Body camera footage leads to lower judgments of intent than dash camera footage

Significance Surveillance video from body cams and dash cams is increasingly used by police organizations to enhance accountability, and yet little is known about their effects on observer judgment. Across eight experiments, body cam footage produced lower judgments of intent in observers than did dash cam footage, in part, because the body cam (vs. dash cam) visual perspective reduces the visual salience of the focal actor. This research informs public policy regarding the interpretation of video surveillance methods of police conduct.


Experiment 2
Sample recruitment and exclusions. We posted an advertisement for this experiment to US residents on Amazon's Mechanical Turk in exchange for $0.50 and kept the experiment open for 24 hours between January 31, 2017, and February 1, 2017. We did not collect any demographic information. We left the survey open until we collected 115 responses. Responses from those participants who completed the entire survey constitute the 105 observations provided in the data file (Experiment 2 Data).
Stimuli. The videos for this experiment were filmed simultaneously by 3 Zoom QD HD Handy-cam digital video recorders. The cameras were worn at chest level by both actors. The actors were of similar height, build, and ethnicity. Both actors were male, and both had the same hair color. The dash cam was positioned 3 meters away from the incident and was 1.5 meters off the ground. The video depicted the two actors bumping into each other. Thus, there were three versions of the video: Two videos were captured by the body cam of each wearer, and one video captured the dash cam perspective. All videos were filmed indoors under florescent lighting.
Procedure. All participants read "In the next video you will see an interaction between two people." All participants then viewed videos of Person A and Person B and were asked to identify both to verify that their web browser showed the videos correctly. In the Body Cam condition, Person B was identified as "Person B body cam" and footage from their body cam was shown on a loop. In the dash cam condition, Person B was identified by a label on the video that read "Person B." To control for possible biased reaction toward either of the two actors, participants were randomly assigned to a second condition in which the first actor was Person A (and the second actor was Person B) or the first actor was Person B (and the second actor was Person A), as shown in Figure S2. After verification that the videos worked on their browsers, participants viewed either a body cam or dash cam video (Actor 1 or 2), and were asked to make judgments on the intentionality of Person B.

Experiment 3
Sample recruitment and exclusions. We posted an advertisement for this experiment to US residents on Amazon's Mechanical Turk in exchange for $0.50 and kept the experiment open for 24 hours between February 23, 2017, and February 24, 2017. We did not collect any demographic information. We collected 228 responses. Responses from those participants who completed the entire survey constituted the 220 observations provided in the data file (Experiment 3 Data).
Stimuli. The videos for this experiment were recorded with 3 Zoom QD HD Handy-cam digital video recorders. The cameras were worn at chest level by both actors. The actors were of similar height, build, and ethnicity. Both actors were female, and both had the same hair color. The dash camera was positioned 3 meters away from the incident and was 1.5 meters off the ground. The video depicted the two actors bumping into each other. Thus, there were three versions of the video: Two videos were captured by the body cam of each wearer, and one video captured the dash cam perspective. All videos were filmed outdoors under natural lighting. Procedure. Experiment 3 followed the same procedure as Experiment 2.

Experiment 4
Sample recruitment and exclusions. We posted an advertisement for this experiment to US residents on Amazon's Mechanical Turk in exchange for $0.50 and kept the experiment open for 24 hours between July 10, 2018, and July 11, 2018. We left the survey open until we collected 402 responses. Complete responses from those participants constitute the 348 observations (Age = 37.16, SD = 12.26, 63% Female, 83% White) provided in the data file (Experiment 4 Data).
Stimuli and Procedures. Two videos from Experiment 1 (the window breaking video and the police shooting 2 video) were used as the stimuli for this experiment. In the control conditions, participants read, "You will see videos of police officers from either their body cam or dash cam. You will then make judgments about these videos." In the perspective-taking conditions, the participants read instructions adapted from past research on perspective taking (1): "In preparing for this task, take the perspective of the police officer in each video. Try to understand what they are thinking. What are their interests and purpose in the situation? Try to imagine what you would be thinking if you were in their shoes." We measured intentionality after each video with two questions (Video 1: "The officer intentionally broke the car window"; "The officer intended to break the car window"; 1 = Strongly Disagree, 7 = Strongly Agree; r = 0.71; Video 2: "The officer intentionally shot the suspect"; "The officer intended to shoot the suspect"; 1 = Strongly Disagree, 7 = Strongly Agree; r = 0.62). Participants then answered manipulation check questions ("In the videos, I put myself in officer's shoes"; "In the videos, I took the perspective of the officer"; 1 = Strongly Disagree, 7 = Strongly Agree; r = 0.75). We also collected ratings of participants' perceived objectivity ("While watching the videos, I was objective in my judgments about the officer"; "I maintained objectivity while watching each video"; 1 = Strongly Disagree, 7 = Strongly Agree; r = 0.79) and their motivation for being accurate ("I wanted to be accurate in judging what happened in each video"; 1 = Strongly Disagree, 7 = Strongly Agree) and fair ("I wanted to be fair in judging what happened in each video"; 1 = Strongly Disagree, 7 = Strongly Agree).
Extended Results. To check our manipulation, we submitted each variable to a 2 (perspective-taking) X 2 (body cam vs. dash cam) between-participant ANOVA. The perspective-taking manipulation influenced participants' self-reported tendency toward taking the perspective of the officer, and also increased their perceived objectivity, but did not impact their perceived accuracy or fairness (Ps > 0.50). Specifically, those in the perspective-taking condition indicated that they took the officer's perspective (M = 5.81, SD = 1.13) to a greater extent than those in the control condition (M = 4.89, SD = 1.31), t(346) = 7.03, P < 0.001. Further, those in the perspective-taking condition reported being more objective in their judgments (M = 5.83, SD = 0.96) than those in the control condition (M = 5.25, SD = 1.09), t(346) = 5.28, P < 0.001.
These variables were submitted to a 2 (perspective taking) X 2 (body cam vs. dash cam) X 2 (incident) measures ANCOVA, with the latter factor being within-subject, along with the seven covariates. We noted the main effect body cam vs. dash cam, F (1, 340) = 31.27, P < 0.001, such that those in the body cam condition (M = 4.74, SD = 1.59) showed lower intentionality judgments than in the dash cam condition (M = 5.38, SD = 1.53). Also, the within-subject factor of incident was significant, F (1, 340) = 8.53, P = 0.004, such that intentionally judgments for the car window video (M = 5.53, SD = 1.51) were higher than the judgments for the shooting (M = 4.60, SD = 1.53). There was no interaction between perspective-taking and incident, F (1, 340) = 0.10, P = 0.75, nor between visual perspective and incident, F (1, 340) = 0.35, P = 0.25, nor between perspective-taking and visual perspective, F (1, 340) = 1.29, P = 0.25. Of the covariates, only the concern about crime variable was significant in the model, F (1, 340) = 6.25, P = 0.004. For ease of interpretation, we combined the intentionality ratings across the 2 videos α = 0.58) and present the full model with this dependent variable in Table S4. Finally, to test the possible interactive effects of concern for crime on the model factors, we ran the following linear model predicting intent, Intent= β0 +(β1P X β2V X β3I X β4C)+ ; where P = Perspective Taking Manipulation (0 = Control, 1 = Perspective Taking Manipulation), V = Video Type (0 = Body Cam, 1 = Dash cam), I = Incident (0 = Window Breaking, 1 = Police Shooting), and C = Concern About Crime (continuous). In this model, concern about crime held no interactive effect on any combination of factors (Ps > .80), while the main effect of body cam vs. dash cam remained significant, β = -0.57, SE = 0.23, t = 37.62, P < 0.001. Stimuli and Procedures. The same videos from Experiment 4 were used as the stimuli for this experiment. The valence of the incident was manipulated with verbal descriptions to be either neutral or negative. In the neutral condition, the person who was not the police officer was identified as a suspect. The negative outcome condition informed participants that "a baby was in the backseat, and was injured by the broken glass," and "the person shot was innocent, and a father of two." After each intention judgment, participants answered a manipulation check question ("The outcome of this video was"; 1 = Negative, 7 = Positive).

Extended Results.
To check the manipulation of incident valence, we submitted the manipulation check measure to a 2 (incident valence) X 2 (body cam vs. dash cam) between-participant ANOVA. We found that incident valence had an effect, such that the negative outcome was rated as more negative (M = 3.29. SD = 1.63) than the positive outcome (M = 3.84, SD = 1.53) F (1, 256) = 7.98, P = 0.005. There was no main effect of body cam (vs. dash cam), nor was the interaction significant (Ps > 0.44), thus suggesting that the manipulation was successful.
Next, we ran a repeated measures ANOVA that found that there was a main effect of intentionality judgments on the video that was presented, F (1, 256) = 69.98, P < 0.001, such that the average intentionally judgment for breaking the car window (M = 6.03, SD = 1.41) was higher than the judgment for the shooting (M = 5.22, SD = 1.33). There was no main effect of the incident valence manipulation on the intentionality ratings. The interactions between outcome valence, F (1, 256) = 0.001, P = 0.98, visual perspective as body cam or dash cam, F (1, 256) = 1.33, P = 0.25, and the interaction of these two manipulations, F (1, 256) = 0.07, P = 0.78, were not statistically significant. For ease of interpretation, we combined intentionality ratings across the videos, and presented a single measure of intentionality (α = 0.54).

Experiment 6
Sample recruitment and exclusions. We posted an advertisement to US residents on Amazon's Mechanical Turk in exchange for $0.50 and kept the experiment open for 24 hours between April 29, 2017, and April 30, 2017. We collected 330 responses. Responses from those participants (Age = 34.88, SD = 10.69, 55% Female, 80% White) who completed the entire survey constituted the 308 observations provided in the data file (Experiment 6 Data).
Stimuli and procedures. The videos for this experiment were filmed with two iPhone cameras. The cameras were worn at chest     Df = 346, *P < 0.05, **P < 0.01, *** P < 0.001, n.s. Not Significant