Health Misinformation in Search and Social Media

Abstract

People increasingly rely on the Internet in order to search for and share health-related information. Indeed, searching for and sharing information about medical treatments are among the most frequent uses of online data. While this is a convenient and fast method to collect information, online sources may contain incorrect information that has the potential to cause harm, especially if people believe what they read without further research or professional medical advice.

The goal of this thesis is to address the misinformation problem in two of the most commonly used online services: search engines and social media platforms. We examined how people use these platforms to search for and share health information. To achieve this, we designed controlled laboratory user studies and employed large-scale social media data analysis tools. The solutions proposed in this thesis can be used to build systems that better support people's health-related decisions.

The techniques described in this thesis addressed online searching and social media sharing in the following manner. First, with respect to search engines, we aimed to determine the extent to which people can be influenced by search engine results when trying to learn about the efficacy of various medical treatments. We conducted a controlled laboratory study wherein we biased the search results towards either correct or incorrect information. We then asked participants to determine the efficacy of different medical treatments. Results showed that people were significantly influenced both positively and negatively by search results bias. More importantly, when the subjects were exposed to incorrect information, they made more incorrect decisions than when they had no interaction with the search results.

Following from this work, we extended the study to gain insights into strategies people use during this decision-making process, via the think-aloud method. We found that, even with verbalization, people were strongly influenced by the search results bias. We also noted that people paid attention to what the majority states, authoritativeness, and content quality when evaluating online content. Understanding the effects of cognitive biases that can arise during online search is a complex undertaking because of the presence of unconscious biases (such as the search results ranking) that the think-aloud method fails to show.

Moving to social media, we first proposed a solution to detect and track misinformation in social media. Using Zika as a case study, we developed a tool for tracking misinformation on Twitter. We collected 13 million tweets regarding the Zika outbreak and tracked rumors outlined by the World Health Organization and the Snopes fact-checking website. We incorporated health professionals, crowdsourcing, and machine learning to capture health-related rumors as well as clarification communications. In this way, we illustrated insights that the proposed tools provide into potentially harmful information on social media, allowing public health researchers and practitioners to respond with targeted and timely action.

From identifying rumor-bearing tweets, we examined individuals on social media who are posting questionable health-related information, in particular those promoting cancer treatments that have been shown to be ineffective. Specifically, we studied 4,212 Twitter users who have posted about one of 139 ineffective ``treatments'' and compared them to a baseline of users generally interested in cancer. Considering features that capture user attributes, writing style, and sentiment, we built a classifier that is able to identify users prone to propagating such misinformation. This classifier achieved an accuracy of over 90%, providing a potential tool for public health officials to identify such individuals for preventive intervention.