Determinants of violent and property crimes in England and Wales: a panel data analysis

Abstract

We examine various determinants of property and violent crimes by using police force area level (PFA) data on England and Wales over the period of 1992–2008. Our list of potential determinants includes two law enforcement variables namely crime-specific detection rate and prison population, and various socio-economic variables such as unemployment rate, real earnings, proportion of young people and the Gini Coefficient. By adopting a fixed effect dynamic GMM estimation methodology we attempt to address the potential bias that arises from the presence of time-invariant unobserved characteristics of a PFA and the endogeneity of several regressors. There is a significant positive effect of own-lagged crime rate. The own-lagged effect is stronger for property crime, on an average, than violent crime. We find that, on an average, higher detection rate and prison population leads to lower property and violent crimes. This is robust to various specifications. However, socio-economic variables with the exception of real earnings play a limited role in explaining different crime types.

Keywords:

• crime
• dynamic panel
• GMM
• law enforcement
• socio-economic variables

JEL Classification:

• K42
• C23

Acknowledgements

We thank Christopher Baum, Ralph Bailey, Anindya Banerjee, Fiona Carmichael, Matthew Cole, John Fender, Nick Horsewood, Somnath Sen and Rudra Sensarma for their valuable comments. Some of this work was done for Lu Han’s PhD thesis.

The opinions in the article do not reflect the opinion of PNC Bank.

Notes

1. See http://www.bbc.co.uk/news/10457112 for details.

2. See http://www.policyexpert.co.uk/news/uncategorized/north-wales-experiences-higher-crime-but-lower-detection-rates/4639/ for details.

3. Two such potential pitfalls are (1) presence of unobserved variable affecting both the crime rate and the determinants of crime rate, and (2) reverse-causality effect whereby crime in turn affect the variable affecting the crime.

4. There are in total 43 Police Force Areas (PFA) in England and Wales. They are Avon and Somerset, Bedfordshire, Cambridgeshire, Cheshire, Cleveland, Cumbria, Derbyshire, Devon and Cornwall, Dorset, Durham, Essex, Gloucestershire, Greater Manchester, Hampshire, Hertfordshire, Humberside, Kent, Lancashire, Leicestershire, Lincolnshire, City of London, Merseyside, Metropolitan Police District, Norfolk, Northamptonshire, Northumbria, North Yorkshire, Nottinghamshire, South Yorkshire, Staffordshire, Suffolk, Surrey, Sussex, Thames Valley, Warwickshire, West Mercia, West Midlands, West Yorkshire, Wiltshire, Dyfed-Powys, Gwent, North Wales, and South Wales.

5. Unlike in the UK, there is a large literature in the US on crime using panel data though the focus is on property crime. It is beyond the scope of this study to summarize those papers. These include papers by Doyle et al. (1999), Gould et al. (2002). There is also a growing literature in the rest of Europe (Reilly and Witt, 1992; Pyle and Deadman, 1994b; Edmark, 2005; Saridakis and Spengler, 2009). One of the few papers using a panel approach which analyses violent crime is Fajnzylbera, Lederman, and Loayzab (2002) but it is a cross country study and could potentially suffer from the aggregation bias discussed above.

6. Since we are using PFA level data, instead of individual crime data, our study can still have some aggregation bias. Furthermore, not all of our explanatory variables are available at the PFA level, which can introduce further aggregation bias in our analysis.

7. While a log linear model is standard and also captures interaction effects we consider the possibility of non-linearities by looking at including a quadratic term for detection. In five out of six crime categories (results available from the authors on request), the quadratic term is insignificant.

8. Carmichael and Ward (2000, 2001) use conviction rate as a proxy for the probability of apprehension, we argue that the deterrence effect of detection rate on potential offenders should be at least as strong as that of conviction rate. This is because, for real offenders, being detected is often regarded as the first step of punishment. Even though the conviction depends on a number of exogenous variables such as the evidence presented by the police, who the judges are etc., the potential criminals still would try to avoid detection in first place. Very few offenders will be so confident, when arrested, that they will not be convicted by the court.

9. In forward orthogonal deviation approach we subtract the average of all the future available observations from the contemporaneous observation. This approach, unlike the usual first difference approach, does not introduce first order serial correlation in the model.

10. We are assuming that the detection rate and prison population are not strictly exogenous, rather they are predetermined.

11. While testing for weak instruments is not usual in dynamic GMM models because of the plethora of instruments used, we do a F-test a la Staiger and Stock (1997) and it performs well for a majority of cases further validating our results.

12. Criminal damage and other offences have been excluded from the analysis because a large proportion of sample does not have data on these variables.

13. The serial numbers of Criminal Statistics are Cm847, Cm1322, Cm1935, Cm2134, Cm2410, Cm2680, Cm3010, Cm3421, Cm3764, Cm4162, Cm4649, Cm5001 and Cm5312. Crime in England and Wales, 2001/02, 2002/03, 2003/04, 2004/05, 2005/06 2006/07, 2007/08 and 2008/09.

14. The category of ‘violence against person’ has been broadened to include harassment, cruelty to or neglect of children, assault on a constable and common assault. The ‘sexual offenses’ category has been added with soliciting or importuning by man. ‘Theft and handling’ has included the new sub-categories of vehicle interference and tampering. ‘Fraud and forgery’ has been expanded to include bankruptcy and insolvency offenses and vehicle/driver fraud. The definitions of ‘robbery’ and ‘burglary’ have not been affected by the change in counting rules.

15. Detection rates and prison population might be correlated which is why we do a variance inflation test which shows that multi-collinearity is not an issue.

16. ‘Local authorities’ are lower government level than government regions but similar to county level. Data by local authorities typically contain six metropolitan counties, 27 nonmetropolitan counties, 56 unitary authorities and the region of London, which has 32 London boroughs and the city of London.

17. Regional Trends is a publication that summarized various aspects of the UK by government regions. It summarizes the ‘real average weekly earnings’ from the Annual Survey of Hours and Earnings published by the Office for National Statistics.

18. See http://www.statistics.gov.uk

19. Carmichael and Ward (2001), analysing county (corresponding roughly to PFA) level data finds a positive relationship between youth unemployment and various property crimes.

20. Another potential difference could be different unemployment measures used. While we report total unemployment and do not break it up by gender or age group like studies by Carmichael and Ward or Witt, Clarke and Fielding, our results in earlier versions were not sensitive to such breakup of overall unemployment rate.

21. Although in our contemporaneous specification we did instrument for the detection rate and prison population with lagged values, starting from lag 2, of these variables (see the results in main findings section).

22. Machin and Meghir (2004)’s data ends in 1996.

23. We considered the possibility that changes in the benefit system may have reduced economic incentives for the unemployed but over our time period, it seems that the unemployment benefit system remained largely unchanged, see http://www.eu-employment-observatory.net/resources/reviews/UK-UBRvw2011.pdf

24. See for example a short report in http://www.guardian.co.uk/uk/2010/apr/22/crime-falls-despite-recession-figures?CMP=AFCYAH