Exploring trends in youth homicide with cluster analysis: new methodological pathways to policy tools Emily k. Asencio University of Akron Robert Nash Parker University of California
Alternative Approach: Hierarchical Cluster Analysis What if we do not know what variables to classify or cluster on? Like Matching in experiments, classification by known variables is a weak design Miss important factors Inadvertently introduce bias Cluster Analysis Creates possible groupings based on similarity or difference in the trends across the cities Run analysis starting with 91 Cities each in their own cluster; data reduction exercise
Hierarchical Cluster Results Aged Cluster 1: New York; Dallas; Los Angeles; Houston; Ft. Worth; Denver; /San Antonio; Philadelphia; /San Diego; Atlanta; Corpus Christi;/ San Francisco; Detroit; Chicago; Birmingham; Cleveland; Dayton ;/ Baltimore; Oakland; Knoxville; Long Beach; Rochester;/ Newark,NJ; Flint; Amarillo; New Orleans; Santa Ana; Seattle;/ Akron; Cincinnati; Memphis; Little Rock;/ Columbus,OH; Charlotte; Stockton; Nashville; San Jose;/ Cluster 2: Louisville; Norfolk; El Paso; Milwaukee; Grand Rapids; Miami; Jacksonville; Ft. Lauderdale; Gary; Lubbock; Jackson (Miss); Portland; Fresno; Shreveport; Boston; Mobile;/ Cluster 3: Kansas City; Richmond; Chattanooga; Virginia Beach;/ Lexington-Fayette; Providence; Pittsburgh; Riverside; Salt Lake City; Columbus,GA;/ Sacramento; Austin; Madison; St. Petersburg; Buffalo; Tacoma;/Omaha; Oklahoma City; Washington DC;/ Honolulu; St. Louis;/ Baton Rouge; Anaheim; Raleigh; Minneapolis; Phoenix; Montgomery; Cluster 4: Tampa; Toledo; Colorado Springs;/ Springfield; Syracuse;/ Wichita; Ft. Wayne;/ Tulsa; Des Moines; Indianapolis;/ Lincoln;/ Tucson; Greensboro; Spokane; Las Vegas; /Albuquerque; Jersey City; Anchorage;/ St. Paul; Worcester
What factors predict cluster membership Set of common predictors from each decade Estimate a logistic regression for each cluster and decade What can results tell us about factors that distinguish clusters?
Common Predictors: Poverty Rate Percent female headed households Percent housing owner occupied Percent Young African American males Unemployment
Results Cluster 1 (New York,Dallas, Los Angeles) 1980: %AA Males 1990: Owner Occupied 2005: Owner Occupied Cluster 2 (Louisville,Norfolk,El Paso) 1980: Poverty 1990: Poverty 2005: No significant effects
Results Cluster 3 (Kansas City(MO),Richmond,Chattanooga) 1980: No significant effects 1990: unemployment 2005: no significant effects Cluster 4(Tamps,Toledo,Colorado Sprgs) 1980: Poverty 1990: Unemployment; Poverty; Owner Occupied; Young AA Males 2005: No significant effects
Conclusions Attempt to use cluster analysis is mixed Clusters have unusual features Pattern of preliminary results hard to discern If better models of the cluster memberships could be developed, cities could see how similar they are to other members Non cluster cities could look for similarities to one of the clusters More work on this needs to be done!