8 Coordination in organizations: A game-theoretic perspective

a-realistic context (subjects choose numbers or letters). Subjects are also paid substantial financial incentives to help ensure that they really do prefer (H,H) to (L,L). The key design elements abstract context and financial incentives fit together to achieve the main purpose, which is to operationalize a setting about which game theory makes a prediction. The purpose of psychology experiments7 is generally to express an informal intuition or regularity in a dramatic, lifelike way or yield data with which to construct a new theory. Because subjects are often unpaid (e.g., participating for course credit), an abstract task might bore or confuse them, so a concrete one is better. Because their purpose is different from the theory-testing motive of economists, psychologists might study coordination differently. For example, we could ask two subjects to perform a mildly aversive task, like pedaling a bicycle at either of two programmed rates, slow or fast, and pay them each a larger prize if they both pedal fast. (Tasks like this were used in early "social loafing" experiments; see Kidwell & Bennett, 1993.) Simultaneous pedaling resembles an assurance game, but we cannot be sure it is. It is not an assurance game if either subject is fanatic enough about exercise to simply prefer pedaling faster; then (fast) is a dominant strategy so (slow,slow) is not an equilibrium. It is also not an assurance game if one subject (or both) dislikes pedaling so much that he would rather forego his prize than match the other player's fast pedaling; then (slow) is a dominant strategy and (fast, fast) is not an equilibrium. So to be sure that pedaling is really an assurance game, we must be sure that both players do not like pedaling too much or like it too little (and also be sure that players know this about others) or else control these preferences and beliefs.8 One way to control them is to use an abstract task like choosing letters or numbers and pay financial rewards, as in Figure 8.2. The advantage of a task like pedaling is its improvement in realism, external validity, or generalizability to organizations (compared to abstract number choosing for money). But this apparent improvement can be reasonably disputed on two grounds. First, judgments of external validity should be grounded in a statistical understanding of what organizational situations are common. Are pedaling-like tasks really more common than choosing abstract symbols that yield monetary rewards? The answer is hard to know. Consider a law firm. Attorneys on a litigation team who spend long hours in court, exerting fatiguing effort, may be like subjects pedaling bikes. But other lawyers research legal details in a library or compile abstract symbols into briefs; perhaps their work is more like choosing letters that yield financial payoffs than like bike pedaling. Second, it is unclear that experiments should be like a random sample of realistic organizational situations. Scientists routinely oversample unusual situations (neglecting mundane ones). Neuroscientists study victims of unusual traumas, biologists study strange species, economists sfudy deCoordination in organizations 183 pressions and market crashes, and astronomers study supernovas. These phenomena are all rare and hence highly unrealistic, but they prOvide an efficient way to learn about general processes. While hesitating to claim the same status for the simple experiments described previously, we simply want to question (1) how to determine which experiments are most realistic and (2) whether realism is desirable. A final thought about context: Economic experiments are happily unrealistic because economics and game theory make no special predictions about behavior in more or less realistic situations. (If the purpose of the experiment is theory testing, the experiments need only be as realistic as the theory.) This does not mean that game theorists insist that context does not matter to outcomes. It simply means that there is no systematic theory of precisely how context matters. Obviously, one way to build up such a theory is to start with context-free baseline experiments (like the 2 x 2 results described earlier) and then experiment incrementally with contextual changes (like oneand two-way communication). Further changes of organizational interest can be added easily.