A Transparency Statement Transforms Community-Police Interactions

Abstract

Over the last three decades, billions of dollars have been invested in community-oriented policing approaches that are meant to promote positive interactions between officers and civilians. However, trust in law enforcement remains low. Our qualitative analysis of over 500 hours of naturalistic observations suggested this is because police officers can make civilians feel threatened even when they are not actively investigating a crime. Here we use a pre-registered field experiment (N = 232 community members interacting with police officers) to show that civilians are less threatened and report greater trust when officers add a 10-word “transparency statement” (“I'm walking around trying to get to know the community”) to the beginning of interactions communicating their benevolent intent. Corroboration of these conclusions comes from natural language processing analyses of the interaction transcripts and from ambulatory sensors that monitored community members’ sympathetic nervous system activation. Further, five online experiments isolated the conditions under which transparency statements were most impactful (total N = 2,241). This multi-method investigation highlights an under-appreciated reason why authority figures such as police so often fail to promote positive relationships with the community: a lack of transparency about the reasons for their behavior.

Full text

1
1 Title: A Transparency Statement Transforms Community-Police Interactions 2
Authors: Kyle S. H. Dobson*†1, Andrea G. Dittmann*†2, David S. Yeager1
3
4
1 Population Research Center and Department of Psychology, University of Texas at Austin, 5
Austin, TX, 78712. 6
2 Emory University, 1300 Clifton Rd, Atlanta, GA 30322 7
† Authors contributed equally. 8
9
*Corresponding authors. Email: kyle.dobson@austin.utexas.edu or 10
andrea.dittmann@emory.edu. 11
12 13

2
Abstract 14
Over the last three decades, billions of dollars have been invested in community-oriented 15
policing approaches that are meant to promote positive interactions between officers and 16
civilians. However, trust in law enforcement remains low. Our qualitative analysis of over 500 17
hours of naturalistic observations suggested this is because police officers can make civilians feel 18
threatened even when they are not actively investigating a crime. Here we use a pre-registered 19
field experiment (N = 232 community members interacting with police officers) to show that 20
civilians are less threatened and report greater trust when officers add a 10-word “transparency 21
statement” (“I'm walking around trying to get to know the community”) to the beginning of 22
interactions communicating their benevolent intent. Corroboration of these conclusions comes 23
from natural language processing analyses of the interaction transcripts and from ambulatory 24
sensors that monitored community members’ sympathetic nervous system activation. Further, 25
five online experiments isolated the conditions under which transparency statements were most 26
impactful (total N = 2,241). This multi-method investigation highlights an under-appreciated 27
reason why authority figures such as police so often fail to promote positive relationships with 28
the community: a lack of transparency about the reasons for their behavior. 29
30 Main Text 31
Since 1994, the U.S. federal government has invested more than $14 billion in more than 13,000 32
law enforcement agencies to support community policing efforts 1. In community policing, 33
officers are asked to be proactive and prevent crimes by questioning community members, 34
solving community problems, and cultivating cooperative relationships, even when they are not 35
actively interrogating a suspect in a potential crime 2. By 2013, 90% of medium-to-large police 36
departments in the U.S. claimed to be using community policing to at least some extent 3, 37
38
Community policing is meant to be a more humane and respectful alternative to the ubiquitous 39 proactive policing model. Proactive policing typically involves an “investigatory stop,” in which 40
police intrusively question people (e.g., “What are you doing here?”), pat them down, search 41
them, or handcuff them when suspected of minor illicit activity 4. One well-known manifestation 42
of this is the “stop and frisk” policy adopted in New York City in the 1990s 5. Proactive policing 43
has become the dominant approach to policing, but it has a high cost. It is a major contributor to 44
the escalation of interactions that end in accusations of wrongful arrest or police brutality 6. 45
46
Although community policing was designed to replace proactive policing, the community 47
policing approach has so far been largely ineffective. Empirical evaluations show no average 48
reduction in crime, and no increase in citizens’ trust in the benevolence of police 7. As a result, 49
community policing has failed to prevent a precipitous decline in the public’s trust in the 50
legitimacy of law enforcement, which is at an all-time low8-9. This loss of trust and legitimacy, in 51
turn, makes citizens more hostile to police and less likely to comply with officers’ requests for 52
the information they need to solve a case—all of which make society less safe from harm 10. 53

3
Interestingly, the effects of community policing, though not beneficial on average, are highly 54
heterogeneous. A systematic review and meta-analysis 7 found that several studies showed 55
significant benefits in line with policymakers’ expectations 11. However, several studies showed 56
significant increases in crime and reductions in trust, contrary to the aims of the approach 12. This 57
raises the possibility that it is not community policing itself, but rather something about how it is 58
implemented, that accounts for the disappointing results in prior evaluation studies. 59
60
Here we present a new method for improving the interactions that are the heart of the community 61
policing model. We show that this method is easy for police officers to implement with little 62
training, and we document that it decreases community members’ sense of threat and increases 63
their trust. Our results suggest that community policing could move closer to fulfilling its 64
promise if officers adjusted the subtle dynamics of their conversations. 65
66
Our research begins with the observation that officers have discretion over their conversational 67
dynamics, but they often fail to leverage that discretion in a way that satisfies the trust-building 68
goals of community policing. A qualitative study of police officers during ride-alongs (i.e. 69
routine officer patrols attended by researchers13) found that civilians with whom officers were 70
interacting still worried that they were being approached for enforcement reasons. In part, this 71
was because patrol officers continued to use aggressive questioning methods that were tailored 72
for suspects (e.g., “What are you doing here? How long have you been here? Where are you 73
going next?”)—even when officers were attempting to cultivate trusting relationships with 74
community members. This study13 furthermore found that community members perceived 75
officers’ interest in them as harassment. As a result, they anxiously tried to disengage from the 76
conversation from the start, only to feel increasingly threatened as the interaction proceeded. At a 77
more general level, community members have learned to fear that officers are accusing them of a 78
crime and are likely to search or handcuff them during an interaction 13. Officers in turn perceive 79
attempts at disengagement or evasion as a suspicious sign that the community member has 80
something criminal to hide. This leads officers to escalate their tactics to interrogate or control 81
community members, resulting in unnecessary arrests or physical harm. The result is a negative 82
cycle of threat and distrust that undermines the goals of community policing, even when officers’ 83
initial intent might have been positive. 84
85
Building on this prior work, we conducted preliminary ethnographic research to generate 86
hypotheses about how to disrupt or even reverse this negative cycle. We conducted over 500 87
hours of field observations, interviews, and ride-alongs with police officers in law enforcement 88
agencies (see Box 1). We observed two units of police officers that were representative of the 89
heterogeneity in the effectiveness of community policing tactics 7—one traditional patrol unit 90
and another unit from a pilot program expressly developed to build relationships with the 91
community. In both units, we observed interactions that were initiated with benevolent intent, in 92
which officers were genuinely attempting to build relationships with their communities. 93

4
However, the two units differed in the tactics they employed to achieve this goal. Some officers 94
peppered community members with questions (e.g., “What are you guys doing here?”, “What are 95
you up to?”), largely consistent with proactive policing tactics 4. When we interviewed 96
community members after those interactions, citizens described feeling threatened, uncertain if 97
they were going to be accused of a crime, and confused as to why the officers chose to talk with 98
them (see left panel in Box 1). Other officers, by contrast, were transparent. From the outset, 99
they explained that their department had a new community policing initiative, and that their job 100
was to get to know the community. Interactions between these officers and community members 101
proceeded very differently. Community members appeared at ease and willing to talk with these 102
officers (see right panel in Box 1). 103
104
Our preliminary qualitative observations led us to generate a hypothesis about how to intervene 105
to improve police-community interactions (see Fig 1A). We theorized that a few words of 106
clarification, in which the officer made was transparent that their questioning was not coming 107
from a proactive policing model, but rather from a community policing model, might put the 108
community member at ease and evoke positive engagement at the outset of the interaction. 109
Importantly, because the first words in a conversation set the stage for everything that follows, if 110
officers briefly clarified from the start what their positive goal is, it could prevent the escalating 111
threat that has an outsized effect on the ultimate outcomes of the interaction. 112
113
We call this a transparency statement. It involves a clear and truthful statement of the officer’s 114
community policing goals at the start of the interaction. Concretely, it is the difference an officer 115
stating "Hi, I'm Officer [Name], how's it going? Can I talk to you for a minute?” which is not 116
transparent but is the default for many officers, and an officer adding the ten-word statement “I'm 117
walking around trying to get to know the community” before asking to talk with the person, 118
which is transparent. See Fig. 1B. Note that without the transparency statement, the community 119
member is free to assume that the officer’s intentions are to repeat the past: aggressively 120
interrogate a suspected criminal. But with the transparency statement, the officer has presumably 121
not left their intentions open to a negative alternative interpretation that could erode trust. To 122
directly test this hypothesis, we conducted the present experiments. 123
124

5
125
Box 1. Themes from Phase 1 that Yielded the Transparency Statement Hypothesis
Lack of Transparency Was Associated with
Threat and Fear of Being Accused of a Crime
Transparency Was Associated with
Positive Emotion and the Development of Rapport
Over a period of four weeks, we conducted four-hour ride-
along observations with patrol officers in a metropolitan area
in the Midwest three-to-four times per week. This was an
area with high-profile accusations of police misconduct. We
observed 18 total patrol officers interact with ~150 different
civilians. Two key themes emerged:
Theme 1: Officers did not explain their intentions at the
outset of the interaction. Officers frequently did not explain
the reason for their interactions at the start of the
conversation. Instead, intentions were only clarified toward
the end, if at all. This was too late because, by this point, the
civilians already appeared anxious and avoidant.
Ex 1: A White female officer drove through a city park,
rolled down the window, and yelled at a civilian sitting on a
bench:
Officer: “You okay?”
Civilian: “Yeah… why?” Civilian stays seated, is standoffish,
seems confused.
Officer: “Just wanted to check—it’s cold!” Officer drives
away with no further conversation.
Theme 2: Officers’ lack of transparency led civilians to
wonder whether they were being accused of a crime.
Although officers said that their interactions were received
positively by the community, when we interviewed civilians,
they felt confused by the conversations and were worried
they would be accused of crimes.
Ex 2: Two White, male officers drove into a local park and
approached four Black residents having a picnic. They began
questioning the residents, saying “How are you?” and “What
are you up to in the park today?” This line of questioning was
met with terse answers, and did not lead to a back-and-forth
rapport. After the officers left, the research team talked with
the civilians. They described feeling anxious and unsure why
the interaction was happening, and they said they felt
threatened.
Ex 3: A Black male and an Asian-American male officer
responded to a 311 call to investigate graffiti outside a
building where there is an Alcoholics Anonymous (AA)
meeting going on. The officers walked around but did not
explain their presence to anyone. The AA meeting attendees
were distracted and staring at the uniformed officers. The AA
meeting leader eventually approached the officers and said,
“Yeah, we’re having an AA meeting right now, so there are a
bunch of alcoholics in there who are afraid you are coming
for them.”
Over a period of three years, we observed a new community
policing unit within a large city in the Midwest which was
known for accusations of police violence against community
members. Within this pilot program, we observed 17 officers,
four sergeants, one lieutenant, and one commander, for more
than 500 hours. Two key themes emerged:
Theme 1. Officers proactively made their intentions
transparent immediately, at the outset of the interaction. The
officers in this community policing unit were aware that their
very presence could make civilians anxious and put them on
edge. Therefore, they proactively explained why they were
approaching them, before they said anything else.
Ex 1: A Latino male officer was approaching a civilian who
was involved in prostitution. This is an arrestable offense but
the officer was not, at that time, planning to arrest the civilian.
Instead the officer’s goal was to clean up a certain corner in the
neighborhood. The officer said: “I told them I wasn’t arresting
them, but ‘I know what you’re doing and you know what
you’re doing, and you shouldn’t be doing that here.’ After I got
real with them, the person dropped their guard, calmed down,
and got out of the area.”
Theme 2. When officers began with a transparency statement,
civilians appeared to feel positively about the police officers’
actions, and a rapport was formed.
Often, the officers were attempting to do positive things for the
community, such as look after a wayward young person. The
officers were attuned to the fact that their interest in helping
could be construed by the community as an attempt to
criminalize or harass the young person. Therefore, they
explained their intentions clearly. Then, when they offered help,
it was often met with positive emotion and the development of
rapport with the community members.
Ex 2: A Latina female officer and her partner had been talking
with a local youth who needed help a few days previously. The
officer described one day when they were on patrol and “we
saw his aunts and uncles so we get down [from the car] and
they’re like ‘Oh [expletive]!’ Immediately, right? We get the
perception of like, ‘The police are here. What happened?’ So,
we tell them, ‘Hey, we're actually here to see [youth’s name].
We met him a few days back and we wanted to bring him a
little something. We promised we were going to come back and
we just haven't been able to.’ So [the aunts and uncles] are like,
‘Really?’ And it was really cool because one of the ladies, she
was like, ‘This is what [city] police should be about.’”
126

6
127 Experiment with live interactions 128
129
Here we report a pre-registered, blinded, controlled, behavioral field experiment that directly 130
tested the causal effect of a transparency statement in live interactions between a community 131
member and a uniformed officer. In Experiment 1, N = 232 adults, who lived in a community 132
known for recent accusations of officers’ use of excessive force, were told that we were studying 133
natural interpersonal interactions and outfitted them with ambulatory physiological sensors. They 134
were not told who they would interact with, or when. A few minutes later, participants were 135
approached by a fully-uniformed officer carrying a weapon and asked questions for 1-2 minutes 136
(Mage = 22.07, SDage = 5.90; 54% female, 4% non-binary/other; 30% held a four-year college 137
degree; 34% White/European American; 32% Asian/Asian American; 17% Latino/Hispanic 138
American; 9% Black/African American; 8% Other). The study design is depicted in Fig. 1B. 139
Officers were asked to have a conversation with a community member like they typically would 140
when engaging in community policing activities. Half of the time the officer added a 141
transparency statement (described above) when they introduced themselves, and half of the time 142
they did not. The instructions given to officers did not differ in any other way. 143
144 Overview of analyses. We report raw means and standard deviations. Average treatment effects 145
and posterior probabilities are calculated from Bayesian machine-learning regression models 15–
146
17 that included covariates measured prior to assignment to condition. Inferences focused on 147
estimation and uncertainty, following best-practice recommendations 18. All conventional null 148
hypothesis tests for measures reported in the main text were P < .05 (see the SI). 149
150 Manipulation check. Community member participants answered our survey questions about the 151
interaction immediately after it ended (See Fig 1B). They rated the officer’s transparency (three 152
items, e.g., “The officer I talked to stated their intentions for talking to me clearly”). As 153
expected, this analysis showed that community members rated the officer as more transparent in 154
the transparency condition (M = 5.64, SD = 1.21) relative to the control condition (M = 4.40, SD 155 = 1.61), average treatment effect (ATE) = .77 SD [posterior distribution 10th %ile = 0.50, 90th 156
%ile = 1.01], posterior probability that the (ATE >0) > .99. See Fig. 1C. 157
158

7
159
Figure 1. A) The three phases of the present research, B) The behavioral study procedures, and C) average
160
treatment effects (ATEs) of the transparency statement in the behavioral study. Note: * = pre-registered,
161
confirmatory outcome. ATE = average treatment effect. EDA = electrodermal activity, which assessed sympathetic
162
nervous system (SNS) activation (see Fig. 3). B) The non-seeing eye represents assignment of community member
163
participants to experimental condition, with the participants and experimenters both kept unaware of condition
164
assignment. Officers were aware of condition assignment because they delivered the manipulation verbally, but
165
study hypotheses were not shared with them. Participants recruited for the study were seated with the sensors during
166
a baseline period, told to act as they naturally would while sitting outside by themselves. Participants did not know
167
when or who they would be approached by, but were told the sensors would monitor them during any interactions
168
they had while in the study. C) ATEs calculated in a Bayesian Causal Forest model that used a machine-learning
169
approach with conservative BART priors; covariates were age, race/ethnicity, gender, education level, and English
170
language proficiency of community members. Thick bars represent the inter-quartile range (IQR) of the posterior
171
distribution of treatment effects, and thin bars depict the 10th to the 90th percentiles. Statistical tests for the final
172
outcome—inspiration—appear in the SI. Conventional null hypothesis tests were all P < .05 (see the SI).
173

8
Pre-registered outcomes. The primary pre-registered outcome was community members’ 174
perceived threat from law enforcement (3 items, e.g. “The police officer did not trust you,” “The 175 officer just wanted to make friendly conversation” (reverse-scored)). Community members 176
reported feeling less threatened in the transparency condition (M = 2.92, SD = 1.45) relative to 177
the control condition (M = 3.40, SD = 1.41), ATE = -.20 SD [-.43, -.01], pr(ATE<0) = .92. 178
179
The secondary pre-registered outcome was trust in the officer’s benevolent intentions (2 items, 180
“In the future, how likely is it that this police officer would go out of their way to help you,” “In 181 the future, how likely is it that this police officer would care about you and your welfare,” 1 = 182 Very unlikely, 7 = Very likely). As predicted, community members reported greater trust in the 183
transparency condition (M = 5.66, SD = 1.14) relative to the control condition (M = 5.31, SD = 184
1.28), ATE = 0.17 [0.001, 0.38], pr(ATE>0) = .90. See Fig. 1B. 185
186 Language and emotion during the interaction. How did transparency affect the unfolding of 187
the interactions over time? To answer this, a principled exploratory analysis was conducted. This 188
involved a) ideographic and statistical analysis of transcripts, and b) use of conservative, 189
Bayesian models that reduce false-positive results. 190
191
First, natural language processing algorithms 19 were applied to the conversation transcripts as a 192
whole. A statistical analysis of all of the conversation transcripts and reflections was conducted, 193
utilizing the popular and well-validated Linguistic Inquiry & Word Count (LIWC-22) software 194
19. The LIWC algorithms tend to identify linguistic features of transcripts of which most people 195
are not conscious. Transcripts from the transparency condition showed that the community 196
members used more authentic language by the end of the interaction, ATE = 46 SD [.23, .66], 197
pr(ATE>0) = .99. This refers to language in which people reveal themselves in a genuine or 198
honest way, not censoring or filtering. This difference across conditions is meaningful because it 199
reveals that participants were using the type of language that forms the foundation of positive 200
rapport that makes people feel unthreatened and open to trust. An analysis of word count also 201
found that community members talked more with officers in the transparency condition, ATE = 202
175 words [129, 232], pr(ATE>0) = .99, consistent with the notion that transparency helped 203
officers establish rapport. On the post-experimental survey, participants reflected on thoughts 204
and feelings from the end of the interaction. In the transparency condition civilians’ text included 205
less causal language, ATE = -.27 SD [-.48, -.08], pr(ATE<0) = .97, suggesting transparency 206
reduced participants’ puzzlement. The natural language analysis also found somewhat more 207
positive emotion words in community members’ written reflections, ATE = .14 SD [-.008, .32], 208
pr(ATE>0) = .88. Finally, participants directly reported on the survey which emotions they felt at 209
the end of the interaction with the officer. Participants in the transparency condition were less 210
than half as likely to report threatened emotions (afraid, nervous, upset, hostile, ashamed) 211
relative to the control condition (Raw data Control = 38%; Transparency = 19%), ATE = -.15pp 212
[-.25, -.05], pr(ATE<0) = .975. 213

9
214
Figure 2. Idiographic analysis of transcripts from police-community interactions (left) and post-interaction
215
reflections (right) for participants in (A) the control condition and (B) transparency condition. Note: Phrases
216
are highlighted to illustrate examples of differences in natural language scored via the LIWC-22 algorithm that were
217
meaningfully different across conditions in statistical analyses of all participants’ data (see the main text).
218
219

10
These statistical patterns were supported by an idiographic analysis of participants who were 220
outliers with respect to the two pre-registered outcomes of threat and trust 20. Idiographic 221
analyses in Figure 2, Panels A-B, show how different the conversations were in the absence 222
versus presence of a transparency statement. On the top (Panel A), the civilian and officer never 223
established a rapport. Instead, we see terse answers. The civilian is guarded and self-conscious. 224
This interpretation is confirmed when we read their reflections after the interaction. The 225
participant reported being puzzled about why the officer was approaching them and concerned 226
by the threatening prospect that they were being investigated for a crime. 227
228
At the bottom of Figure 2 (Panel B), the interaction in the transparency condition is quite 229
different. When the officer begins the conversation with a transparency statement, it ends with 230
the spontaneous back-and-forth, almost resembling new friends who are at ease with one 231
another. In fact, the civilian invites the officer to have future interactions with them at their place 232
of work. When the civilian reflected on this interaction, they admitted that although they were 233
apprehensive at the very start, by the end they felt “comfortable and like we were friends.” 234
Overall, these transcripts matched the trends we saw in our qualitative research (Box 1). 235 236 Stress physiology. We next explored whether the colder and more avoidant response from 237
civilians in the control condition and the more active, warm, engaged response of the 238
transparency condition could be detected through an analysis of sympathetic nervous system 239
(SNS) responses. We conducted an analysis of skin conductance responses (SCR), obtained from 240
electrodermal activity (EDA) sensors embedded in a wrist-worn ambulatory device. These 241
sensors captured in vivo, temporal dynamics of the stress responses during community members’ 242
interactions with police officers. Higher SCRs have sometimes been viewed as evidence of 243
negative stress, but in the present case they were likely to be positive signs of active engagement. 244
This expectation was derived from the Biopsychosocial (BPS) model of challenge and threat 21–
245
23, and from analyses of a laboratory study we conducted to validate the ambulatory EDA data 246
(details reported in SI Fig. S2A-B). In summary, we expected the following trends to emerge 247
when participants received the transparency statement: (a) initial effects on higher SCR at the 248
outset of the interaction, and (b) quicker recovery to homeostasis over time, resulting in 249
negative/reversed treatment effects during the recovery epoch. 250
251
Indeed, we found a positive effect of transparency statements on SCR during the initial ten 252
seconds of the interaction, CATE = 0.13 SD [0.04, 0.25], pr(ATE>0) = .97. See Fig. 3A and B. 253
This is consistent with greater approach motivation, and a challenge response that should come 254
from the formation of a new, positive relationship with an authority figure, like a police officer. 255
Next, we found a stronger decline over time in SCR in the transparency condition, consistent 256
with less threat. By the end of the interaction, the treatment effect had reversed, yielding a 257
difference in CATEs for the first 10 seconds vs. the recovery period of .21 SD [.08, .36], 258
pr(CATE for first 10 seconds > CATE for recovery) = .99. In summary, the EDA data in Fig. 3A 259

11
and B were consistent with greater initial engagement, and less of a threat response during 260
recovery, in the transparency condition compared to the control condition. 261
262
Figure 3. Differences in the temporal dynamics of the police interaction across experimental conditions
263
depicted as (A) overall skin conductance response (SCR) and (B) conditional average treatment effects on
264
SCR. Note: SCR = Skin conductance response. Results in panels A and B were generated by a multilevel random
265
intercept model that used Bayesian Additive Regression Trees (BART), a machine-learning algorithm with
266
conservative prior distributions, to estimate the non-parametric effect of time on SCR and on the treatment effects.
267
Panel A presents loess-smoothed means of the posterior distribution of SCR reactivity estimated by the BART
268
model. In panel B, dark lines correspond to the median of the posterior distribution of effects, boxes correspond to
269
the inter-quartile range, and whiskers correspond to the 2.5
th
to 97.5
th
intervals. SCR was indexed by the average
270
second derivative of EDA with respect to time during each 10-second window. A positive second derivative signals
271
more rapid increases in EDA, which signals an impending peak and thus an SCR event, while a negative second
272
derivative signals that an SCR event is ending and therefore the participant’s SNS arousal was in the process of
273
recovering to homeostasis. The second derivative was used to follow recommendations from papers validating SCR
274
data which have argued
24,25
that a better way to use information from a continuously monitored SCR is to calculate
275
rates of change in slopes rather than discretely coding SCR events, which discards observations, especially when
276
there is a low sampling rate, as in the present case. In our pilot that validated our SCR facets, the second derivative
277
metric was three times more effective at distinguishing challenge and threat responses to a lab stressor than a
278
frequently-used skin conductance response event coding algorithm (see the SI).
279 280 Controlled online experiments 281
282
What are the elements of a strong transparency statement, when do they work, and when do they 283
not work? To answer these questions, we conducted five pre-registered, randomized, controlled 284
experiments using online samples of adult volunteers who anticipated an interaction with an 285
authority figure. In these experiments, we manipulated aspects of the introductory statements of 286
the authority figures and examined effects on the focal, pre-registered threat measure from 287

12
Experiment 1. Extended details on the samples and study procedures are presented in the SI. An 288
internal meta-analysis of all studies’ results for the main replication cells appears in Fig. 4. 289
290
Experiment 2 (N = 609) was a preliminary study that replicated our primary study’s findings 291
using an anticipated interaction with a police officer stopping to talk with a community member 292
outside of a grocery store. Participants in the control condition imagined that the police officer 293
started a conversation by saying “Hi, how are you doing?” In the transparency condition, they 294
anticipated interacting with an officer who explained that they were “just trying to get to know 295
the community better.” As with the live interaction, the transparency statement reduced threat in 296
this anticipated situation, ATE = -0.44 SD [-0.67, -0.22], pr(ATE<0) = .99. In the four 297
experiments that follow, the same transparency effect was replicated each time (see the SI). 298
299
Having validated our anticipated interaction method in this randomized design, we asked: What 300
is it about the statement of the officer’s intentions that had a beneficial effect? In Experiment 3 301
(N = 382), we used a three-condition design to compare the transparency statement and the 302
control condition to a new introduction that was more closely aligned with the aims of the 303
aggressive policing approach (i.e. the officer is “trying to find a suspect in the area”). We 304
suspected that the latter introductory statement would not be as effective because it does not 305
disabuse the participant of the significant concern that they will be accused of a crime. We found 306
that the aggressive policing introduction substantially increased threat compared to a control 307
condition, ATE = 0.39 SD [0.25, 0.53], pr(ATE>0) > .99, which is consistent with qualitative 308
accounts of the investigatory stop. This increase in threat was highly different from the replicated 309
reduction in threat from the transparency statement, Difference in ATEs = -0.46 SD [-0.61, -310
0.33], pr(ATEambiguous> ATEtransparency) > .99. Thus, the transparency statement that explicitly 311
conveyed benevolent intent was more effective. 312
313
What if an introductory statement communicated benevolent intent, but did so in a way that still 314
left open the possibility that this was an “investigatory stop?” In Experiment 4 (N = 450), we 315
used a three-condition design to compare an ambiguously positive introduction (i.e., the officer 316
“just wanted to stop and say hi and see how you are doing”) to the transparency statement and to 317
a control condition. The ambiguously positive introduction did not meaningfully reduce threat 318
compared to the control condition, ATE = 0.08 SD [-0.03, 0.21], pr(ATE<0) = .18, and that null 319
effect was much smaller than the threat reduction effect obtained from the transparency 320
statement, Difference in ATEs = -0.39 SD [-0.53, -0.25], pr(ATEambiguous>ATEtransparency) > .99. A 321
follow-up analysis suggested that this was because participants thought the officer in the 322
ambiguously positive condition was concealing their true intentions (see the SI). Thus, 323
transparency statements seem to reduce threat not simply because they are friendly, positive, or 324
respectful, but because they more clearly take off the table the possibility that individuals are 325
going to be accused of a crime by law enforcement. 326
327

13
328 Figure 4. Forest plot for the internal meta-analysis on the key effect of control (i.e., aggressive policing) vs. 329 transparency statement condition on civilians’ ratings of threat of enforcement for Experiments 1-6. All 330
analyses come from frequentist statistical methods. N’s reflect sample size for key transparent vs. ambiguous 331
condition comparison, rather than the total N for each study. Experiment 1: Control N = 130, Transparency N = 102; 332
Experiment 2: Control N = 162, Transparency N = 143; Experiment 3: Control N = 127, Transparency N = 126; 333
Experiment 4: Control N = 150, Transparency N = 150; Experiment 5: Control N = 88, Transparency N = 86; 334
Experiment 6: Control N = 113, Transparency N = 113. Experiments 3-6 only include ambiguous control vs. 335
transparency statement conditions (i.e., Experiment 3 does not include ambiguous transparency condition; 336
Experiment 4 does not include ambiguous transparency condition; Experiment 5 does not include grocery worker 337
conditions; Experiment 6 does not include park ranger conditions). We computed Cohen’s d and the variance of d 338
and then conducted a meta-analysis with the R package metafor using a random-effects approach 15. RE = random 339
effect. Squares show effect size estimates (Cohen’s d’s). The size of each square gives a representation of each 340
study’s sample size. Error bars show 95% confidence intervals (CIs). The diamond represents the point estimate and 341
95% CI averaged across studies. 342
343
Two final experiments used a different approach to test our assumption that transparency 344
statements are beneficial primarily because they make it clear to the community member that 345
they are not under threat of arrest or bodily harm while they are being questioned. We did this by 346
varying the social role of the interaction partner. In Experiment 5 (N = 349), again a three-cell 347
study, participants anticipated interacting with someone who had no authority to arrest or harm 348
them: a grocery store worker. In that interaction, there was no reduction in threat in the 349
transparency condition relative to the control condition, ATE = -0.03 SD [-0.21, 0.14], 350
RE Model
0 0.2 0.6 1 1.2
Observed Outcome
Experiment 6 (n = 226)
Experiment 5 (n = 174)
Experiment 4 (n = 300)
Experiment 3 (n = 253)
Experiment 2 (n = 305)
Experiment 1 (n = 232)
0.90 [0.63, 1.18]
0.51 [0.21, 0.81]
0.37 [0.14, 0.59]
0.92 [0.66, 1.18]
0.50 [0.27, 0.73]
0.36 [0.10, 0.62]
0.59 [0.39, 0.79]

14
pr(ATE<0) = .59, and that null transparency effect with the grocery store worker was much 351
smaller than an effect obtained from an anticipated police officer interaction, Difference in ATEs 352
= -0.38 SD [-0.65, -0.11], pr(ATEgrocery>ATEpolice) = .97. 353
354
In Experiment 6 (N = 451), the final three-cell study, participants anticipated interacting with 355
someone who technically did have authority to arrest or harm the participant, but who was very 356
unlikely to do so because it is not common practice in their role: a park ranger. When 357
participants anticipated interacting with a park ranger, the transparency statement did reduce 358
threat relative to the control condition, ATE = -0.46 SD [-0.62, -0.31], pr(ATE>0) > .99, but this 359
transparency effect was approximately half the size of the effect obtained when participants 360
anticipated interacting with a police officer, Difference in ATEs = -0.37 SD [-0.61, -0.16], 361
pr(ATEpark ranger>ATEpolice) > .99. 362
363
In summary, transparency statements work best—and are most needed—when they shift the 364
default construal of an interaction, from one that could end in harm or arrest, to one that could 365
end in a positive relationship with a powerful authority figure. We note that the goal of these 366
online studies was to establish internal validity with respect to transparency statements. We do 367
not make claims of generalizability based on these samples. 368
369 Discussion 370 371
This research identified a simple and relatively costless method for making community-police 372
interactions less threatening: a transparency statement. In doing so it highlighted the under-373
appreciated role of timing in conversations between institutional authorities (i.e. law enforcement 374
officers) and civilians. Notably, our observations and interviews 13 found that traditional patrol 375
officers made transparency statements only after trust had been compromised. Consistent with 376
this qualitative observation, our study showed that absent a transparency statement at the start of 377
the conversation, the officers’ questions contributed to a sense of threat that was apparent in an 378
avoidant conversational style (i.e. less authentic, more terse) and physiological profile (lower 379
initial SNS activation). When officers made transparency statements in the first two sentences of 380
the conversation, however, it changed the tenor of the ensuing interaction. For example, twice as 381
many community members reported feeling inspired by the end of the interaction when officers 382
made a transparency statement (28%) than when they did not (14%; see SI and Figure 1C). This 383
shift was apparent in transcripts, self-reports, and stress physiology. 384
385
The present studies raise the intriguing possibility that community policing, in general, might be 386
made more effective if officers could be trained to incorporate transparency statements into their 387
routines. Indeed, in Experiment 1, we individually instructed seven officers in less than five 388
minutes per officer—instruction that might be done efficiently with entire departments. Future 389
studies could involve larger, longitudinal evaluations that assess downstream consequences of 390

15
transparency trainings on outcomes such as wrongful arrests or official excessive use of force 391
complaints, along with greater trust in the benevolence of police officers, more information 392
sharing, more solved cases, and safer communities. 393
394
A key consideration for future research is whether transparency statements could play a role in 395
reducing racial and socioeconomic disparities in policing policies 4,6,26,27. Does a history of 396
aggressive policing in communities with higher proportions of people of color 28 cause citizens to 397
profit more from greater transparency 29? Or does it make people immune to transparency’s 398
effects? New experiments can test this directly. 399
400
These findings also have implications for building trust in other organizational relationships with 401
significant power disparities (e.g., school resource officers, border patrol agents, teachers). We 402
speculate that transparency statements will be especially needed when authority figures can harm 403
those in lower power positions, as is the case for police, principals, teachers and others 30. In 404
these settings, transparency statements might assure students of their benevolent intent, and 405
might prevent escalating negative reactions from children to these authority figures. 406
407
Finally, we note that extant criminal justice reform efforts often seek to change officer behavior 408
via broad and abstract cultural changes 31,32. Police departments often resist change, however. 409
Officers remain skeptical of the effectiveness of community policing, or any other reform, as 410
they retain traditionally aggressive tactics. Our results suggest that reforms that are nearer to 411
officer behaviors—such as interactions with community members—represent an alternative but 412
nevertheless viable pathway to police reform that might be both effective and scalable. 413
414 Methods 415 416 Phase 1: Hypothesis Generation via Qualitative Field Observations 417 418
In the first phase, we used qualitative methods to understand how community policing 419
efforts might be improved and to generate the hypotheses about transparency statements that we 420
tested in Phases 2-3. Note that we cannot make strong claims from our qualitative methods about 421
fully capturing the dynamics of transparency statements in community policing units. Instead, 422
here we are simply presenting the systematic approach we took that enabled us to capture 423
ecologically-valid police-community interactions in the field, as recommended by criminology 424
researchers49. These field observations ultimately enabled us to develop a falsifiable hypothesis 425
about the potential positive impact of transparency statements in community policing. 426 427 Research settings. There were two concurrent strands to the field observations we 428
conducted. We began, first, as part of a larger team of researchers observing a new community-429
oriented policing unit that was developed to address the community’s problems by quickly 430
building relationships with community members. In parallel, we observed traditional patrol 431
officers in a different department in the same metropolitan area. We focus here on the most 432
relevant findings about the distinct dyadic interaction tactics law enforcement officers used and 433

16
their relative effectiveness regarding building trust and rapport with their communities, because 434
these were the observations that led us to develop the present hypotheses. Many other 435
observations were conducted, regarding the governance and decision-making of the two police 436
departments, but these data are less relevant to the present transparency hypotheses. 437
438
Community policing unit. The first team we observed was a police department in a 439
large, urban city, and the specific unit was located in a district that included portions or entireties 440
of eight neighborhoods with a mixed population of diverse racial and socioeconomic groups 441
(total population range = 14,318 – 93,727; 6-62% White; 15-83% Hispanic/Latinx; 2-78% 442
Black/African-American; 1-6% Asian/Asian-American; 1-3% Other/Multiple Races; Median 443
Age = 32.2 – 41.9 years old; 13-57% population 25+ with a 4-year college degree or greater; 61-444
94% Native US; 3.5-12.6% unemployed; median household income = $33,515 – $82,908). This 445
city had a population of more than 2 million residents, and the police department had more than 446
10,000 officers at the time of our data collection. The unit we focused on was testing a pilot 447
program that reorganized the duties of their officers. While officers traditionally go from call-to-448
call according to their radios, this unit was given discretion over how they used their time, with 449
the goal to spend it oriented towards building relationships with the community in their assigned 450
area, and coordinating resources that their communities needed. Officers in this unit were 451
recruited from the entire citywide department by the lieutenant in charge of the unit, ostensibly 452
selected on traits that would facilitate relationship-building in communities (e.g., genuine interest 453
in relationship-building, prior history of attempting to build relationships within the constraints 454
of a patrol officer’s role). This phase provided us with an understanding of how officers who 455
were ostensibly skilled at building relationships with community members attempted to quickly 456
build trust with civilians. Examples of this process appear in Box 1, right panel, Exs 1-2. 457
458
Traditional patrol unit. In parallel, we partnered with another police department in the 459
same metropolitan area to observe more traditional interactions of police whose responsibilities 460
are dictated by radio calls and less focused on developing skills related to connecting with their 461
communities. The employees we observed were patrol officers from a different police 462
department in a moderately-sized, suburban city that was nearby to the city where the 463
community policing unit was located. This city was also relatively racially and 464
socioeconomically diverse (58% White; 12% Hispanic/Latinx; 17% Black/African-American; 465
9% Asian/Asian-American; 7% Other/Multiple Races; Median Age = 36.2 years old; 67% 466
population 25+ with a 4-year college degree or greater; 82% Native US; 12% in poverty; median 467
household income = $82,335). This city had a population of approximately 75,000 residents, and 468
the police department had approximately 150 officers at the time of our data collection. After 469
discussions with department leadership, we began a four-week long period of four-hour ride-470
alongs approximately three to four times per week. Per their supervisors’ instruction, these 471
officers engaged with members of their community as they naturally would if we were not 472
observing them. We observed these interactions and took observational notes on each interaction 473
we observed, and we also obtained direct reflections from both officers and the community 474
members with whom they interacted. We obtained both structured and unstructured reflections 475
from officers after the interaction with the community member ended. First, officers reflected 476
alone via structured question prompts (e.g., “do you think the civilian understood your intentions 477
for talking with them?”; “how connected to the ‘real you’ were you in the interaction?”). Then, 478
we discussed officers’ experiences utilizing unstructured questions. Simultaneously, while the 479

17
officer left the area to reflect alone, we approached the community members they talked to for an 480
optional, consented debriefing discussion of their interaction utilizing a mix of structured and 481
unstructured questions. This phase provided us with an understanding of how officers who were 482
not trained in the community-based policing attempted to quickly build trust with civilians. An 483
example of this process appears in Box 1, left panel, Ex 2. 484
485
Participants. Participants for both phases of the qualitative field observations were the 486
police officers in the two departments we observed (70% Male; 47% White; 30% 487
Hispanic/Latinx; 20% Black/African-American; 3% Asian/Asian-American; approximate age 488
range: 20-60 years old). In total, more than 35 police were observed across three years, which 489
included interviews, ride-alongs, observed unit meetings, and field observations of community 490
events, resulting in more than 500 hours of data observations. In the community policing unit, we 491
observed 17 officers, four sergeants, one lieutenant, and one commander. Throughout the three 492
years of our observation, more than five officers left the unit due to injuries or to pursue other 493
opportunities in the department. In the traditional patrol unit, we observed 18 patrol officers and 494
approximately 148 civilian interactions. 495
496
Data and qualitative analysis. Data from both phases come largely from field 497
observations of ride-alongs with the police participants. For the ride-alongs, police were asked to 498
act as they naturally would while we observed from a distance. Field observations of community 499
events, biweekly debriefing meetings, and hour-long interviews were only collected from the 500
community policing unit. Biweekly department meetings in this phase typically consisted of 501
strategic discussions about how to approach community problems together in a roundtable with 502
the whole unit. Interviews occurred in four rounds and consisted primarily of officer views of 503
their role and their community while in the pilot program. Interviews occurred at the recruiting, 504
beginning, middle, and end stages of the officers’ involvement in the pilot program. As 505
mentioned, for the traditional patrol unit, observations of the ride-alongs were accompanied by 506
both officers’ structured and unstructured reflections alone and with the researcher(s) 507
accompanying them on the ride-along. For all field notes, shorthand notes were recorded in small 508
notebooks or on smartphones and expanded after the conclusion of the observation, as soon as 509
possible. Recorded conversations were transcribed by a professional transcription service. 510
511
Field notes and interview transcripts were analyzed by the two lead researchers by 512
discussing the day’s data collection immediately or soon after it finished, and writing informal, 513
conceptual, theoretical memos to summarize the emerging themes. Transcripts and notes that 514
were flagged as indicative of key themes were then analyzed line-by-line to understand recurring 515
patterns and to check assumptions against the data. Once the key themes were identified and 516
refined (presented in Box 1), then exemplary quotations were selected to illustrate those themes 517
to readers. These quotations were then vetted with a third researcher, to reach agreement on the 518
fit between the data and our interpretations of the data. After reducing the qualitative data to the 519
key themes and focal examples (Box 1), we used these insights to design the randomized 520
experiments and test the hypotheses directly. Additional thematic examples appear in the SI. 521 522 523 524 525

18
Phase 2: Hypothesis Testing in Behavioral Experiment 1 526 527 Participants. The study was pre-registered on the Open Science Framework (OSF; 528
tiny.cc/PreRegExp1). We planned to collect data from a total of 200 participants, which afforded 529
the ability to detect an effect of 0.4 SD at 80 percent power using conventional null hypothesis 530
rejection analyses. Following the pre-registered stopping rule, data collection was terminated on 531
the first day that we met or exceeded our sample size. No data were examined and no hypotheses 532
were tested prior to stopping data collection. 533
We ultimately collected data from 239 U.S. residents from a sample of community 534
members in a large urban city from streets nearby, Craig’s list ads, emails throughout 535
departments to the local university, and word of mouth. After excluding seven individuals who 536
either opted to have their data deleted after learning the full details of the study (n = 3) or had 537
incomplete data (n = 4), we were left with a final sample of 232 (Mage = 22.07, SD = 5.90; 54 538
percent female, 42% male, 4% non-binary/genderqueer/other; 34% Whites/European Americans, 539
32% Asian/Asian Americans, 17% Latino/a/Hispanic Americans, 9% Black/African Americans, 540
5% Multiethnic, 3% Biracial). 541 542 Material development. The study design was created with informal police consultants, 543
who were current and retired police officers from three departments. We developed our 544
procedure to realistically reflect what police do in officer-initiated interactions with their 545
communities, while maintaining alignment with our definition of transparency (i.e. clarifying 546
intentions for approaching civilians). 547 548 Civilian participants. This study was conducted during June and July of 2021. Civilian 549
(i.e., community member) participants were told that the study was about the natural interactions 550
people had with others. Civilian participants were seated individually at a table around a 551
university building in one of five predetermined locations that were not in-view of each other 552
(e.g., a patio table; a bench). Civilian participants were told that they should do what they would 553
naturally do if they were sitting by themselves outside (e.g., using their cell phones or 554
computers). The research assistant left the participants alone with an iPad and stated that 555
everything would be audio-recorded by the iPad during the experiment. Participants did not 556
know who they would be interacting with but were told to act naturally if (and when) anyone 557
approached them. 558
To mask the anticipation of a police officer approaching while also allowing for informed 559
consent, participants were told in their consent form that due to the outdoor nature of the 560
research study, they may be approached by a research assistant, unsheltered (i.e., homeless) 561
person, student, police officer, or others in the area. Once the conversation started, the officers 562
never stated that they were a part of the research study. 563 564 Police officer partners. The police officer partners (N = 7) for this field experiment were 565
employees of the police department with jurisdiction over the community of recruited civilian 566
participants. That is, actors were not used, and all officers in the study were on-duty employees 567
during their regularly scheduled work shifts, not trained confederates; this decision was made to 568
increase the ecological validity of the experiment. 569
Prior to conducting this research, the research team met with police department 570
leadership, who approved the partnership. The seven police officers were recruited by the 571

19
leadership because they were considered by leaders to be representative of the diversity of 572
officers using community policing efforts in this police department. Training of the officers 573
lasted approximately five minutes per officer. The officers overall spent approximately 60 hours 574
on data collection in the role of an experimenter. 575
Police partners were told to initiate interactions with civilians in one of two ways: either 576
providing a transparency statement (i.e., giving explicit reasons for why they are approaching the 577
participant, at the outset of the interaction) or not (our control condition; i.e., using the 578
ambiguous, direct, and short language that simply asks if they can talk to the civilian 579
participants). Police partners were instructed to keep their conversations short (i.e., to five 580
minutes or less, aiming for 2 minutes), but were otherwise left to their own natural conversation 581
style after the introduction. They did not follow an exact script. Notably transcripts confirmed 582
that 100% of the officers in the transparency condition made a transparency statement, and 0% of 583
the officers in the control condition did so at the outset of the conversation. 584
For example, in the transparency condition (N = 100), officers initiated interactions in the 585
following ways: e.g., “Hello, I’m just out and about walking around, talking to people in the 586
community. Is it okay if I talk to you for a minute?” “Hey, my name is [Name], and I’m just 587
taking a walk, trying to get to know my community better. You mind if I sit and talk to you for a 588
second?” 589
In contrast, in the control condition (N = 132), officers initiated interactions in the 590
following ways: e.g., “Hey, can I talk to you for a minute?” “Hey, I’m Officer [Name], can I talk 591
to you for a second?” 592
593 Assignment to condition. Leadership in the police department assigned officers to one 594
of several shifts for the experiment. Researchers, blind to the identities or skills of the officers, 595
assigned officers’ shifts to experimental conditions, so that in some shifts officers used 596
transparency statements, and during other shifts they did not. Civilians, in turn, were assigned to 597
different shifts independently of the officers’ condition assignments. There were an equal 598
number of weekend shifts across conditions. Within experimental conditions, shift was not 599
meaningfully associated with the pre-registered outcomes of threat and trust (Threat: Control 600
group intraclass correlation [ICC] = .04; Transparency group ICC = .02; Trust: Control group 601
ICC = 0; Transparency group ICC = 0). As a result, supplementary, non-pre-registered mixed-602
effects analyses that clustered standard errors by shift did not produce different results (Threat: b 603
= -0.47, t = -2.09, p = .037; Trust: b = 0.37, t = 2.23, p = .025). Furthermore, no participant 604
demographics (i.e., age, gender, race, native U.S. status, education level) differed significantly 605
by condition, t’s ≤ 1.106, p’s ≥ .27 (see Table S2 in SI) or weekday (see Table S3 in SI). 606 607 Measures 608
Unless otherwise acknowledged, all measures were on a seven-point Likert scale, ranging 609
from 1=Strongly Disagree to 7=Strongly Agree. For all measures, civilian participants read that 610
their individual data would not be shared with the police (i.e., that data collection was 611
confidential). Full lists of the scale items and additional information can be found in the SI. See 612
Table S4 in the SI for correlations among all primary variables. 613
Manipulation check: Perceptions of transparency of intentions. As a manipulation 614
check, we measured civilian participants’ perceived clarity of the police partners’ intentions (i.e., 615
“The officer I talked to stated their intentions for talking to me clearly”; “The officer I talked to 616

20
stated their intentions for talking to me immediately”; “I knew why the officer was talking to me 617
from start to finish”; α = .757). 618
Primary measure 1: Threat of enforcement. Civilians’ self-reports of threat were 619
measured using three items that operationalized threat as the expectation of harm in the 620
interaction—i.e., the inverse of trust. Specifically, these questions asked how likely (1=Very 621 unlikely to 7=Very likely) it was that the police officer they talked to “just wanted to get to know 622
you,” (reversed) “just wanted to make friendly conversation,” (reversed) or “did not trust you” α 623
= .768. 624
Primary measure 2: Trust in benevolent intent. Trust is viewed as the first step 625
towards repairing relationships between law enforcement and civilians in the community 626
policing model 12. Civilian participants’ trust was measured using three items adapted from the 627
benevolence subscale of a validated trust scale 33, using the same likelihood scale as threat, 628
regarding future interactions with the police officer they were approached by: “This police 629
officer would care about you and your welfare;” “this police officer would go out of their way to 630
help you; this police officer would not do anything to hurt you.” α = .630). 631
Natural Language Analysis 632
Audio recordings of conversations. Audio was recorded by an iPad device positioned 633
on the table where the civilian participant was seated. Audio began recording when the research 634
assistant set the participant up alone during the pre-interaction waiting period through when the 635
interaction ended, when the research assistant returned to turn the recording device off. All audio 636
recordings were later transcribed and turned into text files by research assistants. 637
Natural language processing: Conversation transcripts. Text analyses of transcripts of 638
the conversations between civilians and officers were divided by speaker role (i.e., officer vs. 639
civilian), condition (i.e., transparency vs. control), and by portion of the conversation (i.e., first, 640
middle, or last third). Transcripts were analyzed with the Linguistic Inquiry and Word Count 641
(LIWC) Software 19. 642
LIWC is one of the most dominant text analysis software in the social sciences, and relies 643
on a “word counting” method that both tallies counts of parts of speech like prepositions, 644
adverbs, and pronouns 34, and also includes four summary dimensions validated in previous 645
research: Analytic 35, Clout 36, Authenticity 37, and Emotional Tone 38. We analyzed the end of 646
the civilian side conversation transcripts and obtained their score on each of these summary 647
dimensions. We expected that the transparency statement would increase civilians’ use of 648
authentic language, because the transparency statement puts civilians at ease and enables rapport 649
to be built. We did not anticipate differences on any of the other dimensions, but analyzed them 650
along all summary dimensions to document the specificity of the effect of transparency 651
statements on authentic language. 652
Natural language analysis: Civilian reflections on the interaction. On the post-653
interaction survey, participants answered an open-ended question asking about the thoughts (one 654
question) and feelings (a second question) they experienced at the end of the interaction. These 655

21
reflections were also scored with the LIWC-22 algorithm. Analyses focused on two patterns that 656
emerged from the idiographic analysis. The first is causal language, or signaling people’s 657
tendency to ponder the cause of the officer’s interactions (i.e. “I was wondering why he was 658
talking to me”). We expected the transparency statement to reduce causal language in the 659
reflections, because the primary goal if the transparency statement is to answer the question of 660
why the officer is talking to them. The second was positive emotion (e.g. comfortable, happy). 661
We expected the transparency statement to increase positive emotion at the end of the 662
interaction, because the statement should provide the foundation for rapport that could result in a 663
positive back-and-forth dialogue. 664
Emotion at the end of the interaction. On the post-interaction survey, civilian 665
participants completed the Short Form version of the Positive and Negative Affect Schedule 666
(PANAS-SF39), selecting an emotion they felt at the end of the interaction. In the standard 667
PANAS-SF measure, there are several emotion categories that correspond to operationalizations 668
of threat (vs. challenge) appraisal 40,41. Emotions were coded for threat (1 = alert, afraid, nervous, 669
upset, hostile, or ashamed; 0 = not). In addition, one emotion (inspired) was explored, due to its 670
alignment with community policing’s goals. 671
Skin conductance responses (SCR). To distinguish between challenge (approach) and 672
threat (avoidance) stress responses, skin conductance responses (SCR), were obtained from 673
electrodermal activity (EDA) sensors embedded in a wrist-worn ambulatory device: the 674
Empatica E4. The Empatica E4 wristband measures physiological responses to stress 675
simultaneously via two sensors placed on the bottom of the wrist 42,43. The E4 sensors have a 676
resolution of 1 digit ~900 pico Siemens, a range from 0.01microSiemens to 100 microSiemens 677
and they apply alternating current (8Hz frequency) with a maximum peak to peak value of 678
100microAmps (at 100microSiemens). 679
EDA assesses changes in the amount of conductance of the skin (i.e. SCR) due to sweat 680
produced by eccrine glands; increased SCRs correspond to increased sympathetic nervous 681
system (SNS) activation innervated by acetylcholine23. 682
We instructed participants to push a button our ambulatory devices whenever someone 683
started to talk to the participant. Pushing this button recorded a timestamp in the ambulatory 684
device data. Because of this participant-led action, some participants forgot to push the button to 685
indicate the interaction started. This is a potential limitation because it caused missing 686
psychophysiological data (the SCR subsample was N = 177). However, this missing data was 687
unavoidable to retain the realism of the unexpected officer-initiated interactions. 688
SCR theoretical predictions. According to the BPS model, SNS arousal during a 689
stressful interaction can follow one of two patterns: a more positive, challenge-type response or a 690
more negative, threat-type response. A challenge-type response corresponds to approach 691
motivation and evokes an initial orienting response with the onset of the stressor (e.g. the 692
approach of the officer), which involves a rapid increase in SNS activity20-22. But a challenge-693
type response would involve a rapid return to homeostasis as the interaction is experienced as not 694
a potential threat23-25. In contrast, more negative, avoidance-oriented, threat-type responses often 695

22
involve blunted initial SNS reactivity, as individuals seek to avoid potentially harmful threats (as 696
is evident from the transcript of the control group participant in Fig. S2A in SI). When 697
individuals cannot escape the stressor, though, those experiencing threat-type responses remain 698
suspicious, vigilant, and alert to impending harm, thus leading their SNS activation to increase 699
over time23-25. 700
As shown in the SI, we validated these expectations with respect to the specific SCR 701
facets extracted from the ambulatory EDA data by using a pilot laboratory study (N = 100) that 702
involved an acutely stressful interaction (the Trier Social Stress Test; TSST48) and gold-standard 703
measures of challenge and threat-type responses (Total Peripheral Resistance, TPR20. In the pilot 704
study shown in Fig. S2A-B in the SI, participants showing challenge-type responses (low TPR) 705
showed an initial increase in SCR at the start of the TSST, as expected, and a sharper decline 706
during recovery, compared to those with threat-type responses (high TPR). 707
EDA processing. To calculate skin conductance responses (SCRs), the EDA data were 708
extracted from the devices after each session and processed in Python using established 709
algorithms for processing EDA data and removing artifacts. To check EDA quality, the 710
algorithm first marked the lost signal and calculated the ratio of lost versus overall signal. The 711
signal was considered lost if the value was below 0.001µS. If the ratio was greater than 0.9 for 712
each five-second window, the algorithm classified the window as of bad quality 44. The 713
algorithm then checked any anomalies in EDA data in each second window. The data were 714
considered anomalies if the maximum increase of a signal value was greater than 20% and the 715
maximum decrease was greater than 10% 45. The segment was marked as of bad quality if this 716
condition was not met in each second window. Finally, the algorithm calculated signal power in 717
second difference from the SC signal in a window of five minutes with four minutes overlap 46. 718
The algorithm set the four-minute overlap to achieve a resolution of smoothed processed data per 719
minute 47. 720
Second derivative of EDA. The SCR metric of interest was calculated as the average 721
second derivative of EDA trends over 10-second intervals. Here is the reason why. When the 722
skin conductance response is moving toward a “spike” then that is an indication of greater 723
engagement with a stressor, and that is therefore the primary physiological response that was 724
relevant to the BPS model predictions. The way to assess whether, in a particular moment, a 725
person is approaching a spike is to assess the second derivative of the time function. The first 726
derivative, of course, is equal to the slope—whether SCR is increasing, and to what extent—but 727
this does not indicate whether a spike is approaching because it does not reveal anything about 728
the rate of change in the slope, and therefore it cannot indicate peaks or troughs in the data. This 729
is what the second derivative provides: it is the change in the slope. When the second derivative 730
is positive, it means that the slope is increasing—for example, that it is the concave part of the 731
time function. When the second derivative is negative, it means that the slope is decreasing—i.e., 732
that it is in the convex part of the time function. Note that a negative second derivative does not 733
mean that SCR is no longer increasing—only that the rate of change is slowing. Likewise, a 734
positive second derivative means that the slope for SCR is increasing—even if it is still in the 735
“downhill” part of the time function. The primary advantage of the second derivative is that it 736

23
can be a leading indicator of a coming spike in the SCR, and therefore it can make fuller use of 737
the data (as compared to waiting for the spike to appear in the data). For example, a 20-second 738
interval of data sampled might have one spike but 20 second derivatives. 739
Exclusion of PPG data. The E4 wristband also has photoplethysmography (PPG) 740
sensors, but the PPG data were not used in the present paper. The PPG sensor shines light on the 741
skin to detect the degree to which light refracts. Fluctuations in blood flow are estimated based 742
on assumptions about blood flow, using the default Empatica software algorithms. Heart rate 743
(HR) and interbeat intervals (IBI) are received as output along a time-series, with estimations at 744
approximately 64 Hz. However, when validating the measurement of the PPG sensors using the 745
gold standard for cardiovascular activity (i.e., electrocardiogram: ECG), our pilot study showed 746
that the E4’s IBI algorithm was only able to extract 33.7% through the TSST (Trier Social Stress 747
Test26) and 17.8% in readings from a controlled lab setting in schools. To our knowledge, heart 748
rate measurement of E4s have not been validated in naturalistic settings like ours in a published 749
paper. In naturalistic, outdoor settings like ours in Experiment 1, there is likely light leak and 750
movement artifacts that created error in measurement that accounts for differences in 751
measurement accuracy recorded in traditional lab settings that have had better success with PPG 752
sensors 42,43. Data from our pilot experiment left us less confident in any measurements of heart 753
rate. However, data from our pilot experiment (see below) gave us confidence in using E4 EDA 754
measurements of skin conductance. 755
Phase 3: Hypothesis Confirmation with Online Replication Experiments 756
Experiment 2 757
Participants read a vignette about a police officer stopping to talk to them after they left a 758
grocery store. A grocery store was chosen because it mirrors the kind of mundane situation in 759
which civilians are approached by officers engaging in community policing. Also, a grocery 760
store is a place where an officer could plausibly be investigating a crime, and so it was a 761
situation in which the officer’s intent could be ambiguous to the participant. 762
Participants. We preregistered our experiment on OSF (tiny.cc/PreRegExp2). We aimed 763
for a sample size of approximately 300 per condition and recruited a sample of 649 U.S. 764
residents from Amazon’s MTurk, of these 609 completed our dependent measures. We made no 765
exclusions (Mage = 37.00, SD = 12.28; 52% female, 1% non-binary/other; 72% Whites/European 766
Americans; 8% Asian/Asian Americans; 6% Latino/Hispanic Americans; 11% Black/African 767
Americans; 1% Native American, 1% Other). Analyses were powered to detect an effect of 0.29 768 SD at 80% power. 769
Design and procedure. Participants were randomly assigned to one of two conditions: 770
ambiguous control or transparency. Participants were asked to imagine walking in and out of a 771
grocery store. For participants in the ambiguous control condition, the police officer simply 772
started a conversation by saying “Hi, how are you doing?” In the neutral transparency condition, 773
the officer explained that they were trying to get to know the community better (see below). 774
1. You leave the store. 775

24
2. On your way out, you notice a police officer's car pull next to where you were walking. 776
3. The police officer walks out of the car and looks your way. 777
4. The police officer walks towards you and starts a conversation, saying, 778
"Hi, how are you doing?" [Control Condition] 779
780
“Hi, how are you doing? My boss is asking me to talk to people to get to know the 781
community better.” [Neutral Transparency Condition] 782
Threat and trust. Participants were assigned to either answer items about perceived 783
threat (e.g., How likely is it that… “The police officer is going to accuse you of something.”, 784
“The police officer does not trust you.”; α = .85) or perceptions of trust from the benevolence 785
subscale of a trustworthiness scale (i.e., How likely is it that this officer… “...is very concerned 786
about my welfare.” “...would not knowingly do anything to hurt me.” “...will go out of his/her 787
way to help me.” “...has a strong sense of justice. “...tries hard to be fair in dealing with others.” 1 788
α = .84). See Table S8A-B in the SI for correlations among all primary variables. 789
Experiment 3 790
Participants. We preregistered our study on OSF (tiny.cc/PreRegExp3). Given the effect 791
size obtained in Experiment 2 (d = 0.50), we sought to obtain a sample large enough to detect 792
this size effect, but across three conditions. Thus, we recruited a sample of 382 U.S. residents 793
from Amazon’s MTurk. We made no exclusions (Mage = 38.52, SD = 12.54; 51% female, 0.3% 794
non-binary/other; 77% Whites/European Americans; 11% Asian/Asian Americans; 8% 795
Latino/Hispanic Americans; 10% Black/African Americans; 2% Native American, 1% 796
Hawaiian/Pacific Islander, 2% Other). Analyses were powered to detect an effect of 0.32 SD at 797
80% power. 798
Design and procedure. Participants were randomly assigned to one of three conditions: 799
ambiguous control, community-oriented transparency, or aggressive policing. All participants 800
read a vignette about a police officer stopping to talk to them after they left a grocery store. The 801
ambiguous control and community-oriented transparency were identical to the two conditions of 802
Experiment 2, while in the aggressive policing condition the officer explained that they were 803
trying to find a suspect (see below). After reading this vignette, participants answered four items 804
about the amount of threat they perceived from the police officer as in Experiment 2. 805
1. You leave the store. 806
2. On your way out, you notice a police officer's car pull next to where you were walking. 807
3. The police officer walks out of the car and looks your way. 808
4. The police officer walks towards you and starts a conversation, saying, 809
"Hi, how are you doing?" [Control Condition] 810
811

25
“Hi, how are you doing? My boss is asking me to talk to people to get to know the 812
community better.” [Neutral Transparency Condition] 813
“Hi, how are you doing? "Hi, how are you doing? My boss is asking me to get to know 814
the community better to find a suspect we're looking for.” [Ambivalent Transparency 815
Condition] 816
Threat. We measured threat using the same scale as in Experiment 2. See Table S9 in the 817
SI for correlations among all primary variables. 818
Experiment 4 819
Participants. We preregistered our study on OSF (tiny.cc/PreRegExp4). We sought to 820
recruit a similar size sample to that of Experiment 3, so we recruited a sample of 450 U.S. 821
residents from Prolific Academic. We made no exclusions (Mage = 31.59, SD = 11.05; 50% 822
female, 2% non-binary/genderqueer/other; 60% Whites/European Americans; 17% Asian/Asian 823
Americans; 8% Latino/Hispanic Americans; 10% Black/African Americans; 0.2% Native 824
American, 2% Biracial, 3% Multiethnic). Analyses were powered to detect an effect of 0.29 SD 825
at 80% power. 826
Design and procedure. In Experiment 4, we asked participants to read a scenario where 827
an officer walked up to them while they were wearing masks outside in a park with their friends.1 828
Participants were randomly assigned to one of three conditions: control, transparent, or 829
ambiguously positive. The officer either asked, “Hi, how are you doing?” (control condition), 830
gave a transparent and concrete reason (“I’m trying to get out of my car every day to talk to 831
people and get to know the community better”), or an ambiguously positive reason (“Just wanted 832
to stop and say hi and see how you are doing”). 833
Imagine the following happened to you... 834
1. You walk into a park with one of your friends. You are both wearing your masks. 835
2. You sit down in the grass and talk with your friend for about 30 minutes. 836
3. You notice a police officer park their car nearby and see them get out of their car. 837
4. The police officer walks around for about 10 minutes, eventually looks your way, and 838
walks up to you. 839
5. The police officer starts a conversation, saying, 840
841
“Hi, how are you doing?” [Control Condition] 842
“Hi, how are you doing? Just wanted to stop and say hi and see how you’re doing.” 843
[Ambiguously Positive Condition] 844
1 In Experiment 4, we replicated our effect in the midst of social distancing mandates during the pandemic and
unrest from protests against the police after the murders of George Floyd and Breonna Taylor, in 2020.

26
“Hi, how are you doing? I'm trying to get out of my car every day to talk to people and 845
get to know the community better.” [Concrete Transparency Condition] 846
Threat. We measured threat using the same scale as in Experiments 2-3. 847
Perceived authenticity. In addition to the threat items measured in previous 848
experiments, we added a measure of perceived authenticity (“How [ ‘authentic,’ ‘sincere,’ ‘fake’ 849
(reversed)] do you think this office is being with you?”; α = .935). 850
Additional demographics. Additional control variables for education, political ideology, 851
social connection to police, and being a US native. In our sample, 39% of participants reported 852
knowing at least one person who was or had previously been employed in the field of law 853
enforcement. Our sample included 20% Republicans, 47% Democrats, 24% Independents, 3% 854
Other, and 7% No Preference. See Table S10 in the SI for correlations among all primary 855
variables. 856
Experiment 5 857
Participants. We preregistered our study on OSF (tiny.cc/PreRegExp5). We sought to 858
recruit a similar size sample to that of Experiment 3, so we recruited a sample of 349 U.S. 859
residents from Amazon’s MTurk. We made no exclusions (Mage = 37.55, SD = 11.36; 53% 860
female, 0.3% non-binary/other; 76% Whites/European Americans; 9% Asian/Asian Americans; 861
7% Latino/Hispanic Americans; 12% Black/African Americans; 2% Native American, 1% 862
Hawaiian/Pacific Islander, 0.3% Other). Analyses were powered to detect an effect of 0.33 SD at 863
80% power. 864
Design and procedure. The design and procedure were identical to that of Experiment 3, 865
except that we manipulated target and removed the aggressive policing condition used in 866
Experiment 3. This resulted in a 2 (Target: Police, Worker) × 2 (Transparency: Ambiguous, 867
Transparent) design. 868
Imagine the following happened to you... 869
1. On your way to the grocery store, you see another person walk in before you. 870
2. You walk behind them into the grocery store and head in to start your shopping. 871
3. You leave the store. 872
4. On your way out, you notice a {police officer’s car / grocery store worker who is 873
moving carts} pull next to where you were walking. 874
5. The {police officer walks out of the car / grocery store worker parks the carts} and 875
looks your way. 876
6. The {police officer / grocery store worker} walks towards you and starts a 877
conversation, saying, 878
879

27
"Hi, how are you doing? My boss is asking me to talk to people to get to know the 880
community better." [Transparent Police Officer / Grocery Store Worker] 881
"Hi, how are you doing? " [Ambiguous Police Officer / Grocery Store Worker] 882
Threat. We measured threat using the same scale as in Experiments 2-4. See Table S11 883
in the SI for correlations among all primary variables. 884
Experiment 6 885
Participants. We preregistered our study on OSF (tiny.cc/PreRegExp6). We sought to 886
recruit a similar size sample to that of Experiment 5, so we recruited a sample of 451 U.S. 887
residents from Amazon’s MTurk. We made no exclusions (Mage = 37.86, SD = 12.87; 58% 888
female, 0.2% non-binary/other; 71% Whites/European Americans; 8% Asian/Asian Americans; 889
5% Latino/Hispanic Americans; 8% Black/African Americans; 1% Native American, 5% Bi- or 890
Multi-racial, 1% Other). Analyses were powered to detect an effect of 0.29 SD at 80% power. 891
Design and procedure. Participants read a similar vignette to Experiment 5 that instead 892
took place at a park. Participants read that a transparent (or ambiguous) police officer (or park 893
ranger) walked up to them. 894
1. You head to a nearby park. 895
2. While at the park, you notice a {police officer / park ranger}'s car pull next to where 896
you were walking. 897
2. The {police officer / park ranger} walks out of the car and looks your way. 898
3. The {police officer / park ranger} walks towards you and starts a conversation, saying, 899
"Hi, how are you doing?" [Ambiguous Police Officer / Park Ranger] 900
“Hi, how are you doing? My boss is asking me to talk to people to get to know the 901
community better.” [Transparent Police Officer / Park Ranger] 902
Threat. We measured threat using the same scale as in Experiments 2-5. See Table S12 903
in the SI for correlations among all primary variables. 904
Data Availability 905
All data supporting the findings in this manuscript are available on the Open Science Framework 906
and can be found at the following links: http://tiny.cc/DataExp1, http://tiny.cc/DataExp2; 907
http://tiny.cc/DataExp3; http://tiny.cc/DataExp4; http://tiny.cc/DataExp5; 908
http://tiny.cc/DataExp6 909
910 Code Availability 911
All code for analyses supporting the findings in this manuscript are available on the Open 912
Science Framework and can be found at the following links: http://tiny.cc/SynExp1, 913
http://tiny.cc/SynExp2-6 914

28
References 915
1. Department of Justice. Justice department announces $139 million for law enforcement 916
hiring to advance community policing. (2021). https://www.justice.gov/opa/pr/justice-917
department-announces-139-million-law-enforcement-hiring-advance-community-policing 918
2. Skogan, W. G. & Hartnett, S. M. Community Policing, Chicago Style. (Oxford University 919
Press, 2000). 920
3. Harrell, E. Contacts between police and the public, 2018 – Statistical tables. Statistical 921
Tables 14 (2018). 922
4. Jones-Brown, D. & Maule, B. A. Racially biased policing. In Race, Ethnicity, and Policing 923
140–174 (New York University Press, 2010). 924
5. Meares, T. L. The law and social science of stop and frisk. Annu. Rev. Law. Soc. Sci. 10, 925
335–352 (2014). 926
6. Hall, A. V., Hall, E. V. & Perry, J. L. Black and blue: Exploring racial bias and law 927
enforcement in the killings of unarmed black male civilians. American Psychologist 71, 175–928
186 (2016). 929
7. Gill, C., Weisburd, D., Telep, C. W., Vitter, Z. & Bennett, T. Community-oriented policing 930
to reduce crime, disorder and fear and increase satisfaction and legitimacy among citizens: A 931
systematic review. J Exp Criminol 10, 399–428 (2014). 932
8. Gallup. Amid pandemic, confidence in key U.S. institutions surges. Gallup.com 933
https://news.gallup.com/poll/317135/amid-pandemic-confidence-key-institutions-934
surges.aspx (2020). 935
9. MacDonald, J. M. The effectiveness of community policing in reducing urban violence. 936
Crime & Delinquency 48, 592–618 (2002). 937
10. O’Brien, T. C. & Tyler, T. R. Rebuilding trust between police & communities through 938
procedural justice & reconciliation. Behavioral Science & Policy 5, 34–50 (2019). 939
11. Peyton, K., Sierra-Arévalo, M. & Rand, D. G. A field experiment on community policing 940
and police legitimacy. Proc Natl Acad Sci USA 116, 19894–19898 (2019). 941
12. Blair, G. et al. Community policing does not build citizen trust in police or reduce crime in 942
the Global South. Science 374, eabd3446 (2021). 943
13. Rios, V. M., Prieto, G. & Ibarra, J. M. Mano suave–mano dura: Legitimacy policing and 944
Latino stop-and-frisk. Am Sociol Rev 85, 58–75 (2020). 945

29
14. Authors. Manuscript blinded for review. Blinded for review (2022). 946
15. Hahn, P. R., Murray, J. S. & Carvalho, C. M. Bayesian regression tree models for causal 947
inference: Regularization, confounding, and heterogeneous effects. Bayesian Analysis (2020) 948
doi:10.1214/19-BA1195. 949
16. Yeager, D. S. et al. A national experiment reveals where a growth mindset improves 950
achievement. Nature 573, 364–369 (2019). 951
17. Woody, S., Carvalho, C. M. & Murray, J. S. Model interpretation through lower-dimensional 952
posterior summarization. Journal of Computational and Graphical Statistics 30, 144–161 953
(2021). 954
18. McShane, B. B., Gal, D., Gelman, A., Robert, C. & Tackett, J. L. Abandon statistical 955
significance. null 73, 235–245 (2019). 956
19. Boyd, R., Ashokkumar, A., Seraj, S. & Pennebaker, J. The Development and Psychometric 957
Properties of LIWC-22. (2022). doi:10.13140/RG.2.2.23890.43205. 958
20. Yu, D. et al. Exploring idiographic approaches to children’s executive function performance: 959
An intensive longitudinal study. Journal for Person-Oriented Research 6, 73 (2020). 960
21. Blascovich, J. & Mendes, W. B. Social psychophysiology and embodiment. In Handbook of 961
Social Psychology (eds. Fiske, S. T., Gilbert, D. T. & Lindzey, G.) 194–227 (Wiley, 2010). 962
22. Jamieson, J. P. Challenge and threat appraisals. in Handbook of Competence and Motivation: 963
Theory and Application (eds. Elliot, A. J., Dweck, C. S. & Yeager, D. S.) (Guilford Press, 964
2017). 965
23. Blascovich, J. & Tomaka, J. The biopsychosocial model of arousal regulation. In Advances 966
in Experimental Social Psychology (ed. Zanna, M. P.) vol. 28 1–51 (Academic Press, 1996). 967
24. Tronstad, C., Kalvøy, H., Grimnes, S. & Martinsen, Ø. G. Waveform difference between 968
skin conductance and skin potential responses in relation to electrical and evaporative 969
properties of skin. Psychophysiology 50, 1070–1078 (2013). 970
25. Bach, D. R., Flandin, G., Friston, K. J. & Dolan, R. J. Time-series analysis for rapid event-971
related skin conductance responses. Journal of Neuroscience Methods 184, 224–234 (2009). 972
26. Voigt, R. et al. Language from police body camera footage shows racial disparities in officer 973
respect. Proc Natl Acad Sci USA 114, 6521–6526 (2017). 974
27. Pierson, E. et al. A large-scale analysis of racial disparities in police stops across the United 975
States. Nat Hum Behav 4, 736–745 (2020). 976

30
28. Braga, A. A., Brunson, R. K. & Drakulich, K. M. Race, place, and effective policing. Annu. 977
Rev. of Soc. 45, 535–555 (2019). 978
29. Sewell, A. A. & Jefferson, K. A. Collateral damage: The health effects of invasive police 979
encounters in New York City. Journal of Urban Health 93, 42–67 (2016). 980
30. Granot, Y., Tyler, T. R. & Durkin, A. Legal socialization during adolescence: The emerging 981
role of school resource officers. Journal of Social Issues 77, 414–436 (2021). 982
31. Wood, G., Tyler, T. R. & Papachristos, A. V. Procedural justice training reduces police use 983
of force and complaints against officers. Proc Natl Acad Sci USA 117, 9815–9821 (2020). 984
32. Weisburd, D. et al. Reforming the police through procedural justice training: A multicity 985
randomized trial at crime hot spots. Proc. Natl Acad Sci USA 119, e2118780119 (2022). 986
33. Mayer, R. C. & Davis, J. H. The effect of the performance appraisal system on trust for 987
management: A field quasi-experiment. Journal of Applied Psychology 84, 14 (1999). 988
34. Boyd, R. L. & Schwartz, H. A. Natural language analysis and the psychology of verbal 989
behavior: The past, present, and future states of the field. Journal of Language and Social 990
Psychology 40, 21–41 (2021). 991
35. Jordan, K. N., Sterling, J., Pennebaker, J. W. & Boyd, R. L. Examining long-term trends in 992
politics and culture through language of political leaders and cultural institutions. Proc. Natl. 993
Acad. Sci. U.S.A. 116, 3476–3481 (2019). 994
36. Kacewicz, E., Pennebaker, J. W., Davis, M., Jeon, M. & Graesser, A. C. Pronoun use reflects 995
standings in social hierarchies. Journal of Language and Social Psychology 33, 125–143 996
(2014). 997
37. Markowitz, D. M., Kouchaki, M., Gino, F., Hancock, J. T. & Boyd, R. L. Authentic first 998
impressions relate to interpersonal, social, and entrepreneurial success. Social Psychological 999
and Personality Science 19485506221086136 (2022) doi:10.1177/19485506221086138. 1000
38. Cohn, M. A., Mehl, M. R. & Pennebaker, J. W. Linguistic markers of psychological change 1001
surrounding September 11, 2001. Psychol Sci 15, 687–693 (2004). 1002
39. Watson, D., Anna, L. & Tellegen, A. Development and validation of brief measures of 1003
positive and negative affect: The PANAS scales. Journal of Personality and Social 1004
Psychology 8. 1005

31
40. Blascovich, J. & Mendes, W. B. Social psychophysiology and embodiment. In Handbook of 1006
Social Psychology (eds. Fiske, S. T., Gilbert, D. T. & Lindzey, G.) 194–227 (John Wiley & 1007
Sons, 2010). 1008
41. Blascovich, J. & Mendes, W. B. Challenge and threat appraisals. in Feeling and Thinking: 1009
The Role of Affect in Social Contagion 59–82 (Cambridge University Press, 2001). 1010
42. van Lier, H. G. et al. A standardized validity assessment protocol for physiological signals 1011
from wearable technology: Methodological underpinnings and an application to the E4 1012
biosensor. Behavior Research Methods 52, 607–629 (2020). 1013
43. Milstein, N. & Gordon, I. Validating measures of electrodermal activity and heart rate 1014
variability derived from the Empatica E4 utilized in research settings that involve interactive 1015
dyadic states. Frontiers in Behavioral Neuroscience 148 (2020). 1016
44. Kocielnik, R., Sidorova, N., Maggi, F. M., Ouwerkerk, M. & Westerink, J. H. D. M. Smart 1017
technologies for long-term stress monitoring at work. in Proceedings of the 26th IEEE 1018
International Symposium on Computer-Based Medical Systems 53–58 (IEEE, 2013). 1019
doi:10.1109/CBMS.2013.6627764. 1020
45. Society for Psychophysiological Research Ad Hoc Committee on Electrodermal Measures. 1021
Publication recommendations for electrodermal measurements: Publication standards for 1022
EDA. Psychophysiol 49, 1017–1034 (2012). 1023
46. Wijsman, J., Grundlehner, B., Liu, H., Penders, J. & Hermens, H. Wearable physiological 1024
sensors reflect mental stress state in office-like situations. In 2013 Humaine Association 1025
Conference on Affective Computing and Intelligent Interaction 600–605 (IEEE, 2013). 1026
doi:10.1109/ACII.2013.105. 1027
47. Smets, E. et al. Large-scale wearable data reveal digital phenotypes for daily-life stress 1028
detection. npj Digital Med 1, 67 (2018). 1029
48. Kirschbaum, C., Pirke, K.-M. & Hellhammer, D. H. The ‘Trier Social Stress Test’–A tool for 1030
investigating psychobiological stress responses in a laboratory setting. Neuropsychobiology 1031
28, 76–81 (1993). 1032
49. Mastrofski, S. & Parks, R. B. Improving observational studies of police. Criminology 28, 1033
475–496 (1990). 1034
1035
1036

32
Acknowledgements 1037
1038
This research was approved the University of Texas at Austin’s Institutional Review Board 1039
(protocol #20210303). The authors thank J. Bowling, R. Chen, M. Moore, and H. Yi for 1040
assistance in the data collection, C. Bryan, K. DeCelles, J. Jamieson, M. Mooijman, J. 1041
Pennebaker, and E. Seo for their critiques and comments on this work, and M. Clapper for 1042
management of the data collection and partnering police departments. 1043
1044 Funding. This research was supported by the Dispute Resolution Research Center (DRRC), the 1045
William T. Grant Foundation (grant # 202102), the Eunice Kennedy Shriver National Institute of 1046
Child Health and Human Development (grant #s P2CHD042849, PRC, and R01HD084772), 1047
National Science Foundation (grant # 1761179), and an Advanced Research Fellowship from the 1048
Jacobs Foundation to D. Yeager. Any opinions, findings, and conclusions or recommendations 1049
expressed in this material are those of the author(s) and do not necessarily reflect the views of 1050
the National Science Foundation, the National Institutes of Health, and other funders. 1051 1052 Author contributions. The conceptualization of the research hypotheses and the design of the 1053
study were initiated and led equally by K. Dobson and A. Dittmann. Dobson and A. Dittmann 1054
supervised all data collection and processing by research assistants, conducted all field 1055
observations and interviews, produced the datasets, scored transcripts using natural language 1056
processing transcripts, and drafted the supplemental materials. D. Yeager conducted the 1057
Bayesian analyses presented in the main text, wrote the first draft of the manuscript, and edited 1058
the supplemental materials. All authors edited and approved the final manuscript and 1059
supplemental materials. 1060
1061 Competing interests. The authors have no competing interests to declare. 1062
1063 Data and materials availability. Preregistration details, code for analyses, anonymized data, and 1064
experiment materials can be found at http://tiny.cc/TransparencyProject 1065
1066

1
Supplemental Materials 1 2 This online supplement contains the following information: 3
• Table S1: An open science disclosure table 4
• An overview of the Bayesian Causal Forest statistical analysis method. 5
• Additional background on community policing 6
• Additional thematic results from qualitative field observations 7
• A pilot laboratory experiment validating the wrist-worn EDA sensor’s data (Figures S1 8 and S2) 9
• Additional results for Experiment 1 10
• Additional results for Experiments 2-6 11
• Tables S2 to S12: Descriptive statistics, balance tests, and additional analyses 12 13 14 15 Table S1. Open Science Disclosure Table for Experiments 1-6. 16 17 Study Registration Materials Data Syntax
1 http://tiny.cc/PreRegExp1 http://tiny.cc/MatExp1 http://tiny.cc/DataExp1 http://tiny.cc/SynExp1
2 http://tiny.cc/PreRegExp2
http://tiny.cc/MatExp2-6
http://tiny.cc/DataExp2
http://tiny.cc/SynExp2-6
3 http://tiny.cc/PreRegExp3 http://tiny.cc/DataExp3
4 http://tiny.cc/PreRegExp4 http://tiny.cc/DataExp4
5 http://tiny.cc/PreRegExp5 http://tiny.cc/DataExp5
6 http://tiny.cc/PreRegExp6 http://tiny.cc/DataExp6
18 19 20 21

2
Overview of Statistical Approach: Bayesian Causal Forest (BCF) Analysis 22
All analyses in the main text were conducted with a conservative, Bayesian, machine-learning 23 method called Bayesian Causal Forest (BCF). BCF has been found, in multiple open 24 competitions and simulation studies, to identify average treatment effects (ATEs) with minimal 25 bias, while also detecting true sources of treatment effect heterogeneity, if they are present, while 26 not lending much credence to noise 1–3. BCF builds on and has in several cases surpassed the 27 popular Bayesian Additive Regression Trees (BART, 4) approach. Both Bayesian regression tree 28 models and BCF in particular are consistently top performers in empirical evaluations of 29 methods for causal inference 2–5. 30
Here we provide more details about how the BCF model was estimated for the skin conductance 31 response (SCR) outcome, because this is the outcome with the most complex model since it 32 involves repeated measures. A simpler model that excludes the repeated measures and person-33 level random effect was fit for the person-level outcomes of threat, trust, natural language, and 34 emotions. 35
In the BCF analysis, the SCR for participant j at 10-second interval  is denoted by , and is 36 modeled by 37
 =+,+[()⋅
]+. 38
where  is the vector of person-level covariates (police officer ID, age, race/ethnicity, gender, 39 native English speaker status) and  is the person-by-time-interval variable. The treatment 40 effect is moderated by time, , in that time interacts with the student treatment indicator . We 41 allow for person-level intercept random effects, . The person-by-time-interval error term is  42 and is assumed to be normally distributed with variance . 43
Here,  is a nonparametric function which allows for nonlinearities and interactions between 44 covariates in affecting the expected outcome. Furthermore,  is a nonparametric function that 45 allows the treatment effect to vary across time. This model is meant to mimic that of the 46 repeated-measures (i.e. multilevel) linear analysis in terms of the specification of the control and 47 treatment modifier variables, but relax the strict assumption of linearity and additivity between 48 the covariates and the expected value of the outcome. 49
Using a nonparametric Bayesian approach in this manner has several advantages. First, it allows 50 the data to speak and better inform us about the relationships between the covariates and the 51 outcome. It allows us to uncover (possibly unanticipated) sources of heterogeneity in the 52 treatment effect over time while requiring few prior assumptions. The prior we use results in 53 posterior estimates that are inherently conservative, making it unlikely that we will dramatically 54 over- or under-estimate the effect of the intervention at different time points. 55
Prior specification 56
To complete our Bayesian model, we must specify prior distributions for the unknown values in 57 the equation above. These include the nonparametric functions (⋅) and (⋅), the random effect 58 , and the error variance . 59

3
The prior for the functions (⋅) and (⋅) is taken from the Bayesian Causal Forests model (BCF; 60 1). Under this model, both functions have a sum-of-trees representation, as first defined for 61 Bayesian methods in Chipman, George, and McCulloch (4). Each tree consists of a set of internal 62 decision nodes which partitions the covariate space, and a set of terminal nodes, or leaves, 63 corresponding to each element of the partition. The prior for each of (⋅),
(. ) and (⋅), is 64 comprised of three parts: the number of trees, two parameters controlling the depth of each tree, 65 and a prior on the leaf parameters. Use of this sum-of-trees term allows for detection of 66 nonlinearity and interactions between covariates. 67
The key feature of the BCF model is that the prior for (⋅), which explains heterogeneity in the 68 intervention effect, is regularized more heavily compared to the control function (⋅) in order to 69 shrink toward homogeneous effects. The prior for (⋅) uses fewer trees, with each tree being 70 regularized to be shallower (that is, contain fewer partitions). Details are on prior specification 71 are given in Hahn, Murray, and Carvahlo (1) and Chipman, George, and McCulloch (4). 72
The random effect  is given a Gaussian prior with the standard deviation having a prior of a 73 half -distribution with 3 degrees of freedom, as recommended by Gelman (7). Finally, the error 74 variance is given an inverse chi-squared prior with 3 degrees of freedom and scale parameter 75 informed by the data. 76
Posterior Inference 77
After conditioning on observed data, we can update the prior distribution to obtain a posterior 78 distribution. To calculate the posterior distribution for quantities of interest, we implement a 79 Markov chain Monte Carlo (MCMC) sampling scheme. Since the primary component of the 80 model is the sum-of-trees functions (⋅) and (⋅), the MCMC scheme relies on a Bayesian 81 backfitting algorithm 6. The rest of the parameters in the model are conditionally conjugate, 82 making their posterior sampling relatively efficient. 83
The BCF estimate of  was approximately additive. Therefore, to give an interpretable estimate 84 of this conditional intervention effect, we created an additive summary of the fitted (⋅) function 85 using splines, and looked at the partial effect of changing the 10-second time interval. This 86 additive summary captures > 98% of the predictive variance in the posterior of , so this additive 87 summary is a faithful recapitulation of the fitted CATE function. 88
89 90 91

4
Additional Background on Community Policing 92 93 Community-oriented policing is a strategy in which police officers prioritize initiating non-94 enforcement interactions and building cooperative relationships with community members. 95 These interactions happen in a variety of places, including unexpected interactions (e.g., on 96 streets, in businesses, or in public recreation areas) and expected interactions (e.g., community 97 meetings and recreational events). Expected interactions are community- and/or police-organized 98 events that may, for example, involve a panel of police officers asking the community what the 99 problems of in the area are and how the police can mitigate those problems. Or, for example, 100 expected interactions may involve meet-and-greet events where the police and communities offer 101 free resources (e.g., food, drinks) while also providing information to educate others about 102 resources from police or the community. Unexpected interactions that are community-oriented 103 often happen while police are on patrol and observe idle community members in the field. For 104 example, police may initiate conversations with community members in parks by walking up to 105 them and starting a friendly conversation. They may engage in “positive ticketing” with children, 106 where the goal is to reinforce positive behavior (e.g., wearing a helmet while on a bike) with a 107 reward (e.g., free ice cream coupons), hoping to associate positivity with police presence at a 108 young age. Or, police may walk into businesses to check on the well-being of the employees in 109 those businesses, asking if there are any problems that law enforcement can address or having 110 friendly conversation to build a relationship. In our experiments, we focused on the unexpected 111 community-oriented interactions that officers initiate with the public. The percentages of 112 community-oriented interactions in American law enforcement is unclear, and so is the 113 proportion of unexpected to expected officer-initiated, non-enforcement interactions. However, 114 work has suggested that approximately 90 percent of agencies have endorsed community 115 policing as a strategy. The strategic logic of community policing is that gaining trust and 116 perceived legitimacy from community members will increase cooperation. Cooperative 117 community-police teams are expected to facilitate more efficient disclosure of community 118 problems to law enforcement to allow police to serve the public more effectively. 119 120

5
Additional Themes from Qualitative Field Observations 121
Transparency of intent was a useful tactic for connecting. Officers in the community policing 122 unit had all developed different tactics to connect from their assignments prior to joining the unit. 123 The most concrete form of connection was transparently expressing their intentions. Typically, 124 police prefer to keep information close to the vest and refrain from making themselves more 125 vulnerable. However, we found that they consistently felt comfortable—at a minimum— 126 expressing their intentions clearly to members of the community. There are many reasons these 127 officers could have felt comfortable transparently expressing themselves. For example, these 128 officers were chosen by leadership based on their previous ability to build trusting relationships 129 in their prior units. Additionally, leadership in this unit created a strong culture of openness that 130 comprised biweekly meetings where the officers would share their thoughts and feelings. The 131 leaders of these meetings would often appraise the expressions of their officers, citing examples 132 and concrete tactics they learned. For example, the lieutenant of the community policing unit, a 133 Hispanic male with over 20 years of law enforcement experience, described how he had learned 134 the tactic of being transparent prior to his current role, reflecting on the importance of your 135 “approach” to a member of the community: 136
“Interesting thing about dealing with Hispanic families and potential gang recruitment. 137 Hispanic families are more protective of males. Females aren’t really ever involved in 138 gang activity anyway, but they are doing most of the actual work. Very protective of 139 sons... So, we approach them and we say, ‘your son is being targeted and we want to 140 keep these things from happening to him, we want to protect him’ rather than saying, 141 ‘your son is part of a gang’ which makes them more defensive.” 142
The lieutenant understands how his intentions may be misinterpreted, and how he makes sure to 143 communicate his intentions clearly to try to prevent these misinterpretations. Expecting that 144 people may be defensive of their families—due to a belief that police do harm—the lieutenant 145 starts the conversations by explaining his genuine intent and his occupational role as a protector. 146 He speaks about how he wants to “protect” the son who might be recruited by a more harmful 147 organization (i.e., a gang). He minimizes the enforcer role which otherwise could make the 148 mother feel he is going to ruin her son’s life by enforcing the law. 149
Similarly, on a ride along with this officer and his partner, a female Latina officer, it was clear 150 how effective they were at initiating interactions that could even have resulted in enforcement 151 action (i.e., issuing citations) that remained conversational and did not escalate. Their use of 152 transparency of their intent is particularly clear in an interaction with two middle-aged Latinx 153 males who were visibly intoxicated and who had open containers in sight: 154
“Two Latinx males in their 40s-50s are sitting on curb of [street in their area] with open 155 containers clearly in sight. A Latino male and Latina female officer roll down the 156 window, and start by saying “Cuidado!” (Spanish for “be careful!”). They go on to 157 explain to them that they cannot be sitting on the curb like that, and that it is dangerous 158 because cars are always coming and going through this alley, and it is hard to spot them, 159 and they wouldn’t want them getting hit. They also say that they shouldn’t be sitting there 160

6
with open containers. The two men are both very respectful, and respond, ‘We are getting 161 up and going right now, no worries’ and start to pack up and head out immediately. 162 163 As we continue on drive after this interaction, the female officer explains that this is a 164 time that they could have issued citations, and that maybe other officers would have 165 issued citations, but that they like to, at least initially, give people the benefit of the doubt. 166 This strategy, they believe, is more effective, because if they don’t necessarily write out 167 citations from the get-go, and they give people a warning the first time or two, this can be 168 another way to build trust and rapport. They will issue citations if they need to if it is 169 really becoming a persistent issue, but they have found that if they don’t issue citations 170 right away, that typically going forward, when they encounter people like these two men, 171 they will stop their behavior right away without being prompted.” 172
In sum, their focus is primarily on preventing harm, and they do so by explaining their actions 173 and decisions to the community—engaging in transparency of their intentions. While the two 174 men were technically breaking a law, their approach was to explain why they should move 175 because they are not in a safe location, and issue a verbal warning—rather than write up a ticket 176 on the first offense that they witnessed. By utilizing transparency, they were able to signal to 177 community members that they are trustworthy. 178
Transparency of intent was not the default approach of all officers. Many officers we 179 observed initiated interactions without making their intentions explicitly clear. For example, two 180 officers in the community policing unit did a drive by of a known problem corner in the area, and 181 initiated an interaction with a group of men they presumed might have been part of drug deals 182 that were known to occur in the area: 183
“The citizens on this block are at their wits end about the drug sales happening here. (I 184 look at the homes on this block. They are two-story brick homes—slightly bigger than the 185 small bungalows I’ve seen on other streets. But every other one is boarded up with wood 186 planks. Some houses are on lawns that are kept, but they are adjacent to houses that 187 appear abandoned with overgrown weeds.) 188
We get to the end of the block and there are 4 Black males ages 18-30. We pull over and 189 stop our car. Two men approach our car and the other two hang back all looking at us. 190 The two officers speak from the window. The officer driving asks, ‘Why do you hang out 191 here? Don’t you know that this block is hot?’ The other officer adds, ‘This is a hot spot 192 and you don’t want to be here.’ 193 194 One of the men has tattoos on his neck and arms says something to the effect of the block 195 being safe now, that the party that shot in the area isn’t here anymore. 196 197 The officers say they don’t want to be caught in the wrong place at the wrong time. This 198 man says he’s not worried about it. And so the officers say have a good day.” 199 200 Unlike some officers, these officers never made it clear that their intent was simply to check in 201 on these individuals – they began the interaction by peppering the group of young Black men 202 with questions. In response, the men replied defensively and spoke curtly, in abbreviated 203

7
sentences. The initiation never progressed beyond a few cursory rounds of back and forth. As 204 such, the officers quickly decided to move on. 205 206 Similarly, on another ride along, a White female officer initiated an interaction with four Black 207 men in their 20s-30s. While she stated her intent for the interaction to the researcher, she did not 208 state it to the group of men. She explained that she was about to do a business check, but when 209 she saw the group of four men, she decided to change course and speak to these men first. While 210 they were not committing a crime, she wanted to do a “well-being check” and to find out why 211 they were loitering: 212 213 “She interacted with everyone all at once from her car. She pulled up by four individuals, 214 in the parking lot, as if she was about to pull into a space but still perpendicular or 215 angled to the spaces so she could face the men, who were standing outside of the business 216 without any visible bags or belongings. 217 218 She starts by asking ‘You guys good? How are you doing?’ without any context or lead-219 up. 220 The four individuals start moving away from the business immediately, with only a couple 221 speaking-up. They respond ‘We’re good’ as they move away from her car and towards 222 [the street]. 223 224 She asks, ‘You guys hear about the shooting?’ 225 226 They mutter something along the lines of ‘No… shooting…?’ seemingly unsure about 227 what she is referring to. 228 229 She continues, saying, ‘Yeah, someone died and another is in critical condition--they 230 opened 15 rounds.’ 231 232 One of the men, the most talkative out of the bunch with her, says ‘15 rounds?!’ in what 233 seems like disbelief. 234 235 She continues, saying ‘Yeah… so be careful.’ They are already on their way about 236 midway through the parking lot by the time she says this.” 237 238 Akin to the previous example, while the officer clearly did not intend to enforce the law against 239 these four men, she never stated this clearly and explicitly to the group of men. Throughout the 240 brief interaction, most appeared reserved and unwilling to interact with her, and even the 241 individual who did engage with the officer spoke in short sentences. Mirroring this verbal 242 discomfort, the group also physically moved away from the officer quickly before she could 243 even conclude the interaction. 244 245

8
Pilot Study Validating the ambulatory SCR Second Derivative Metric 246
To validate the algorithm and scoring method used for the present study’s ambulatory SCR data, 247 we conducted a laboratory experiment using a standardized social stress procedure and gold-248 standard physiological measures. This was needed because no previous study had validated 249 ambulatory SCR metrics within the biopsychosocial model of challenge and threat. 250
Participants. Participants were prescreened and excluded for physician-diagnosed hypertension, 251 a cardiac pacemaker, BMI > 30, and medications with cardiac side effects (e.g., 18). A total of 252 166 students were recruited from a university social science subject pool (120 females, 46 males; 253 76 White/Caucasian, 12 Black/African-American, 17 Latinx, 65 Asian/Asian-American, 2 254 Pacific Islander, 4 Mixed Ethnicity, 7 Other; Mage = 19.81, SD = 1.16, range = 18–26; 32% 255 reported their mothers did not have a college degree). Due to logistical limitations (i.e. the 256 number of available wearable devices and participant schedules), 100 of these participants were 257 able to wear the ambulatory skin conductance sensors during the pilot experiments, in addition to 258 the gold-standard measures of cardiovascular challenge/threat responses. 259
Procedure. After intake questions, application of sensors, and acclimation to the lab 260 environment, participants rested for a 5-min baseline cardiovascular and skin conductance 261 recording, which occurred approximately 25-min after arrival at the laboratory. Participants then 262 completed the Trier Social Stress Test (TSST; 26), which is depicted in Figure S1 below. The 263 TSST asks participants to give an impromptu speech about their personal strengths and 264 weaknesses in front of two evaluators (see Fig. S1). Evaluators are presented as members of the 265 research team who are experts in nonverbal communication and will be monitoring and assessing 266 the participant’s speech quality, ability to clearly communicate ideas, and nonverbal signaling. 267 Throughout the speech (and math) epochs of the TSST, evaluators provide negative nonverbal 268 feedback (e.g., furrowing brow, sighing, crossing arms, etc.) and no positive feedback, either 269 nonverbal or verbal
7
. At the conclusion of speeches, and without prior warning, participants are 270 asked to do mental math (counting backwards from 996 in increments of 7) as quickly as 271 possible in front of the same unsupportive evaluators. Incorrect answers were identified, and 272 participants were instructed to begin back at the start. This procedure is widely used to induce 273 the experience of negative, threat type stress responses
8,9
. After completion of the TSST task, 274 participants rested quietly for a 3-min recovery recording. Prior to leaving the lab all participants 275 were debriefed and comforted. 276 277 Figure S1. Procedure for the Trier Social Stress Test (TSST). 278
279
Physiological measures. The following measures were collected during baseline and throughout 280 the TSST: electrocardiography (ECG), impedance cardiography (ICG), and blood pressure (BP), 281

9
along with skin conductance from an ambulatory (i.e. wrist-worn) sensor. ECG and ICG signals 282 were sampled at 1000 Hz, and integrated with a Biopac MP150 system. ECG sensors were 283 affixed in a Lead II configuration. Biopac NICOO100C cardiac impedance hardware with band 284 sensors (mylar tapes wrapped around participants’ necks and torsos) were used to measure 285 impedance magnitude (Zo) and its derivative (dZ/dt). BP readings were obtained using 286 Colin7000 systems. Cuffs were placed on participants' non-dominant arm to measure pressure 287 from the brachial artery. BP recordings were taken at 2-min intervals during baseline, throughout 288 the stress task, and recovery. BP recordings were initiated from a separate control room. ECG 289 and ICG signals were scored offline by trained personnel. One-minute ensemble averages were 290 analyzed using Mindware software IMPv3.0.21. Stroke volume (SV) was calculated using the 291 Kubicek method 10. B- and X-points in the dZ/dt wave, as well as Q- and R-points in the ECG 292 wave, were automatically detected using the maximum slope change method. Then, trained 293 coders blind to other variables examined all placements and corrected erroneous placements 294 when necessary. 295 296 The key analysis of gold-standard physiological measures sought to categorize individuals into 297 threat vs. challenge groups using two key indices: total peripheral resistance (TPR) and stroke 298 volume (SV). This suite is commonly used to threat- versus challenge-type stress responses (for 299 a review see 11. TPR is the clearest indicator of threat-type responses and was therefore the focal 300 outcome measure in this research; even so, we sought to identify only “true” challenge/threat 301 responses, and so we also attended to SV, a commonly-accepted measure of “challenge” 302 response. TPR assesses vascular resistance, and when threatened, resistance increases from 303 baseline 12. TPR was calculated using the following validated formula: (MAP / CO) ∗ 80 13. SV 304 is the amount of blood ejected from the heart on each beat (on average per minute). Increases in 305 SV index greater beat-to-beat cardiac efficiency and more blood being pumped through the 306 cardiovascular system, and are often observed in challenge states, as the body spreads more 307 oxygenated blood to the periphery 8. Decreases in SV, on the other hand, are more frequently 308 observed in threat states (even though threat can also elicit little or no change in SV; 32). We 309 computed reactivity scores by subtracting each person’s average levels from the five minutes of 310 the baseline epoch. Thus, all TPR and SV results in the paper account for baseline differences. 311 312 Participants were categorized as showing a clear “threat-type” physiological profile when their 313 TPR reactivity and SV reactivity (compared to baseline) were in the highest/lowest quartiles, 314 respectively. They were categorized as showing a clear “challenge-type” physiological profile 315 when their TPR and SV reactivity were in the lowest/highest quartiles, respectively. This 316 “extreme groups” method was used in order to have very clearly different groups and to avoid 317 ambiguity about the expected SCR values of more mixed challenge/threat profiles. 318 319 SCR values were processed using the same, previously-validated scoring algorithm that was used 320 in the primary police interaction study. The raw data from the devices were read into Python and 321 the same processing scripts were run, which included the same decision rules for implausible 322 biological values, outlying cases, and measurement artifacts, as well as the same function to 323 calculate the second derivative at each moment. Overall, with this data, we could assess what 324 patterns in SCR second derivative should be expected in unambiguous challenge/threat states. 325 326

10
Pilot study results. We found that the TSST validation study (Figure S2A-B) yielded 327 differences in SCR across challenge/threat groups that were highly parallel to the differences 328 across treatment/control groups in the police interaction study (Figure S2C-D). In Figure S2A, 329 we present the minute-by-minute levels of SCR (second derivative) on the y-axis and minutes, 330 across three study epochs, on the x-axis. Among those who were known, through TPR and SV 331 measures, to show a challenge-type responses, we see an initial orienting response—higher 332 SCR—which matches the prediction from BPS models that those with a challenge response 333 should show greater initial SNS reactivity at the onset of a stressor relative to those with more of 334 a threat-type response. In Figure S2B, we plot the “effect” of challenge vs. threat-type responses 335 (i.e. the difference between the two kinds of responses) on SCR levels by minute. Again, we see 336 clear evidence that challenge-type responses are associated with more positive initial SCR 337 responses, and this fades as the interaction proceeds. In Figure S2C-D, we reproduce the results 338 from the main study, and show that the experiment’s results were highly parallel. As a result of 339 this validation study, we have stronger evidence for interpreting higher initial SCR levels (in 340 S2C) as consistent with more positive, challenge-type stress responses, and lower initial SCR 341 levels as a threat-type response. 342
Figure S2. Validity of SCR second derivative measure with respect to challenge/threat 343 measures (A,B) and comparison to the present police interaction study (C,D).
Note: In (A) and
344
(C), lines are the posterior medians; In (B) and (D), the boxes are the IQRs and the lines are the 95% intervals.
345
346

11
Experiment 1 Additional Results 347
Primary Outcomes: No moderation by participant demographic characteristics. 348 There was no moderation by participant demographics on our primary outcomes of threat or trust 349 (see Table S5). 350
Three-item version of trust. When including the third item assessing trust (which 351 showed unexpectedly weak correlations with the other two items), we found that civilians in the 352 transparency condition reported directionally more trust (M = 5.59, SD = 1.12) than those in the 353 control condition (M = 5.30, SD = 1.19), p = .053,
η
2 = .016. In effect, there were no treatment 354 effect on the single item that showed very low reliability and appeared to confuse participants, 355 but there was a meaningful effect (reported in the main text) on the two other items combined. 356 357 Inspiration. A final exploratory analysis examined the effect of the transparency 358 statement on one of the primary goals of community policing: to inspire citizenry with respectful 359 interactions. We found that just 14% of participants in the control condition reported feeling 360 generally inspired at the end of the conversation with the officer, but this number was twice as 361 high, 28%, in the transparency condition, pr(ATE>0) = .92. Although this finding was from an 362 unplanned analysis, it was consistent with transcripts (such as the one depicted in Fig. 2B in the 363 main text) and suggests a need for future research on the broader trust-building effects of 364 transparent community policing. 365 366 Experiment 1 Results without Covariates 367 368 Manipulation Check: Perceptions of transparency of intentions. Similar to the results 369 presented in the main text, civilians in the transparency condition perceived that the officer’s 370 intentions were significantly clearer (M = 5.64, SD = 1.21) than those in the control condition (M 371 = 4.40, SD = 1.61, p < .001,
η
2 = .154). 372 373 Primary Measure 1: Threat of enforcement. Similar to the results presented in the 374 main text, and in support of Hypothesis 1, civilians in the transparency condition felt 375 significantly less threatened (M = 2.92, SD = 1.45) than those in the control condition (M = 3.40, 376 SD = 1.41, p = .01,
η
2 = .027). 377 378 Primary Measure 2: Trust in benevolence. Also similar to the results presented in the 379 main text and in further support of Hypothesis 1, civilians in the transparency condition felt 380 significantly more trust (M = 5.69, SD = 1.14) than those in the control condition (M = 5.31, SD 381 = 1.28, p = .019,
η
2 = .023). 382 383 384

12
Experiments 2-6 Moderation by Demographics 385 386 Primary Outcomes: Moderation by participant demographic characteristic. There 387 was no moderation by participant demographics on threat or trust (see Tables S6). 388 389

13
Table S2. Experiment 1: No Baseline Differences Between Conditions. 390 391 Control Transparency Statement
Condition Test Statistic
(t or X2) p
M or % SD M or % SD
Age 22.16 5.79 21.95 6.07 0.27 .79
Gender
(1 = Male) 42% 42% -0.02 .99
Race
(1 = White) 35% 32% 0.44 .66
Native U.S.
(1 = Native) 78% 82% -0.82 .41
Education level 3.66 1.32 3.47 1.32 1.11 .27
392 393 394

14
Table S3. Experiment 1: No Baseline Demographic Differences by Weekday. 395 396 Tuesday Wednesday Thursday Friday Saturday Sunday
b t or Z p b t or Z p b t or Z p b t or Z p b t or Z p b t or Z p
Age 0.65 0.44 .66 1.81 1.29 .20 1.83 1.26 .21 2.52 1.61 .11 0.16 0.09 .93 1.26 0.73 .47
Gender
(1 = Male) 0.17 0.35 0.73 0.16 0.32 .75 -0.16 -0.33 .75 0.17 0.32 .75 0.31 0.54 .59 0.90 1.51 .13
Race
(1 = White) -0.54 -1.01 .31 -0.20 -0.41 .68 0.15 0.30 .77 0.30 0.57 .57 -0.73 -1.15 .25 -0.61 -0.94 .35
Native U.S.
(1 = Native) 0.34 0.57 .57 -0.05 -0.09 .93 1.08 1.58 .11 0.50 0.77 .44 0.08 0.12 .91 -0.74 -1.17 .24
Education
level 0.03 0.08 .94 0.64 2.07 .04* -0.06 -0.18 .85 -0.13 -0.37 .71 -0.04 -0.11 .92 0.40 1.06 .29
397 398

15
Table S4. Experiment 1 Means, Standard Deviations, and Correlations. 399 Variable Mean S.D. 1 2 3 4 5 6 7 8 9 10 11
1. Condition
(1 = Transparent)
2. Transparency 4.95 1.57 0.39***
3. Threat 3.19 1.44 -0.17* -0.33***
4. Trust
(3-item) 5.43 1.17 0.13+ 0.23*** -0.54***
5. Trust
(2-item) 5.47 1.23 0.15* 0.23*** -0.55*** 0.92***
6. Threat (PANAS) 0.26 -- -0.20** -0.21*** 0.36*** -0.44*** -0.39***
7. Inspired (PANAS) 0.20 -- 0.16* 0.24*** -0.22*** 0.35*** 0.34*** -0.30***
8. Race
(1 = White) 0.34 0.47 -0.03 0.06 0.02 0.16* 0.13* -0.02 0.04
9. Gender
(1 = Male) 0.42 0.49 0.00 0.10 -0.12+ 0.02 0.07 -0.11+ 0.00 0.02
10. Education level 3.57 1.32 -0.07 -0.10 0.05 -0.03 -0.01 0.08 -0.13* 0.04 0.02
11. Age 22.07 5.90 -0.02 -0.03 -0.01 0.08 0.10 0.00 0.07 0.18* -0.05 0.52***
12. U.S. Native
(1 = Native) 0.80 0.41 0.05 0.08 0.07 0.00 -0.02 0.11+ 0.07 0.25*** -0.01 -0.19*** -0.02
Note: * p < .05, ** p < .01, *** p < .001. 400 401 402 403 404 405

16
Table S5. Experiment 1: No Evidence of Moderation of the Transparency Statement Intervention on Threat and Trust by 406 Individual Differences. 407
Threat Trust
Moderator tested b t p b t p
Age -0.02 -0.62 .54 -0.01 -0.50 .61
Gender
(1 = Male) 0.16 0.41 .68 -0.07 -0.22 .82
Race
(1 = White) -0.46 -1.17 .24 -0.27 -0.81 .42
Native U.S.
(1 = Native) -0.41 -0.89 .38 -0.18 -0.45 .65
Education level -0.24 -1.69 .09 -0.03 -0.27 .79
Note: Each row represents the test of a Transparency Statement Intervention × Moderator interaction in a separate regression model that also 408 includes in it the condition variable and the moderator. b=unstandardized regression coefficient. Results from multiple linear regressions models 409 utilizing null hypothesis tests and two-tailed p-values. 410 411 412

17
Table S6. Experiments 2-6: No Evidence of Moderation of the Transparency Statement Intervention on Threat by Individual 413 Differences. 414
Moderator
tested
Experiment 2 Experiment 3 Experiment 4 Experiment 5 Experiment 6
b t p b t p b t p b t p b t p
Age -0.004 -0.26 .80 -0.007 -1.83 .07 0.006 0.36 .72 0.03 1.67 .10 0.008 0.53 .60
Gender
1 = Male 0.33 1.08 .28 0.07 0.49 .62 0.13 0.36 .72 -0.61 -1.59 .11 0.23 0.64 .52
Race
1 =White -0.76 -2.11 .04* 0.08 0.42 .67 -0.20 -0.57 .57 -0.50 -1.14 .25 -0.20 -0.51 .61
Education
level 0.03 0.26 .80 0.008 0.12 .90 -0.02 -0.19 .85 -- -- -- -- -- --
Note: Each row represents the test of a Transparency Statement Condition × Moderator interaction in a separate regression model that also 415 includes in it the condition variable and the moderator. b=unstandardized regression coefficient. 416 417

18
Table S7. Experiment 1 treatment effects for LIWC text analyses. 418 419 Control Transparency Statement Condition ATE from BCF
analysis in SD units
M SD M SD
Authentic 46.26 37.34 69.44 28.71 .43
Analytic 17.65 23.75 15.99 15.82 -.02
Clout 62.23 35.56 43.42 33.42 -.39
Tone 74.72 32.21 69.28 28.72 -.07
Note: The treatment effect on clout was unexpected. Post-hoc, this effect can be interpreted as a decrease in language meant to signal 420 rank and hierarchy, which would be reasonable for people in the transparency condition if they, in fact, felt less of a need to show 421 dominance and power in the interaction with the officer. 422 423

19
Table S8A-B. Means, Standard Deviations, and Correlations for Experiment 2. 424
Threat (n = 305)
Variable Mean SD 1 2 3 4 5
1. Condition
(1 = Transparent)
2. Threat 3.92 1.39 -0.25***
3. Race
(1 = White) 0.72 -- -0.04 -0.09
4. Gender (1 = Male) 0.47 -- -0.01 0.08 -0.10+
5. Education level 5.75 1.17 0.08 -0.08 0.02 -0.01
6. Age 37.00 12.28 0.07 0.00 0.17*** -0.09 0.02
Trust (n = 304)
Variable Mean SD 1 2 3 4 5
1. Condition
(1 = Transparent)
2. Trust 4.61 0.71 0.10+
3. Race
(1 = White) 0.77 -- 0.00 0.09
4. Gender (1 = Male) 0.46 -- 0.05 -0.06 0.02
5. Education level 5.68 1.14 -0.05 -0.03 0.04 0.08
6. Age 37.24 12.82 -0.11+ 0.15* 0.08 -0.19*** 0.16***
Note: * p < .05, ** p < .01, *** p < .001. 425 426
427

20
Table S9. Means, Standard Deviations, and Correlations for Experiment 3. 428
Variable Mean SD 1 2 3 4 5 6
1. Condition
(1 = Aggressive Transparency)
2. Condition
(1 = Community Transparency)
-0.495***
3. Threat 3.66 1.30 0.302*** -0.308***
4. Race
(1 = White) 0.77 --
-0.010 0.040 0.006
5. Gender (1 = Male) 0.49 --
0.039 0.089 0.133** -0.058
6. Education level 5.83 1.23 0.0003 -0.010 0.046 -0.048 -0.089
7. Age 38.52 12.54 -0.037 0.063 0.001 -0.069 0.027 0.059
Note: * p < .05, ** p < .01, *** p < .001. 429 430
431
432
433
434
435
436
437
438
439

21
Table S10. Means, Standard Deviations, and Correlations for Experiment 4. 440
Variable Mean SD 1 2 3 4 5 6 7
1. Condition
(1 = Ambiguous Positivity)
2. Condition
(1 = Transparency)
-0.500***
3. Threat 4.28 1.53 0.107* -0.183***
4. Authenticity 4.11 1.59 -0.116* 0.144** -0.797***
5. Race
(1 = White) 0.49 --
0.006 0.026 -0.086 0.068
6. Gender (1 = Male) 0.50 --
0.016 -0.041 0.034 -0.002 0.049
7. Education level 4.24 1.44 -0.059 0.036 -0.011 -0.014 0.149** 0.098*
8. Age 31.52 11.14 0.113* -0.055 -0.033 0.054 0.088 0.080 0.114*
Note: * p < .05, ** p < .01, *** p < .001. 441 442
443
444
445
446
447
448
449
450

22
Table S11. Means, Standard Deviations, and Correlations for Experiment 5. 451
Variable Mean SD 1 2 3 4 5 6
1. Condition
(1 = Ambiguous Police)
2. Condition
(1 = Transparent Worker)
-0.340***
3. Condition
(1 = Transparent Police)
-0.332*** -0.335***
4. Threat 3.53 1.38 0.363*** -0.194*** 0.044
5. Race
(1 = White) 0.76 --
-0.101 0.010 0.030 -0.107*
6. Gender
(1 = Male) 0.46 --
0.015 -0.017 0.054 0.0485 -0.088
7. Age 37.55 11.36 -0.033 0.089 -0.078 -0.051 0.128* -0.093
Note: * p < .05, ** p < .01, *** p < .001. 452 453
454
455
456
457
458
459
460
461
462

23
Table S12. Means, Standard Deviations, and Correlations for Experiment 6. 463
Variable Mean SD 1 2 3 4 5 6
1. Condition
(1 = Ambiguous Police)
2. Condition
(1 = Transparent Ranger)
-0.33***
3. Condition
(1 = Transparent Police)
-0.33*** -0.33***
4. Threat 3.61 1.41 0.32*** -0.18*** -0.21***
5. Race
(1 = White) 0.71 --
-0.10* 0.12* -0.09+ -0.08+
6. Gender
(1 = Male) 0.42 --
0.10* -0.03 -0.09+ 0.17*** 0.00
7. Age 37.86 12.87 -0.03 -0.02 0.03 -0.16*** 0.17*** -0.14***
Note: * p < .05, ** p < .01, *** p < .001. 464

24
465 References 466
1. Hahn, P. R., Murray, J. S. & Carvalho, C. M. Bayesian regression tree models for causal 467
inference: Regularization, confounding, and heterogeneous effects. Bayesian Anal. (2020) 468
doi:10.1214/19-BA1195. 469
2. McConnell, K. J. & Lindner, S. Estimating treatment effects with machine learning. Health 470
Serv. Res. 54, 1273–1282 (2019). 471
3. Wendling, T. et al. Comparing methods for estimation of heterogeneous treatment effects 472
using observational data from health care databases. Stat. Med. 37, 3309–3324 (2018). 473
4. Dorie, V., Hill, J., Shalit, U., Scott, M. & Cervone, D. Automated versus do-it-yourself 474
methods for causal inference: Lessons learned from a data analysis competition. Stat. Sci. 34, 475
43–68 (2019). 476
5. Hahn, P. R., Dorie, V. & Murray, J. S. Atlantic Causal Inference Conference (ACIC) Data 477
Analysis Challenge 2017. ArXiv Prepr. ArXiv190509515 (2019). 478
6. Hastie, T. & Tibshirani, R. Bayesian backfitting (with comments and a rejoinder by the 479
authors. Stat. Sci. 15, 196–223 (2000). 480
7. Kirschbaum, C., Pirke, K.-M. & Hellhammer, D. H. The ‘Trier Social Stress Test’–A tool for 481
investigating psychobiological stress responses in a laboratory setting. Neuropsychobiology 482
28, 76–81 (1993). 483
8. Yeager, D. S., Lee, H. Y. & Jamieson, J. P. How to improve adolescent stress responses: 484
Insights from integrating implicit theories of personality and biopsychosocial models. 485
Psychol. Sci. 27, 1078–1091 (2016). 486
9. Dickerson, S. S. & Kemeny, M. E. Acute stressors and cortisol responses: A theoretical 487
integration and synthesis of laboratory research. Psychol. Bull. 130, 355–391 (2004). 488
10. Sherwood, A., Royal, S. A., Hutcheson, J. S. & Turner, J. R. Comparison of impedance 489
cardiographic measurements using band and spot electrodes. Psychophysiology 29, 734–741 490
(1992). 491
11. Mendes, W. B. & Park, J. Neurobiological concomitants of motivational states. In Advances 492
in Motivation Science vol. 1 233–270 (Elsevier, 2014). 493

25
12. Blascovich, J. & Mendes, W. B. Social psychophysiology and embodiment. In Handbook of 494
Social Psychology (eds. Fiske, S. T., Gilbert, D. T. & Lindzey, G.) 194–227 (John Wiley & 495
Sons, 2010). 496
13. Sherwood, A. et al. Methodological guidelines for impedance cardiography. 497
Psychophysiology 27, 1–23 (1990). 498
14. Jamieson, J. P., Nock, M. K. & Mendes, W. B. Mind over matter: Reappraising arousal 499
improves cardiovascular and cognitive responses to stress. J. Exp. Psychol. Gen. 141, 417–500
422 (2012). 501
15. Viechtbauer, W. Conducting meta-analyses in R with the metafor package. J. Stat. Softw. 36, 502
(2010). 503
504