Setting health research priorities using the CHNRI method: V. Quantitative properties of human collective knowledge

Introduction The CHNRI method for setting health research priorities has crowdsourcing as the major component. It uses the collective opinion of a group of experts to generate, assess and prioritize between many competing health research ideas. It is difficult to compare the accuracy of human individual and collective opinions in predicting uncertain future outcomes before the outcomes are known. However, this limitation does not apply to existing knowledge, which is an important component underlying opinion. In this paper, we report several experiments to explore the quantitative properties of human collective knowledge and discuss their relevance to the CHNRI method. Methods We conducted a series of experiments in groups of about 160 (range: 122-175) undergraduate Year 2 medical students to compare their collective knowledge to their individual knowledge. We asked them to answer 10 questions on each of the following: (i) an area in which they have a degree of expertise (undergraduate Year 1 medical curriculum); (ii) an area in which they likely have some knowledge (general knowledge); and (iii) an area in which they are not expected to have any knowledge (astronomy). We also presented them with 20 pairs of well-known celebrities and asked them to identify the older person of the pair. In all these experiments our goal was to examine how the collective answer compares to the distribution of students' individual answers. Results When answering the questions in their own area of expertise, the collective answer (the median) was in the top 20.83% of the most accurate individual responses; in general knowledge, it was in the top 11.93%; and in an area with no expertise, the group answer was in the top 7.02%. However, the collective answer based on mean values fared much worse, ranging from top 75.60% to top 95.91%. Also, when confronted with guessing the older of the two celebrities, the collective response was correct in 18/20 cases (90%), while the 8 most successful individuals among the students had 19/20 correct answers (95%). However, when the system in which the students who were not sure of the correct answer were allowed to either choose an award of half of the point in all such instances, or withdraw from responding, in order to improve the score of the collective, the collective was correct in 19/20 cases (95%), while the 3 most successful individuals were correct in 17/20 cases (85%). Conclusions Our experiments showed that the collective knowledge of a group with expertise in the subject should always be very close to the true value. In most cases and under most assumption, the collective knowledge will be more accurate than the knowledge of an "average" individual, but there always seems to be a small group of individuals who manage to out-perform the collective. The accuracy of collective prediction may be enhanced by allowing the individuals with low confidence in their answer to withdraw from answering.