Despite the clarion call to action on “data driven decision making”, there is a lot of evidence that business decisions are still largely made on intuition.
We hear constantly that “you can’t manage what you don’t measure”, but maybe it is not as simple as that.
I have had several conversations recently about examples where data-driven and AI enabled insights have provided the wrong inferences.
It has become dangerously easy to be complacent in the face of very attractive, plausible and beguiling results from algorithms, analyses and data visualisations.
But we have to remain vigilant.
I guess the root of all these discussions is we must “always challenge our assumptions” and “look for what’s missing”.
A read a great article in Forbes by Pawel Rzeszucinski, “Look For What’s Missing: How To Avoid Survivorship Bias In Data”, that got me thinking on this again.
The most classic example of survivorship bias is still one of the easiest to understand.
A group of WWII statisticians were tasked with analyzing the damage suffered by planes that returned from combat. By doing so, they hoped to determine which areas of the planes required more or less armour in order to improve their performance and lifespan.
Well, there’s the bias: analyzing only the planes that made it back home in spite of the adversities of combat. The planes that didn’t make it back home were not in the “data set”.
On analyzing the location of bullet holes, the first impulse was to focus on protecting those same damaged spots. But as the now famous Abraham Wald wisely observed, the best option was actually to do the exact opposite.
The planes they were analyzing only made it through combat because their most crucial parts were spared damage. It would make sense to add more armour to the areas where there was no evidence of survival, notably the cockpit and the engine.
This is the problem of survivorship bias: The examples that don’t “survive” are not able to explain their side of the story, so all our considerations will be biased toward the “winners”, the ones left to expose their point of view.
In a much more prosaic example, I was discussing benchmark indicators of invoice processing in accounts payable and the time to process invoices from receipt to “ready to pay”. The challenge is that “invoice receipt date” was the date that an invoice hit the ERP or invoice processing platform. This is a useful measure but did not reflect the “end to end” experience where supplier invoices were emailed into a mailbox, where they waited patiently to be “processed” into the ERP. This sometimes took days.
The real metric was substantially longer than that being reported, resulting in misleading implications being interpreted.
There are a host of examples of this “survivorship bias” in the data that we use every day.
This whole debate reminds me that we have to keep our brains well and truly “switched on”, whatever sophisticated technology and models we are using to crunch our data.
We must always consider “what we don’t see” and whether we have data on the full “universe” of the business problem, or just a “survivor” subset, which inherently creates dangerous data bias for decision making.
My mind keeps getting drawn to the cartoon, paraphrased, that has an exchange “Are you excited about the implications presented by Artificial Intelligence?” with the response “I’m more concerned about the implications of the decline in Human Intelligence!”.
To be fair, it is as much about “attention” and “complacency” as “intelligence”, but it makes for a good cartoon 😉!
Maybe “intuition” isn’t all bad in decision making, and at least provides the basis of a good “A/B” test!
“You can’t manage what you don’t measure” is too simplistic. Understanding and “thoughtful measurement” are also key elements.
Thanks for reading . . .