This summer has witnessed unprecedented protest and unrest in the United States. Armed Conflict Location and Event Data Project banded together with Bridging Divides Initiative at Princeton to create the US Crisis Monitor. They just came out with a report showing a preponderance of peaceful protests under the banner of Black Lives Matter. Although we must be cognizant of their class origins and political bias, their data show that there were 10,666 peaceful demonstrations and 618 riots between May 24 and Aug 22. In what follows, we’ll work with their county-level data.

The spatial cross-section of the protests and riots contain interesting information on the social basis of the unrest. Since many cities have seen sustained demonstrations, and, in some cases, rioting, we would be throwing away valuable information if we only looked at whether a county had at least one protest/riot this summer. Far better to use the entire information contained in the data. That requires making distributional assumptions. It turns out that the negative binomial distribution, that served us so well for police shootings, provides a good fit for protests/riots as well.

The spread and density of peaceful demonstrations is quite impressive. Every single state has seen multiple protests. There have been demonstrations at 1,485 locations in the US, and 980 of these locations have witnessed multiple demonstrations.

Riots are considerably much more concentrated. Only 203 places have seen rioting. Of these, 96 have seen multiple riots over the course of the summer.

Besides the distributional assumption, there are two further issues of statistical housekeeping. First, throwing everything in a single multiple linear regression model is *not* kosher. The reason is that in order to identify the effect of feature X on response y, we need to find a set of covariates Z, that blocks the flow of spurious information through all backdoor paths from X to y. Controlling for such a set of covariates Z then allows us to identify the fixed effect of X on y. For multiple X’s, the Z’s may not coincide. If we were to throw them all in the same regression, we’re almost guaranteed to control for the wrong variables (such as mediators or common effects), thus leading to misidentification of the fixed-effects. So multiple linear regressions with lots of predictors are almost guaranteed to yield spurious results — scholars need to drop the practice. (This is a serious problem with complicated predictive models for the 2020 presidential election, such as those of Andrew Gelman and *FiveThirtyEight*.) In order to get around this problem, we’ll estimate the fixed-effects of various features controlling only for the variables in the null model. While not perfect, it is dramatically superior to the kitchen-sink approach.

Second, we need a null model of protests/riots, controlling for which we may estimate the fixed-effects of various features/predictors. This can be either simply county population, since we expect more populous cities to have a higher frequency of protests. Or it can be race-specific county populations, since for obvious reasons, we expect higher participation by African-Americans. We shall test for both.

Okay, so much for preliminaries. Let’s begin with the null model of peaceful demonstrations this summer. Table 1 documents our estimates.

Table 1. Null model of protest frequency. | ||||

b | std err | Z | P | |

Log population | 1.994 | 0.041 | 48.77 | 0.000 |

Log black population | -0.038 | 0.052 | -0.73 | 0.468 |

Source: US Crisis Monitor, author’s computations. Negative Binomial model fitted with IRLS. Intercept included but not shown. All features have been robustly standardized to have mean 0 and variance 1. Estimates in bold are significant at the 5 percent level. |

Our first result is that, after controlling for county population size, the size of the black population does *not* predict higher protest frequency (*P* = 0.468). The implication is that the protests have not been dominated by the African-American community. Whites have been there all along, standing shoulder to shoulder with blacks in their fight for racial justice. This should come as no surprise to anyone who’s been participating or watching. Note also that the elasticity of population size against protest frequency is 2 — suggesting that protests have been especially frequent in the most populous cities.

Since the size of the black population does not predict protest frequency, we throw it out and only control for county population size. Table 2 documents our estimates for selected features of interest after controlling for county population size (included but not shown). Note that each feature is estimated separately, as explained above.

Table 2. Predictors of protest frequency. | ||||

b | std err | Z | P | |

College Graduation Rate | 0.362 | 0.020 | 18.43 | 0.000 |

Log Median Rent | 0.186 | 0.027 | 6.95 | 0.000 |

Log White HH Income | 0.084 | 0.031 | 2.75 | 0.006 |

Log Black HH Income | -0.141 | 0.030 | -4.77 | 0.000 |

Log Per Capita Income | 0.206 | 0.024 | 8.67 | 0.000 |

Trump Swing | -0.156 | 0.022 | -7.10 | 0.000 |

Overdose Death Rate | -0.096 | 0.023 | -4.17 | 0.000 |

Source: US Crisis Monitor, author’s computations. Negative Binomial models fitted with IRLS. Intercept and features of the Null Model included but not shown. All features have been robustly standardized to have mean 0 and variance 1. Estimates in bold are significant at the 5 percent level. |

College graduation rate is very strongly associated with protest frequency across the three thousand US counties. It has the largest elasticity: one standard deviation higher college graduation rate predicts +0.36 more protests on average. The implication is that more educated counties have seen a higher frequency of peaceful demonstrations. There is thus a strong class bias in protest frequency. This interpretation in reinforced by the elasticities of per capita income (b = +0.21, *P* < 0.001) and median rent (b = +0.19, *P* < 0.001). However, with race-specific median household income (estimated jointly), we uncover an even more interesting pattern. While the median household income for whites is positively associated with protest frequency (b = +0.08, *P* = 0.006), the median household income for blacks is *negatively* associated with protest frequency (b = -0.14, *P* < 0.001). The interpretation is obvious: while affluent whites are overrepresented in BLM protests, affluent blacks are underrepresented.

We know that overdose deaths have been concentrated in the white working-class and the white working-class is who put Trump in the White House. Here we find that the Trump swing — the change in GOP vote between 2012 and 2016 — is negatively associated with protest frequency (b = -0.16, *P* < 0.001). This suggests a strong partisan bias in the distribution of protests — Democratic strongholds have seen many protests, Trump country has not seen all that many. We also find that overdose death rate is negatively associated with protest frequency (b = -0.10, *P* < 0.001). This is congruent with class-partisan bias of protest frequency. Finally, we note than manufacturing towns are *not* significantly less likely to have witnessed protests this summer (OR = 0.85, *P* = 0.108).

The overall picture is that affluent, college-educated Democrats have been preponderant at BLM protests. Again, these results are hardly surprising to anyone who has been paying attention.

We now move on to riot frequency. Again, have a negative binomial model of type 2. Table 3 displays our null model.

Table 3. Null model of riot frequency. | ||||

b | std err | Z | P | |

Log Population | 2.245 | 0.181 | 12.42 | 0.000 |

Log Black Population | 1.141 | 0.260 | 4.38 | 0.000 |

Source: US Crisis Monitor, author’s computations. Negative Binomial model fitted with IRLS. Intercept included but not shown. All features have been robustly standardized to have mean 0 and variance 1. Estimates in bold are significant at the 5 percent level. |

Unlike protest frequency, we find that the size of the black population predicts riot frequency, even after controlling for county population size. It is possible that aggressive protests in black communities are more likely to be declared a riot by the police, the media and perhaps even by the researchers who compiled the data. But the strength of the association is very strong. We therefore control for race-specific population size when we estimate the strength of association with our selected features. Table 4 reports our estimates.

Table 4. Predictors of riot frequency. | ||||

b | std err | Z | P | |

College Graduation Rate | 0.485 | 0.080 | 6.02 | 0.000 |

Log Median Rent | -0.010 | 0.106 | -0.10 | 0.922 |

Log Black HH Income | -0.787 | 0.112 | -7.03 | 0.000 |

Log White HH Income | 0.213 | 0.117 | 1.82 | 0.070 |

Log Per Capita Income | 0.081 | 0.098 | 0.83 | 0.408 |

Trump Swing | -0.291 | 0.094 | -3.09 | 0.002 |

Overdose Death Rate | -0.127 | 0.089 | -1.42 | 0.156 |

Source: US Crisis Monitor, author’s computations. Negative Binomial models fitted with IRLS. Intercept and features of the Null Model included but not shown. All features have been robustly standardized to have mean 0 and variance 1. Estimates in bold are significant at the 5 percent level. |

We find that college graduation rate is very strongly and positively associated with riot frequency. More educated places have seen a greater frequency of rioting, even after controlling for race-specific population size. With an elasticity of almost one-half (b = +0.49, *P* < 0.001), college education is the strongest predictor of riot frequency. On the other hand, per capita income and median rent are not significantly associated with the frequency of rioting. It is possible that this may have to do with the low frequency of the riots, so that it is hard to obtain statistically significant results. In other words, we may lack the power to find statistical significance even if the associations are true in the data generating process.

Interestingly, while the median income of white households is only marginally associated with riot frequency (b = +0.21, *P* = 0.070), the median income of black households is strongly *negatively* associated with riot frequency (b = -0.79, *P* < 0.001). The interpretation is that educated whites and less affluent blacks are overrepresented among rioters. Put another way, affluent black professionals are not only underrepresented among protestors, they are especially underrepresented among rioters.

We find that overdose death rates are not significantly associated with riot frequency (b = -0.13, *P* = 0.156), although the sign is congruent with the association we found for protests. Meanwhile, the Trump swing is strongly and negatively associated with riot frequency (b = -0.29, *P* = 0.002). This attests to the partisan bias in rioting. Finally, whereas 3 percent of manufacturing counties have witnessed at least one riot, 7 percent of non-manufacturing counties have. The odds ratio is extremely significant (OR = 0.38, *P* < 0.001). Again, this hardly comes as a surprise. These are not exactly Hard Hat riots.

The overall picture is that college-educated white and working class black Democrats are overrepresented among rioters.

These patterns not only shed light onto the social basis of the unrest, they also explain the class-partisan polarization in perceptions of BLM and public safety. Biden’s in a pretty tough spot. He keeps denouncing the violence on the streets, but hardly anyone outside the professional class is buying. Other professional-class Dems, who dominate the media, want to push-back against the narrative equating their party to BLM and BLM to street violence, by appealing to the overwhelmingly peaceful nature of the protests. But if facts swayed American voters, we’d have another Clinton in the White House now, wouldn’t we?

“But if facts swayed American voters, we’d have another Clinton in the White House now, wouldn’t we?”

No we absolutely would not. We would have a Bernie Sanders in the white house. The best thing that ever happened to HRC was the over the top right wing smears because it gave average dems license to ignore her horrible record. Pointing out her various flaws was like talking to a stone wall, it just didn’t register that a democrat could do awful things, when in reality I have a hard time coming up with anything a democrat has done recently that most of their base wouldn’t call horrible.