Actually in a paper I reported normalized co-integrating vector coefficient instead of OLS. The reviewer commented that the power of it is very small. Can any one explain this to me? Actually I followed it from a research paper
Normalization is arbitrary. You normalize on the coeffcient of the variable that is assumed to be endogeneous. The Engle Granger (OLS) method provides a normalization a priori on the coefficient of the left hand side variable. The Johansen (ML) approach does not provide a normalization. Instead, you can normalize the coefficients after the estimation is done. So, for example, if you are interested in a money demand function, you should divide all the long run parameters by the coefficient of real money balances. Hence, normallization could lead to a better interpretation of the results.
Normallization does not lead to a change in power. Probably, the point is that the Engle Granger method is less reliable than the Johansen approach: the former distinguishes between endogenous and exogeneous variables a priori (i.e. without testing), cannot account for the fact that you may have multiple cointegrating vectors and estimates the long run relationship by ignoring short run dynamics.