Telling a Similar Story Twice? NCVS/UCR Convergence in Serious Violent Crime Rates in Rural, Suburban, and Urban Places (1973–2010)

Abstract

Objectives

This study examines UCR and NCVS serious violence crime trends in urban, suburban, and rural areas, and assesses the extent to which discrepancies in the two data series are due to victim reporting or police crime-recording practices. Particular attention is paid to the dynamics of the rural data series.

Methods

NCVS data for 1973–2010 are used to estimate subnational rates of serious violence and comparable rates for crimes that victims said were reported to police, and these estimates are compared to subnational UCR data. Time-series cointegration analysis is used to assess convergence in the NCVS and UCR series along with descriptive comparative analyses.

Results

The degree of convergence in UCR and NCVS trends was found to vary across areas; however this was not due to differences in rates of reporting to police. Suburban and urban UCR and NCVS trends converged with and without adjustment for police reporting. Little evidence of NCVS/UCR series convergence was found in rural areas even after victim reporting was taken into account.

Conclusion

The recording and production of crime data by the police appears to contribute to subnational differences in the convergence between the UCR and NCVS series. The findings suggest rural crime trend analysis should not be based solely on UCR data. To illustrate the difference between conclusions based on UCR and NCVS rural violence trends, we find that poverty rates have a large, significant association with rural violence as measured in the NCVS, but are unrelated to UCR rates.

Appendices

Appendix 1

The Potential Role of Non-Discretionary NCVS-UCR Procedural Differences

UCR rates are based on where the crime incident occurred, while NCVS rates are classified according to where the respondent lives. This dissimilarity in the meaning of the location variables between the systems can be consequential if, for example, rural residents are more likely to suffer serious violence while visiting other types of locations, especially if this pattern has changed over time.

The NCVS asks victims whether the incident occurred in the same city or town in which they reside, and if the answer is no, they are then asked whether it occurred in the same county in which they reside. All victims are also asked whether the incident occurred at home; within one mile of the home; 1–5 miles; 5–50 miles or more than 50 miles. Using this information, we classified victimizations that occurred in the same city or county, or at home or within one mile as having occurred in the ‘same type of area’ in which the respondent resides. For the period 1993–2010, we found that 92 % of urban residents, 82 % of suburbanites, and 84 % of rural residents experienced their victimizations in the ‘same type of area’ in which they lived. Importantly, we observed no trend in these crime-specific proportions over the 1993–2010 time frame and so inconsistencies in changes in crime rates between the systems are unlikely to be affected by changes in where incidents occur relative to the type of place in which the victim lives.¹¹

An Exploration of Subnational Differences in Crime Composition

Other issues that may affect the differences in our findings across rural and non-rural places are the possibility of varying crime compositions across areas, and differences in reporting and police recording of incidents that may be associated with crime composition. As noted above, the likelihood that victims will notify the police is partly a function of the characteristics of the victimization incident and discretion in police recording of incidents may vary according to the characteristics of crime events as well. If the types of incidents that involve the most discretion occur more frequently in rural relative to urban places, then this pattern may help account for discrepancies between rural NCVS and UCR estimates. For instance, if rural victims experience proportionately fewer incidents of stranger (versus nonstranger) violence, and victim reporting to police and police report-taking are more likely for stranger incidents, this may help account for the differences in our findings across places. Likewise, a similar pattern might emerge if rural places experience fewer robberies and if this crime is subject to less discretion but more likely to be reported to the police.

To assess the issue of how compositional differences in crime types might affect UCR-NCVS convergence over time, we first examined the 1973–2010 trends in crime composition across the three types of areas. The NCVS data show that the composition of crime types (i.e., rape, robbery, and aggravated assault) is remarkably stable across all three types of areas over time (see Table 6). In each of the areas, the modal type of serious violence is aggravated assault and the degree to which such incidents compose the violence rate varies by less than three percentage points over time periods within each area. Robberies compose a greater proportion of violent crimes for urban residents than elsewhere, but here too the percentage of violent crimes that are robberies in urban, suburban, and rural areas varies little over time. The stability of crime composition over time within each of the areas suggests that differences in the UCR-NCVS rural findings are not due to unique changes in types of crime over time in rural areas.

Table 6 Crime composition by type of crime, area, and time period: NCVS

Second, because rural crime victims experience lower rates of stranger violence than victims elsewhere, we also examined how the proportion of stranger incidents over time varied across areas (see also Table 6). Generally speaking, the NCVS rural data do show lower proportions of stranger violence than elsewhere. In all three places, the proportion of serious violence committed by strangers declined over time because stranger violence rates declined faster than non-stranger rates. However, the decline in the percent of violence committed by strangers was proportionately greater in rural areas and this pattern is not what would be expected to prompt an apparent increase in consistency between the UCR and NCVS rural rates in the later part of the series. Instead, we would have expected decreasing consistency over time as incidents in rural areas became increasingly less stranger-dominated.

Third, we estimated a series of additional Johansen’s models comparing the rural UCR rates to rural NCVS rates of serious stranger violence and non-stranger violence. We estimate these additional models because the disaggregated NCVS trends may uncover nuances obscured by a focus on total rates of violence. More importantly, this analysis could help to determine whether the lack of cointegration in rural places is due to a discretionary processes associated with the victim-offender relationship. Estimates from Johansen’s cointegration tests on the correspondence between the NCVS non-stranger rates and UCR rates yield no evidence of cointegration (r = 0, $\lambda$ _Trace  = 11.89; 5 % critical value = 15.41) and the same conclusion emerges from the estimates of the stranger violence models (r = 0, $\lambda$ _Trace  = 9.52; 5 % critical value = 15.41). Thus, the results of the multivariate models in combination with the descriptive trends suggest that variation in crime type—or compositional differences—does not appear to account for the divergence between the series in rural places. Much of the descriptive and time-series evidence presented thus far suggests that discretionary processes relating to informal networks in rural places may not be the main culprit behind the NCVS/UCR divergence in rural places. However, it is important to note that since the UCR trends cannot be disaggregated according to the victim-offender relationship the scope of this comparison is necessarily limited.

As a final check on the potential influences of crime composition on our findings, we estimated survey weighted multivariate logistic regression models of victim reporting to police and police report taking for the years 1994–2010 (see Table 7).¹² The key predictor variable of interest was rural location, controlling for year of incident, type of crime (i.e., rape, robbery, and aggravated assault), stranger versus non-stranger incident, and location of the incident (i.e., more than one mile away from home versus one mile or less). These analyses found that victim reporting to the police was not significantly different in rural versus other areas during this period; that police report-taking was significantly less likely in rural areas; and that these patterns were independent of the influences of year, crime type, victim–offender relationship, and location of the incident. The results also suggest that both reporting to the police and police report-taking for serious violence increased some during this period.

Table 7 Logistic regressions of reporting to the police and police report-taking: pooled NCVS 1994–2010

Appendix 2

Consideration of Potential NCVS Contributions to Divergence

We also considered whether some of the observed divergence in the UCR and NCVS-PRV rates may be due to aspects of the NCVS system that could vary by subnational location (see Biderman and Lynch 1991). For example, if there is more measurement error in the NCVS data in rural places than in other areas, this may also contribute to divergence in the data trends. Potential differences in measurement errors include variations in sampling errors and nonresponse errors. The possibility that rural, urban, and suburban residents may interpret the NCVS serious violence questions differently is unlikely given the similarity in the effects of the redesign on the three groups of respondents and comparable levels of police reporting.

The issue of differences in sampling errors may be a concern because the standard errors of the NCVS and NCVS-PRV rates will be a function of both the rate in which victimization occurs, as well as the sample size on which the rate is based. If the serious violent victimization rates and sample sizes are much smaller in rural areas than elsewhere, the underlying trend may be harder to detect, particularly because NCVS sample sizes declined some over time. To investigate the potential impact of this issue we examined population proportions across the years and three types of areas, as well as sample sizes and weights. During the earlier years of the series (from approximately 1973 to 1985) the rural population was equal to or slightly larger in size than the urban population, and the suburban population constituted the largest proportion of the US population. Thereafter however, the proportion of the US population living in rural areas began to decline, from about 31 % in 1985 to 16 % in 2010. The proportion of the US population living in urban areas declined some in the 1970s, but overall was similar in 2010 (32 %) to what it was in 1973 (31 %). The proportion of the US population living in suburban areas gradually increased over time from 38 % in 1973 to 52 % in 2010.

Accompanying these changes in population proportions in each area were declines in NCVS sample size. For example, in 1973 there were about 72,000 households in the sample (resulting in about 220,000 personal interviews) while in 2010 there were about 41,000 households in the sample (resulting in about 167,000 personal interviews). Examination of non-weighted cases and the sampling weights across time and place reveals that survey nonresponse was unrelated to type of place.¹³ It also suggests that the overall declines in sample size occurred rather equally in urban, rural, and suburban places except during the period 1997–2005. During this period sample sizes in rural areas declined disproportionately in an effort to reduce survey costs, and corresponding sample weights were adjusted upward to make up for the disproportionate decrease in cases.¹⁴ Though we are unable to detect any bias during this period in the rural NCVS trends, this does suggest that some caution is warranted when interpreting differences in the annual levels of the UCR and NCVS-PRV rates for this period.¹⁵ The cointegration analyses of the trends in these rates are not affected by this issue.