Here's an example for why these stats are inherently unreliable when used to infer causes.

Let's say the whites in the 'compliant' sample said "yes, officer" and "no, officer" more often than the blacks did, while blacks said "yeah" and "no" instead. Well, that one difference could account for a non-insignificant amount of difference in force used against subjects. Saying "yes, officer" instills a greater deal of safety and comfort into the situation and distills tension the officer may be feeling. It is reasonable to assume that the samples could contain unaccounted for variables like this.