Quantifying effects of demographic biases on estimation of cultural ecosystem services using social media in Japan

• CES evaluations using social media can be influenced by biases in user’s attributes. • Three CES indicators were compared across demographic groups of Twitter users. • Differences in sentiment and travel costs among groups were limited. • The geographic distribution of the number of georeferenced posts differed by groups. • For inclusive evaluations of CESs, considering user biases is crucial. Quantifying cultural ecosystem services (CESs) has been difficult using traditional methods such as questionnaires, so increasing numbers of studies are utilizing big data obtained from social media for evaluation of CESs. Although data obtained from social media are often considered to be biased and non-representative, the actual effects of biases on evaluations of CESs are rarely quantified. In this study, we sampled posts from a microblogging service, Twitter, and investigated the effects of demographic biases on three indicators capturing aspects of social media which can be used for evaluations of CESs (spatial distribution, sentiment of post, and travel cost), focusing on the Japanese national parks as a case study. We found that Twitter users obtained from the randomly sampled 1% posts had a different distribution of demographic attributes (age, sex, and residential prefecture) from the Japanese population. More importantly, we found that activities on the platform such as estimated frequency of georeferencing and posting are different among demographic groups. This indicates that spatial distribution of CESs commonly evaluated by the number of georeferenced posts can be affected by additional demographic biases which cannot be captured by statistics based on all users on social media platforms. On the other hand, among the three indicators, only spatial distribution was strongly affected by the demographic attributes of users, indicating that the potential effects of biases are different for each indicator tested. These results clearly show that consideration of demographic attributes and their potential effects would be necessary to obtain representative and inclusive evaluations of CESs.