Large-scale macroeconometric models are now regularly used for forecasting and policy analysis. Such models are only approximations to the true structure of the economy, however, and the way they are used typically ignores the uncertainty inherent in their construction. It is important to take this uncertainty into account, especially if the models contain non-linear relations between the endogenous variables. Such non-linearity will lead to biased forecasts and inappropriate policy conclusions unless proper solution methods are employed, involving 'stochastic' simulations. It is commonly assumed that such non-linearity is unimportant and can safely be ignored. In Discussion Paper No. 86, Paul Fisher and Research Fellow Mark Salmon investigate the wisdom of ignoring non-linearity in model simulations.

Given the potential implications of these results for policy analysis and forecasting, it is not surprising that a number of researchers have attempted to measure the significance of non- linearity in their models, typically through 'Monte Carlo' simulations designed to assess the potential bias. Most model builders nevertheless continue to employ deterministic simulation methods, arguing that these experiments indicate that non- linearity is of little practical importance.

Fisher and Salmon contend that these experiments may have misleading conclusions. Mariano and Brown have shown that the forecast bias in a non-linear model may be decomposed into two parts, one due to non-linearity and the other to inconsistent parameter estimates. These two terms may take opposite signs and so offset each other. The overall bias may therefore be small even though the non-linearity may in fact be important.

The practical importance of non-linearity is in general difficult to quantify in large macroeconometric models, and simple illustrations using constructed non-linear examples appear to carry little weight with practitioners. Fisher and Salmon argue that the best way to determine whether non-linearity is important is by examining the non-linear effects present in existing large- scale models. They used the models of the UK economy constructed by the London Business School and the National Institute of Economic and Social Research. Their analysis and simulation experiments reveal that the apparent absence of non-linear bias found in earlier studies may also be due to deficiencies in the simulation methodologies employed.

The importance of non-linearities shown by these experiments suggests a need for simple tests for their presence. Fisher and Salmon develop two such tests, both of which confirm the importance of non-linearity in the LBS and NIESR models. The authors conclude that model proprietors should employ efficient stochastic simulation methods to ensure that forecasts are unbiased and that policy analysis can be conducted appropriately.