inconsistency problem, especially if the estimated model still includes specific effects and
lagged endogenous variables12.
On the other hand, this approach neglects from the start the possibility of a levels
relation among the relevant variables. In other words these approach discards the
possibility of a long run effect of monetary policy on deposits and credit. This is at odds
with the usual approach in the literature, which postulates a levels relationship for the
money (or credit) demand, in which the (real) stock of money (or credit) is modelled as a
function of GDP, say, and the levels of some relevant interest rates.
During the last five years or so a very important strand of the literature on models of
panel data has been concerned with the analysis of the consequences of using panels with
a large time-series dimension. The main results available so far concern unit root and
cointegration tests on panel data as well as the asymptotic properties of some well known
estimators when the variables in the model are integrated of order one, I(1) and T (the
time dimension) and N (the cross section dimension) are large. 13
In the empirical section below we basically estimate loan-supply functions, which are
generalisations of (3.4) or (4.1). These equations must be seen as cointegrating relations,
which in the limit can be estimated for individual banks. Being static relations, the
estimated coefficients should be read as the long run effects. By introducing interaction
effects as in (4.1), this approach also allows testing whether the long run effect of
monetary policy differs across banks, according to size, liquidity or capitalisation ratios.
Under the assumption of cointegration Phillips and Moon (1999) have shown that the
Pooled OLS estimator (POLS) is consistent when T and N tend to infinity and has a
normal limit distribution, provide the condition (N / T) → 0 is met. Notice that
asymptotic normality is specific to panel data, as it does not hold in pure time-series data.
The rate of consistency depends on the initial assumptions about the model. The condition
( N / T ) → 0 indicates that this asymptotic theory results are likely to be useful in practice
when N is moderate and T is large. We can expect such data configuration when we have
panels with a large time-series dimension and where the relevant cross-section dimension
is not very large.
12 See Alvarez and Arellano (1998) for a survey on the asymptotic properties of various estimators, in
dynamic panels, with stationary regressors.
13 In what concerns the asymptotic properties of the estimators important papers are Phillips and Moon
(1999), Kao and Chiang (2000), Pedroni (1996), Pesaran, Shin and Smith (1999), Binder, Hsiao and Pesaran
(2000), Pesaran and Shin (1995). Interesting surveys on the subject are Phillips and Moon (2000), Baltagi and
Kao (2000) and Banerjee (1999).
14