Trusting the Math

Many of us have tracked the incredible news of the Flint, Michigan (USA) water crisis that began in 2014. Having grown-up in several Flint suburbs and having had a college internship with General Motors at a now defunct plant in downtown Flint, I've pained over how this event has transpired. While the primary goal was to save money in supplying water to residents, the eventual outcome resulted in flaring cases of lead poisoning from old water pipes. Now the crisis has turned into a "search and destroy" mission to find and replace contaminated homes and businesses plagued with newly surfaced lead. Yet from a professional view, I was impressed at the 2017 developments in how the city of Flint adopted machine learning models to reduce risks with missing lead pipes that sit dormant throughout the city. The paper explaining the models can be found here.

I write this simply to amplify a superb example of what can happen when data science works and when human judgment (understandably, yet painfully, in this case) opts to ignore what prediction tells us. In Flint's case, a couple of details to consider (sourced via the above linked article published by The Atlantic):

1. The employed models used to help crews dig only where pipes were likely to need replacement achieved a 70 percent rate of accuracy through 2017 (based on 8,833 homes inspected - of those, 6,228 homes had their pipes replaced). This is the part of the story where the math was working.
2. Now for the part where emotion and decisions over-ruled the math.Yet under pressure from residents to not be skipped, Flint's Mayor asked their contractor to abandon the guidance of the ML model and "... dig across the city’s wards and in every house on selected blocks, rather than picking out the homes likely to have lead because of age, property type, or other characteristics that could be correlated with the pipes." I sincerely empathize with this tough decision. Consider the resident's point of view: to see your neighbor's yard dug up from inspection for lead pipes only to see the crews pack up and not check yours?! How would you react? I'd probably call City Hall myself and demand inspection. This is a hard moment to ignore for leaders.

I encourage you to continue reading the above article from The Atlantic as it chronicles the subsequent ups and downs of Flint's digging strategy. In context of this dilemma, I can attest that following the math...trusting algorithms...using probabilities to guide one's thinking is really tough work. But it's invaluable when your organization's culture is data-focused. Meaning what?

Collections teams for banks trust predictive models to prioritize who to chase for delinquent payments. Call center agents don't have the band-width to go after 100% of the late-payers, so models often do the ranking for them. In fact, some banks will not only rank "likely payers if we call them," they'll also apply another model to determine what script or dialogue to use with the customer to maximize compliance in owning up to back payments. It's applied data science that often help risky decisions.

In an even more serious setting, oncologists are sometimes unsure if suspect cells are benign or malignant. Image analyses can guide the teams to escalate/de-escalate the scrutiny of a questionable cell sample from a biopsy based on deep learning models that brought in thousands of prior images of similar/dissimilar cell samples. What the humans cannot detect or prioritize, machines actually can aid our next steps.

I'm a big fan of balancing humility in decision-making with the ability to listen to data from past or recent history. Like Peter Drucker said, "Culture eats strategy for breakfast." In this case, trusting the math absolutely begins with leadership - not with the technology or the subject matter expertise. We have to want to improve our decision-making. Let both policy and sound data guide your thinking when it's available.