Talking to your boss about data
Here’s what two MIT business analytics students learned in the new From Analytics to Action class.
By Brian Eastwood |
March 23, 2017
“Analytics is more people-related. In the real world, there are different players and stakeholders.”
One of the biggest challenges in data analytics is presenting results in a way that’s meaningful to people who aren’t data scientists. As MIT Sloan Master of Business Analytics student Souhail Halaby pointed out, there’s a model that shows that the winner of the Super Bowl can predict the next year’s stock market performance.
While this is interesting, it’s little more than the butterfly effect, Halaby said. “You need human intuition to determine whether the correlation is important,” he said.
Halaby, and more than a dozen other students who enrolled in the new master’s program, learned how to apply a human touch to mathematical models in a five-day class offered during the Sloan Innovation Period last October. The class, taught by organization studies professor Ray Reagans, will be offered to more students and expanded to a half-semester beginning in the fall of 2017.
Here’s what students learned in the first version of the class.
‘Stakeholders have to trust your data’
In the class, students used the R statistical computing software environment and the Julia dynamic programming language to build a visualization of kidney transplant data. Potential kidneys vary in quality, blood type, and location, and patients awaiting a transplant have to consider their options — accept a kidney now from a donor of average health, or wait longer for a kidney from a healthy donor?
Patients are only one group interested in the results of this model. So, too, are the doctors who treat patients and the hospital executives who want to improve outcomes for transplants. Each of these stakeholders comes to the table with a different point of view and set of expectations for what the model will tell them. But they also bring a level of subject matter expertise that a data scientist does not possess, and they are accustomed to using that expertise — and not a data model — to make a decision.
Business analytics student Afshine Amidi said an important takeaway from the fall class was the importance of a model’s interpretability, especially when presenting it to people who do not have a background in data analytics.
“It’s important for stakeholders to see what your model is about. What data are you using to produce your result?” Amidi said, adding that storytelling can be an effective way to pique an audience’s interest in how a model solves a problem. “Your stakeholders have to trust your data. You have to convince them.”
‘Making the case for innovation’
Making that happen requires considering how a model will be adopted and implemented, said Reagans, one of five professors (along with two doctoral students) who taught the class this fall. (The others were Dimitris Bertsimas, Emilio Castilla, Jack Dunn, Roberto Fernandez, Thomas Kochan, and Jerry Kung.)
“When making the case for their innovation, the students needed to understand how people in an organization are likely to think about the problem the students are solving,” he said. Since the typical stakeholder may not be used to thinking about problems in terms of data and algorithms, the class offered insight into addressing and overcoming resistance to behavior change.
One strategy is to build informal networks within an organization in order to improve the odds of getting buy-in for an initiative, Reagans said. Another is to anticipate stakeholders’ cognitive biases, such as knee-jerk reactions or the silo effect, and try to overcome them by asking the right questions and establishing a framework for making decisions.
‘Analytics is more people-related’
While mathematical models are nothing new, predictive models powered by machine learning have only emerged in the past few years, Halaby noted. Analytics sits at the intersection of mathematical and predictive models.
Machine learning platforms such as Google DeepMind and IBM Watson make it possible to create powerful and accurate models, but they aren’t intuitive, Halaby said. “There are layers upon layers of computational code that even a programmer doesn’t understand.”
The real power is in the ability to run a wide, deep, and up-to-date data set through a simple model such as a linear regression and generate rapid iterations of the same calculation. This will produce descriptions that are “descriptive and intuitive to the people you need to present to,” he said.
“Analytics is more people-related,” Halaby added. “In the real world, there are different players and stakeholders. The implication of that is how you present the data, how you use the mathematical principles to get the results you want.”