Kriminalitet i Danmark: En analyse af kriminaliteten i Danmark på baggrund af udvalgte parametre

Abstract

This thesis investigates the relationship between population characteristics and crime based on population data from Danmarks Statistik covering etchnicity, age, gender and unemployment over an 11-year period.

The characteristics are investigated for correlations with the crime-rate within a population. Traditional statistics and Machine Learning are both used to illuminate interaction effects, predictability of crime, importance of predictor variables and to test whether these methods can correctly predict a population groups crime-rate in a test-set. It is found that all characteristics impact the level of crime in a simple additive regression. In models that include interactions effects, we conclude that gender and age have significant interaction effects. Unemployment is statistically significant when included in a model with full interaction effects. This conclusion changes when unemployment is included in models where only the partial interaction effects of unemployment itself is included. Unemployment itself is not statistically significant, but the interaction effects between unemployment and gender & ethnicity are significant.

We also apply machine learning techniques, including Random Forest and logistic regression as a machine learning technique. Random Forest finds that gender is the most important characteristic. Unfortunately, the Random Forest model is assessed as flawed. Based on theory, we expect that logistic regression would be superior, as the Random Forest model suffers from lack of predictor variables.

The results are put into context by comparison with General Strain Theory (“GST”), a criminological theory that attempts to explain drivers of criminal delinquency through stimuli throughout life. Our models generally support the claims of General Strain Theory. Specifically, our findings support the claims of GST that gender, age and ethnicity are all significant predictor variables of crime. However, partially modifying the claims of GST we find that unemployment is not a significant predictor of crime independently when the impacts of the prior variables are included. The only statistically significant impacts of unemployment are the interaction effects with other variables.