Predictability of gun crimes: a comparison of hot spot and risk terrain modelling techniques

Abstract

The current study was designed to assess the possible differences in the accuracy and precision of two methodological mapping techniques as predictors of future gun crimes in Little Rock, AR: (1) risk terrain modelling (RTM) and (2) nearest neighbour hierarchical (Nnh), a traditional hot spot technique, which relies on past crime to predict where future crime is likely to occur. Data from the Little Rock Police Department, the Little Rock Treasury Department and the 2000 census were used to examine Nnh hot spot and RTM methods of gun crime prediction. The RTM incorporated measures of crime generators and crime attractors, while Nnh hot spots were constructed from 2008 gun crime data. The two measures were compared using their predictive accuracy index (PAI) and recapture rate index (RRI) values. Six of the seven social and physical environmental measures in the RTM significantly predicted future gun crime locations and the Nnh hot spots predicted 7% of future gun crime. PAI and RRI values suggested the RTM was more precise than the Nnh hot spot technique and the Nnh hot spot technique was more accurate than the RTM technique. Relying on one spatial prediction technique may create problems with accuracy and reliability. Multiple techniques may be needed to fully assess the phenomenon. Accuracy is a potential limitation of RTM when compared to other techniques, however, RTM is more reliable than Nnh hot spots due to the inclusion of the environmental backcloth. The findings were discussed in relation to crime prediction and policing efforts.

Keywords:

• risk terrain modelling
• hot spot
• crime generators and attractors

Notes

1. Numerous factors are incorporated into the decision-making process (Cornish and Clarke 1986).

2. Their hotspot technique relied on raster data output for comparison. The current study incorporates vector data output that identified gun crime clusters.

3. The purpose of this study is not to generalise the findings to any particular population, but to compare the two analytical methods.

4. An important limitation of using official crime statistics, such as LRPD data, is that only crimes reported to the police are included in the data. This may create systematic bias in the results due to crimes that involved a gun going unreported.

5. The average street length in Little Rock was about 430 ft, but the nature of RTM used in the current study relies on a smaller cell size that allowed for the incorporation of awareness space.

6. The current study conducted five smoothing iterations to take into account surrounding block group values.

7. Caplan and Kennedy (2013) released a RTM specific programme known as RTMDx to aid in the operationalisation and building of the ‘best’ model. The programme was not utilised in the current study because it operates on point data alone when creating the ‘best’ model. It would be possible to take the current study's point data and input it into RTMDx to help operationalised measures but the ‘best’ model would be misspecified. That is, because the model only used the point data and not the block group data, the ‘best’ model would not reflect the entire risk factor list. The social characteristics at the block group level could alter how the point data were operationalised but RTMDx cannot currently account for that possibility.

8. Moran's I determines whether there is any correlation, positive or negative, and compares the values of nearby areas to look for similar rates (Anselin et al. 2000). Moran's I tests against the null hypothesis of clustering, that incidents are randomly distributed. The z-score identifies the significance of the correlation and the Moran's I signifies the direction (positive or negative). The values range from –1, dispersion, to +1, highly clustered incidents. A corresponding significance value is given with the Moran's I value. For the use within RTM, a value near 0 is wanted to signify independence among the distribution of gun crimes.

9. Each independent variable map layer was trimmed to the Little Rock boundaries so the 300 ft × 300 ft grid would line up correctly when the variables were combined.

10. The results to the binary logistic regression analysis were robust across zero inflated negative binomial regression, ordinal logistic regression and penalised likelihood logistic regression.

11. Gun crimes were collapsed into a binary measure due to the small number of block groups in the sample that contained more than one gun crime.

12. Lum (2008) discussed how Ripley's K function tests for statistical significance of clustering by comparing observed data points to random data points. The comparisons between observed and expected are based on Monte Carlo simulations. Levine (2006) stated that Monte Carlo simulations must be run to achieve statistical significance of the findings. Monte Carlo simulations consist of randomly distributed points that were run with the same parameters defined by the user to determine if the findings of the actual data were significant when compared to random simulations. The output gives the approximate confidence intervals (minimum and maximum) of the simulations and the observed value of the actual data.

13. The researchers acknowledge the work completed by Weisburd et al. (2004, 2009), but the focus of the current paper is not street segments.

14. Chainey et al.'s (2008) assertion that KDE was the best hot spot technique was met with criticism due to various reasoning (see Levine 2008, Pezzuchi 2008).

15. Their study examined five hot spot techniques: STAC Ellipses, STAC Convex Hulls, Nnh Ellipses, Nnh Convex Hulls and KDE.

16. Nnh is a hot spot technique that is part of the CrimeStat III programme and the output was in ellipse form (see Levine 2010). The researchers are aware of the arbitrariness of ellipses in regards to the distribution of crime within the ellipses.

17. The search radius was set to random distance with a significance level set to .001%. Similar to Ripley's K function, Nnh uses Monte Carlo simulations to determine if the results of the analysis were by chance.

18. Pawn shops were found to be non-significant predictors for gun crimes, excluding them from the analysis. The non-significance can be due to the few pawn shops that were located in Little Rock, AR.

19. The technique was also run at the .05 level to determine if more hot spots would be identified but no new hot spots formed at the .05 level.

20. On the other hand, RTM does identify many areas as risky when in reality those areas are not dangerous for criminal activity.

21. All significant measures predicted gun crime in the predicted direction, positively, other than fast food establishments, which predicted a reduction in gun crime.