Convergence in household credit demand across euro area countries: evidence from panel data

Abstract

This article contributes to the literature on the convergence of financial systems in the euro area by estimating household credit demand in individual countries. Using the ARDL framework advocated notably by Pesaran et al. (1999), the article provides evidence on the convergence of long-run credit demand determinants (interest rates, investment and house prices) in the largest euro area countries, while short run-dynamics remain heterogenous across countries. The article also demonstrates that the equation uncovers demand rather than supply behaviour.

Acknowledgements

The authors thank Claudine Guibert for providing excellent research assistance in compiling the database, as well as seminar participants at the Banque de France, the Bank of England and Maatstricht University, in particular Jean-Pierre Urbain, for useful comments. Any views expressed in this paper are the authors’ own and do not necessarily reflect those of the Banque de France or the Eurosystem

Notes

¹Another issue is whether one should consider real or nominal interest rates, but the debt burden depends on nominal interest rates.

²See Banerjee et al. (1993).

³Here, the equivalence with Equation 1 requires that and ; for j = 1, 2, … , p − 1; for j = 1, 2, … , q − 1.

⁴It is equivalent to multiplying both members of the VAR specification by the matrix:

⁵ Y_t is I(0), like v_t , if θ = 0 that is if π _yx  − ω′Π _xx  = 0 with ω ≠ 0.

⁶The long-term coefficients are estimated by maximization of a concentrated likelihood function, through an iterative procedure (‘Newton–Raphson’ or ‘back-substitution’ algorithm), introducing the θ _i , π_i and σ _i ² coefficients (pooled estimation). Using the estimated θ _i coefficients (as derived from the previous algorithms), the short-run coefficients (including the π_i , σ _i ² and the intercepts) are then estimated separately for each country by OLS. The PMG estimator for the short-run coefficients is the average over all countries.

⁷This formulation is less restrictive than a Pooled model, which specifies constant coefficients.

⁸ , which is equivalent to the ‘Within’ operator W_N ; where I is the identity matrix, and .

⁹ ADF_t statistic used to take into account both endogeneity of regressors and serial correlations of residuals.

¹⁰ T → ∞ followed by N → ∞.

¹¹See tables in Pesaran et al. (2001); for more than one regressor.

¹²In principle, one should also have verified before that ΔY does not cause ΔX in Granger's sense. We assume that it is the case in our empirical study.

¹³None of the empirical results includes a deterministic trend, as it turns out not to be significant over the sample period.

¹⁴The results are based on the Gauss routine from Pesaran et al. (1999)

¹⁵The statistic of the Pearson test of significance of the correlation coefficient r is follows a Student distribution with (n − 2) degrees of freedom, where n is the number of observations.

¹⁶Indeed, the DFE estimation is also supported by the data on the basis of a Hausman test, comparing MG and DFE, but we rely in the rest of the article on the PMG model, which is the most general specification. In addition, the existence of differences in the adjustment coefficient also favours the PMG model.

¹⁷Results for CUSUM tests are available upon request from the authors.

¹⁸The results are based on Kao's modified NPT (CNPT) routine by J. Hlouskova and M. Wagner.

¹⁹For illustration purposes, we present results implementing Banerjee et al.'s (1998) approach, with a simple test of significance (Student t) of the error-correction mechanism π _i from an OLS estimation of the ECM model.

²⁰In addition, the absence of cointegration for panel data has different implications than in the pure time series context, since the estimator remains consistent in the former case, while it creates spurious results in the latter case (Entorf, 1997; Kao, 1999).

²¹ where α _xx and β _xx are two matrices k × r of full column rank r.

²² and are two (k + 1) × (r + 1) dimensional matrices, while α _yx , α _xx , β _yx , β _xx , are, respectively, 1 × (r + 1), k × (r + 1), 1 × (r + 1) and k × (r + 1).