14 July 2017

Traffic Stop Racial Disparities

Welcome back. I came across a recent study that was sort of a good news, bad news report on racial relations in the US.
 

Oops...side-view mirror image of
a traffic stop. (multiple websites)
Seeking to improve community relations, the Police Department of Oakland, California, cooperated with a team of Stanford University researchers on an analysis of language transcribed from body-camera video of traffic stops. 

The investigation found words uttered by police officers, though respectful to everyone, were significantly more respectful to whites than to blacks.

Body-Camera Video Transcriptions
The researchers used transcriptions of 183 hours of body-camera video from 981 routine traffic stops of black (682) and white (299) drivers made by 245 different police officers during April 2014. In all, they obtained 36,738 officer utterances for analysis.

The investigation was completed in three steps, which the researchers labeled three separate studies.

Study 1: Human Ratings of Officer Utterances
To test whether human raters could reliably judge differences in respect from transcriptions of officers’ language, the researchers randomly sampled 312 utterances spoken to blacks and 102 utterances spoken to whites.

They enlisted 70 participants (39 female, average age 25) to each rate 60 of the 414 utterances on five overlapping dimensions: how respectful, polite, friendly, formal and impartial the officer was in each exchange. Participants read the officer’s utterance together with the immediately preceding driver’s utterance if there was one.

Each of the 414 utterances was rated by at least 10 participants, and the scores were averaged to calculate a single rating on each dimension. The five dimension ratings were then combined statistically.

Study 1 determined that human raters could reliably glean key features of police officer treatment from their language. The results showed officers’ utterances toward black drivers rated lower than utterances toward white drivers on every dimension.

Study 2: Developing Computational Models
To extend the value of study 1 toward a more general solution for analyzing body-camera language, the researchers developed computational linguistic models of respect and formality.

The models were based on theories of politeness, power and social distance in respectful language, for example, apologizing (sorry for stopping you), softening of commands, saying thank you and using formal (sir, ma’am) instead of informal titles. The utterances from study 1 were used to tune the models.

Study 2 determined that the model-derived ratings agreed with the human ratings from study 1 about as well as human ratings agreed with each other.

Study 3: Assessing Racial Disparities in Respect

Having determined that humans can reliably rate officer speech and that the same ratings can be reliably modeled, the researchers applied the models to the full dataset of 36,738 utterances from the transcribed interactions.

The results showed strong evidence that utterances spoken to whites were consistently more respectful, even after controlling for contextual factors of the interaction, such as the severity of the offense or outcome of the stop.

Wrap Up
The bad news, of course, is that racial disparities in showing respect haven’t gone away. But the good news is that, not only did the Oakland Police Department cooperate in the study, it began implementing new training programs with the earliest results in 2014.

More good news is that models and methodology from the study can be applied elsewhere as well as to gauge the effectiveness of interventions other than traffic stops.

The work demonstrated the value of body camera video transcriptions as an important source of data to address the limitations of methodologies that rely on citizens’ recollection or direct observation. The computational linguistic models of transcribed datasets have the potential to allow useful information to be extracted while maintaining objectivity and privacy and ultimately improve police-community relations.

Thanks for stopping by.

P.S.
Stanford study of body camera videos in Proceedings of the National Academy of Sciences: www.pnas.org/content/early/2017/05/30/1702413114.full
Example media reports on Stanford study:
www.cnn.com/2017/06/05/health/police-language-race-oakland-study/index.html
www.pbs.org/newshour/rundown/police-respect-whites-blacks-traffic-stops-language-analysis-finds/
www.nytimes.com/2017/06/06/us/police-race-bias-study.html?_r=0

No comments: