Capturing Moral Intuitions in a Kidney Matching Market
http://jpdickerson.com/pubs/freedman18adapting.pdf
The recent paper “Adapting a Kidney Exchange Algorithm to Align with Human Values” documents the process of creating a kidney exchange market which prioritizes patients in a way that aligns with human values.
A kidney exchange market takes the form of a large graph. The nodes in this graph consist of a patient in need of a kidney paired with a living donor who would like to give the patient their kidney but unfortunately cannot due to medical incompatibility. Edges between nodes indicate donation compatibility. A kidney exchange clearing algorithm seeks to find a set of disjoint cycles in this graph each of which corresponds to a chain of kidney donations.
Although the theory behind such an algorithm is fairly straightforward there are a large number of practical considerations. For example it’s illegal to compel someone to donate their kidney so in order to avoid a donor backing out after their patient has received a kidney all swaps in a cycle must take place simultaneously. This places a very serious practical constraint on the size of cycles which can be used – even a three or four cycle is pushing it.
Even with such practical considerations taken into account, there are still rather fraught moral questions related to what precisely our algorithm should seek to accomplish. A simple goal might be to maximize the overall number of patients who receive kidneys. However since unmatched nodes from one iteration of the algorithm continue on to the next until the patient dies or the donor gets fed up this would almost certainly result in hard to match patients building up in the market while easy to match patients all get matched with each other immediately.
Moreover, there are a number of reasons we might wish to prioritize one patient over another. Perhaps we might place more value on giving a kidney to a young patient with decades left to live than on giving one to a very elderly patient who will only have the use of it for a few years. Or we might wish to deprioritize patients whose addictions are likely to destroy the new kidney the same way they did the old (or we might not! This is controversial)
In order to nicely capture these intuitions the authors of this paper began by gathering from doctors a list of which traits it is acceptable to prioritize a patient on the basis of. They then asked the same doctors to prioritize patient profiles which differed along the relevant traits. Based on these they statistically estimated what weight to give each trait and built a prioritization model. They then used this model to create a market clearing algorithm which followed the given prioritization scheme. Based on simulation results they were able to determine that the prioritized algorithm achieved virtually the same number of matches as the algorithm which only attempted match maximization. This is an encouraging result which suggests that prioritization schemes do not have a major implementation cost.