Area Differences and Time Trends in Crime Reporting: Comparing New York with Other Metropolitan Areas

Abstract

Police measures of crime are shaped by victims’ decisions to notify the police. To obtain a better understanding of US crime trends, this study uses the National Crime Victimization Survey to examine geographic differences and temporal trends in crime reporting in New York and other metropolitan areas for the period 1979-2004. We find that net of crime characteristics and survey methodology, the New York metropolitan area showed fewer increases in crime reporting than did other metropolitan areas. These divergent trends suggest that the real differences in the drop of nonlethal violence between New York and other areas may have been smaller than those indicated by police-based crime statistics. We also find that from the early 1990s to 2004, New York showed a sharp decrease in the likelihood of victims perceiving that “police wouldn’t help.” This trend suggests that police reforms in New York City have not resulted in more victims using police-related reasons to explain their nonreporting behavior. Instead, researchers need to develop a broader theoretical framework (not an exclusive focus on police actions) to understand how police- and nonpolice-related factors may explain the geographic variation in the trends of reporting observed in this study.

Keywords:

• victimization
• police notification
• crime-reporting trends
• National Crime Victimization Survey (NCVS)
• New York City

Notes

1. In the NCVS, most variables had little (<1%) or no missing data. The variable with the largest percentage of missing data was household income (13%). Unknown information about offenders was another reason for missing data in the violence sample (<5%). To assess bias that might result from missing data, we compared incidents with and without missing data and found that these incidents had similar characteristics, except that incidents with missing income had a slightly higher percentage of nonwhite victims, and that incidents with missing offender information had a slightly higher percentage of robbery; the differences were statistically significant, but the differences were small. Because the probability of having missing data was not associated with the probability of calling the police (we found no statistically significant difference in the reporting rates), deleting the incidents with missing data would not bias the regression estimates (Little, 1992). Indeed, we compared three analysis strategies (a complete data analysis, an analysis with dichotomous variables indicating cases with missing data, and an analysis based on multiple imputation in which missing values in household income were replaced with five sets of plausible values based on variables with full information) and found that the conclusions were essentially the same. We, therefore, reported the results from the complete data analysis (also see Baumer & Lauritsen, 2010).

2. The final weights for the incidents were the product of two weights: the census-generated incident weight and the redesign weight. Baumer and Lauritsen (2010) detailed the procedures used to create the weights. Conceptually, the redesign weight was used to weight down police-known incidents in the NCS. This adjustment removed the NCS–NCVS differences in the likelihood of reporting caused by the redesign.

3. In the analysis, we examined the full sample of victims, both nonreporters and reporters. Reporters were coded 0 on the dependent variable; that is, we assume that victims who reported to the police would expect the police to take their reports seriously and to offer needed help. In unreported analysis, we also analyzed the data by focusing exclusively on the nonreporters. The analysis yielded similar conclusions regarding the patterns of temporal changes in victims’ perceptions of the police. When the models were reestimated with multiply-imputed income, the interpretation of the results did not change.

4. Because the NCVS uses a stratified, multistage cluster sampling design, our analysis used survey design variables (pseudostratum and secucode, both provided by the Census Bureau) to account for clustering of the data in the complex sampling design. All models were checked for multicollinearity and no problems were found. In the data, all bivariate correlations were lower than .6 and all variance inflation factors were lower than 4. To avoid collinearity between the linear and nonlinear time trends, time was centered in our analyses at the midpoint of the observation interval (the year of 1992).

5. In order to conserve space, Table 1 displays only the results for the key independent variables. The results for the control variables are not listed but are available upon request.

6. The conclusions were the same whether we analyzed the full sample of victims or we restricted the analysis to only the nonreporters (see footnote 3).

Additional information

Notes on contributors

Min Xie

Min Xie is an assistant professor in the School of Criminology and Criminal Justice at Arizona State University. She received her BA and MA in information management from Beijing University, China, and her PhD in criminal justice from University at Albany, SUNY. Her research interests include theories of criminal victimization, race, ethnicity and gender issues, multilevel and longitudinal models, and spatial data analysis. Her work has appeared in Criminology, Journal of Quantitative Criminology, Justice Quarterly, and Homicide Studies. In 2012 she was co-recipient of the Ruth Shonle Cavan Young Scholar Award from the American Society of Criminology. She recently received funding for her research on immigration and violent victimization from the National Institute of Justice, the American Society of Criminology, and the Bureau of Justice Statistics.