Measuring model misspecification: application to propensity score methods with complex survey data