It is argued that standard impulse response analysis based on vector autoregressive models has a number ofshortcomings. Although the impulse responses are estimatedquantities, measures for sampling variability such asconfidence intervals sometimes are not provided. Ifconfidence intervals are given, they often are based onbootstrap methods with dubious theoretical properties. Theseproblems are illustrated using two German monetary systems. Proposals are made for improving current practice. Specialemphasis is placed on systems with cointegratedvariables.