More robust estimation of average treatment effects using kernel optimal matching in an observational study of spine surgical interventions