Don’t make data-driven decisions. Make data-informed decisions, or so my friend Hank Childers advises.
It’s a nice distinction that recognizes both the value of evidence and its limitations.
For those who like to rely on evidence we live in tricky times. The increasing availability of evidence about just about any topic is accompanied by an at-least-equally increasing supply of disinformation, in direct proportion to the profit to be made by biasing things in the profit-maker’s favor.
Which is one reason I’m skeptical of the long-term reliability of IBM’s Watson as a medical diagnostician.
There’s a program called Resumix. What it does is scan resumes so as to find the best skill-to-task matches among the applicants. It’s popular among a certain class of recruiter because reading resumes is an eye-blearing chore.
Worse, the recruiter might, through fatigue, miss something on a resume. Worse still, the recruiter might inadvertently practice alphabetical discrimination, if, for example, the resumes are sorted in name order: Inevitably those at the front and back of the stack will receive more attention than those in the middle.
But on the other side of the Resumix coin is this: Most applicants know how to play the Resumix game. Using techniques similar to how those who write websites learn how to get the attention of search engines, job-seekers make sure Resumix sees what it’s supposed to see in their resumes.
If Watson becomes the diagnostician of choice, do you think there’s any chance at all that those who stand to profit from the “right” diagnosis won’t figure out how to insert what Watson is looking for in the text of the research papers they underwrite and pharmaceutical ads they run?
It’s one thing for those who developed and continue to refine IBM’s Watson division … for whom, by the way, I have immense respect … to teach it to read and understand medical journals and such. That task is merely incredibly difficult.
But teaching it to recognize and discard utter horse pucky as it does so? Once we try to move beyond the exclamation (“That’s a pile of horse pucky!”) to actual definition, it isn’t easy to find one that isn’t synonymous with “I don’t want that to be true!”
Well, not, that isn’t right. A useful definition is easy: Horse pucky is a plausible-sounding narrative that’s built on a foundation of misinformation and bad logic.
Defining horse pucky? Easy. Demonstrating that something is horse pucky, especially in an age of increasingly voluminous disinformation? Very, very hard. It’s much easier to declare that something is horse pucky and move on … easier, but intellectually bankrupt.
So imagine you’re leading a team that has to make an important decision. You want the team to make a data-informed decision — one that’s as free from individual biases as possible; one that’s the result of discussion (solving shared problems), not argument (one side winning; the other losing).
Is this even possible given the known human frailties that come into play when it comes to evaluating evidence?
No, if your goal is perfection. Absolutely if your goal is improvement.
While it’s fashionable to disparage the goal of objective inquiry because of “what we now know to be true about how humans think,” those doing the disparaging are relying on evidence built on the philosophical foundations of objective inquiry … and are drawing the wrong conclusions from that evidence.
Here’s the secret:
The secret of evidence-informed decision-making: Don’t start by gathering evidence.
Evidence does, of course, play a key role in evidence-informed decision-making (and what would you do without KJR to give you profound insights like this?). But it isn’t where you start, especially when a team is involved.
Starting with evidence-gathering ensures you’ll be presiding over an argument — a contest with winners and losers — when what you want is collaboration to solve a shared problem.
Evidence-gathering follows two essential prerequisite steps. The first is to reach a consensus on the problem you’re trying to solve or opportunity you’re trying to chase. Without this, nothing useful will happen. With it, everyone agrees on what success will look like when and if it eventually happens.
The second prerequisite step is consensus on the process and decision-making framework the team will use to make its decision. This means thinking through the criteria that matter for comparing the available alternatives, and how to apply evidence to evaluate each alternative for each of the criteria.
Only then should the team start gathering evidence.
Informed decisions take hard, detailed work. So before you start, it’s worth asking yourself — is this decision worth the time and effort?
Or is a coin-toss good enough?
You want the team to make a data-informed decision — one that’s as free from individual biases as possible; one that’s the result of discussion (solving shared problems), not argument (one side winning; the other losing).
Is this even possible given the known human frailties that come into play when it comes to evaluating evidence?
No, if your goal is perfection. Absolutely if your goal is improvement.
This is exactly why I think Watson could be useful on a panel of doctors/diagnosticians. Is Watson ever going to be perfect? No. Neither are any of the human doctors. The biggest difference is that Watson should be able to identify everything that led it to make its decision, whereas humans can just as easily be misled by profit motives, biases, prejudicial or inaccurate medical journal articles – yet be unaware of or unwilling to acknowledge these biases.
BTW, I still don’t think Watson will ever (or should ever) be the “diagnostician of choice”: I think it should be a tool used by diagnosticians and doctors, the final diagnosis being made by a human, who remains accountable.
The only analogy I can come up with right now is using spell-check or autocorrect: they are tools, and useful tools (as I grow older my spelling is becoming worse, so sad) but it is up to the people composing the documents to insure they’re using the right words and terms in the context, and overriding spell-check and autocorrect when necessary.
As always insightful. It reminds me of my favorite Isaac Asimov story: “The Machine That Won the War.”
( http://www.olivenri.com/machine_won_files/The_Machine_that_Won_the_War01.pdf )
Someone developed a medical diagnosis program years ago. You picked from a short list of complaints then start answering questions. It used the basic “20 questions” logic tree format.
The output was a list of possible diagnoses, probability of each, and tests to confirm or refute each. Based on the probabilities and costs, the physician (or patient) decided what to do next.
Head-to-head tests showed that the app was right significantly more often than the doctors, and virtually never missed a rare possibility.
Doctors universally hated them, on the basis that diagnosis is “an art” that you requires years of experience and intuition. Apparently they couldn’t interpret the results of the head-to-head. :-/
I completely agree with Sara that automated diagnosis should be included, but not authoritative. As a patient, I would love to get something like: “There is an 83% chance it is condition X with a $50 blood test to confirm; and a 0.3% chance of this fatal condition with a $5 test to confirm.” Much better than, “It looks like condition X, let’s get you a prescription.”
That quote by John Tukey is pure gold.
And this article is an excellent reminder of what we need to do; clearly define the problems we are facing first.
How many groups get stuck at step 1 – getting consensus re: the problem/opportunity? Those are the times a leader who clearly defines the problem/opportunity can really help. Sometimes the leader (or the organization culture) can take care of step 2
I get it that Watson’s strength is having a sea of information poured into it and being able to find stuff in the sea. There may be a place for regurgitated medical articles in diagnosis, but I believe a better data set is case histories. As a comentor mentioned above, the twenty questions approach with the knowledge of a few medical textbooks was very helpful. The weakness of human doctors is that they can’t call up the relevant part of Harrison’s Principles of Internal Medicine in the face of a particular case. It’s the Tukey quote – figure out what you are trying to do. In the very infrequent cases that have hardly ever happened before, the medical literature may pay off. But you’d probably have to look back to when we published reports of unusual cases. I see a better match for Watson in law, finding relevant precedents.
A different comment from the above. One of my young relatives has taken a job with a company that analyzes a company’s masses of data to find answers on how to improve the company.
1) I know this is a big deal now. 2) Do you remember the term “garbage in – Garbage out?”
He told me one of his firm’s customers was Radio Shack. This was last November. I told him I thought it was great that his company had other customers. At the time I couldn’t imagine there was much that would help RS in their company data. Since then we have the wonderful interviews with RS leaders (In Wired, I believe) in which we learned that the company had a succession of incredibly poor leaders that drove it into the ground just when it had a major new opportunity in the Maker movement.