Examining Systematic Crime Reporting Bias Across Three Immigrant Generations: Prevalence, Trends, and Divergence in Self-Reported and Official Reported Arrests

Abstract

Objective

Mounting evidence reveals that foreign-born, first generation immigrants have significantly lower levels of criminal involvement compared to their US-born, second and third-plus generation peers. This study investigates whether this finding is influenced by differential crime reporting practices by testing for systematic crime reporting bias across first, second, and third-plus generation immigrants.

Methods

This study draws on data from the Pathways to Desistance Study, a longitudinal investigation of the transition from adolescence to young adulthood among a sample of serious adolescent offenders. Self-reported and official reports of arrest are compared longitudinally across ten waves of data spanning 7 years from adolescence into young adulthood for nearly 1300 adjudicated males and females.

Results

This study reveals a high degree of correspondence between self-reports of arrest and official reports of arrest when compared within groups distinguished by immigrant generation. Longitudinal patterns of divergence, disaggregated by under-reporting and over-reporting, in self- and official-reports of arrest indicated a very high degree of similarity regardless of immigrant generation. We found no evidence of systematic crime reporting bias among foreign-born, first generation immigrants compared to their US-born peers.

Conclusions

First generation immigrants are characterized by lower levels of offending that are not attributable to a differential tendency to under-report their involvement in crime.

Appendix

The longitudinal analysis of individuals is complicated by a number of issues. First, observations gathered over time are nested within each individual violating the independence assumption of ordinary least squares regression which affects the estimation of standard errors and increases the likelihood of obtaining a type 1 error. Second, because individuals display at least a modest level of behavioral stability over time, individuals are more similar to themselves across observations than they are to other individuals (Osgood 2009). Third, the variable number and spacing of observations across respondents results in an unbalanced model. HLM is flexible and can accommodate these factors (Raudenbush and Bryk 2002; Snijders and Bosker 1999).

Equations are specified using either Poisson (count dependent variables) or Bernoulli (dichotomous dependent variables) extensions of HLM. In the interest of space, we detail here the equations using a Poisson model; modeling decisions are consistently applied. We allow for overdispersion in each model as the addition of the overdispersion parameter has been shown to result in more accurate significance tests (Osgood 2000). To examine offending trajectories over time we use a two-level hierarchical model where offending is measured at level 1 and includes repeated measures of self or official reports of arrest for individuals across 10 waves. The level 1, within individual equation is:

$$\eta_{it} = \log \left( {\lambda_{it} } \right)$$

$$\eta_{it} = \pi_{0i} + \pi_{1i} \left( {age} \right)_{it}\,+\,\pi_{2i} \left( {age^{2} } \right)_{it}\,+\,\pi_{3i} \left( {age^{3} } \right)_{it}\,+\,\pi_{4i} \left( {timesecure} \right)_{it}$$

where $\eta_{it}$is the log of the average arrest rate for individual i at age t. To capture the potential of non-linear arrest (or divergence) trajectories, the equation is specified to follow a cubic function of age. The age terms were grand mean centered to allow for more stable estimation due to collinearlity between the age terms. Grand mean centering yielded an interpretation of $\pi_{1i}$ as the average rate of growth across the entire observation period. We observe standard growth parameter models prior to the addition of predictors to assess whether a cubic model is necessary to capture accurately the longitudinal trend in our outcome of interest excluding higher order polynomials when not significant. The proportion of time spent in a secure facility is included as a time-varying covariate and is grand-mean centered. The subscript i attached to variables at level 1 indicates that these variables can take on different values for each individual. The subscript t indicates that the variables can take on different values over time.

Individual level characteristics were entered into the equation at level 2. Coefficient effects at this level indicate how much variation in the intercept (i.e., initial offending level) and slopes (i.e., growth in the offending trajectory) is explained by between-individual characteristics. The level 2, between-individual equations are:

$$\pi_{0i} = \beta_{00} + \beta_{01 \ldots 0k} \left( {controls} \right)_{i} + r_{0i}$$

$$\pi_{1i} = \beta_{10} + r_{1i}$$

$$\pi_{2i} = \beta_{20} + r_{2i}$$

$$\pi_{3i} = \beta_{30} + r_{3i}$$

$$\pi_{4i} = \beta_{40}$$

where variation in the log of the crime rate at the age coded as zero (the mean age for the sample) is explained by an array of sociodemographic and behavioral controls. The dichotomous measures were left in their original form (not centered) allowing for the generation of offending “profiles” for specific demographic groups. We allow for variation between individuals on the dependent crime/arrest variables ($\pi_{0i}$) and age ($\pi_{1i}$, $\pi_{2i}$, $\pi_{3i}$) parameters as indicated by the error terms ($r_{0i}$), ($r_{1i}$), ($r_{2i}$) and ($r_{3i}$). The inclusion of an error term on the age slope indicates that the estimated rate of crime/arrest is allowed to vary across individuals and that the model allows for gradual change in arrest over time across individuals (see Horney et al. 1995).