Workshop on resilience Paris 14 June 2007 SVAR analysis of short-term resilience: A summary of the methodological issues and the results for the US and Germany Alain de Serres OECD Economics Department 1
Stylised facts Euro area United States Other English-speaking countries 1 5.0 Real GDP growth Output gap 2.0 4.0 1.5 1.0 3.0 0.5 0.0 2.0-0.5 1.0-1.0-1.5 0.0 1998 99 2000 01 02 03 04 05 06 07 2008 1998 99 2000 01 02 03 04 05 06 07 2008-2.0 Growth domestic sectors 2 Growth in globalized sectors 3 6.0 12 5.0 10 8 4.0 6 3.0 4 2 2.0 0 1.0-2 -4 0.0 1998 99 2000 01 02 03 04 05 06 07 2008 1998 99 2000 01 02 03 04 05 06 07 2008-6 2
Empirical examination of resilience Defined resilience as the capacity to absorb shocks and to recover quickly following an adverse one. Considered SVAR methodology as a well-suited framework to analyse cross-country differences in the response of output and key components of demand to various shocks that can be identified under certain conditions. First step: estimation of individual system for G7 countries plus Spain 3
SVAR model 8 variables VARs including a mixture of foreign and domestic variables Foreign real GDP, oil prices, 2 components of aggregate demand (externally- and internally-focussed), real GDP, CPI inflation, government net lending, nominal interest rates Shocks identified on basis of restrictions on the matrix of contemporaneous feedback effects on the variables Bernanke (1986), Blanchard and Watson (1986), Sims (1986) Identification scheme allows for structural interpretation of shocks, though interpretation not always straightforward. 4
SVAR model: additional restrictions Block exogeneity assumption: domestic variables assumed to have no impact neither contemporaneously nor with lags on the set of foreign variables (exception for US model) Approach used in earlier studies: Cushman and Zha, 1997: Canada: variables entered in loglevels with introduction of common linear time trend Dungey and Pagan, 2000: Australia: variables de-trended a priori with a linear trend Buckle et al., 2002: New Zealand: Data HP-filtered 5
Adjustment time for impulse response profiles: average response of output across shocks 6
Adjustment time for impulse response profiles: average response of internal demand across shocks 7
Methodological issues Uniform application of HP filter to de-trend all variables: Distortions in the time-series properties led to risk of spurious regression / correlation results Main findings regarding differences in degree of resilience questioned on two grounds Criterion used : number of periods before IRF crosses the zero line following a shock Absence of formal statistical test : could not say whether differences were statistically significant 8
Problems with the HP-filter Biased and inconsistent estimates Uniform smoothness parameter HP filter distorts time-series properties of the variables with risk of spurious correlations 9
Alternative treatment of data Specification in levels, first-differences or VECM form? Even in case of non-stationary variables, estimated coefficients are consistent and the asymptotic distribution of individual parameter is standard, i.e. normal distribution (Sims, Stock and Watson, 1990). Impulse response functions are also consistent estimators of the true IRF except in the long run (where they do not converge with a probability of one (Phillips, 1996)). In presence of I(1) variables, specification in first-diff. or VECM may lead to more efficient estimates in a finite sample if restrictions imposed are the correct ones. If not, then system is mis-specified with high risk of biases. Bottom line: common practice of transforming models into stationary representations by firs-differencing or using co-integration operators is often unnecessary even if data are likely to be integrated. Argument made by Dungey and Pagan (2000) to justify use of linear trend 10
Re-estimation of the SVAR system with linear time trend Guay and Pelgrin (2006) re-estimate SVAR for G7 countries with similar specification (same 8 variables) Quarterly data over period 1975q1-2004q4 Variables de-trended on the basis of a linear time trend: common trend and prior de-trending as in Dungey and Pagan (2000) Block exogeneity restriction is dropped Lag selection on basis of Schwartz criteria: Two lags chosen in almost all countries Bootstrapping method used to correct for small-sample biases (due to presence of persistent series) 11
Results for United States and Germany: 2 polar cases For both countries, IRFs are generally more protracted than those estimated on the basis of HP-filtered data Structural shocks induce more persistent effects on key variables for Germany than for the United States Difference in resilience observed in original study not purely artefact of HP-filtering 12
Main results for GDP: United States 13
Main results: Germany 14
Number of quarters before IRF for output crosses the zero line: four shocks SHOCK TO: World output External Domestic Oil prices demand demand United States 5 11 11 26 Germany 7 40 20 30 15
Are the differences statistically significant? 1st test: Evaluate the difference between the impulse responses of the two countries to a given shock and for a given horizon (k, k+1, k+2, ) 2 nd test: Also based on the difference in the IRFs of the two countries but compares the cumulative impact of a given shock in two countries over a given horizon Both cases: Null hypothesis of no difference between the impulse responses of both countries. No difference in resilience Comparison of IRFs as a test of relative resilience can be rationalised under the assumption that the exogenous dynamic process of the structural shocks disturbing economies is the same for the two countries 16
Difference in GDP IRFs: Germany US impact over different horizons (first test) 17
Difference in GDP IRFs: Germany - US 18
Difference in cumulated impulse responses (4Q) Output and domestic demand 19
Difference in cumulated impulse responses (4Q) Lower confidence level 20
Difference in cumulated impulse responses (12Q) Lower confidence level 21
Conclusions SVAR systems estimated on data using alternative de-trending methodologies point to differences in resilience: particularly the case if compare US and Germany (but also vis-à-vis Japan) Using two statistical tests based on differences in IRFs between countries, difficult to find differences that are statistically significant at conventional confidence levels (Guay and Pelgrin, 2006) Possible explanations: SVAR estimated over (relatively) long period: countries seen as resilient today (US, UK, Canada) were not so resilient in the early 1980s (and vice-versa for Germany) 8 variables SVAR implies a large number of estimated parameters, even with relatively short lags: affects precision. Is VAR modelling approach well-suited for investigating the issue of resilience? 22