The system GMM estimator in dynamic panel data models which combines two sets of moment conditions, i.e., for the differenced equation and for the model in levels, is known to be more efficient than the first-difference GMM estimator. However, an initial optimal weight matrix is not known for the system estimation procedure. Therefore, we suggest the use of `a suboptimal weight matrix' which may reduce the finite sample bias whilst increasing its asymptotic efficiency. Our Monte Carlo experiments show that the small sample properties of the suboptimal system estimator are much more reliable than any other conventional system GMM estimator in terms of bias.