Spanish-language disinformation spreads via WhatsApp, but its private nature makes it inherently difficult to study. Using Google Trends data, I find that WhatsApp searches predict pro-Trump swings at the DMA level, even after adjusting for political and demographic factors.
Caveats: Google searches are not equivalent to downloads or usage, media markets are large geographies, and the use of aggregate-level data obscures variation within media markets.
Caveats #2: Causal inference is difficult. Change scores and fixed effects help reduce confounding, but do not eliminate it entirely. More granular data on WhatsApp usage and voting behavior would allow us to implement more credible designs.
The patterns here are interesting from a descriptive perspective, but I’ve included relevant covariates so I don’t waste peoples’ time with something that can easily be explained by ethnic composition, age distributions, etc.
However, I’ve received questions about unobserved confounding, so out of curiosity, I estimated sensitivity analysis models (Cinelli and Hazlett 2019) to place some bounds on this. Major takeaway: you need a confounder 4x as strong as % Hispanic for the estimate to go to zero.
This is no substitute for better data or designs, but it gives us some sense of how fragile these patterns might be. Given the prominence of % Latino in post-election analyses, I found the result above surprising. Bottom line: more research is needed!
Updated figure. I'm grateful for the comments, and will continue assessing the robustness of these patterns. I hope we'll eventually be able to test this relationship using more/better data, but this first-cut analysis has greatly improved thanks to many of you.
• • •
Missing some Tweet in this thread? You can try to
force a refresh