I626 IEEE TRANSACTIONS ON SYSTEMS. MAN, AND CYBERNETICS, VOL. 23. NO. 6, NOVEMBERIDECEMBER 1993 Distributed Detection in Teams with Partial Information: A Normative-Descriptive Model A. Pete, K. R. Pattipati, and D. L. Kleinman Abstract-This paper considers a hierarchical team faced with a binary detection problem, wherein decision makers (DM’s) have access to different subsets of noise-corrupted information about the true state of the environment. A normative model is developed that aggregates individual expertise of DM’s at dif- ferent levels of hierarchy. The resulting team expertise is char- acterized in the form of a Team Receiver Operating Charac- teristic (ROC) curve, thereby replacing the team by an equivalent single decision-making node. The normative model is tested against human teams in a laboratory experiment. The team objective is to minimize the cost of errors in the final de- cision at the primary DM, where the cost structure and the information structure are treated as independent variables. Discrepancies between normative predictions and experimental results are attributed to inherent limitations and cognitive biases of humans. These human characteristics are quantified and the normative model is augmented with psychologically in- terpretable (descriptive) factors. The resulting normative- descriptive model yields accurate predictions of both the formance and strategy variables of human teams. I. INTRODUCTION A. Motivation N LARGE-SCALE systems that involve humans, I chines, computers, databases and communication per- ma- net- works (e.g., electric power networks, military command and control, air traffic control, product distribution and supply), problem scope and complexity often require that the information acquisition, processing, and decision- making functions be distributed over a team of human de- cision makers (DM’s), who may be geographically sepa- rated. A key problem in such team decision-making con- texts is the aggregation of individual capabilities into a team expertise; in other words: how does a team of ex- perts become an expert team? Based on their local knowl- edge and expertise, DM’s process subsets of the total in- formation available to the team and individual assessments are aggregated into a final decision. Indeed, team exper- tise is a result of coupled individual and team level pro- cesses. Manuscript received April 10, 1992; revised January 22, 1993. This work was supported in part by the National Science Foundation under Grant IRI- 8902755 and in part by the Office of Naval Research under ONR Contract N00014-90-J-1753. The authors are with the Department of Electrical and Systems Engi- neering, University of Connecticut, Storrs, CT 06269. IEEE Log Number 9209647. Existing models of binary (dichotomous) team decision problems, such as “Distributed Binary Hypothesis Test- ing” in signal detection theory (SDT) and “Group Judg- mental Accuracy” in social sciences generally assume that all the DM’s (variously referred to as voters, nodes, sen- sors, etc.) have their own conditionally independent mea- surements with respect to the task to be judged, and that these local opinions of the DM’s are to be aggregated into a final decision. In this paper, we consider a more realistic class of decision problems wherein DM’s have uncertain knowledge of different local events (local hypotheses), which are only probabilistically related to the team task (global hypotheses). Some examples are the following: 1) In clinical judgment tasks, a physician is presented a number of “measurements” (such as the results of di- agnostic tests: X-ray, blood, ultrasound, CAT-scan, mag- netic resonance imaging, etc.) to estimate the likelihood of a given disease. Each measurement relates to the con- dition of a specific organ of the human body. If the mea- surement is within prespecified limits, the organ is con- sidered normal; otherwise it is abnormal. A given test (such as blood pressure) yields different results with re- spect to different diseases. Consequently, the results of various diagnostic tests must be interpreted in the context of a particular disease and then combined into either nu- merical judgments (e.g., a rating on the severity of the disease) or categorical classification (e.g., a tumor judged malignant or benign) 111, 121. 2) The process of auditing a client’s financial state- ments is comprised of three sequential phases: i) review and evaluation of the client’s system of internal controls; ii) testing the features of the internal control system (termed the assessment of control reliance); and iii) per- forming the necessary direct tests of account balances [3]. Control reliance is defined as the risk that an intolerable, uncontrollable error is generated in a transaction cycle (e.g., disbursement of vendor payments). A transaction cycle may consist of several processes (e.g., preparing disbursement request forms, computer entry of data, run- ning a program to produce a weekly disbursement file, etc.). In order to determine the control reliance for the whole cycle, an auditor assesses the control risk for all the constituent processes separately and then aggregates them into a final risk assessment. An error in a composite 001 8-9472/93$03 00 @ 1993 IEEE