The economic consequences of euro-area macro-modelling shortcuts

Abstract

Should euro-area economies be modelled in an aggregate (area-wide) fashion or in a disaggregate (multi-country) one? This article tackles that question from both statistical and economic viewpoint. From a statistical viewpoint, aggregation bias criteria are found to signal that the degree of structural heterogeneity among euro-area economies is such that the loss of information entailed by an aggregate modelling approach may be far from trifling. From an economic viewpoint, we investigate the following issue. Are those statistically detectable heterogeneities of any practical relevance when it comes to supporting monetary policy decision-making? To provide an answer to this question, we compute simple optimal monetary policy reaction functions on the basis of either an aggregate model or a disaggregate one, and compare the associated welfare losses. The results suggest that the welfare under-performance of an area-wide-model-based rule is not only nonnegligible, but also robust with respect to a number of sensitivity analyses.

Acknowledgements

We thank the late Albert Ando, Paolo Angelini, Fabio Busetti, Jürgen von Hagen, Andrew Hughes Hallett, Marco Lippi, Daniele Terlizzese, John Williams; two anonymous referees; participants in the CEPR-ZEI Conference ‘Empirical Models of the Euro Economy 2: Sectoral Aspects and Performance,’ Bonn, 2003; participants in the Ecomod 2003 Conference, Istanbul, 2003; and participants in workshops at the Center for Financial Studies (CFS), the Banque de France and University of Rome ‘La Sapienza’ for many helpful comments and suggestions on previous versions of this work. The usual disclaimers apply. The views expressed in this article are those of the authors and do necessarily represent those of the Bank of Italy.

Notes

¹ A (very) partial list of works that have a bearing on this issue includes Ramaswamy and Sloek (1997), Dornbusch et al. (1998), Guiso et al. (1999), Clements et al. (2001), Ciccarelli and Rebucci (2002), Perman and Tavera (2005), Dedola and Lippi (2005), and the articles collected in Angeloni et al. (2003). Not much effort has yet been devoted to trying to identify, with the aid of macroeconometric models, the structural determinants underlying the observed heterogeneities. Fragmentary evidence may be found in van Els et al. (2003).

² From the viewpoint of the debate on robust rules (see the contributions in Taylor, 1999; as well as Levin and Williams, 2003; and Adalid et al., 2005), this article may be viewed as focussing on one particular type of robustness (i.e. robustness with respect to the assumption of aggregability).

³ Labhard et al. (2001) argue that the actual choice of the aggregating function is unlikely to affect the properties of the model in any significant way. The aggregating functions used in this article are briefly described below.

⁴ Both approaches are being pursued in practice, even by the same institutions. For instance, the ECB maintains an aggregate econometric model (Fagan et al., 2001), called AWM (Area Wide Model), and a disaggregate one, called MCM (Multi Country Model). Also, the euro-area projections by the Eurosystem (the institution comprising the ECB and the national central banks of euro-area countries) are the result, of a multi-staged process that involves both country-specific and area-wide models (ECB, 2001).

⁵ See Rudebusch and Svensson (1999) for a strictly backward-looking version and Clarida et al. (1999) for a fully forward-looking one.

⁶Lucas (1976). See, however, Sims (1982), Farmer (1991), Favero and Hendry (1992) and Altissimo et al. (2002) for a number of arguments that mitigate the implications of the Lucas Critique.

⁷ A Wald test cannot reject the restriction (p-value of 18.8%).

⁸ Neither equation includes foreign variables; in other words, the euro-area is modelled as a closed economy, similarly to what is systematically done for the US, and increasingly for the euro-area as well.

⁹ The source of data is the ESA-95 National Accounts for inflation and the output gap, and the BIS data-bank for the short-term interest rate. Inflation is measured by the quarter-on-quarter rate of change of the (seasonally adjusted) households’ consumption deflator. Potential output was estimated by applying a band-pass filter (Baxter and King, 1995) to the (log) GDP, selecting frequency components of 32 quarters and higher, with a truncation of 16 quarters. National variables were aggregated using a fixed-weight procedure. For inflation, 1999 PPP consumer spending shares were used. For the output gap, the weights are given by 1999 PPP real GDP shares. For interest rates, the weights are the PPP nominal GDP shares computed by the OECD. The GDP and consumer spending weights are, respectively 0.43 and 0.44 for Germany, 0.29 and 0.27 for France, 0.28 and 0.29 for Italy.

¹⁰ A Wald test on the restriction cannot reject with p-values of 16.8%, 5.2% and 83.3% for Germany, France and Italy, respectively.

¹¹ For most of the sample period, the exchange rates among Germany, France and Italy, though constrained by the ERM of the EMS, were not fixed; lagged values of all variables included in the model were used as instruments for the exchange rates. As in the AEAM, the interest rate enters the AD equations with a lag.

¹² In keeping with the approach followed in similar literature, neither the AEAM nor the DEAM that we use for our experiments in this article include any constant terms, i.e. they may be taken to provide a description of the functioning of the euro-area economy in the neighbourhood of equilibrium. This amounts to implicitly assuming that the same equilibrium values apply to all countries, a condition that does not necessarily hold in the sample period, particularly regarding the (implied) equilibrium real interest rates.

¹³ For more details, see the mimeo version of this article (Monteforte and Siviero, 2003). As a term of comparison, see the evidence presented in the articles collected in Angeloni et al. (2003).

¹⁴ A thorough comparison may be found in Monteforte and Siviero (2003).

¹⁵ More detailed results may be found in Monteforte (2007).

¹⁶ The simple optimal monetary policy rules we consider have been used in earlier works modelling the monetary policy of the ECB (e.g. Garcia-Iglesias, 2007).

¹⁷ For this purpose, we solve a standard stochastic linear regulator problem (see Chow, 1970; Sargent, 1987; and for an application to the issue of optimal monetary policy design, Rudebusch and Svensson, 1999).

¹⁸ Most of the computations relied on Matlab routines (in turn calling AIM to compute the state-space representation of the models; AIM is available on the website of the Board of the Federal Reserve System). The simulations with random drawings from the distribution of estimated residuals were run in Speakeasy/Modeleasy+. All software codes are available from the authors upon request.

¹⁹ The countries that would benefit the most from adopting the DEAMBR rather than the AEAMBR are Germany and France.

²⁰ Each replication includes 800 realizations of the shocks for the six stochastic equations in the model, one realization per period. Only the last 400 simulated values are used to evaluate the objective function, to prevent the results from being biased by the initial conditions.

²¹ The test is the standard one-sided test for the equality of the means of normally distributed variables. The 1% critical value is 2.326.

²² See Monteforte and Siviero (2003) for details.

²³ See the results in the mimeo version of this article, Monteforte and Siviero (2003), where the sample ends in April 1998.

²⁴ Hughes Hallet and Piscitelli (2002) show that integration may or may not imply convergence, a key factor in determining the result being the size of the economies involved.

²⁵ It would, of course, be possible to introduce the further complication that the speed of convergence is not the same for all countries. For the sake of simplicity we ignore that possibility. Let us just remark that our definition of partial convergence tends to make cross-country heterogeneity disappear more smoothly than it would be possible.

²⁶ Specifically, the hybrid model is such that, in the AS, currrent inflation reacts to both past and future expected inflation; in the AD, the real interest rate is defined as the difference between the nominal interest rate and average expected inflation four quarters ahead. The model was estimated with GMM. Estimation details are available from the authors upon request.

²⁷ Hughes Hallett (2000).