Y Qian et al./ European Economic Review 43(1999)1085-1094 1091 cm2=2C[1-(1-p)6 Under the U-form, the top manager is responsible for coordinating the four tasks. He thus receives four messages through noisy communication, each corresponding to one of the four tasks. When the program is bad(with probabil y 1-p), the reform fails and a new program will be tried in the next period. If the program is good (with probability p), there are three possibilities: (i)With probability 24, coordination is successful for both products A and B. (ii)With probability(1-12)2, coordination fails in both A and B. This will give the same outcome as a bad program. (iii With probability 22(1-2), coordination for good, the top manager will use the same program for m ng that the program is one of the two products is successful. In this case, know product in which the coordination failed and solve only the attribute matching problem in the next period. Hence, the payoff of reform under the U-form 兀u2=p{A/1-)+22(1-2LA2(1-) +6]+(1-22)bxu2}+(1-p)5ru2 where r is the expected payoff of change for one product for a good program, or Using the above recursive formula of T, we obtain ru2=2pA1-(1-i2)26]/(-b)[1-(1-12k ×[1-b[p(1-12)2+(1-p) When a reform program is introduced in period 1, a setup cost C is paid instead of 2C in the M-form) because only the top manager does attribute matching With probability 1-p the program is bad, which is discovered after one period. With probability p(1-12) the program is good but coordination fails for both products. In both cases, a new program is tried in the next period When the program is good and coordination is successful for at least one of the two products, the program will be known to be good. In such a case, it reasonable(and consistent with our assumptions on costs) to assume that no new setup cost needs to be paid in the next period. Indeed, the top manager has already been trained for that program and he has been able to successfully coordinate attribute matching for one product. Under this assumption, we have cu2=C+[p(1-42)2+(1-pkcu2 cu2=C/1-[p(1-12)2+1-p]6}or c .2 "2C/[1!(1!p)d]. Under the U-form, the top manager is responsible for coordinating the four tasks. He thus receives four messages through noisy communication, each corresponding to one of the four tasks. When the program is bad (with probabil￾ity 1!p), the reform fails and a new program will be tried in the next period. If the program is good (with probability p), there are three possibilities: (i) With probability j4, coordination is successful for both products A and B. (ii) With probability (1!j2)2, coordination fails in both A and B. This will give the same outcome as a bad program. (iii) With probability 2j2(1!j2), coordination for one of the two products is successful. In this case, knowing that the program is good, the top manager will use the same program for the product in which the coordination failed and solve only the attribute matching problem in the next period. Hence, the payo! of reform under the U-form is n 62 "pMj4A/(1!d)#2j2(1!j2)[A/[2(1!d)] #dn]#(1!j2)2dn62 N#(1!p)dn62 , where n is the expected payo! of change for one product for a good program, or n"j2A/[2(1!d)]#(1!j2)dn. Using the above recursive formula of n, we obtain n 62 "j2pA[1!(1!j2)2d]/M(1!d)[1!(1!j2)d] ][1!d[p(1!j2)2#(1!p)]]N. When a reform program is introduced in period 1, a setup cost C is paid (instead of 2C in the M-form) because only the top manager does attribute matching. With probability 1!p the program is bad, which is discovered after one period. With probability p(1!j2)2 the program is good but coordination fails for both products. In both cases, a new program is tried in the next period. When the program is good and coordination is successful for at least one of the two products, the program will be known to be good. In such a case, it is reasonable (and consistent with our assumptions on costs) to assume that no new setup cost needs to be paid in the next period. Indeed, the top manager has already been trained for that program and he has been able to successfully coordinate attribute matching for one product. Under this assumption, we have c 62 "C#d[p(1!j2)2#(1!p)]c 62 or c 62 "C/M1![p(1!j2)2#1!p]dN. Y. Qian et al. / European Economic Review 43 (1999) 1085}1094 1091