Y. Qian et al./ European Economic Review 43(1999)1085-1094 2. The model We consider an economy with two regions 'A' andB, and two functions, '1 and 2. There are four tasks: 14, 24, 1B, and 2B, where task ir involves process for region r. Each of the four tasks has many different attributes. A reform program is characterized by its new attributes. At the implementation stage of a reform program, unexpected contingencies appear which we call attribute shocks. Attributes between tasks Ir and 2r(r= A, B)must be matched in order to implement a reform program successfully e consider an infinite horizon with discount factor 8. One(and only one blueprint of the reform program is made available each period. With probability p the blueprint is a good one and with probability 1-p it is bad. Blueprints available over time are stochastically independent. We assume that if a blueprint is good, it will remain good in any region in the future; however, good coordina tion (i.e, attribute matching) in one region cannot be ' copied in another region because of differences in local conditions. If a blueprint tried in one region is good and coordination is successful, then the same blueprint can be used successfully elsewhere, but coordination in another region is still necessary in order to adjust to local conditions. In each period, a manager collects information about the attribute shocks and sends a message to another manager. Each message contains information about all the attributes in one task. We assume that information transmission between any two managers is imperfect so that the probability of each message being correct is i and the probability of being wrong is 1-, where0<i<1 e assume that the noises in information transmission are independent across tasks and over time. Based on the information received, the manager carries out his main job: attribute matching Consider the payoffs for unit A (payoffs for unit B are defined symmetrically) Let the status quo(without change)payoff in tasks 1A and 2A be i, respectively. The benefits from change are defined as follows. Suppose the program is good, then(i)with a change in task 1A but not in task 2A or vice-versa, the payoff is (A +1)/4 if the attributes between 1A and 2A are matched, 0 otherwise; and i with a change in both tasks 1A and 24, the payoff is A/2 if the attributes between 1A and 2A are matched, 0 otherwise. If the program is bad, then the payoff is always 0 when change is implemented. We assume that PA>l, i.e. the expected per period benefit from change, as compared to the status quo, is positive Ve assume that all blueprints are made available for free, but for each manager there is a setup cost associated with coordinating changes. This cost can be interpreted as a training cost: to implement a reform, the managers need to be trained on how to match the attributes of the reform. Because blueprints are free and the setup costs for coordination are not, when a failure occurs in the previous period (either due to a bad program or bad coordination), the2. The model We consider an economy with two regions &A' and &B', and two functions, &1' and &2'. There are four tasks: 1A, 2A, 1B, and 2B, where task ir involves process i for region r. Each of the four tasks has many di!erent attributes. A reform program is characterized by its new attributes. At the implementation stage of a reform program, unexpected contingencies appear which we call &attribute shocks'. Attributes between tasks 1r and 2r (r"A, B) must be matched in order to implement a reform program successfully. We consider an in"nite horizon with discount factor d. One (and only one) blueprint of the reform program is made available each period. With probability p the blueprint is a good one and with probability 1!p it is bad. Blueprints available over time are stochastically independent. We assume that if a blueprint is good, it will remain good in any region in the future; however, good coordina￾tion (i.e., attribute matching) in one region cannot be &copied' in another region because of di!erences in local conditions. If a blueprint tried in one region is good and coordination is successful, then the same blueprint can be used successfully elsewhere, but coordination in another region is still necessary in order to adjust to local conditions. In each period, a manager collects information about the &attribute shocks' and sends a message to another manager. Each message contains information about all the attributes in one task. We assume that information transmission between any two managers is imperfect so that the probability of each message being correct is j and the probability of being wrong is 1!j, where 04j41. We assume that the noises in information transmission are independent across tasks and over time. Based on the information received, the manager carries out his main job: attribute matching. Consider the payo!s for unit A (payo!s for unit B are de"ned symmetrically). Let the status quo (without change) payo! in tasks 1A and 2A be 1 2 , respectively. The bene"ts from change are de"ned as follows. Suppose the program is good, then (i) with a change in task 1A but not in task 2A or vice-versa, the payo! is (A#1)/4 if the attributes between 1A and 2A are matched, 0 otherwise; and (ii) with a change in both tasks 1A and 2A, the payo! is A/2 if the attributes between 1A and 2A are matched, 0 otherwise. If the program is bad, then the payo! is always 0 when change is implemented. We assume that pA'1, i.e. the expected per period bene"t from change, as compared to the status quo, is positive. We assume that all blueprints are made available for free, but for each manager there is a setup cost associated with coordinating changes. This cost can be interpreted as a training cost: to implement a reform, the managers need to be trained on how to match the attributes of the reform. Because blueprints are free and the setup costs for coordination are not, when a failure occurs in the previous period (either due to a bad program or bad coordination), the 1088 Y. Qian et al. / European Economic Review 43 (1999) 1085}1094