I recently performed analysis for a media watchdog organization regarding Big Tech’s political influence. Consequently, I’ve received some (American) right-wing press cut. Not too happy that only one side of the aisle is paying attention to what, in my opinion, is a centrist concern, but I’ll take whatever attention (to the issue) that I can get.
Disclosure: I’m somewhat of a leftist in political theory. But I value constitutional democracy over ideology.
A common belief, backed by data and whistle-blowers’ leaks, is that Big Tech is suppressing conservative voices. In light of this, Chris White, a reporter at the conservative Daily Caller, asked me to evaluate whether Facebook is suppressing conservative pundit Deneen Borelli’s content. They observed, as suggested by the plot below, that her online attention decreased after she spoke at the Conservative Political Action Conference (CPAC). I stated that I while I could not prove any causal link between her CPAC performance and potential Facebook suppression, I could perhaps establish correlation between the perceived change in slope in the plot below and the CPAC event. I proceeded to investigate whether the change in slope at the time of CPAC is statistically significant.
Python code (but not the proprietary data) used to conduct this analysis is available for peer review at https://github.com/badassdatascience/media-bias/tree/master/Daily-Caller/Borelli.
I applied a simple linear regression model relating time to “lifetime total likes”, using information provided by the Daily Caller (and not available to share, sorry!). I also inserted a dummy variable to indicate pre- and during/post-CPAC periods. Additionally, I accounted for autocorrelation with a lag of one (further discussed below). Finally, I used an interaction term between time and the dummy variable; if this variable proves significant in the model than we gain evidence that the slope change correlates with the CPAC event.Before fitting the model to the data, I looked at the autocorrelation:Here we can see that there is statistically significant autocorrelation to roughly twelve lags. I only used one lag in the model, but this one lag is the most important one, and taking all twelve would have left less of the time-series itself to analyze.
Fitting the model shows a statistically significant change in slope (p=0.002 on the interaction term), suggesting a correlation between CPAC’s occurrence and the attenuation of Ms. Borelli’s lifetime total likes:Admittedly, I forget to evaluate effect size. I’ll do that if someone needs that information.
This analysis does not explicitly say Facebook targeted Ms. Borelli. In the name of good scientific practice, when reporting this analysis originally to the Daily Caller I also proposed the alternative theory that she simply saturated her pool of potential viewers.
However, we now know that Facebook changed its algorithm about that time to deemphasize political content, both on the left and the right. This proves the most likely explanation for these observations.
You can read Mr. White’s article detailing the consequences of Facebook’s algorithm change, with reference to this analysis, at Facebook’s Recent Algorithm Changes Are Laying Waste To Conservative And Liberal Outlets.