Studying attention to IPCC climate change maps with mobile eye-tracking

seconds per map. The near-term projection map of each type was always displayed before the corresponding long-term projection map, but the order of projection types was randomised across participants (see S4 Fig for the procedure diagram). Participants in the paired-viewing condition were encouraged to discuss the maps with their partner, while those in the single- viewing condition viewed them independently. For the second part (artwork viewing), participants viewed two large-scale paintings mounted on the wall. They carried the companion device with them and were free to choose the viewing order and spend as much time as they wished on the artworks, moving freely around them. After both sessions, participants filled out the 22-item climate change anxiety scale and optionally provided demographic information and comments. The experiment concluded with a debriefing by the experimenter. 2.5 Data analysis Data analysis progressed through several stages, from raw recordings to detailed analyses. Recordings from the companion device were uploaded to Pupil Cloud, a GDPR-compliant online platform for data processing and visualisation. Fixation detection used an extended I-VT (Identification by Velocity Threshold) algorithm. Data quality was inspected informally by visually checking raw gaze data overlaid with fixations on Pupil Cloud, alongside fixation duration rates relative to recording duration. No valid recordings were excluded. Some spatio- temporal random noise in gaze data, as reported by the manufacturer, was assumed not to affect the primarily descriptive analyses.

To map the XY-coordinates of raw gaze and fixation data, two streams of parallel process- ing were employed on the cloud. This dual approach was designed to take advantage of a newly available algorithm in beta version at the time of data collection, which was later released as a stable version. The first, marker mapper enrichment (MME), utilised fiducial markers (Apriltags) placed around the laptop monitor and paintings to define areas of interest (AOIs). Recordings were manually time-stamped for segmentation, and markers served as anchor points to normalise gaze data. The second, reference image mapper (RIM), employed video recordings and snapshots to create a structure-from-motion model for normalisation. While both methods produced comparable data structures, RIM demonstrated higher accuracy and was used for all subsequent analyses (see S5 Fig for an overview). RIM preserved the height-to- width ratio of stimuli in pixel-based values, unlike MME, which distorted the aspect ratio dur- ing normalisation. Surface-normalised fixation data were used to create heatmaps visualising viewing patterns. The enriched dataset, along with supplementary materials, was uploaded to Kaggle and Google Colab for further analysis using Python-based notebooks (e.g., pandas, matplotlib, sea- born), alongside offline software (e.g., jamovi). Affinity Designer and Affinity Photo were used to refine plot outputs. Data and analyses were also uploaded to the Open Science Framework (osf.io). Primary gaze metrics included total fixation duration (in milliseconds and percent-

age), fixation count, average fixation duration, and proxy saccadic scanpath length. As mobile eye tracking lacked constant participant-to-stimulus distance, scanpath length was calculated using pixel-based on-screen values as a proxy. 3. Results The results section generally reports descriptive values, specifically the mean (M) and standard error of the mean (±SEM), unless stated otherwise. The primary indicators of participant engagement with the maps and artworks were derived from fixation-based metrics, including PLOS ONE Studying attention to IPCC climate change maps with mobile eye-tracking PLOS ONE | https://doi.org/10.1371/journal.pone.0316909 January 10, 2025 5 / 17