A CEPR conference on ‘Model Specification, Identification and Estimation in Empirical Macroeconomics’ was held in Perugia on 10/11 January 1997. The aim of the conference was to bring together academics and central bankers from both sides of the Atlantic to discuss a variety of econometric analyses of the monetary transmission mechanism. While the issues are of obvious policy relevance for central bankers and macroeconomists in general, they remain controversial both at the theoretical level and at the level of measurement. In recent years, there has emerged a body of empirical research relating to the US experience. This has generated debate both about methodological issues – for example, about the appropriate analytical tools – and about establishment of the ‘stylized facts’. The conference organizer was Lucrezia Reichlin (ECARE, Université Libre de Bruxelles and CEPR), and the conference hosts were Banca d’Italia.

The first paper was ‘Modelling Money’ by Lawrence Christiano (Northwestern University), Martin Eichenbaum (Northwestern University) and Charles Evans (Federal Reserve Bank of Chicago). Much of the literature on monetary shocks assumes that monetary policy is endogenous in that such shocks induce policy reactions. Quantitative general equilibrium monetary models, however, assume that this is exogenous. This paper set out to reconcile the two views. The authors showed that there is observational equivalence between the endogenous and exogenous policy rules. On the basis of this result – which does not depend on the uniqueness of the equilibrium – the paper proposed a methodology for estimating an exogenous monetary policy rule which does not induce mis-specification, and in which a comparison is made of the economy’s response to monetary policy shocks under both an endogenous and an exogenous policy rule.

The procedure involved estimating a VAR model with endogenous policy, computing the impulse response function to monetary shocks, and then estimating statistically admissible univariate exogenous representations of monetary policy rules. Finally, an exogenous policy rule is chosen which is able to reproduce the estimated responses to monetary shocks under an endogenous rule. The authors applied this procedure to the United States for the period 1965–85. They found that a contractionary shock induced a rapid hump-shaped output response, a delayed price response, and a reduction of money. The results were sensitive to the monetary aggregate chosen.

In view of the observational equivalence result, the authors argued that economic models should take into account both representations of monetary policy. They suggested this could be done by incorporating differential short-term elasticities for narrow and broad monetary aggregates, and they constructed a limited-participation model of credit-market imperfections which reproduced the result.

Charles Bean (LSE and CEPR) suggested that, since only a small fraction of the variation of the money stock appeared to be caused by monetary policy shocks, the procedure should also be used to confront the effects of non-monetary shocks. But he also questioned whether the approach would always lead to reliable, statistically ‘consistent’ inference, since it was unclear what the monetary ‘shocks’ corresponded to in reality. While they might reflect random optimization errors, or ‘trembles’ by policy-makers, the model did not allow for similar errors by private agents. But the shocks could also plausibly be interpreted as reflecting private knowledge about the state of the economy that is not available to the econometrician, in which case they would no longer be exogenous. Lucrezia Reichlin (ECARE, Université Libre de Bruxelles and CEPR) suggested the need for a better sensitivity analysis with respect to the identification scheme, and Helmut Lütkepohl (Humboldt-Universität zu Berlin) noted that co-integration issues, which were not considered, may alter the results.

Since much of the literature on measuring monetary policy using identified VARs focuses on the United States, it in effect considers only closed-economy models. By contrast, policy discussions on monetary stances in open economies often use monetary conditions indices (MCIs), which include exchange rates. In ‘Measuring Monetary Policy in the G–7 Countries: Interest Rates Versus Exchange Rates’, Frank Smets (Bank for International Settlements and CEPR) sought to reconcile the two approaches. In considering the measurement of monetary policy shocks in an open-economy framework for the G–7 countries in the post-Bretton Woods and EMS-ERM periods, Smets showed that the exchange rate played a significant role.

In a simple model, the paper argued that targeting a weighted average of the real interest rate and the real exchange rate in response to excess demand shocks is an optimal monetary policy reaction function. The optimal weights are based on the elasticities of aggregate demand with respect to the real interest and exchange rates, which may depend in turn on the openness of the country. The model also suggests a number of identification restrictions. In particular, within a four-variable VAR, which includes output, prices, a three-month interest rate and the exchange rate, Smets showed how knowledge about the weight applied in the MCI can be used as an identification restriction to distinguish unilateral monetary policy shocks from policy responses to exogenous exchange rate shocks. His identification scheme encompassed the extreme cases of targeting often used in the VAR literature, and also allowed for a sensitivity analysis of the estimated effects of monetary policy to the assumed weight.

Several results emerged for the G–7 countries, excluding the United States: 1) For all countries, except the United Kingdom, the case of pure interest-rate targeting was generally rejected. In the UK case, there were signs of a more general mis-specification of the model. 2) For almost all countries, the case of pure exchange rate targeting was also rejected. 3) The weight attached to the exchange rate differed across countries, being larger for those – Germany, France and, to a lesser extent, Italy – in the ERM than for the others. 4) Although the role of the exchange rate differed across countries, the estimated effects were generally consistent with theory. A contractionary monetary policy shock leads to an increase in the real interest rate, an appreciation of the real exchange rate, and a fall in output and prices. 5) The sensitivity of the estimated impulse-response functions to the weight is low in Germany and Japan, but larger in other countries.

Daniele Terlizzese (Banca d’Italia) raised questions about identification issues. First, while Smets had assumed a vertical long-run Philips curve, the specification of the paper implied that persistent demand and/or supply shocks would have long-run output effects. Since this did not rule out a long-run output-inflation relationship was there not a contradiction? Second, Smets measured a monetary policy shock as a shock to the inflation target. Again, this appeared contradictory in that the inflation target is a time-series process yet, given the theoretical objective function of the monetary authority, the target should be constant. Jérôme Henry (European Monetary Institute) suggested introducing commodity prices to capture the linkages between exchange rates and prices and, eventually, resolving some strange results. He also asked whether changes in the ERM might have introduced time instability over the period analysed.

‘Measuring Monetary Policy with VAR Models: An Evaluation. Does Germany Differ From the US?’, by Fabio C Bagliano (Università di Torino) and Carlo A Favero (IGIER, Università Bocconi and CEPR), also investigated monetary policy shock issues. These included model specification, particularly the choice of sample period; the measurement of shocks, through comparing those identified in a VAR with ‘computed’ shocks based on observable financial-market variables; and the role of the long-term interest rate in transmitting shocks. The authors examined these questions in a structural VAR for the United States over the period 1965–88. Their identification assumptions were (a) no contemporaneous reaction of macroeconomic to monetary policy variables; (b) recursive contemporaneous relations among the macro variables; either full adjustment of non-borrowed reserves to shocks on the demand for total reserves and for borrowed reserves, or adjustment only in respect of demand for total reserves, or adjustment to neither shock.

On the specification issue, the authors concluded that the model is stable over a period which covers a single monetary regime. Specification and parameter stability tests, however, revealed that the estimated VAR model was mis-specified and unstable over the whole sample period 1965–88, but was stable for 1988–96. On the measurement issue, they compared four measures, one relying on the structural VAR they adopted, one given by the diference between the 30 days’ Fed interest rate and the 30 days’ Fed funds futures contract, a third based on the changes in the three months’ interest rate at the dates of monetary policy announcement, and the fourth being the difference between observed and expected overnight rates at the dates of monetary policy meetings. The first three measures had low correlations, but including them in the VAR yielded similar results in terms of responses to the shocks. The last measure had a very low correlation with the others and yielded very different results. Finally, the authors showed that the long-term interest rate is a significant determinant of the monetary policy reaction function.

Frederic Mishkin (Federal Reserve Bank of New York and NBER) questioned the usefulness of VAR analysis for understanding monetary policy. VAR analysis provides information about the responses to shocks, not about the transmission of monetary policy. Furthermore, monetary policy shocks which may be interpreted as surprises may not correspond to any real policy. The many arguments in favour of a transparent monetary policy also favour no monetary policy surprise: a transparent monetary policy may solve the time-inconsistency problem, reduce political pressure and decrease uncertainty. Mishkin also noted a surprising result of the paper, namely the low correlation among alternative measures of shocks which nonetheless yield similar impulse response functions. He suggested these outcomes may have stemmed from measurement errors, but he also queried the value of the impulse response functions.

Francis Y Kumah (CentER, Tilburg University) presented a paper on ‘The Effect of Monetary Policy on Exchange Rates: How to Solve the Puzzles’. Empirical analyses of monetary policy shocks have generated two theoretical puzzles, namely the forward discount bias and exchange rate puzzles. First, there has been a failure to find, after a negative monetary policy shock, an initial appreciation followed by a gradual depreciation, as would be the case if uncovered interest rate parity held; second, there is the tendency of the domestic currencies (of non-US G–7 countries) to depreciate against the US dollar following domestic monetary tightening.

Kumah investigated whether these puzzles resulted from the measurement of monetary policy shocks based on particular identification schemes. He compared the results of a fully recursive model, a semi-recursive model based on the actual procedure of the Federal reserve system, and a structural VAR model which explicitly modelled international monetary policy interdependence. He found that the latter model yielded results most consistent with theory, while a fully recursive identification scheme generated both the forward discount bias and exchange rate puzzles. Other results were that the structural VAR model revealed significant monetary policy interdependence; and that exchange rate fluctuations were owing to exchange rate shocks and to both domestic and, to a lesser extent, foreign monetary policy shocks.

Fabio Canova (Universitat Pompeu Fabra and CEPR) raised questions about the specification and identification of the structural VAR model. He also drew attention to a high degree of short-run variability in the results. Concerning the choice of variables, he suggested first that measurement errors would be avoided by using foreign prices or inflation, rather than the real exchange rate, and second, that unemployment rates are poor indicators of labour-market characteristics in Europe. On the identification issue, he noted that the foreign interest rate in the model had been related to shocks to non-borrowed reserves and the exchange rate, but not to shocks to total reserves and the Fed Funds rate – a strong assumption. In addition, the exchange rate had been assumed to depend contemporaneously on all variables in the system, possibly contradicting the fact that exchange rate markets are forward-looking. Canova further argued that, since the monetary policy reaction function depends on all variables in the system, the identification of the real sector should be considered more carefully.

Frank Smets also questioned the identification scheme, arguing that it was legitimate to omit one transmission channel for foreign output shocks, since such shocks contemporaneously affect domestic output but not the domestic interest rate. Furthermore, there is no restriction on the contemporaneous relation between the exchange rate and the foreign and domestic interest rates.

In ‘Are the Effects of Monetary Policy Asymmetric?’, René Garcia (Université de Montréal) and Huntley Schaller (Carleton University) analysed the effects of monetary policy on the growth rate of real output according to the state of the business cycle and the probability of a regime-switch. Asymmetries between expansions and recessions may result from a liquidity trap, an L-shaped curve, or finance constraints. The authors found evidence that asymmetries did exist. They also found support for the view that interest rates affect the probability of a regime-switch. Garcia and Schaller argued that VAR models are not appropriate for studying asymmetries. Consequently, they opted instead for Markov switching models which relate the dynamics of output growth to monetary policy shocks according to the state of the economy, and in which the probability of a regime-switch may depend on monetary policy shocks.

The paper’s findings were that: a contractionary monetary policy shock lowered output growth; this effect was stronger in a recession than during expansion; such a policy shock increased the probabilities of moving from expansion to recession and of staying in recession; and successive expansionary monetary policy shocks substantially increased the probability of moving out of recession. These results were robust with respect to a wide variety of alternative assumptions, data classifications and variables.

Helmut Lütkepohl (Humboldt Universität zu Berlin) commented that the results raised serious questions for analyses that do not take into account the possibility of asymmetries. In particular, the impulse response functions derived from linear models may be misleading. He enquired why the authors had not used multivariate Markov switching models, but also raised the possibility of using other types of non-linear models, such as smooth-transition models, in order better to capture the differences between recession episodes and between expansion episodes. He also noted the need to pay more attention to the order of integration and cointegration of the variables. In response, Garcia argued that multivariate models would render interpretation of the states of recession and expansion more difficult. He agreed that smooth-transition models might be a good option in multivariate systems, not least for incorporating unobserved variables such as thresholds, but noted that such models raise serious estimation difficulties.

Lawrence Christiano, who was impressed by the very large increase in the goodness of fit between the VAR models and the Markov switching model, noted that the effect of a positive monetary shock was a fall in interest rates followed by an increase in output. This held in expansions and in recessions. However, the paper showed that the magnitude of this effect differed between recessions and expansions. Ignazio Visco (Banca d’Italia) found the results of the paper counter-intuitive. In particular, the effect of monetary policy on the probability of moving out of recession was larger than that suggested by central bankers. There was also a need to to pay more attention to the interaction between output and prices, especially in recession.

‘Money and Output: A Reliable Predictor or an Intermittent Signal?’ was the title of a paper written by Anthony Garratt (Department of Applied Economics, Cambridge) and Andrew Scott (LBS and CEPR). They analysed the links between monetary variables and output in the United Kingdom over the period 1969–95 by investigating several questions: whether money was an important source of output fluctuations; whether money predicted output; which monetary aggregate was the most important source of output fluctuations; whether the money-output relationship varied over time; and whether the relationship was asymmetric, differing in recessions as compared to expansions.

Using a structural VAR identifying nominal (monetary) shocks as having no permanent impact on output, the authors claimed first, that real rather than nominal shocks explained the largest part of output fluctuations. Since money changes may move the aggregate demand curve, and hence alter prices, they also considered the link between prices and output. However, they found no correlation between the cyclical components of output and of prices with different leads and lags. Second, correlations and spectral analysis of the cyclical components of the variables revealed that, in contrast to long-term interest rates, monetary aggregates did not lead output over the business cycle. This result was robust with respect to different detrending techniques, although Granger causality tests indicated that M0 predicted output over the whole sample, and that personal credit did likewise after 1992.

Third, using a methodology that considered each expansionary and contractionary episode separately, Garratt and Scott found that the relationship between money and output was very unstable across monetary aggregates. Fourth, the relationship between each monetary aggregate and output was very unstable across business-cycle episodes. Furthermore, the results of stability tests suggested that the transmission mechanisms of monetary policy shocks had changed over time. Some regularities were identified, however: M0 and the short- and long-term interest rates have a stable correlation structure and relationship with output, especially in expansions. Finally, while the relationship between money and output was very unstable across expansions and recessions, there was no evidence of asymmetry.

In view of these results, the authors suggested that monetary aggregates were only one source of relevant information for setting interest rates, and that the importance of supply shocks made it significant to assess a wide range of non-monetary variables as well. They also argued that the differential behaviour of monetary aggregates across different business cycles suggested that, in any given business cycle episode, different shocks and different monetary transmission mechanisms might be at work. Consequently, while monetary aggregates might provide some useful information about the current setting of interest rates, extracting this information was an extremely complex business.

Harald Uhlig suggested the paper’s findings be compared with other VAR results and with the empirical evidence for the United States. On the finding of differences from cycle to cycle, he wondered about the possible causes of such changes. Candidates included oil prices, exchange rate regimes, policy regime changes, and developments in payment methods. Uhlig also questioned whether the paper’s purely empirical analysis could support the policy implications drawn by the authors, particularly with regard to the superiority of nominal GDP targeting over monetary targets. He stressed that the fact monetary aggregates behave differently across business cycles does not imply that policy also needs to be different across cycles. Frederic Mishkin argued that monetary aggregates may be useful as a simple signal to the public. The result of this paper (and other papers at the conference), however, revealed that monetary aggregates are not used in this way.

Lawrence Christiano also pondered the meaning of differences between cycles. In particular, should this finding be used to criticize existing monetary models? To the list of potential causes suggested by Uhlig, Christiano added taxation schemes. He disagreed with a suggestion from Scott that, for the operation of monetary policy, central bankers should be interested in elements of a VAR analysis other than the impulse response function of output to monetary shocks. VARs and estimated impulse response functions were an excellent means of trying to discriminate between different monetary models, and research on them could contribute to the design of monetary policy rather than merely provide input into day-to-day monetary policy decisions.

Eugenio Gaiotti, Andrea Gavosto and Giuseppe Grande (all Banca d’Italia) were the authors of ‘Monetary Policy and Inflation in Italy in the 1990s: Interpretations and Evidence’. Their paper investigated the relative importance for inflation of various transmission mechanisms that have featured in the recent debate about Italian monetary policy strategy over different periods. These included the impact on inflation of aggregate demand fluctuations, inflation expectations, the 1992 lira depreciation, the 1995 lira appreciation and expectations of the sustainability of the public debt. The paper compared the results of studies based on the Bank of Italy’s econometric model with those of a recursive VAR estimated over the period 1985–95.

The authors measured monetary policy shocks as the interest shocks in a recursive VAR for the following variables: output gap, wages, inflation, short-term interest rate, import price (or government default premium), exchange rate and inflation expectations (measured by survey observations). They found that inflation reacted positively to shocks on the output gap, wages, import prices and inflation expectations, but negatively to exchange rate appreciations and interest rate shocks. The effect of inflation expectations was statistically significant. Although shocks to fiscal expectations had a temporary impact on the exchange rate (but no significant effect on inflation), no evidence was found in favour of an hypothesis of fiscal dominance.

The authors used the VAR to focus on the determinants of inflation in 1992–5, following the lira’s exit from the ERM. Inflation fluctuations were ascribed to demand shocks, inflation expectations shocks, and exchange rate shocks. From 1992 to mid-1994, the decrease in inflation – despite the adverse effect of the 1992 exchange rate shock – was mostly owing to a fall in demand. The subsequent resurgence in inflation was the result of both positive demand shocks and adverse shocks to inflation expectations. In late-1995, the appreciation of the lira, together with a restrictive monetary policy, slowed down inflation. Thus exchange rate shocks were an important, but by no means the only determinant of inflation.

Guido Tabellini (IGIER, Università Bocconi, and CEPR) remarked first that the relative price adjustment phenomenon, which implies that inflation depends on the real exchange rate, was not considered in the paper. He suggested as well that foreign commodity prices, which are exogenous, be used in place of endogenous import prices. With reference to the identification assumptions, he called for more discussion of the sensitivity of the results to alternative schemes. In his view, the paper also placed too little emphasis on the role of costs and wages, which were the main focus of public debate on the causes of accelerated inflation after 1992, and he wondered about the effects of the break in the wages equation on inflation. Etsuro Shioji (Universitat Pompeu Fabra) also questioned the identification scheme for the exchange rate. The assumption that the exchange rate does not affect prices contemporaneously may be valid for Japan, because most Japanese imports are raw materials, but it is a questionable assumption for countries that import mainly finished products. He urged more precise analysis of the import structure before choosing identification assumptions.

Flavio Padrini (Minister of Treasury, Italy) had doubts about using inflation expectations in a VAR which already contains an endogenous inflation variable. An interest-rate shock might suffice instead to represent an inflation-expectation shock. On this point, Charles Bean argued that an inflation-expectation shock may actually represent another shock which is itself an important determinant of inflation. For Carlo Favero, the long-term interest rate or a spread may be such a determinant. He agreed with Padrini that inclusion of both a variable and its expectation raises impossible identification issues, or is redundant. Lucrezia Reichlin wondered what the relationship was between the conditional expectation of inflation, as given by the VAR, and expected inflation, while Ignazio Visco argued that there should be at least long-run consistency between both.

In responding to Tabellini’s comments, Gavosto explained that import prices were used in preference to commodity prices in order to verify the role of foreign producers’ pricing policies in transmitting exchange rate shocks to inflation. On the wages issue, wage moderation had been observed after 1992, but it was unclear whether this resulted from a structural break in the labour market or from the effect of demand and unemployment. In the VAR, wage moderation is endogenously determined by adverse demand shocks. In any event, he argued, the strong effect of the output gap on inflation in the VAR is partly transmitted through wages.

In ‘The Liquidity Effect and Long-Run Neutrality: Identification by Inequality Constraints’, Ben Bernanke (Princeton University) and Ilian Mihov (INSEAD, Fontainebleau) analysed two issues on which empirical research has reached no consensus: first, the liquidity effect, i.e. the interest rate reduction occasioned by an expansionary monetary-policy shock; and second, the long-run neutrality of money, which implies that monetary-policy shocks have no permanent effect on output.

The authors ascribed the failure to reconcile these phenomena with empirical evidence to the identification assumptions. As an illustration, they showed that, in a bivariate recursive VAR, there is no long-run neutrality of money – a result which is robust with respect to alternative orderings and monetary aggregates. If prices are introduced long-run neutrality may be achieved, but only at the price of other implausible results. This suggests that the results, and particularly the long-run phenomena, are sensitive to the short-run identification assumptions.

Bernanke and Mihov also addressed the issues of parameter instability, endogeneity of money and robustness with respect to identification schemes. Tests for structural shifts showed little evidence of parameter instability. Switching regression models, however, suggested some instability in the second half of the 1980s. Parameter instability in a bivariate VAR of output and money disappeared when commodity prices were introduced.

The paper relied on an identification scheme which modelled monetary policy as endogenous. It considered two sets of variables – macroeconomic variables (real GDP, GDP deflator, commodity prices) and policy variables (total and non-borrowed reserves, Federal Funds rate) – together with a set of relationship assumptions which yielded a significant liquidity effect and long-run neutrality at intermediate horizons, but not at horizons longer than seven years. A robustness analysis was performed via inequality identification constraints. Following the approach of King and Watson (1993), it specified a range of admissible short-run identification restrictions which imply a liquidity effect, then analysed the long-run responses of the system over this range of identification schemes.

The authors assumed three inequality restrictions covering a set of identification assumptions used in the literature: a downward-sloping demand curve for total reserves; an upward-sloping borrowing function; and Fed smoothing of interest rates (or, at least, non-amplification of demand and borrowing shocks). These constraints implied a range of admissible parameters that produce a liquidity effect. The results indicated first, that the size of the liquidity effect depended on the identification scheme; second, that the response of output to a monetary-policy shock was a monotonically increasing function of the liquidity effect; third, that long-run neutrality of money was rejected for horizons larger than ten years, even where accepted for shorter horizons; and fourth, the size of the liquidity effect was reduced after 1982 and subsequently may have become insignificant.

Lucrezia Reichlin questioned the comparison strategy and suggested both other identification schemes and other reasons for the failure to find long-run neutrality of money. In her view, the conference generally had highlighted the need for selection criteria for a general framework in which just-identified models could be compared. Other possible identification strategies included long-run constraints, short-run constraints, shape constraints and constraints that minimized some criterion. She was puzzled by the finding of long-run neutrality at intermediate, but not longer horizons. A possible explanation was the low precision of the long-run estimates. On the long-run neutrality problem itself, Reichlin proposed an alternative identification scheme which consists in imposing the shape constraint of long-run neutrality at some horizons (around six or seven years) which would give more scope for modelling the Fed’s operating procedure.

Overall, the papers presented at the meeting showed that the results of VAR analysis of monetary-policy shocks are sensitive to several factors: the sample periods, and particularly the policy regimes; the business-cycle episodes, where there may be large differences between expansions and recessions; different business-cycle episodes; the identification assumptions; and the monetary aggregate considered. There is thus a need to develop comparison and selection criteria between alternative models, specifications and identification schemes.