Lag Selection in Subset VAR Models with an Application to a U.S. Monetary System

In this paper we consider alternative modeling strategies for specification of subset VAR models. We present four strategies and show that under certain conditions a testing procedure based on t-ratios is equivalent to eliminating sequentially lags that lead to the largest improvement in a prespecified model selection criterion. One finding from our Monte Carlo study is that differences between alternative strategies are small. Moreover, all strategies often fail to discover the true model. We argue that finding the correct model is not always the final modeling objective and find that using subset strategies results in models with improved forecast precision. To illustrate how these subset strategies can improve results from impulse response analysis, we use a VAR model of monetary policy shocks for the U.S. economy. While the response patterns from full and subset VARs are qualitatively identical, confidence bands from the unrestricted model are considerably wider. We conclude that subset strategies can be useful modeling tools when forecasting or impulse response analysis is the main objective.