|
Abstract:
|
Outlier detection for continuous variables in administrative data and establishment surveys can be challenging due to variation in the size of reporting units and time trends across them. Without proper detection and treatment for the outliers, statistics generated from suspect data may lead to inaccurate or spurious conclusions. In this presentation, we will compare several major outlier detection methods for detecting unusual crime counts reported by police agencies in the FBI’s National Incident-Based Reporting System (NIBRS) data. Methods considered include: the ratio-to-median test, the Hidiroglou-Berthelot method, and Shewhart quality control charts. We evaluate the performance of these methods when tweaking model parameters, and test how data adjustment techniques, such as log-transforming the crime counts and incorporating the population served of the reporting agency, impact the error rates. In addition, we cover an in-depth analysis to identify agency characteristics associated with the presence of outliers in NIBRS. We will show how we incorporate the analysis results into the final statistical adjustments to address abnormal reporting patterns.
|
