Four Solutions to the Reference Class Problem

Suppose you want to diagnose, predict future development or treat a particular patient. The reference class problem is this: Which population is most appropriate for these tasks? The reference class problem is hardwired. After spending a huge amount of research time, I get the impression that a satisfying solution is not even close. However, I am not giving up…

In our recent paper, “Four solutions to the reference class problem” which will appear in G. Hofer-Szabo, L. Wronski ( Eds), Making it formally explicit: Probability, Causality and DeterminismJon Williamson and I shed some light on the reference class problem.

 

The abstract reads:

 

“We present and analyse four approaches to the reference class problem. First, we present a new objective Bayesian solution to the reference class problem. Second, we review Pollock’s combinatorial approach to the reference class problem. Third, we discuss a machine learning approach that is based on considering reference classes that are similar to the individual of interest. Fourth, we show how evidence of mechanisms, when combined with the objective Bayesian approach, can help to solve the reference class problem. We argue that this last approach is the most promising, and we note some positive aspects of the similarity approach.”

 

Overall, evidence of mechanisms may constrain a Bayesian network or other graphical models. The augmented Bayesian network then yields better single-case probabilities and consequently better decisions for specific patients. Find out how this exactly works by considering Section 6 of our paper…

 

words by Christian Wallmann

Leave a Reply

Your email address will not be published. Required fields are marked *