Robustness of collective efficacy on crime in a developing nation: association with crime reduction compared to police services

Abstract

This study examines police, collective efficacy and crime in the developing country of Trinidad and Tobago to test the generalizability of collective efficacy as protection against crime and to compare the relative influence of public perceptions of police activity versus collective efficacy on crime. The results support that collective efficacy can be generated in a developing country and that its depressive effects on crime are replicated even in that context. The study further shows that police can contribute to collective efficacy and subsequently lower the levels of crime by minimizing police misconduct and improving the quality of services, although the strongest effect of the police variables was between police misconduct on crime. In sum, residents in Trinidad and Tobago, a developing nation with many challenges, report experiencing collective efficacy, that more of it means less crime, and that public impressions about police services and misconduct may affect both collective efficacy and crime.

Keywords:

• collective efficacy
• police misconduct
• police service quality
• public perceptions of police
• crime

Acknowledgement

Funding for a portion of this research was provided by the Ministry of National Security of Trinidad and Tobago. The points of view expressed in this paper are those of the author alone and do not represent the official policies or positions of the Ministry of National Security or the Trinidad and Tobago Police Service. The author thanks Stephen D. Mastrofski, David B. Wilson, Roger Parks, and Devon Johnson for feedback on this research and earlier versions.

Notes

 1. These statistics are generated based on an analysis of the data from the 43 countries that completed the full questionnaire from the World Values Survey wave 4 (2005–2008), retrieved from the World Values Survey Association at http://www.worldvaluessurvey.org on 30 January 2009. The list of countries include Andorra, Argentina, Australia, Brazil, Bulgaria, Burkina Faso, Chile, China, Cyprus, East Germany, Ethiopia, Finland, Ghana, India, Indonesia, Italy, Japan, Jordan, Malaysia, Mali, Mexico, Moldova, Morocco, Peru, Poland, Romania, Rwanda, South Africa, Serbia, South Korea, Slovenia, Spain, Sweden, Switzerland, Taiwan, Thailand, Trinidad and Tobago, Turkey, Ukraine, United States of America, Vietnam, West Germany, and Zambia. Surveys in the United States and in Trinidad and Tobago were conducted in 2006.

 2. Although the collective efficacy measure used in this study is not the same as the measure for the Chicago studies (I used a second-order factor score), it does rely on similar indicators of informal social control and social cohesion and so conceptually is comparable. To facilitate a reasonable comparison of the levels of collective efficacy in Chicago versus Trinidad neighborhoods, I created a collective efficacy score using a strategy consistent with Morenoff et al. (2001) and Sampson et al. (1997). They averaged responses to ten 5-point likert questions to form their measure of collective efficacy. Their neighborhood mean was 3.89 out of 5. I computed a comparable Trinidad collective efficacy score by averaging responses to six 4-point likert questions. The neighborhood average in Trinidad was 2.94 out of 4. Converting these scores to percentages of the range of possible scores for each study, the Chicago neighborhoods scored 77.8% and the Trinidad neighborhoods scored 73.4%. I also compared the level of variation across neighborhoods in the two locales using the coefficient of relative variation (CRV) outlined by Dunham and Alpert (1988). The coefficient of relative variation is computed by dividing the standard deviation by the mean. This computation provides the standard deviation standardized by the size of the mean. This allows an assessment of the amount of variation across scales, controlling for differences in the size of the means (Dunham and Alpert 1988, p. 513). Both studies show low levels of variation across neighborhoods (CRV for Chicago = .06, CRV for Trinidad = .05).

 3. For the analyses, I excluded a community with only one interview and one interview with missing community identification information. Therefore, the total number of interviewees available for analysis is 2967, and the number of communities is 72.

 4. The CIA (2008) reports that the population Trinidad and Tobago contains approximately 40% Indo-Trinidadian, 37.5% Afro-Trinidadian, 20.5% mixed race, and 1% other, and fewer than 1% unknown.

 5. Since the selection of respondents within neighborhoods is random, the racial characterization of the neighborhoods is computed based upon the percentage of the respondents claiming to be African, Indian, Mixed, or Other. A neighborhood is characterized as predominantly African when 60% or more of the respondents from that neighborhood report being Afro-Trinidadian. Likewise, a neighborhood is characterized as predominantly Indian when at least 60% of the respondents in that neighborhood report being Indo-Trinidadian. The remaining neighborhoods are characterized as mixed.

 6. Of course, another sample might suggest different results regarding which effects are relatively stronger.

 7. I was not able to separately model the two levels. The data showed too little variance at the neighborhood level to simultaneously replicate key variables at the individual and neighborhood levels. Furthermore, there are more parameters of interest than the number of neighborhoods surveyed, prohibiting a two-level model.

 8. Six questions are imitations of questions in the PHDCN. Sampson et al. (1997) used five of the questions in their analysis. Reisig and Cancino (2004) and Morenoff et al. (2001) applied similar measures.

 9. The dataset contains 2967 cases. However, the structural equation model dropped 43 cases from the analysis due to missing data on all independent variables and one case with missing data on all dependent variables. Descriptive statistics are provided for the 2924 cases in the analysis. For the analysis, when some dependent variables for a respondent are missing data, a robust weighted least squares estimator (WLSMV) is used. According to Brown (2006, p. 76), when at least one factor indicator is ordinal, as with this dataset, the weighted least squares or robust weighted least squares estimators are more appropriate than normal theory maximum likelihood.

10. Brown (2006) explains that Cronbach's alpha misrepresents reliability when tau equivalence is not upheld (tau equivalence is achieved when the indicators have equal loadings, but different error variances). Raykov (1997) found that when tau equivalence is not upheld and the item loadings of a factor differ by more than .2 and when one or more loadings fall(s) below .6, alpha does not perform well, seriously underrepresenting reliability. Raykov provides an alternative estimate of reliability that reconciles the problems within the context of the confirmatory factor analysis measurement model. The formula is the true score variance divided by the total variance. When measurement errors are not correlated, the equation can be expressed as the squared sum of the unstandardized factor loadings divided by the sum of the squared sum of the unstandardized factor loadings and the sum of the unstandardized measurement error variances (Brown 2006, p. 338). This formula was used to compute Raykov's reliability estimate using Mplus. For the second-order factor, collective efficacy, all of the root observed indicators are used (Raykov 2009).

11. For comparison purposes, the mean population density of the PHDCN Chicago neighborhoods was 7530 per square kilometer (Sampson and Raudenbush 1999). The mean poverty level in the Chicago neighborhoods was 21% and the mean percentage black was 36%. I did not find any reference to actual estimates of residential stability in Chicago within the series of works by Sampson and colleagues that rely on PHDCN data. However, Silver and Miller (2004) report that Chicago residents had an average of .91 moves within the five years prior to the survey, suggesting residents were considerably more mobile in Chicago than in Trinidad and Tobago.

12. Model fit indicators provide slightly different perspectives on model fit (e.g., absolute fit, close fit, comparative fit relative to the null model) and are differentially impacted by various aspects of the analysis, such as ‘sample size, model complexity, estimation method (e.g., ML [maximum likelihood], WLS [weighted least squares]), amount and type of misspecification, normality of data, and type of data.’ (Brown 2006, p. 86) As a result, some estimators may not always agree on how well the model fits the data. Given the somewhat inconclusive nature of the available model fit statistics, Brown recommends using model fit as only one indication of the appropriateness of the model, applied in combination with information such as statistical significance, absence of negative residual variance (Heywood cases), and direction and size of the parameter estimates relative to the predicted direction and size. In this case, where the inconsistent statistic is the WRMR, it is worth noting that one of the designers of the MPlus software, a well-regarded statistician, reports that the WRMR is not a well-studied it statistic and has not performed as well as expected (Muthen 2008). CFI and TLI values greater than or equal to .95 represent a good fit, and between .90 and .95 are acceptable model fit ranges. RMSEA values less than .05 depict good fit, less than .08 represents a reasonable fit, and RMSEA values .08 to less than .1 are considered mediocre fit. (Brown 2006) WRMR values less than or equal to 1.0 represents a good model fit (Yu 2002).

13. Trinidad residents were not asked whether they had the capacity to move, so interpreting residents’ longevity as immobility may be problematic. However, in some of the neighborhoods where residents are squatters or in areas with extreme poverty, this may be a reasonable assumption.