4.3 Empirical findings
The basic model
The conceptual approach of our VAR analyisis is as follows. First, a benchmark
model for the traditional macroeconomic variables Y, P, IS and M is estimated.
Second, when the dynamics of the system are found to be plausible at the global
level, this is considered as a confirmation of our global approach, and the asset price
variables HPI and MSW will be added one by one. The basic specification is given
by the following vector of endogenous variables (with the corresponding Cholesky
ordering):12
Xt = (yp IS m)'t
The Cholesky ordering of the basic specification follows the principle that monetary
variables should be ordered last, since they are supposed to react faster to the real
economy than vice versa (Favero, 2001). Variables are taken in log-levels except for
the short-term interest rate, and a constant and a linear time trend are added to
the model. The usual criteria are applied to determine the lag length.13 Most of the
criteria point at a lag length of 2, which is also sufficient to avoid serial correlation
among the residuals and seems to be appropriate in order to estimate a parsimonious
model where possible.14 While this is true not only for the benchmark specification
but also for the following models we will continue with 2 lags for the whole analysis.
Figure 3 shows the complete impulse responses from the basic specification. Out-
put declines with an interest rate shock and increases with a liquidity shock, which
is in line with our expectations, but both effects are not significant at the 5% level.
Prices move upwards through an innovation to the output variable which might
give support to the consideration of the output gap in assessing inflationary pres-
sures. The particularly interesting reaction of prices to a global liquidity shock is
only slightly significant after a few periods, but the significance (and the level of
the impact) increases over time. We interpret this in favour of the hypothesis that
12 Lower case variables denote logarithms.
13 Explicitly, the Likelihood Ratio test, the Final Prediction Error, the Akaike information crite-
rion, the Schwarz criterion and the Hannan-Quinn criterion are used.
14To test for autocorrelation of the residuals, we performed the Lagrange Multiplier test.
16