Bayesian Networks for Risk Management Without Data
Recorded on March 9, 2018.
In this webinar, we will discuss how to estimate and manage risk without having any measurements or observations from the problem domain under study. To illustrate these challenges, we will explore the seemingly straightforward risk of getting a speeding ticket, a hopefully rare event. As it turns out, though, it is a risk that is rather difficult to estimate and manage properly.
More specifically, we wish to determine the value of a radar detector for reducing the risk of getting a speeding ticket. One's intuition probably suggests that getting a warning about the police conducting speed checks would reduce one's probability of getting caught speeding. However, beyond this fundamental causal relationship, i.e., warning → slowing down, additional effects may come into play. For instance, one can speculate that drivers with radar detectors can drive faster "safely" because they know that they can slow down as they approach a radar gun. Of course, this assumes a high degree of reliability of the radar detector being used.
Managing a Risk We Cannot Measure
Unfortunately, no data is available on the efficacy of radar detectors in real-world traffic conditions. Furthermore, and for good reason, there is no publicly available data on the speed checking patterns of police and their corresponding ticketing thresholds. This leaves most drivers with a strictly qualitative judgment to make about whether or not to use a radar detector.
However, a more quantitative approach would be necessary for someone who needs to make an institutional decision, such as a policy maker, a vehicle fleet manager, or an insurance company. For instance, we might wish to answer the following questions:
- Will prohibiting radar detectors effectively reduce speeding?
- Would a fleetwide deployment of radar detectors reduce the number of violations, the amount of fines, and reduce insurance premiums for the fleet?
- How could one formally trade off saving time due to increased speeds with risking the consequences of speeding violations? What is the expected value of using a radar detector?
- What is the highest "under-the-threshold" speed with and without radar detector?
- How do worst-case penalty scenarios compare with and without using radar detectors?
- From the police perspective, given the presence of radar detectors, what is the true rate of unobserved speeding in traffic?
Bayesian Networks as a Reasoning Framework
The purpose of our webinar is to present Bayesian networks as a framework through which we can formally reason about this problem domain even without any data. This absence of data, however, does not mean our reasoning needs to be purely qualitative. We will employ the Bayesia Expert Knowledge Elicitation Environment (BEKEE) to elicit assessments from webinar participants from their personal driving experience. We will see that collecting these individual opinions can parameterize a Bayesian network, which we can subsequently analyze with regard to the above questions.
Importantly, a Bayesian network allows us to distinguish between observational and causal inference, which is essential when computing the impact of either mandating or prohibiting radar detectors, i.e., performing an intervention in the domain. Also, given that Bayesian networks are inherently probabilistic, performing inference produces distributions. This is of critical importance in our context, as we need to assess "occasional and unpleasant" vs. "rare and catastrophic" events.
BayesiaLab Feature Focus
- Bayesia Expert Knowledge Elicitation Environment (BEKEE)
- Utility Nodes
- Policy Learning of Static Bayesian Networks
Please note that the subject matter of this webinar was selected strictly for illustrative purposes, expecting that most webinar participants would be familiar with this problem domain. However, all reasoning about speeding is strictly hypothetical. We do not advocate speeding under any circumstances. Please obey all traffic laws in your jurisdiction at all times.