One of the sessions being done at JSM this year is a retrospective on Deming's statistical legacy. The ASA has various associations with Deming. Each year there is a Deming lecture. But I think we rarely talk about Deming. He is often thought of as a management consultant rather than as a statistician or a philosopher. To the extent he is perceived as a statistician at all, his contributions are often perceived as a few practical applications, things like the control chart. Many of his other practical ideas that have been adapted, like the plan-do-study-act cycle, are thought of as outside the field of statistics entirely and having nothing to do with statistical thinking.
I think Deming's most fundamental contribution was quite different, a core philosophical challenge to the foundations of statistics which I think recent decades have generally sustained and statistics as a profession hasn't really fully recovered from. We haven't changed so much. As Winston Churchill put it, every now and then a man [sic] stumbles on the truth, but most of us manage to pick ourselves up and keep going anyway. But our field is starting to change, and I think Deming's real legacy is starting to have an impact, although still a very slight one. So I think it's worth thinking about the challenge.
1. Science vs. math. I think Deming's first and most foundational challenge challenge is to think of the field as a science of variation, and of how to acquire knowledge in the face of variation, rather than a branch of mathematics. Mathematics deduces consequences from assumed truths. Science attempts to generalize and predict from what we observe. I think the best explanation of the practical difference is Lawrence Summers' [sic] distinction between what he called the "smart people," econometricians and the like who had lots of fancy mathematics to back up their theories and were the obviously right people to follow when he was a grad student if one wanted to be thought smart and to advance ones career, and the "stupid people," sociologists and such, who had only a few empirical observations they were unable to fully explain and were therefore obviously wrong. As Dr. Summers recounted, our world today looks much more like what the stupid people predicted decades ago than what the smart people predicted. Deming would say this means the stupid people weren't so stupid. And maybe the smart people weren't so smart.
2. Standard statistical assumptions often don't hold. Based on his background in physics, Deming recognized the complex world of social and biological systems is highly interactive, non-linear, and heavy-tailed. Exchangeability assumptions in particular - assumptions that orders of operation don’t matter - can be particularly untenable. Thus the thinking that we now associate with the control chart was once an inquiry into establishing the conditions under which statistical assumptions and associated inferences approximately hold. This occurs if you can show - or intervene so as to make - the underlying process stable over time. This work is today most associated with achieving uniform product for quality purposes. But that's simply an application. The underlying work is epistemological, aimed at the question of under what conditions can we use statistical methods to obtain reliable knowledge. Deming's key insight is that these conditions are rarely the natural state of complex systems. Intervention is usually required to achieve them.
In this respect, Deming criticized statisticians, much as Dr. Summers criticized econometricians, for being too beholden to mathematics, being too quick to confuse elegant mathematics for intelligence or truth, too quick to assume that the world works to make the calculations easier. We have been too quick, in short, equate beauty and truth. But Deming's warning, as Lao Tzu's was, is that the truth is often not beautiful. (And as George Box put it, statisticians and artists both suffer from being too easily in love with their models.)
Today key societal mistakes can be traced to exactly the assumptions Deming criticized. Prior to the financial collapse, models were too quick to assume defaults would be independent, and too quick to extrapolate the recent past into the distant future. We have been too quick to assume that drugs wouldn't result in evolution of the organisms they target. Much of our difficulties lie in the fact that we simply assumed stable conditions with no basis for believing them, other than an assurance -- what we now recognized was a false assurance -- that that's how science works. We believed the world existed to make our jobs easier, that because we were at the top of the mathematical food chain and academic social hierarchy, our theories HAD to be the right ones. We now see that that was hubris.
Today the study of complex systems, thinkers like Nicholas Nissim Taleb, and others are able to articulate and explain why we were wrong. But when Deming started, the underlying mathematics and concepts of complex systems hadn't been invented. Deming was virtually alone, describing things in simple, qualitative terms, with only people like Poincare to rely on. Not having elegant mathematics to back him up, he was often classified as one of the stupid people and ignored, good for the sorts of lower-level statisticians who do applications like industrial statistics, but not someone someone doing theory should take seriously.
In my own JSM talk, on estimands, we will be talking about the idea that even in a randomized clinical trial, the traditional gold standard of statistical application, post-randomization dynamic events ("intercurrent "events) can influence the results, sometimes severely, and cannot simply be assumed to be uninformative. Once the instant of randomization completes, patients become subject to processes occurring over time which can potentially be confounding. This is a revolution in thinking. Once upon a time, the objectors were ignored. No longer. And this, occurring just within the last decade or so, is a truly watershed change.
And it's a change that, I would humbly suggest, is directly tracable to Deming's legacy. It is Deming who introduced the idea that we simply cannot avoid treating the phenomena we want to learn about as dynamic, interactive systems, systems that can at best approximate the conditions, often with significant and skilled management and intervention required, under which statistical inference becomes approximately reliable. It was Deming who distinguished between enumerative and analytic studies, emphasizing that most applied research seeks to make predictions about the future, not simply to extrapolate to an existing frame. And it is Deming who taught that to have any hope of using the past to help us understand the future, we need to begin by trying to understand the dynamics, the mechanisms, by which these systems work, conceived as systems, if we want to learn anything or get anywhere at all.
------------------------------
Jonathan Siegel
Deputy Director Clinical Statistics
------------------------------