IMPROVING SURROGATE DECISION MAKING

Purpose: To investigate the accuracy of surrogate decision makers and the extent to which combining their predictions yields gain in accuracy.

Method: Surrogate decision-making becomes necessary when a person is no longer capable of expressing treatment preferences, and when no living will, detailing treatment preferences, is available. The members of 42 Swiss families predicted treatment preferences of a randomly selected family member. The scenarios described three medical complications occurring in eight health states, namely, the current health state and seven hypothetical, critical health states. After rendering individual predictions, the members of each family also produced predictions as a "natural" group. The target person recorded his or her treatment preferences in the same scenarios. We used the dimensions of a receiver operating characteristics (ROC) analysis such as sensitivity and specificity to determine the predictive accuracy of individuals' predictions, relative to the group predictions. In addition, we aggregated individuals' predictions into the verdict of a "statistical" group, in which each prediction received equal weight.

Results: Overall, individual surrogates correctly predicted 67% of the target persons' treatment preferences. Their discrimination rate (area under the ROC-curve, AUC) was 75%, with a sensitivity of 75% and a specificity of 57%. The natural groups and the statistical groups significantly outperformed the typical individual family members and did not differ from each other. The natural (statistical) group correctly predicted 71% (72%) of target persons' preferences, with a sensitivity of 78% (78%) and a specificity of 65% (64%). This increase in accuracy is also reflected in a much greater AUC of 82% (83%).

Conclusion: Individual surrogate decisions suffer from bounded overall accuracy and low specificity. These limitations can be overcome by aggregating across individual predictions. This can be achieved either in terms of statistical aggregation (in which individual predictions are mechanically aggregated) or in terms of natural aggregation (in which the group is required to achieve a consensual verdict). People are reluctant to delegate decisions to statistical ("actuarial") decision aids (Kleinmuntz, 1990; Arkes et al., 2007). Our studies show that one effective way to reconcile the power of aggregating opinions with family members' desire for autonomy is simply asking them to reach a consensual verdict.