Title: Explaining NSW long term trends in property and violent crime
1Explaining NSW long term trends in property and
violent crime
Steve Moffatt and Lucy Snowball NSW Bureau of
Crime Statistics and Research
2Purpose of research
- Determine the general structure of trends and
seasonality - Explain some exogenous influences on crime
trends, particularly those useful for forecasting - Forecasts for state and regions
- Test scenarios
3Background property crime
- Long term rise (1990s) followed by fall in
property crime recorded incidents since 2000 - Motor vehicle theft, steal from motor vehicle,
dwelling, retail store, person - Robbery
- Break and enter
- Receiving/handling stolen goods
- Fraud (stabilised after rise)
4Property crime (theft robbery) NSW 95-07
5Background violent crime
- Steep rise (1990s) followed by flattening rise
since 2001 in violent crime recorded incidents - Assault
- Sexual assault
- Harassment
- Other offences against the person
-
- Stable or falling murder, attempted murder,
manslaughter, blackmail, extortion
6Violent recorded crime NSW 95-07
7Background Summary
- Fall in property crime incidents
- Coincided with continuation of upward trend in
violent crime incidents - Demand for short term forecasting at state and
local area level - Previous trend research has focused more on
property crime - Few clues on why violent crime trend persisting,
recent focus on alcohol related assaults
8Predictors
- Seasonality and month characteristics
- Police and Justice
- Police activity, incapacitation, deterrence
- Alcohol and drug use
- Economic cycles
9General Models
First equation
Second equation
- Trends (quadratic, cubic)
- Seasonality (months, weekends)
- Police and Justice (POIs by status)
- Exogenous influences (economy, drugs)
10Model characteristics
- Violent offences model in levels (ARMA)
- Quadratic trend
- Property offences in differences (ARIMA)
- Cubic trend
- Lagged dependent variable or POI variables by
status - MA(1) error term
11Property crime POI trends
12Violent crime POI trends
13Model results (Violent offences)
14Forecasts Violent offences
15Forecasts Violent offences
16Forecasts Violent offences
17Model results (Property offences)
18Forecasts Property offences
19Forecasts Property offences
20Forecasts Property offences
21Model selection and forecast accuracy
- Stationarity of dependent variable
- Most appropriate trend
- MLE ARMA/ARIMA
- Log likelihood and Wald Chi Sq
- Error tests and RMSE for forecast
22Accuracy vs. Parsimony
- Over fitting (including non significant
variables) improves forecast accuracy - However reduction in significance of model
- Fit for purpose
- Overfitted models useful for forecasting
- Parsimonious models useful for determining which
factors influence long term trends
23Conclusions
- Can achieve well fitting models for violent and
property crime with good forecasting power - Majority of trend explained using structure
(quadratic or cubic), seasonal (month) terms - Weekend dummy and summer months a good proxy for
alcohol consumption - POIs (clear-up variables) act as a control for
autocorrelation
24Next steps
- Report state level trends, seasonal components
and influences to NSW Police - Project models from state level to regional level
- Demand at local area command level
- Panel data sets for regions
- Develop models for other crimes, particularly
high volume offences that are resilient to police
activity - Malicious damage
- Assault (domestic violence related and
non-domestic violence) - Harassment