I just read the Birdwatch white-paper which Community Notes is based on. I realized that in the Appendix, specifically table 10 which relates to RQ2, the model for Tweet agreement on a 5-point scale has limited explanatory power. Modestly speaking.
Using an unweighted OLS with an adjusted R-squared of 0.03 but with statistically significant variables (few of them) and a statistically significant F-statistics just means the model is "better" than nothing. In the lens of statistics this model is useless for explainability or to use the estimated dependent variable 'Tweet agreement' in a bivariate association as erroneous.
The percentage of standard deviation explained is:1 - sqrt(1 - r_squared)*100 = 2.02%
In summary, the model can explain about 4% in the variability of Tweet agreement and about 2% of the standard deviation of its errors.
Have you thought about redoing the analysis/paper with statistically rigorous methods?
jakorostami
commented
6 months ago
Hi.
I just read the Birdwatch white-paper which Community Notes is based on. I realized that in the Appendix, specifically table 10 which relates to RQ2, the model for Tweet agreement on a 5-point scale has limited explanatory power. Modestly speaking.
Using an unweighted OLS with an adjusted R-squared of 0.03 but with statistically significant variables (few of them) and a statistically significant F-statistics just means the model is "better" than nothing. In the lens of statistics this model is useless for explainability or to use the estimated dependent variable 'Tweet agreement' in a bivariate association as erroneous.
The percentage of standard deviation explained is:
1 - sqrt(1 - r_squared)*100 = 2.02%In summary, the model can explain about 4% in the variability of Tweet agreement and about 2% of the standard deviation of its errors.
Have you thought about redoing the analysis/paper with statistically rigorous methods?
Kindly