Dealing with limited overlap in estimation of average treatment effects