Trusting Parametric Risk Analysis

Just finished a recent (2019) book focusing on issues with introducing AI in our lives. Its “A Human’s Guide to Machine Intelligence” by Kartik Hosangar, a Wharton professor. A quick read; not a lot of jargon. Gets one thinking about the challenges that are happening right now with putting “algorithms” in the figurative driver’s seat (i.e., decision chair) of our lives and work. I particularly enjoyed his review of research in how people view and trust/distrust the algorithms. For those of us in the project management world, these issues are critical to understand as AI comes on board.

The book had immediate import to an “algorithm” trust/acceptance issue I have faced. In 2006 I began helping clients implement a parametric cost and schedule risk analysis tool in their quantitative risk analysis (QRA) practice (now offered as ValidRisk). The method was also established as an AACE Recommended Practice in 2008 (RP 42R-08). A parametric model is an algorithm. It can be considered entry level AI, albeit resulting from manual regression of modest databases rather than machine learning from big data. But from an algorithm trust/acceptance point of view, Dr. Hosangar’s book applies.

The issue I experienced is that in the early years, the parametric model was faced with immediate resistance from peers. The idea that an algorithm could predict cost growth better than subjective judgment (or monte-carlo simulation applied to said judgment) seemed to trigger an almost visceral negative reaction in some. Not being a psychologist, I ascribed it to existential dread that the method might threaten their livelihood. Most risk analysts have little experience in or exposure to analytics per se (understanding a distribution not being enough).

Dr. Hosangar has improved my understanding of the psychology. He provides a good overview of the issue of trust in algorithms. For example, in chapter 7 he describes research about how we are “much more inclined to trust algorithms with estimates and predictions that are objective”; i.e., that do not involve a decision that involves emotion and bias. If an algorithm says you need 25% contingency, but opinion has always said 10% (optimism bias) and 25% contingency threatens the project’s chance of approval, it’s a situation fraught with emotion.

He also talks about the inability to “reconcile the idea of algorithms with the possibility that they can fail, even while accepting that humans are fallible”. i.e., we lose trust in an algorithm that fails even once, but accept that our judgment fails repeatedly. In the parametric model case, research shows that subjective QRA methods do very poorly when there are significant systemic risks, but, rather than trust an algorithm, the known failure is accepted.

A piece of practical advice Dr. Hosangar gives is that users “being allowed to make even tiny tweaks to an algorithm increased the chances that a person would trust it”. In the case of parametric risk models, this points to the need to build in the capability to “calibrate” the industry model; i.e., to study one’s own cost growth data and use that to “tweak” the model, no matter how inconsequential the calibration study’s findings.

The good news is that awareness of “analytics” has increased in the last 3-4 years. The reception of the parametric risk modeling method has improved. As algorithms, be they parametric models or ML/AI applications, increasingly find their way into our controls, estimating, risk analysis and other project work, I think many would benefit from understanding the psychology of trust that this book addresses.