(1/4) This point Mummolo makes in the back-and-forth about the Knox et al. article is relevant to policing experiments, too. It is why the key conclusions in the famous Mazerolle et al. (2013) QCET experiment are probably wrong. https://twitter.com/jonmummolo/status/1275790514873401344
(2/4) Their key finding was that procedural justice in one encounter has sizable indirect effects on general attitudes (see below), but they use a non-randomized mediator (encounter-specific perceptions) that is a collider variable. Some of the omitted variables are on pages 5-6.
(3/4) Elwert & Winship (2014: 44-45) discuss the problem, noting that only the total effect in experiments is causally identified. In Mazerolle et al. (2013), the total effects of procedural justice on general attitudes are very small and mostly non-significant (see Table 1).
(4/4) Methodologists in other fields have developed methods to try to recover indirect effects in experiments with non-randomized mediators:
https://journals.sagepub.com/doi/10.1177/1088868314542878
https://journals.sagepub.com/doi/abs/10.1177/0002716209351526
https://journals.sagepub.com/doi/10.1177/1088868314542878
https://journals.sagepub.com/doi/abs/10.1177/0002716209351526