The Relational Contact Mechanism

Contact, or an extended, face-to-face social interaction, between two antagonistic groups can promote tolerance and acceptance under certain circumstances, such as when groups have equal status, a common goal, cooperation, and authority support (Allport Reference Allport1954). Since the development of contact theory, it has been generalized to apply to many different types of groups (Pettigrew and Tropp Reference Pettigrew and Tropp2006), including the police (Rosenbaum et al. Reference Rosenbaum, Schuck, Costello, Hawkins and Ring2005; Viki et al. Reference Viki, Culmer, Eller and Abrams2006). Moreover, the theory has held even when Allport’s (Reference Allport1954) conditions are not met (Pettigrew and Tropp Reference Pettigrew and Tropp2006). I apply particular mechanisms of contact theory developed by Pettigrew (Reference Pettigrew1998) to the context of areas of limited statehood and citizen–state contact.

Pettigrew (Reference Pettigrew1998) develops contact theory by exploring the processes through which contact achieves improved perceptions and behavior toward an out-group or toward a social group with which an individual does not identify. In particular, Pettigrew (Reference Pettigrew1998) stresses that improvements in perception occur when contact takes on a relationship-building form. Relationship building means developing a mutual understanding or connection with the out-group member (police, in this case). The first interaction or contact is important for the development of a connection as it sets the stage for future interactions. Pettigrew (Reference Pettigrew1998) posits that perceptions change to become positive on the first interaction if (1) interactions lead to positive updating of negative priors and (2) if they facilitate social bonds.^{Footnote 5}

First, he suggests that contact with out-group members provides new information to the person that could update his or her prior beliefs about the out-group. Citizens learn new information about the out-group member as they interact with and observe the behavior of the out-group member. Individuals’ perceptions change when the out-group member’s behavior is starkly inconsistent with their prior negative beliefs about the out-group. As mentioned above, priors about the police (the out-group) tend to be negative. But if the police act in a manner that is inconsistent with people’s priors—they act friendly instead of abusive—then citizens may update their beliefs about the police to be more positive. In contrast, if the police continue to be predatory, then no updating occurs or the existing negative perceptions harden (Chong and Druckman Reference Chong and Druckman2010).

In addition to information updating, interactions could generate social bonds between in-group and out-group members (Pettigrew Reference Pettigrew1998). Social bonds form as initial anxiety about the interaction subsides over the course of the visit. Pettigrew (Reference Pettigrew1998) suggests that longer encounters and repeated encounters allow social bonds to form and deepen. Other studies have found that repetition may not be needed and that a single encounter can lead to empathy. Broockman and Kalla (Reference Broockman and Kalla2016), for example, find that during door-to-door canvassing, single ten-minute conversations reduced prejudice of transphobia for at least three months. This research suggests that extended interactions facilitate dialogue and empathy, which could humanize both actors and subsequently reduce anxiety.

Moreover, prior to the visits, opinions about the police may have formed in the abstract, but the door-to-door visits put a literal face to the police. Social bonds, after all, only form with an actual person. After the visit, when people think about the police, they are likely to think about the individual police officer that they met and about the interaction they had with that person. As such, the nature of that (one) interaction—whether it is information updating or leads to social bonds—matters for generating positive opinions about the police.

Yet, even if police interactions facilitate positive perceptions of individual police officers, these positive effects may not extend to the entire organization. However, Pettigrew (Reference Pettigrew1998) argues that the perception of the individual police officer generalizes to the entire group if the police officer is a “typical representative” of the group. In this case, a “typical” police officer is one who wears a police uniform. The uniform is used to suppress individuality and designate membership in a group (Joseph and Alex Reference Joseph and Alex1972). When wearing a uniform, the police officer adopts the uniform identify of the police. Thus, as long as police officers wear uniforms, they may be considered “typical” members of the group.

Importantly, “typical appearances” of police officers differ from the “typical behavior” of police officers. As mentioned above, citizens may understand the typical behavior of police officers to be abusive. As such, atypical, non-predatory behavior of police officers during social interactions could facilitate changes in opinion precisely because it is atypical. The uniform, however, demonstrates that this “new” behavior could be typical of other police officers in the organization.

In short, when individuals living in areas of limited statehood build relationships with police officers, they may be more likely to perceive the police positively because the encounter elicits perceptional change about the individual police officers with whom they met, and this perceptional change also generalizes to the police more broadly. This leads to the following hypothesis:

Hypothesis 1. When citizens living in areas of limited statehood engage in relationship-building interactions with police officers, they are more likely to develop positive perceptions of the police force.

Strengthening the Relational Contact Mechanism

Although Hypothesis 1 extends to any police officer, it is possible that certain types of police officers may be more likely to elicit positive perceptions than others. In particular, when police officers share demographic characteristics with the population or, more bluntly, when they look like the population, they may be more easily able to develop relationships with locals. The idea is similar to Pitkin’s (Reference Pitkin1967) concept of descriptive representation, in which political representatives, by virtue of looking like the population, are accorded some form of legitimacy. Mosher (Reference Mosher1982, 12) applies theories of descriptive representation to state agencies, suggesting that a state agency is “passively representative” if the composition of the police force mirrors the population in terms of race, social class, ethnicity, religion, gender, occupation, or other such characteristics.

Given the public’s negative priors about the police, the inclusion of police officers who look like citizens into the police force could signal that the police are willing to serve the minority (Meier Reference Meier1975; Pitkin Reference Pitkin1967). The visible, shared demographic characteristic of being female makes the female officer easily identifiable as someone who may have concern for other women or members of the minority group. Women may feel that female officers are less likely to behave in an abusive or corrupt manner toward them. Thus, by interacting with female officers, women’s priors about the police are positively updated. Within policing, scholars have found, for example, that when female officers are added, women are more likely to report crimes and perceptions of the police improve (Karim Reference Karim2017; Keiser, WilkinsMeier, and Holland Reference Keiser, Wilkins, Meier and Holland2002; Leger Reference Leger1997; Natarajan Reference Natarajan2016).

Additionally, social bonds may form more quickly when the police officer and the citizen share demographic characteristics. This is because it may be easier for those with shared demographic characteristics to find common ground during a social interaction. For example, women may be quicker to find common ground with other women than with men because they are more likely to share womanly experiences, such as pregnancy or childcare. Women’s conversations may differ from men’s, as women talk more about home and family, and use more emotive words (Haas Reference Haas1979; Hirschman Reference Hirschman1994). Females may develop more ways to express affection and interpersonal concern in their social interactions with members of their own sex (Hall and Braunwald Reference Hall and Braunwald1981). As such, conversations between women may be less anxiety-prone and more prone to empathy formation.

In theory, any type of shared demographic characteristic could lead to improved opinions about police authority, but visible characteristics such as race, ethnicity, and sex are more likely to do so because the similarity is identifiable during a social interaction. For the purpose of this paper, I focus on sex as the shared demographic characteristic, because there is less research on the effects of the integration of women into the security forces than on the integration of different ethnic groups into the security forces.^{Footnote 6} This means that when female police officers interact with local women, local women’s views about the police should improve. This leads to the following hypothesis:

Hypothesis 2. When women living in areas of limited statehood engage in relationship-building interactions with female police officers, they are more likely to develop positive perceptions of the police than if they interact with male police officers.

A question may arise as to whether female police officers are perceived as typical police officers. Because historically police officers have been men (including in Liberia), there is reason to believe that they are not perceived as representatives of the police. At the same time, their novelty could elicit attitudinal change among the population. However, it is not clear how attitudes of local men and women might change. On the one hand, the novelty of women in the police force could be looked on favorably, particularly by women, as mentioned above. At the same time, both men and women may have prior gender stereotypes about appropriate roles for women. As the provision of security is a predominantly masculine role, it is possible that women’s inclusion may lead to resentment of the police because female inclusion violates social norms. As such, the different sets of possibilities suggest that the sign of the effect could be indeterminate. For this reason, I test the effect of female visits on local females (H2), but I also include analyses on the effect of female visits on both male and female locals as a theory-building exercise.

The Procedural Justice Mechanism

In addition to the relational contact mechanism and the performance-based legitimacy mechanism, the procedural justice mechanism could also help explain how attitudes toward state agencies change as the state expands into areas of limited statehood. The mechanism posits that perceptions of fairness/unfairness during procedural interactions with state agents legitimize/delegitimize the state (Mcloughlin Reference Mcloughlin2018; Sunshine and Tyler Reference Sunshine and Tyler2003; Tyler Reference Tyler2006). A procedural interaction is a transactional one, in which either the state is carrying out its duty or citizens are seeking goods and services from the state. Procedural interactions could take the form of experiences with the court system, court judgments, police stops, traffic tickets, or any other interactions that involve some form of procedural decision made by a state agent (Huo and Tyler Reference Huo and Tyler2002). It could also take the form of goods and service provision, such as medical care at a public hospital or receiving social security or an ID card from the state.

The procedural fairness mechanism could be classified as another model of relational state building. In fact, Fisk and Cherney (Reference Fisk and Cherney2017, 267) classify it as such because procedural justice requires interactions between citizens and state agents. Yet, unlike the relational contact mechanism, the goal of the interaction is not to foster a connection between two individuals; rather, the state agent is responsible for a decision that impacts the citizen’s life; it has some degree of power over the citizen. Additionally, the state agent is responding to a citizen rather than seeking out the citizen to build a relationship before it provides a good. Thus, whereas the relational contact mechanism is a proactive and non-hierarchical model of relationship building, the procedural justice mechanism is a reactive, hierarchical model of relationship building. To improve opinions, the procedural justice mechanism requires state agents to be perceived as fair in their decision making during such interactions (McLoughlin Reference McLoughlin2015; Rothstein Reference Rothstein2009; Tyler Reference Tyler2006). This leads to the following hypothesis:

Hypothesis 3. When citizens living in areas of limited statehood perceive police unfairness, they are less likely to develop positive perceptions of the police force.

Experimental Context: Grand Kru County

The field experiment in Liberia was conducted in one of Liberia’s most isolated counties: Grand Kru County.^{Footnote 7} Civilians who live in the county have very little interaction with the state and generally do not perceive the police favorably. The county has historically been neglected by the state, and at the end of the war, it was considered a “forgotten county.”^{Footnote 8} In 2014, there were only 19 LNP officers assigned to the county, which had a population of 57,106 people. In 2011, only 16% of the population of the county had interacted with a police officer, which was one of the lowest percentages out of all the counties in Liberia (Vinck, Pham, and Kreutzer Reference Vinck, Pham and Kreutzer2011).

Although the county has limited state presence, like other rural counties in Liberia and other areas of limited statehood, the county possesses alternative sources of governance such as chieftaincies, secret societies, and other traditional structures that compete with the state’s authority.^{Footnote 9} Moreover, even though the UNMIL mission was present in Liberia, it had a minimal footprint in Grand Kru County. As such, Grand Kru County is representative of rural counties in which there is some minimal international presence, making it representative of most rural subunits in countries with peacekeeping missions (Ruggeri, Dorussen, and Gizelis Reference Ruggeri, Dorussen and Gizelis2018). Finally, Grand Kru’s relative homogeneous ethnic balance of Kru and Grebo made it possible to implement the experiment without worrying about the possible confounding effects of ethnic heterogeneity. It is, thus, representative of other sub-Saharan countries, of which many subunits are ethnically homogeneous (Gershman and Rivera Reference Gershman and Rivera2018). Overall, the county provided the conditions necessary for the experiment: an area of limited statehood with competing governance structures where the state is expanding and where perceptions about the police are generally negative.

The Ethics of Doing Fieldwork (with Police Officers) in Areas of Limited Statehood

Lake and Cronin-Furman (Reference Lake and Cronin-Furman2018) have recently called on scholars who engage in fieldwork to incorporate a section on ethics in the research design section of papers. Although IRB approval is the minimal requirement necessary to meet ethical standards,^{Footnote 10} this field experiment necessitated more careful ethical considerations because of the nature of the manipulation, the research context, and safety concerns of the enumerators. First, it would have been unethical to introduce state expansion through the field experiment into a county and then discontinued the services after the implementation of the experiment, as it would have raised expectations about the police that the state could never meet. To remedy this, the experiment was rolled out just before the LNP’s plan to expand police into the rural counties began. Moreover, the field experiment required my research team (including police officers) to spend three months in the field, often walking long distances and potentially getting sick when they would be far away from a health clinic. Thus, as a team, we created a safety plan and determined the conditions under which the research would be delayed or stopped. Finally, it was important to ensure that introducing law enforcement to communities (that have little prior interactions with them) would not bring harm to them or to the police officers. This was important for the experiment itself as one of the mechanisms required that the police behave in a non-predatory manner. The police officers in my team were trained in community policing techniques. Additionally, one enumerator was assigned to each police team, which meant that they monitored not only the correct implementation of the experiment but also the behavior of the police. Although the enumerators’ presence perhaps compromised external validity, in this case it was necessary to ensure that officers did not harm citizens and that citizens did not harm police officers.

Treatments

The above sections provided three different mechanisms for how expansion of police presence into areas of limited statehood might affect perceptions of the police. I experimentally test the relational contact mechanism (H1 and H2) and observationally test the procedural justice mechanism (H3). The treatment for the relational contact mechanism was a 20–30 minute, face-to-face household visit by police officers. The treatment used to test H2 was a 20–30 minute, face-to-face household visit by female police officers compared with a 20–30 minute, face-to-face household visit by male police officers. In total, two male officers and two female officers (four total) were selected for the study and memorized a scripted visit, which ensured that the household visits were similar in content.^{Footnote 11} Officers were required to wear their uniform and introduce themselves as the LNP, so that there was no confusion about who they represented. Mixed-sex teams were not possible for two reasons. First, testing H2 required all female (and all male) teams of police officers. Second, there were insufficient funds to add a third treatment arm.

Officers for the project were chosen based on similarity in their personalities. This was done to ensure that specific personality traits would not drive the treatment effect. The LNP in Monrovia provided 15 Kru/Grebo-speaking officers for the study. I then filmed each of the 15 officers giving the speech to an audience, as if they were engaging in community policing. The videos were then shown to a group of Kru/Grebo people in Monrovia who rated the officers on different characteristics.^{Footnote 12} Four officers were chosen (two female and two male) based on similarity in personality scores both across and within teams. The officers also engaged in one month of intensive training, including a pilot test in a different county, to ensure that they behaved the same way during the visits.

At the end of the visit, the officers gave the household cards with phone numbers of the local police. These cards were later used to identify whether individuals within the households had been treated. The officers spent between one and three continuous days in each village and were required to remain in their designated neighborhoods during the entire time.

Experimental Design

The research design used an intent-to-treat design. When the police officers visited the households, they treated the entire household. That is, they spoke to and interacted with all the members of the household who were present at the time of the visit. Individuals randomly selected for the endline survey were 18 and over.^{Footnote 13}

According to the 2008 census, 43 villages included 80 households or more. From these, 15 were randomly selected for the study and constitute a representative sample of the county. The Online Appendix provides descriptive comparisons of the villages selected and not selected for the study. Within each village, a randomly selected number of houses were visited by female, male, and no police officers (see Figure 1). Within each village, randomization occurred at the neighborhood level. When the team, which included four police officers and three enumerators, arrived in the villages, they asked for a list of the neighborhoods in the village. On average, in the 15 villages selected for treatment, each village included four neighborhoods, and each neighborhood contained about 56 households. Neighborhoods were randomly assigned to the female police officer treatment, male police officer treatment, and the control group. The target for each treatment group was to reach 25 households. They used a random walk technique to randomly choose houses in the randomly assigned neighborhood. The total sample size consisted of 750 households, 375 visits by male officers, and 375 visits by female officers. Table 1 provides the research design for the experiment.

FIGURE 1. Research Design

TABLE 1. Research Design for Treatment

Outcomes

Three weeks after the officers’ visits, the Center for Applied Research and Training (CART) surveyed the treated households. The survey included questions that ranged from maternal healthcare to security. The enumerators presented themselves as an independent agency doing a survey on health and security in the county. The survey was conducted three weeks after the treatment for both theoretical and practical reasons. Theoretically, because there were no systematic repeat visits, the treatment effect was likely to decay quickly. The goal was not to capture decay or longer-term effects but rather to better understand whether opinions changed at all in the short term when exposed to a first-time relationship-building encounter. Second, the surveys needed to be completed before the rainy season, as it would have been too dangerous for the enumerators to travel during that time. The enumeration team surveyed the exact same households as those that were visited by police officers. They were able to recognize the households based on maps that were drawn by the enumerators who were a part of the police team, and by a code that was written discreetly adjacent to the household. All treated households were located.

To avoid spillover, the control group survey was conducted at the same time as the treatment was implemented, which minimized the number of treated household members communicating to those in the control group about the police visits. The time lag in survey implementation, however, presents several problems. First, it may have caused a different group of people to be available for the survey due to farming and travel schedules. However, the enumerators used the same sampling method for both sets of surveys to ensure that all household members had an equal chance of being selected for the survey. Second, individuals who were surveyed earlier may have told other members of the community about a survey, thereby priming them about the survey prior to the implementation of the treatment group survey. To account for this suspicion, the Online Appendix provides models that include those who were suspicious about the survey as a control as well as a control for whether other people were nearby at the time of the survey.

Specific to this study and consistent with a dyadic and monadic understanding of legitimacy, I assessed two types of questions: preferences for the police over non-state and external actors (dyadic) and positive perceptions of the police (monadic). To address the former, I included questions about preferences for the police to respond to two security crisis situations: a community dispute (hala hala) and mob violence. I coded whether respondents preferred the police over UNMIL, community leaders, community watch teams, traditional leaders, or others.^{Footnote 14}

Spillover and H3

In addition to lagging the time of the surveys, to prevent spillover, the treatment was assigned at the neighborhood level. Neighborhoods are generally far apart in rural villages, which means that those in a control neighborhood would not have interacted with police officers in a treated neighborhood. The police officers also spent the night in their designated neighborhood, so that they would not be seen in other parts of the village at night.

Despite these precautions, there is evidence of a violation of the stable unit treatment value assumption (SUTVA). Specifically, 45 individuals in the control group had knowledge about the police visits. The spillover group, however, allowed me to observationally test H3—the procedural justice mechanism. The procedural justice mechanism requires the police to make decisions about procedural services and requires that the public perceive the procedure in a fair way. This did not occur with the randomization process. The spillover group felt they did not get equitable treatment from the police. They felt that they were treated unfairly because they did not get a household visit; they felt as if they were missing out on an important government service.^{Footnote 17}

Models

In addition to presenting the results in a regression framework, I present the Average Effect Size (AES), which measures across all questions within each cluster (prefer police, abuse, and effectiveness), following the procedure proposed in Clingingsmith, Khwaja, and Kremer (Reference Clingingsmith, Khwaja and Kremer2009) and Kling, Liebman, and Katz (Reference Kling, Liebman and Katz2007). The AES across J related dependent variables is given by $\tau = \sum\limits_{j = 1}^J {{{{\pi _j}} \over {{\sigma _j}}}}$, where π _j is the average treatment effect on each dependent variable and σ _j is the standard deviation of dependent variable j in the control group. To test the null hypothesis of no average effect, the effects π _j are jointly estimated using a seemingly unrelated regression. The J dependent variables are stacked to compute a variance–covariance matrix for testing the statistical significance of τ, the AES.^{Footnote 18} AES coefficients are interpreted in terms of standard deviations from the control group mean. All models are presented with standard errors clustered at the neighborhood level. In addition to AES, I present a logit model to test Hypothesis 3, which treats the data as observational.

Response Bias and Internal Validity

It is possible that respondents may have felt the need to say positive things about the police out of fear of their return. The enumerators collected data on the reactions of the participants at the beginning, during, and at the end of the visit. On average, the enumerators recorded that about 24% of the participants were anxious at the beginning of the visit, but that 99% of the participants were appreciative of the visit after it was over and that 96% wanted the police to return to their house in the future. This suggests that by the end of the visit, participants in the study were not scared of the police and that they likely did not answer questions on the survey due to fear of the police. Furthermore, the interactions with the police decreased anxiety about the police over the course of the visits, providing evidence of the social bond mechanism that is necessary for relationship building.

Additionally, to verify that the relational contact mechanism contributed to improved perceptions and not the content of the speech, which included information about police professionalization, I conducted a survey experiment in a different set of villages in Grand Kru County. Enumerators read the same script to randomly selected citizens during the implementation of a survey and measured the same responses. The results show that this information actually led to a decrease in preferring the police to respond to security crises and a decrease in perceptions of police effectiveness (see Online Appendix). There was a small decrease in perceptions of abuse but not to the same degree as the police visits. These findings confirm that it was not the content of the speech, but rather the relational dynamics of the visit that helped change perceptions.

Randomization Checks and Other Threats to Inference

The data for the randomization checks come from the surveys implemented posttreatment. A baseline survey was not possible because it would have aroused suspicion among community members.^{Footnote 19} Members of communities are not used to receiving visits by strangers. About 62% had never interacted with community outsiders in their own village. Due to the novelty of strangers visiting villages, the study took great care to disassociate the police visits from the survey—adding another survey would have made it difficult to decouple the surveys from the police visits.

Despite randomizing treatment at the neighborhood level, balance was mixed (see Online Appendix). Compared with those that did not experience visits by police, those that experienced visits by female officers tended to be less Christian, own land, farm, and have a higher number of household members. Balance was better between those that experienced visits by male officers and the control group. Those that experienced visits by male officers were less likely to be traditional leaders and more likely to have a larger household. Imbalance may be because the randomization was done at the neighborhood level. Some neighborhoods may have had higher numbers of land owners, farmers, Christians, larger families, or traditional leaders. However, the imbalanced covariates are unlikely to influence perceptions of the state security forces, as they are not common variables in predicting police perceptions. Nevertheless, they are included in the models below.

Finally, there may be an issue of multiple testing as I include multiple dependent variables, two treatments, and assess the heterogeneous treatment effects of respondents’ sex. To account for this, I provide adjusted p values in the Online Appendix. I also present AES of the dependent variables, which mitigates Type I and II errors (Clingingsmith, Khwaja, and Kremer Reference Clingingsmith, Khwaja and Kremer2009; Kling, Liebman, and Katz Reference Kling, Liebman and Katz2007).