Each of the three methods is efficient in its nature. Depending on the aim of the study design, the following should be noted carefully and cautiously in selecting and applying the method upon which choice can be made.

#The FMOL computes estimators which employ semi-parametric corrections to eliminate potential problems caused by the long run correlations between cointegrating equations and stochastic regressors innovations. The resultant FMOLS) estimator is asymptotically unbiased and has fully efficient mixture normal asymptotics which allow for standard Wald tests using asymptotic Chi-square statistical inference. Much suitable for application where the series are cointegrated at first difference I(1).

#The CCR is closely related to FMOLS, but employs stationary transformations of the data to obtain least squares estimates to remove some long run dependence between the cointegrating equation and stochastic regressors’ innovations. CCR removes 2nd order bias, and estimates follow a mixture normal distribution that is free of non-scalar irascible parameters and permits asymptotic Chi-square testing.

#The DOLS constructs an asymptotically efficient estimator that eliminates the feedback in the cointegrating system. In addition, the DOLS augments the cointegrating regression with lags and leads to ensure that the resulting cointegrating equation error term is orthogonal to the entire history of the stochastic regressor innovations.

it depends on whether you are using single or multiple equation system

It all depends which system of equations has been adopted