3 Investment, credit and output growth - a VAR analysis
fication in levels, as the alternative - an estimation in first differences - seems to have even
more severe shortcomings. The time series in the first differences have a much higher variance
in the beginning of the sample than towards the end of the sample. The intuition of this
phenomenon is that at this very early stage of development, the time series start to grow
from very low levels. Thus, positive as well as the negative growth rates will have a much
larger amplitude than in the later part of the sample, where they have reached a higher level.
Proceeding with the VAR in levels, we need to keep in mind, however, a potential bias in our
results if the time series are not clearly cointegrated. Except for the bivariate combination
of services and bank lending, we can reject the null of no cointegration at least in one of the
three approaches (Engle/Granger, Johansen, Trace/Max-Eigenvalue Statistic).
3 Investment, credit and output growth - a VAR analysis
In the subsequent analysis, we take two different approaches of modeling the link between
financial development and growth. One of the key issues in a VAR framework is the identifi-
cation of structural shocks. In our first approach, we apply the concept of generalized impulse
responses. This approach has the benefit that the impulse response functions are indepen-
dent of the ordering of the variables in the VAR. Its drawback, however, is that the structural
shocks are ultimately not identified. We simulate a system shock, where the contemporaneous
reactions of the other variables are already included.
In the second approach we follow the structural identification proposed in Tornell and
Westermann (2005). In this paper, the identification is based on a theoretical two-sector
growth model that also guides the analysis in the later sections of this paper. We employ a
Cholesky decomposition, where output cannot contemporaneously react to domestic lending
in the same period. The intuition is that output results from investment that is financed by
domestic credit in the period t-1. This also applies to sectoral output. As lending, on the
other hand, can react to changes in output in the same period, we have a recursive system
that can be used to identify shocks from each variable, following the standard Cholesky
procedure. The advantage of this approach is that a structural interpretation can be given
to the impulse response functions in the context of this model. A drawback is that we need
to limit the analysis to a bivariate system. In our view, neither of the two approaches may