To assess the linkages between public encounters with the police and attitudes towards legal authorities, we draw on both waves of data from The Crime, Safety and Policing in Australia Survey (Murphy et al. 2010a, b). In 2007, a nationally representative mail survey of adults in Australia was conducted on the extent of respondents’ experiences and beliefs about crime and policing in their community. Participants were drawn from Australia’s publicly available electoral roll. All electors aged 18 + in Australia are required by law to register their name and home address on the roll; the roll thus consists of a representative subject pool. Selection was stratified by State and Territory jurisdiction. 5700 residents were sent a survey booklet by mail—after several reminders and weeks, an adjusted response rate of 40% was achieved (n = 2120). Two years later, a follow-up panel survey was undertaken in 2009 aiming to examine whether attitudes and experiences of crime and policing had changed over the two-year intervening period. After a series of reminders, a total of 1190 usable responses were returned. Considering the adjusted response rate (for respondents who had died or moved address between waves 1 and 2), an attrition rate of 35% was achieved.Footnote 8
Some 46% of the respondents in the final sample were male, 52% had post-secondary education; respondents were on average 54 years of age in 2007 and had an average self-reported household income of approximately AUD$80,000. Using the 2006 Australian Census data as a benchmark, we conclude that older and more educated respondents in the sample were slightly over-represented, but overall the sample was largely representative of the Australian population. For instance, 49% of the population are male, 46% had post-secondary education, and the average household income is approximately AUD$54,000.
Measures and Measurement Model
Following previous work (Bradford et al. 2009; Jackson et al. 2012) we unpack perceived trustworthiness into trust in procedural fairness and trust in police effectiveness. We also consider one part of people’s judgements about the legitimacy of the police: namely, a normatively grounded sense of duty to obey the police. All questions were measured on a five-point Likert scale ranging from “strongly disagree” to “strongly agree” (except the indicator of perceived police effectiveness, which ranges from “very poor job” to “very good job”). The exact wording for each question asked, their assigned latent construct, and their descriptive statistics can be found in Table 1.
Unless otherwise mentioned, all measures were coded in such way that higher values indicate more positive evaluations of the construct measured. In order to assess the scaling properties and empirical distinctiveness of our measures, we use confirmatory factor analysis (CFA) with categorical (ordinal) indicators, focusing on the two components of trust and felt duty to obey the police commands. As expected, questions tap into the three theorized constructs sensibly, indicating that our empirical indicators can be empirically distinguished in the three premised dimensions. A full analysis of the empirical distinctiveness of the three latent constructs and their equivalence across waves can be found in the Appendix (A.1 and A.2, respectively).
Measuring Attitudes Towards Encounters with the Police
Most previous work on the impact of police contact on perceived trustworthiness assumed a single dimension of satisfaction with the encounter. Skogan (2006), for instance, used six questions evaluating respondents’ perceived politeness, helpfulness, and fairness of police officers to create an indicator of positive versus negative encounter. In the same vein, Slocum et al. (2016) asked respondents the extent to which they were satisfied with the encounter and created an indicator with three groups—dissatisfied, neutral, or satisfied with the encounter. However, there might be more than one underlying dimension of contact evaluation, and we distinguish between respondents’ perceived fairness in the procedures used in the contact and their satisfaction with the outcome’s favorability (Bradford et al. 2014; Tyler and Fagan 2008).
In the wave 2 survey, respondents were asked how many times they had contact with police in the previous 12 months. Some 38% (\(n = 440\)) of wave 2 respondents had at least one encounter (i.e., encounters that happened at some point in between the first and the second waves of data collection)—57% of those contacts were citizen-initiated, while 43% were police-initiated (mostly involving some type of police stop). The sub-sample of respondents who did experience an encounter with the legal officials at some point in between waves 1 and 2 were further asked five follow-up questions evaluating the process and four follow-up questions evaluating the outcome favorability of such encounter—these questions can be found in Table 1.
In order to confirm that process and outcome evaluations are indeed empirically distinguishable as suggested in the procedural justice literature, we first use CFA to assess the scaling properties of the nine indicators simultaneously. Considering only the 440 respondents who experienced at least one encounter with the police between waves 1 and 2, we fit two models with one and two factors and all indicators set as categorical. Despite being correlated (r = 0.66), the two theorized dimensions—process and outcome evaluations—seem to be empirically distinguishable given that the two-factor solution has the best model fit. We then fit a third two-factor CFA model after dropping one of the questions measuring process evaluation as it was weakly correlated with the latent constructFootnote 9—we thus use four questions to measure process evaluation and four questions to measure outcome evaluation in all subsequent analyses. Standardized factor loadings, model fit statistics, and a full account of the measurement models can be found in the Appendix A.3.
After confirming that we can treat process and outcome evaluations as two distinct constructs, for each dimension we need to somehow classify respondents’ most recent encounter with the police as positive or negative, so that their scores for trustworthiness and legitimacy can be compared with the scores of the respondents who did not have any recent contact with police. One solution for that is handpicking the ‘positive’ and ‘negative’ categories based on responses on the Likert scale. However, cutoff decisions for this formative approach are arbitrary. Instead, we adopt a data-driven approach and fit latent class models on items concerning respondents’ evaluation of their most recent encounter regarding both process and outcome. This approach is preferable as it permits the modeling of unobserved heterogeneity underlying the two dimensions of contact evaluation (see Na et al. 2015; Nylund et al. 2007).
For each process evaluation and outcome evaluation, we fit three models with two, three, and four latent classes. In both cases, the three-class solution emerged as the preferred solution—see details on deciding the number of classes in the Appendix A.4. Roughly, the three classes indicate negative, neutral, and positively experienced contact with police. Our interpretation is that these classes represent encounters that went ‘worse than expected’ (negative), ‘as expected’ (neutral—note that this category includes encounters rated ‘OK’), or ‘better than expected’ (positive). It is important to be clear what we mean here. Considering that Australia is by and large a high-trust country (see Table 1; see also Hinds and Murphy, 2007), we assume that expectations of the behavior of officials who represent key institutions in society are broadly positive. Most people expect to be treated relatively well by police officers, and for police to achieve positive outcomes more often than not. Encounters that go ‘OK’ are therefore experienced as ‘neutral’, since that is what is expected. It is only when they are better than expected that the experience shifts into the positive. Given this assumption, it is not surprising that the second class—the ‘neutral’ group, when encounters go as expected—is composed of respondents who mostly answered “agree” with the statements (as opposed most pertinently to “strongly agree” to every question posed to them). By contrast, an encounter with police wherein respondents classify most of the indicators as anything lower than the fourth point in the Likert scale (e.g., “neither agree nor disagree” with a given statement, or worse) we assume to indicate a negative (i.e., worse than expected) contact, since ‘neither/nor’ indicates at best uncertainty about whether police behaved in line with expectation. Mostly ticking the fifth point to answer the questions (i.e., “strongly agree”) would indicate an encounter that went better than expected (i.e., positive)—see results in probability scale in Figs. 1 and 2.
Looking at the derived classes, in relation to process evaluation, 54% of the sub-sample of respondents who did have recent contact with police had ‘negative’ experiences, 37% had ‘neutral’ experiences, and 9% had ‘positive’ experiences. Regarding outcome favorability, 34% had ‘negative’ experiences, 50% had ‘neutral’ experiences, and 16% had ‘positive’ experiences (Figs. 3, 4).
In order to assess the relationship between police-citizen encounters and public attitudes toward the police, we use a similar strategy as the one commonly found in the literature—i.e. we compare positive, neutral, and negative encounters with the group of respondents who did not experience contact with police as the baseline category. Unlike almost all previous studies, however, we now focus on changes in attitudes towards the police over time, and distinguish perceived quality of police behavior in encounters in terms of process and outcome. We thus fit two autoregressive structural equation models (SEM), one assessing the association between process evaluation and changes in trustworthiness and legitimacy and the other between outcome evaluation and changes in trustworthiness and legitimacy.Footnote 10 This dynamic panel model permits the modeling of change in attitudes towards the police because of the inclusion of lagged dependent variables as covariates.Footnote 11 It also allows us to investigate the extent to which prior attitudes are associated with positive or negative perceptions of contact with legal officials.
Based on the theorized diagram shown in Fig. 1, we fit two autoregressive SEMs: (a) one with the four ‘contact’ groups (i.e., no contact; negative contact; neutral contact; positive contact) indicating process evaluation (testing the hypotheses 1A, 1B, and 1C) and (b) another with the four ‘contact’ groups indicating outcome evaluation (testing the hypotheses 2A, 2B, and 2C). In both models, all hypotheses are tested keeping the group of respondents who had no encounter with the police between waves 1 and 2 as the reference group—i.e. three dummy variables indicating negative, neutral, and positive contact are displayed. All coefficients are standardized, which allows for comparisons.
Each model includes three aspects of interest. First and foremost, arrows departing from each of three dummies indicating negative, neutral, and positive contact reflect expected changes in attitudes depending on the type of contact with police in relation to people with no contact. This is the crucial aspect of the models and is used to test hypotheses 1A, 2A, 3A, 1B, 2B, and 3B. Second, both models include a set of autoregressive parameters—the arrows departing from attitudes towards the police before an encounter to attitudes after an encounter. Those parameters speak to the stability of the variables—as psychological constructs, perceived police trustworthiness and legitimacy are expected to be highly stable over time. Finally, both models include multinomial logistic paths regressing contact evaluation on T1 measures of trustworthiness and legitimacy. This aspect of the models account for different odds of having an (un)satisfactory contact with police given different prior attitudes. For the multinomial logistic paths only, we use the ‘neutral contact’ group as the reference category—we are thus estimating the association between prior attitudes and the odds of having a positive or a negative encounter in comparison with a neutral encounter.Footnote 12 Both models include age at T1 (difference between 2007 and year of birth), gender (1 = male), and national identity (1 = Australian non-Aboriginal) as control variables.