Intervention(s)
In this experiment, participants will be paid to scroll through a randomized feed of short videos and report the emotions they experience while watching each video. The "intervention" consists of within- and across-person randomization of videos into these feeds, which will allow us to estimate the effects of different types of videos on emotions.
Via a previous data collection, we have access to the TikTok, Instagram, and YouTube viewership histories of about 390 young Americans. We will sample videos for this experiment through the following procedure. (1) We randomly sample a subset of the creators who appear in this data (by randomly picking videos and adding their creators to our subsample). (2) We scrape videos posted by these creators in the week preceding the rollout of the experiment. (3) We scrape the metadata of these videos (transcripts, captions, thumbnails, etc.) and classify features of the videos using LLMs. Of particular interest is whether the video prominently features a creator whom viewers would consider 'enviable' on one or more dimensions (appearance, success, wealth, etc.), and whether the video is about politics. Politics will take primacy in our codings, so that a video classified as political cannot be classified as enviable. (4) From the videos scraped and classified this way, we will randomly sample a subset of videos, maintaining equal weight on all the sampled creators, and stratifying so that we end up with 33% each of politics, enviable, and other videos. We will sample a number of videos such that, given our expected sample size of participant-video pairs, each video in our data is seen by 5-10 participants. We will exclude advertisements and non-English-language videos.
Once we have this set of videos, participants in the experiment will see a feed of videos randomly sampled from this set, in a random order, stratified to be 33% politics, 33% enviable, and 33% other.