160 likes | 247 Vues
Improving the Quality of Investigative Decision Making. Craig Bennell, Ph.D. Carleton University Ottawa, Ontario, Canada cbennell@connect.carleton.ca. Two-Alternative Decisions. Is this suspect lying? Will this negotiation end in failure? Are these crimes linked?
E N D
Improving the Quality of Investigative Decision Making Craig Bennell, Ph.D. Carleton University Ottawa, Ontario, Canada cbennell@connect.carleton.ca
Two-Alternative Decisions • Is this suspect lying? • Will this negotiation end in failure? • Are these crimes linked? • Was this bite mark made by an adult? • Is this suicide note fake? • Is this allegation of rape false? • To improve decision-making we need a valid measure of performance...
EXAMINE THE EVIDENCE COUNT THE FREQUENCIES OF DECISION OUTCOMES SET A DECISION THRESHOLD CALCULATE PERCENTAGE CORRECT How Should We Measure Performance?
ROC ANALYSIS (Bennell, 2005)
ROC ANALYSIS An AUC of .60 indicates that there is a 60% chance that a randomly selected positive case will be higher on evidence X than a randomly selected negative case (Bennell, 2005)
The Accuracy of Diagnostic Systems: The Polygraph CQT and GKT • The Control Question Test • Relevant (R) and Control (C) questions • For each R-C pair a score of -3 to +3 is given • Greater than +5: truthful, Less than -5: deceitful • The Guilty Knowledge Test • Multiple choice questions • One alternative relates to guilty knowledge (GK) • Elevated responses to GK questions: deceitful • How accurate are these procedures?
Median AUC for GKT = .88 Median AUC for CQT = .85 Accuracy scores for 52 lab studies examining the validity of the polygraph (National Academy of Science, 2003)
Identifying Accurate Predictors: The Fake Suicide Note • What discriminates between real and fake notes? • Structure variables • Content variables • Combinations of the two • How can we tell if some are better than others? • How can we determine how much better?
4 3 2 1 1 - Structure model (AUC = .69) 2 - Content model (AUC = .74) 3 - Optimal model (AUC = .82) 4 - Inclusive model (AUC = .85) Optimal: Average sentence length (shorter in genuine) Positive affect (more in genuine) Probability of false alarms (Jones & Bennell, 2005)
Setting Appropriate Thresholds:Linking Serial Burglaries • Many behaviours can be used • Inter-crime distance is very useful (AUC=.80) • The closer two crimes are to one another the more likely it is that they are linked • When should we decide that crimes are linked? • What is an appropriate threshold?
3.50km (85% H, 45% FA) 2.50km (72% H, 25% FA) 1.50km (58% H, 11% FA) 1.10km (49% H, 10% FA) (Bennell & Canter, 2002)
Comparing Decision Makers:Bite Mark Identification • It is often unclear as to the source of a bite mark • Non-accidental (adult) • Accidental (child) • How well can people perform on this task? • Is expertise required?
Senior forensic dentists (AUC=.69) Student dentists (AUC=.69) Junior forensic dentists (AUC=.68) General dentists (AUC=.62) Social workers (AUC=.63) Police officers (AUC=.62) (Whittaker et al., 1998)
Conclusions • ROC analysis provides a way of calculating accuracy scores that are not dependent on arbitrary thresholds • ROC analysis also provides a way of identifying thresholds that meet the demands of specific situations • The result for police investigations will likely be improved decision-making accuracy and utility