Quantifying and Reducing Bias in Maximum Likelihood Estimation of Structured Anomalies