Y. Qian et al./ European Economic Review 43(1999)1085-1094 We define the expected net payoff under the M-form and U-form, respectively, U2=丌u2-cu2 Comparing these two expressions, one can easily see the basic trade-off between the M-form and the U-form organizations: the M-form benefits from advantages in coordination because of better use of local information but forgoes economies of scale which give the U-form lower costs in implementing reforms. Therefore the M-form will be more efficient than the u-form when communication quality is below a critical value, or when the setup cost is not too higl We next compare the trade-off between a big bang approach to reforms and a gradual approach under the M-form organization. Under the gradual ap- proach, a reform is tried first in one region and later extended to another region, conditional on the success of its implementation in the first region. If the program is a good one, the first period payoff is(A 1)/2. In the second period, the same program is then used in another region with a payoff of A in each period. However, if the program is bad, the experimenting region A will get 0 payoff and the non-experimenting region B will get In this case, a new experiment in region A will take place again in the next period. Therefore, the expected payoff of the M-form with experimentation is given by 兀m1=P(4/2(1-8)++6421-}+(1-pl+brm 兀m=[pA(1+0)+(1-0)/{2(1-δ[1-(1-p)]}, The setup cost in the first period is C because only region A,s manager does attribute matching. If the program is good, region B will use the same program in period 2 and another cost C will be paid in period 2 because region B manager needs to match attributes according to local conditions. Region B can thus imitate region A,'s success but cannot copy it since local coordination is still required to introduce a successful blueprint. With probability 1-P, the pro- gram is bad and a new blueprint must be tried. We are then back to the situation of period 1. Hence we get m1=C+OLpC +(1-p)cml] (1+p)C/1-(1-p)6] One can show that on the benefit side, the gradual approach always reduces the expected benefits from change as soon as reform brings a higher expected outcome than the status quo. This is due to the delay in the full implementationWe de"ne the expected net payo! under the M-form and U-form, respectively, M2 "n.2 !c .2 and ;2 "n 62 !c 62 . Comparing these two expressions, one can easily see the basic trade-o! between the M-form and the U-form organizations: the M-form bene"ts from advantages in coordination because of better use of local information but forgoes economies of scale which give the U-form lower costs in implementing reforms. Therefore, the M-form will be more e$cient than the U-form when communication quality is below a critical value, or when the setup cost is not too high. We next compare the trade-o! between a big bang approach to reforms and a gradual approach under the M-form organization. Under the gradual ap￾proach, a reform is tried "rst in one region and later extended to another region, conditional on the success of its implementation in the "rst region. If the program is a good one, the "rst period payo! is (A#1)/2. In the second period, the same program is then used in another region with a payo! of A in each period. However, if the program is bad, the experimenting region A will get 0 payo! and the non-experimenting region B will get 1 2 . In this case, a new experiment in region A will take place again in the next period. Therefore, the expected payo! of the M-form with experimentation is given by n .1 "pMA/2(1!d)#1 2 #dA/2(1!d)N#(1!p)M1 2 #dn.1 N or n .1 "[pA(1#d)#(1!d)]/M2(1!d)[1!(1!p)d]N. The setup cost in the "rst period is C because only region A's manager does attribute matching. If the program is good, region B will use the same program in period 2 and another cost C will be paid in period 2 because region B's manager needs to match attributes according to local conditions. Region B can thus imitate region A's success but cannot copy it since local coordination is still required to introduce a successful blueprint. With probability 1!p, the pro￾gram is bad and a new blueprint must be tried. We are then back to the situation of period 1. Hence we get c .1 "C#d[pC#(1!p)c .1 ] or c .1 "(1#pd)C/[1!(1!p)d]. One can show that on the bene"t side, the gradual approach always reduces the expected bene"ts from change as soon as reform brings a higher expected outcome than the status quo. This is due to the delay in the full implementation 1092 Y. Qian et al. / European Economic Review 43 (1999) 1085}1094