Assessing administrative data quality

Abstract

Administrative datasets, characterized by large scale and high volume, often exhibit data quality issues, including reporting delays, missing entries, and irregular reporting. Outlier detection as a precursor to imputation is a critical step for reliably identifying missing reports and ensuring accurate downstream estimates. This paper proposes a novel two-step method for detecting irregularities in administrative reported data by combining clustering with robust outlier detection using the median and the median absolute deviation. Through 10,000 simulations across 10 distinct scenarios, we evaluate the proposed method against established approaches, including the mean ± SD, boxplot, and ratio-to-median methods. The results show that the proposed method matches or outperforms traditional one-step methods in eight scenarios and ranks second best in the remaining two. We further apply the method to monthly counts from 136 NIBRS agencies, examining the frequency and severity of flagged observations, comparing results with the ratio-to-median approach, and assessing sensitivity to the primary tuning parameter k. Overall, the proposed framework provides a robust and interpretable approach to identifying reporting irregularities in administrative data, with clear implications for improving data quality and downstream statistical estimation.