Tyler VanderWeele’s contribution to the Discussion of ‘Experimental evaluation of algorithm-assisted human decision-making: application to pretrial public safety assessment’ by Imai et al

Imai et al. (2023) are to be commended for their thoughtful evaluation of the Public Safety Assessment (PSA) in judicial decision-making. Rigorous designs for evaluating how predictive algorithms can assist human decision-making are very much needed. While deep learning algorithms have advanced considerably, in many contexts it is important not to abandon human decision-making. Such contexts include settings in which there are special considerations which are not easily coded (e.g., an experienced physician’s intuition; or specific details of a criminal case), settings in which there are incommensurable ends so that it is not possible to specify a univariate objective function (e.g., admissions trade-offs between diversity and academic preparedness), settings in which moral considerations are not easily embedded as constraints, and settings in which human decision-making or autonomy is itself partially constitutive of human flourishing (e.g., choosing one’s spouse). However, even in such settings predictive algorithms can be of assistance in human decision-making. They can provide additional information or estimates (difficult for the human mind to carry out) that can be used, amongst other considerations, in decision-making. They can help identify and eliminate choices that are clearly suboptimal when competing ends are in play. They can help identify a set of the most promising possibilities when the options are vast and yet human choice is itself important (e.g., online-dating match-algorithms).

Using an approximately randomised design, Imai et al. (2023) evaluate the effects of providing PSA information on judicial decision-making, and on the fairness of those decisions. However, an important issue in all algorithm-assisted human decision-making contexts concerns exactly what information to provide, i.e., what information will most aid decision-makers? This point is not discussed by Imai et al. (2023), perhaps reasonably, because their purpose is to evaluate the PSA information and system already in place. However, other information may have been more useful to judges. For example, one could imagine presenting instead the actual predicted probabilities of ‘failure to appear’ and ‘new criminal activity’ (rather than the somewhat convoluted and difficult-to-interpret PSA). Elsewhere, I have proposed such an approach in medical decision-making in which clinicians are provided predicted outcome-probabilities for all treatment options but then make, in conjunction with the patient, the actual decision, accompanied by randomised trials to evaluate the effects of providing this additional information (VanderWeele et al., 2019; cf. Wu et al., 2020). Such predicted probabilities might well have been more valuable to the judges than the PSA. It is also conceivable that randomised trials could evaluate the effects of providing this information versus PSA. Again, further work on optimising algorithms for human decision-making contexts, on evaluating these algorithms, and on choosing which information to present, will be important in future research.