Donald E. Brown, Keynote Presentation
Discrete Choice Models for Incident Prediction
Terrorist and criminal incidents take a number of forms, for instance, suicide bombings, explosive devices, sniper shootings and arson. Conventional approaches to mathematically understanding and predicting the locations of such attacks are to model them as point patterns and apply techniques from the extensive literature of point pattern analysis. Such methods, including density estimation and clustering techniques, typically focus on the spatial relationships of the attacks. While these approaches do occasionally reveal important spatial consistencies they often fail as predictive models. Their failure stems from an incorrect model of the underlying causal mechanism which is not a physical or natural phenomenon but a choice process by a motivated attacker. To capture this choice process requires an approach to understanding the preferences of the attackers. This presentation shows how this can be done using utility functions. These functions then form the basis for predictions of future attacks. This presentation discusses this approach and shows its usefulness to several problems in incident prediction.

Howard S. Burkom, Keynote Presentation
Adaptations of Data Modeling and Process Control for Prospective Biosurveillance
Syndromic surveillance involves the monitoring of available data sources for early warning of outbreaks of unspecified disease or of specified disease before the confirmation of identifying symptoms, with the objective to complement physician sentinel surveillance with false alarm rates acceptable to the public health infrastructure. Data sources include clinical data such as counts of syndrome-specific emergency department visits or physician office visits, and nonclinical data such as over-the-counter remedy sales and school/work absentee rates. A common approach among system developers has been to adapt chart-based methods from the field of statistical process control. Major obstacles to this approach are the evolving and often nonstationary input data streams, the uncertainty of the nature of the signal to be detected, and the presence of systematic or periodic behavior in the data background. Thus, robust detection performance, measured by timeliness and sensitivity at controlled alert rates, requires a combination of modeling and process control suitable to the characteristics of the monitored data. This presentation investigates such combination methods using standard evaluation techniques as well as focused techniques specific to the biosurveillance context. A discussion of generalized exponential smoothing methods will be given along with a performance comparison against a recently published regression modeling approach.

David A. Dickey
Introduction to Time Series and Intervention
In this shortened version of my workshop presentation, I will briefly talk about time series structures, show the time series representation for the exponentially weighted moving average operator that we have seen in some of our syndromic surveillance papers, and show some examples of intervention analysis, including a new example detailing the impact of the 9/11 terrorist strike on passenger volume at our local RDU airport.

Ryan Gill
Logistic Joinpoint Methods for Occupational Cohort Data
We describe logistic joinpoint methods for modeling incidence data. The model is well-suited for data where temporal changes in the functional pattern are suspected. We report the current progress and future plans in applying the methods to mortality data for the Louisville VC cohort of chemical workers.

Michael Last
Detecting Jumps in Piece-wise Locally Stationary Time Series
Finding jumps in piece-wise stationary time series is a problem considered by many authors. This model is too rigid for some applications, such as seismology. I will discuss a method for detecting jumps in piece-wise locally stationary time series, and look at applications to work in seismology and speech recognition.

Carey E. Priebe, Keynote Presentation
Scan Statistics on Enron Graphs
We introduce a theory of scan statistics on graphs and apply the ideas to the problem of anomaly detection in a time series of Enron email graphs.

Francisco Vera
Feature Based Density Estimation
Ideas to extend upon Donald E. Brown's criminal prediction model.