10 Steps to Optimize Your Crime Analysis
8 likes2,467 views
The document outlines 10 steps to optimize crime analysis, emphasizing the importance of data quality and analytic techniques. Key steps include examining geocoding accuracy, exploring predictive accuracy with kernel density estimation, and utilizing tools like the near repeat calculator and risk terrain modeling. Additionally, it highlights the value of open data and forming research partnerships to enhance analysis capabilities.
10 Steps to Optimize Your Crime Analysis
- 1. 10 Steps to Optimize Your Crime Analysis 340 N 12th St, Suite 402 Philadelphia, PA 19107 215.925.2600 info@azavea.com www.azavea.com/hunchlab
- 2. About Us Robert Cheetham President & CEO cheetham@azavea.com 215.701.7713 Jeremy Heffner HunchLab Product Manager jheffner@azavea.com 215.701.7712
- 3. About Azavea • Founded in 2000 • 32 people • Based in Philadelphia – Boston office – Minneapolis office • Geospatial + web + mobile – Software development – Spatial analysis services
- 4. Clients & Industries • Public Safety • Municipal Services • Public Health • Human Services • Culture • Elections & Politics • Land Conservation • Economic Development
- 5. web-based crime analysis, early warning, and risk forecasting
- 6. 10 Steps to Optimize Your Crime Analysis
- 7. 10 Steps to Optimize Your Crime Analysis Crime Analysis Data Analytic Use Cases Quality Techniques • Geocoding • Kernel Density Map • Open data Predictive Accuracy • Dates & Times • Find research • NNI and Gi* partners • Polygon Hierarchies • Near Repeat Calculator • Randomized controlled trial • Risk Terrain Modeling
- 8. Data Quality
- 9. 1. Examine Geocoding Accuracy • Geocoding – Process of turning addresses into geographic coordinates – Examine accuracy • Correct locations – Geocoding method » Commercial geocoder » Street center line » Parcel database – POIs and landmarks – Incorrect clustering at: » Precinct locations, zip code and city centroids • High geocoding success rates – Ratcliffe suggests at least 85% (lowest acceptable) » http://bit.ly/ratcliffegeocoding – Examine unsuccessful geocodes for patterns
- 10. 2. Examine Dates & Times • Dates & Times – Event-related times • Actual occurrence time – From / to time interval • Report time • Officers responded time
- 11. 2. Examine Dates & Times • Dates & Times – Examining accuracy • Data entry defaults • Data validation on input • Clustering by time cycles – Day of week – Day of month – Day of year – Hour of day – Minute of hour
- 12. 2. Examine Dates & Times
- 13. 3. Examine Polygon Hierarchy • Polygon Hierarchy – A set of geographic areas that nest within each other used to organize resources (i.e. divisions, districts, PSAs, beats) • What makes a good hierarchy? – Perfectly nested polygons • No sliver polygons – Areas should be periodically rebalanced based on changing crime levels – Consider that splitting areas based on streets means one side of street is in one district / other side is in a different district.
- 15. 4. Test Predictive Accuracy of KDE • Kernel Density Estimation – A smoothing technique that generates hotspot maps
- 16. 4. Test Predictive Accuracy of KDE • When we look at a hotspot map what are we assuming? – That crimes will happen in the hotspots again. – But… • How predictive is it? • How much historic data should we use? • What search radius should we use? • What is the density cutoff for a hotspot? Source: Chainey, http://www.popcenter.org/conference/conferencepapers/2010/Chainey-Gi-hotSpots.pdf
- 17. 4. Test Predictive Accuracy of KDE • Predictive Accuracy Index – Spencer Chainey, Jill Dando Institute • http://www.palgrave-journals.com/sj/journal/v21/n1/full/8350066a.html – Incorporates: • Desire for a high hit rate – Lots of crime incidents in a prior ‘hotspot’ • Desire for a small geographic area – Less to patrol, etc.
- 18. 4. Test Predictive Accuracy of KDE • Predictive Accuracy Index Steps 1. Generate kernel density map for historic period 2. Measure predictive validity against future time period 3. Higher number better • Example from Chainey
- 19. 4. Test Predictive Accuracy of KDE • Predictive Accuracy Index – Caveats • Number is relative to crime type and geography • Best for comparing different techniques (or parameter variations of techniques) for the same predictive period
- 20. 4. Test Predictive Accuracy of KDE • Predictive Accuracy Index – Caveats • Number is relative to crime type and geography • Best for comparing different techniques (or parameter variations of techniques) for the same predictive period Is there something similar but better than kernel density?
- 21. 5. Test NNI and use Gi* • Gi* – Spencer Chainey, Jill Dando Institute • http://www.popcenter.org/conference/conferencepapers/2010/Chainey-Gi-hotSpots.pdf – LISA statistic • Local indicator of spatial association – Compares local averages to global averages – Generates map visually similar to KDE
- 22. 5. Test NNI and use Gi*
- 23. 5. Test NNI and use Gi* • How do we know hotspots exist though? – Calculate nearest neighbor index (NNI) • Determines if clustering exists – NNI ~ 1: data is randomly distributed – NNI < 1: data is clustered – NNI > 1: data is uniformly distributed • Helps to answer if we have enough historic data for statistical significance • If data is not clustered neither Gi* nor KDE should be used
- 24. 5. Test NNI and use Gi* • Summary of Steps – Test for clustering with nearest neighbor index – Calculate crime counts within a grid – Run Gi* statistic – Set color ramp breakpoints based on fixed statistical significance levels
- 25. 5. Test NNI and use Gi* • Summary of Steps – Test for clustering with nearest neighbor index – Calculate crime counts within a grid – Run Gi* statistic – Set color ramp breakpoints based on fixed statistical significance levels Remember our friend the predictive accuracy index?
- 26. 5. Test NNI and use Gi* • Is it really better than KDE? – Example from Chainey
- 27. 6. Run the Near Repeat Calculator • Near Repeat Pattern Analysis – Measures ‘contagion’ effect of crime incident – How does one burglary change the risk that another burglary will occur nearby in the coming days? • Common in Some Types of Crime – Burglary – Theft from Vehicle – Gun Crime – Robbery – Bicycle Theft
- 28. 6. Run the Near Repeat Calculator
- 29. 6. Run the Near Repeat Calculator
- 30. 6. Run the Near Repeat Calculator • Near Repeat Calculator – http://www.temple.edu/cj/misc/nr/ • Papers – Near-Repeat Patterns in Philadelphia Shootings (2008) • One city block & two weeks after one shooting – 33% increase in likelihood of a second event Jerry Ratcliffe Temple University
- 31. 7. Conduct a Randomized Trial • Randomized Controlled Trial – An experiment where study subjects (e.g. locations) are randomly assigned to different treatment protocols – Academics do this regularly • But why do this yourself? – Proves/disproves a technique’s efficacy for your department – Successfully mimicking a published trial gives you the skills to experiment based on local anecdotal evidence
- 32. 7. Conduct a Randomized Trial • Philadelphia Foot Patrol Experiment – Jerry Ratcliffe, Temple University • http://bit.ly/phillyfootpatrol – Concentrated patrol in 60 violent crime hotspots – Outlines full methodology • Experimental design • Evaluation – Result was a net reduction of 53 violent crimes
- 33. 8. Generate a Risk Terrain Model • Risk Terrain Modeling – Joel Caplan & Les Kennedy, Rutgers University • http://www.rutgerscps.org/rtm/ – Forms a combined risk surface of several spatial risk factors that correlate with a particular type of crime – Describes the environmental context within crime occurs
- 34. 8. Generate a Risk Terrain Model • Steps to Building a Model 1. List potential risk factors • Literature review • Departmental experience 2. Assemble GIS data sets for each factor 3. Operationalize each factor and test for correlation 4. Combine correlated factors into combined risk terrain
- 35. 8. Generate a Risk Terrain Model Gun shootings example Source: Rutgers, http://www.rutgerscps.org/rtm/irvrtmgoogearth.htm
- 36. 8. Generate a Risk Terrain Model Gun shootings example Source: Rutgers, http://www.rutgerscps.org/rtm/irvrtmgoogearth.htm
- 37. 8. Generate a Risk Terrain Model Gun shootings example Source: Rutgers, http://www.rutgerscps.org/rtm/irvrtmgoogearth.htm
- 38. 8. Generate a Risk Terrain Model • Risk Terrain Modeling – Risk Terrain Modeling Manual • http://www.rutgerscps.org/rtm/ – Online training • http://www.rutgerscps.org/rtm/webinar.html
- 39. Use Cases
- 40. 9. Open Up Data (Appropriately) • Increase transparency and community engagement – Open data movement – Increases trust – Allows novel uses for crime data – Increases perceived value of good data • Not a new idea – NIJ guide released in 2001 • https://www.ncjrs.gov/pdffiles1/nij/188739.pdf
- 41. 9. Open Up Data (Appropriately) • Public crime mapping sites – Omega Group • http://www.crimemapping.com/
- 42. 9. Open Up Data (Appropriately) • Data portals – Chicago Data Portal • http://data.cityofchicago.org/ • Raw incident data from 2001 to present
- 43. 10. Find Research Partners • Benefits of Conducting Research – Lowers the risk of trying something new – Supplements limited resources • Labor • Software • Knowledge/techniques/statistical rigor – Encourages cutting edge analysis • Types of Partners – Academic crime researchers – Nonprofits – Commercial entities
- 44. 10. Find Research Partners (A brief plug)
- 45. Q&A
- 46. About Us Robert Cheetham President & CEO cheetham@azavea.com 215.701.7713 Jeremy Heffner HunchLab Product Manager jheffner@azavea.com 215.701.7712 www.azavea.com/hunchlab