Algorithms and Health Misinformation: A Case Study of Vaccine Books on Amazon

Abstract

This study examines how vaccine-related books appear on Amazon, focusing on search and recommendation algorithms. We collected vaccine related books that appeared on the first 10 search result pages by Amazon for seven consecutive days and content coded each book. We also collected Amazon’s recommendations for each vaccine book and mapped the network of recommendation among these books. First, we found that the number of vaccine-hesitant books outnumbered vaccine-supportive books two to one. Of these vaccine-hesitant books, 21% were written by physicians and medical experts. Second, although we did not find evidence that their search algorithm systematically favored any particular type of book, the three top ranked books across the seven days were all vaccine-hesitant ones. Lastly, using a network model, we found that books sharing similar views of vaccines were recommended together such that when a user views a vaccine-hesitant book, many other vaccine-hesitant books are further recommended for the user. The three most frequently recommended books were vaccine-hesitant ones. The potential consequences of blindly applying commercial algorithms to a complicated health messages such as vaccines are discussed.

Acknowledgements

The authors wish to thank the editor, and two anonymous reviewers for critically reading the manuscript and suggesting substantial improvements.

Supplementary Material

Supplemental data for this article can be accessed on the publisher’s website.

Notes

¹ Overall, the relevant books overlapped in the range of 72 ~ 96% between a given pair of days during the data collection period. In particular, 45 books appeared throughout all seven days.

² Conceptually recommendations should be mutual for two books according to the “bought together” logic. Yet, the algorithm selectively displays “bought together” recommendations for popular items which can result in an asymmetric recommendation network.

³ Intercoder reliability measures (Krippendorff’s alpha) ranged between 70 ~ 80% initially.

⁴ The visualization for our model’s goodness of fit can be provided upon request.

⁵ Some of these were different editions of the same book.

⁶ The ratio of vaccine hesitant books to vaccine supportive books for each day was 3.1 (day1), 2.8 (day2), 3.0 (day3), 2.8 (day4), 2.4 (day5), 2.5 (day6), and 2.9 (day7).

⁷ The small value in the ranking variable indicates the better ranking, thus a negative coefficient.