Challenges For Causal Inference On Digital Platforms
summary
A randomized audit uses 1% control / 4% treatment users on Twitter (April–Aug 2020) to define an amplification ratio and measures millions of political tweets across seven countries and 6.2 million shares of US news. Result: political content is generally amplified and right‑leaning sources show modestly higher amplification; German center parties outperform extremes. Separate within‑study comparisons for four user settings show standard observational adjustments fail—three of four methods get the causal sign wrong—while RCTs show the algorithmic timeline raises engagement.
AI-written summary · about
chapters (21)
- 0:00Intro
- 0:14Causal stories about algorithmic moderation on social platforms
- 1:13Two projects about causal inference on social platforms
- 2:30Twitter's Home timeline
- 4:10Who is benefitting from the algorithmic timeline?
- 6:15Experimental setup starting in 2016
- 8:14Idiosyncrasies of the experimental setup
- 9:30Hodge-podge causal effects
- 10:59Defining and measuring algorithmic amplification
- 13:56Scope of algorithmic audit
- 14:34Algorithmic amplification for each party
- 16:32Right versus left of the spectrum
- 25:37Discussion
- 32:57User choice and controls
- 34:44Understanding causal effects of user agency
- 36:46Experimental setup: Within-study comparison
- 37:45Scope of study
- 40:08Persistent confounding
- 43:55Catch 22
- 44:55Causal effects of algorithms in social systems
- 46:14Broader directions
related talks
32mPlatform-supported Auditing Of Social Media Algorithms For Public Interest
39mCausal Effect Estimation in Practice: Lessons Learned from E-commerce & Banking
43mBeyond Just Prediction: Causal Thinking in Machine Learning
30mPredictive Analytics in Social Media and Online Display Advertising
38mHow to Enhance Customer Targeting in Marketing
41mOn the Path to Causal Inference
46mFrom correlation to causation: measuring ad effectiveness at scale
36mMeasuring Media Impact- Practical Geo-Lift Incrementality Testing