitor the process. This hazardous idea combines a naive view of production with a misunderstanding of what SPC does. Real-world processes do not remain unchanged.”
Michel Baudin‘s comments: I have consulted for major companies in aerospace and automotive that produced world-class quality products without using SPC. Working on projects with current and former quality managers convinced me that it was because they knew better.
They didn’t think they needed SPC for any purpose other than humoring outside auditors, and, given their quality performance, I wasn’t going to challenge them on it. Reflecting on this some years later, it dawned on me that the key is the ratio of true to false alarms generated by the system.
Yes, processes change, but there are other means of monitoring and controlling their variability.
SPC ignores mathematics
“Some say that SPC glosses over mathematical theory.”
Michel Baudin‘s comments: To be more specific, it is the SPC literature post-Shewhart that glosses over the math and only gives cookbook recipes for setting limits. It tells the reader to use control chart constants without explaining where they come from. Wheeler’s books and papers are no exception. The only author I found who discusses the math is Shewhart.
When you dig into this math, you find that the constants are based on the Gaussian distribution. It doesn’t mean you can’t use them with a range of other distributions. You just have to know that the meaning of limit crossings will be different.
“[Shewhart] turned things around by starting with generic, fixed-width decision limits which would always result in a suitably small probability of a false alarm (less than 2%)”
Michel Baudin‘s comments: Whether a 2% probability of a false alarm is suitable depends on the rate of true alarms. If you get 10 times more true than false alarms, it’s OK. If all the alarms you get are false, it’s not.
“[…] if you’ve been taught that SPC ignores mathematics, read The Secret Foundation of Statistical Inference.”
Michel Baudin‘s comments: The article Wheeler links to contains statements that might have applied in 1930 but no longer do. I have addressed these already in an earlier post: Strange Statements About Probability Models.
In particular, he writes, “In statistical inference, the assumption of independent and identically distributed random variables is a necessary condition.” This is not true, as I learned in my first experience working with real data, using kriging to estimate an orebody containing several metals from measurements on a borehole grid. The values from neighboring boreholes were not independent.
In this document Wheeler also admits to using the Gaussian distribution to set limits, which he then applies to data that do not follow this distibution:
“When the techniques we develop under the assumption of a normal distribution turn out to be robust in practice, we do not need to give any thought to whether or not the data appear to come from a normal distribution. Thus, with robust techniques, the worst-case assumption of normally distributed random variables is used as a starting point, but it does not become a prerequisite that has to be verified in practice.”
SPC is obsolete
“Some may say that SPC is a World War II technique in need of updating.
This hazardous idea ignores the very nature of mathematics. Calculus is a 17th-century idea. Does it need updating? The Pythagorean theorem is 2,500 years old. Does it need updating?”
Michel Baudin‘s comments: Wheeler here ignores several developments since World War II, in manufacturing processes, automatic controls, information technology, and in probability/statistics.
The World War II vintage methods are loaded with tricks to work around the limitations of contemporary information technology. For example, Harold Dodge introduced Range charts because operators at Western Electric refused to perform the calculations required to estimate sample standard deviations. Such issues are no longer a concern.
It’s not about ignoring old theories but about not artificially limiting yourself to old theories. Slide rules still work; we just don’t use them. And yes, calculus has had many updates, and I would not recommend studying it from the original publications by Newton or Leibniz. Ditto for geometry.
“In fact, the essential concept behind SPC dates back to Aristotle, who told us that we can discover the causes that affect a system by looking at those points where the system changes. It is this ancient idea that Shewhart’s process behavior charts formalize by using the foundations of modern statistical analysis. Rigorous mathematical ideas do not become obsolete.”
Michel Baudin‘s comments: Is that the central concept of SPC? I thought it was telling assignable causes from common causes using control limits. Shewhart discussed control charts, not “process behavior charts,” a term introduced by Wheeler for XmR charts, which were not part of Shewhart’s charts. And what we call “modern statistical analysis” today didn’t exist 100 years ago; the discipline was still in its infancy. Mathematical ideas do not become wrong, but they lose their usefulness when devices based on them are replaced with more powerful ones.
And where did Aristotle say this anyway?
“Shewhart proposed fixed-width decision limits and a small, variable risk of a false alarm. As Shewhart said, as long as we know the risk of a false alarm is small, we really don’t need to know the exact risk—we will still end up being right most of the time when we identify a potential signal.”
Michel Baudin‘s comments: I searched both of Shewhart’s books in vain for “fixed-width decision limits” and “false alarms.” If he did propose this, where did he?
That the risk of a false alarm is “small” does not guarantee being right “most of the time” when issuing alarms. As stated above, it depends on the ratio of true to false alarms, which decreases as the process stabilizes.
Inspection removes the need for SPC
“Some say that if they do 100% go/no-go gauging, they don’t need to waste time measuring the parts. This hazardous idea confuses the role of inspection with what is needed to operate the process.”
Michel Baudin‘s comments: Envision a chaku-chaku line, a U-shaped cell with machines that all unload part automatically, so that the operators’ job is to pick up the unloaded part, take it to the next machine, load it, and start the machine cycle. Working this way, an operator can attend to 20 machines. You would like to check some dimensions after some of the operations, without disrupting the flow.
If a machine can be programmed to take measurements before unloading the part, you use this capability. Otherwise, a 100% go/no-go gauge chcck integrated in the operator’s tasks can be more effective for rapid problem detection than taking samples offline to a CMM. And it works only if problems are sufficiently rare, on the order of 0.1% of defectives or less. If the operation produces 10% of defectives, the chaku-chaku line won’t work anyway.
It’s not a universal solution, but then, nothing is, and it is only for geometry. I haven’t heard of go/no-go gauges for electrical or chemical characteristics.
Summary
“Collect your data. Organize them rationally based on the principles of rational sampling and rational subgrouping. Place them on a process behavior chart and start learning what your process can tell you. As you identify assignable causes of exceptional variation and make them into controlled inputs for the process, you’ll find your process operating on-target with minimum variance. And that is world-class quality.”
Michel Baudin‘s comments: Unsurprisingly, there Wheeler goes again. “Process Behaviorn Chart” is his name for the XmR chart, which he thinks is all you need to achieve world-class quality.
It is not the experience of people who have actually achieved world-class quality. Focusing on assignable causes implies leaving common causes alone, which amounts to troubleshooting, not improvement. And there just isn’t any one-size-fits-all procedure to analyze data. You need to use everything you know about the backstory of the data, and use the best technology available to follow your nose. Sometimes you need advanced analytics, and sometimes visualization suffices.
In a recent case, a company collected 6 numerical characteristics at Final Test, as mandated by the customer. Loaded into one Excel workbook for each lot, this data lay fallow on the company’s server. As there were no problem reports from the customer, there was no emergency but, to me, this data was like Forrest Gump’s box of chocolates: I didn’t know what I was going to find.
So I asked Claude for code to convert all these Excel workbooks into a dimensional table, with the lot header data and the measurements. Based on what I knew about the data, I asked Claude for several visualizations of the overall distributions and relationships among the characteristics, which revealed heterogeneity.
To dig further, I asked Claude to plot all the points against their test dates, which revealed multi-week gaps in the production history, and visibly different distributions each time production restarted. An XmR chart, with part sequence numbers on the x-axis instead of test dates, would have hidden these gaps. And no control limit calculation was needed, as the shifts were visibly obvious.
The weeks-long interruptions in production are due to the customer’s ordering process and are not immediately changeable, but the engineers are now reviewing the line startup procedures to make the distributions consistent across production runs.
In retrospect, I could just have asked Claude to analyze this data for me, but I wasn’t ready to give up this much control.
June 2, 2026 @ 1:28 am
I read your response here and pondered for a while about if, or how, to write a comment or two… I would never say there is only one “view of the world” e.g. the Wheeler view, but neither would the Baudin view be the one. (Certainly, not my view of the world neither.)
I won’t deny that Wheeler has been very influential on how I work with data and I think my use of data improved accordingly.
Do I only use XmR charts? Certainly not. Did I ever understand that Wheeler thought I should only use XmR charts? Also not. I once spoke to him (face to face at the end of a seminar) about a problem I was working on and the end result was to use a time series plot where context would guide us to the answer (and it did).
If I could pin Wheeler down to one word from what I learnt over the years it would be “context”. Things start, and end, in context. He made me realise this more than anybody else. My colleagues joke about that word here at work. Whichever technique we use to make sense of the data is our choice. I try to use the simplest that serves to achieve a given objective.
While we have “modern-day” processes that have data very rapidly collected in computer systems, and we certainly can have many quality/process characteristics collected in parallel … and we can soon be drowning in data, there are also cases (many) that are not so dissimilar to how things were in previous decades.
Not all problems are complicated. Many cases (still) do have univariate data. Many cases do still have data that come along slowly. Not all companies have data scientists and competent statisticians at hand to work on their problems using e.g. machine learning techniques …
We also have to communicate results to managers and the simpler we can do this the better (hence why I often use time series plots and histograms) …Your experiences / your clients’ processes may well support your view on the lack of need for SPC in the modern day (or its irrelevance) but I find your view to be as narrow as you suggest Wheeler’s is. (If not “as narrow” then certainly in that direction.)
I have an ongoing project that uses approaches (with data) that I suspect you’d be happy to see being used. Do I ditch the “older” stuff and consider it below me, or somehow obsolete? No. Both have their place. Shouldn’t we keep doors open in life rather than close them?
June 2, 2026 @ 2:17 am
Thank you for your thoughtful comment.
In his writings, Wheeler talks about context, but his conclusion always seems to be plotting an XmR chart. When he did this with Atlantic hurricanes, it prompted to see what I could learn from the same NOAA database, and it led me to plot their paths instead, and two posts on this blog, Always the hurricanes blowing, Parts I and II.
Yes, not all problems are complicated, but some are. Also, the communication of results is a key part of data science that has seen many innovations, through the work of Edward Tufte, Jean-Luc Doumont, and others.