Financial development, life insurance and growth: Evidence from 17 European countries

Abstract

This study constructs a simple model to demonstrate that life insurance and financial development simultaneously affect economic growth. We provide empirical evidence on the model’s critical prediction. By analysing panel data for 17 advanced European countries from 1980 to 2015, the results show that the effect of private credit on real economic growth is negative in both the long and short run. The negative finance–growth nexus may be due to excessive financing in European countries. The financial crises that occurred during the study period may also have contributed to the negative effects. We find that an increase in the consumption of life insurance is a viable and long-term policy since life insurance penetration promotes long-term economic growth but is not obvious in the short term. Finally, life insurance development is a panacea in the finance–growth nexus since it not only helps moderate long-term real growth volatility but also absorbs the side effect of private credit on real economic growth.

Appendices

Appendix 1: Variable definitions

See Table 6.

Table 6 Variable definitions

Appendix 2: Methodology

Panel unit root tests

This paper examines the stationarity of variables using the panel unit root test of Pesaran (2007) that allows for cross-sectional dependence. Before adopting the panel unit-root test, a cross-sectional dependence (CD) test developed by Pesaran (2004) is used to examine the null hypothesis of cross-sectional independence among individuals in the panel:

$$CD = \left[ {\frac{TN(N - 1)}{2}} \right]^{1/2} \overline{\overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\frown}$}}{\rho } },$$

(12)

where $\overline{\overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\frown}$}}{\rho } } = \left( {\frac{2}{N(N - 1)}} \right)\sum\limits_{i = 1}^{N - 1} {\sum\limits_{j = i + 1}^{N} {\overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\frown}$}}{\rho }_{ij} } }$, in which $\overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\frown}$}}{\rho }_{ij}$ is the pair-wise cross-sectional correlation coefficients of residuals from the conventional Augmented Dicky-Fuller (ADF) regression. T is the sample period and N is the cross-section dimension.

Next, consider the following cross-sectionally augmented Dickey–Fuller (CADF) regression:

$$\Delta y_{it} = \alpha_{i} + \kappa_{i} t + \beta_{i} y_{it - 1} + \gamma \overline{y}_{t - 1} + \theta_{i} \Delta \overline{y}_{t} + \varepsilon_{it} ,\quad t = 1, \ldots ,{\text{T}} \quad {\text{and}}\quad i = 1, \ldots ,N$$

(13)

where $\overline{y}_{t} = {{\sum\nolimits_{i = 1}^{N} {y_{it} } } \mathord{\left/ {\vphantom {{\sum\nolimits_{i = 1}^{N} {y_{it} } } N}} \right. \kern-\nulldelimiterspace} N}$ is the cross-sectional mean of $y_{it}$. The purpose of including the cross-sectional mean in the above equation is to control for contemporaneous correlation among $y_{it}$. The null hypothesis of the test can be expressed as $H_{0} :\beta_{i} = 0$ for all i against the alternative hypothesis $H_{1} :\beta_{i} < 0$ for some i.

The test statistic (CIPS) provided by Pesaran (2007) is given by:

$$CIPS(N,T) = \frac{{\sum\nolimits_{i = 1}^{N} {t_{i} (N,T)} }}{N},$$

(14)

where $t_{i} (N,T)$ is the t statistic of $\beta_{i}$ in Eq. (13). To avoid the problem of extreme statistics caused by small sample observations, Pesaran (2007) constructs a truncated version of the CIPS, denoted as CIPS^*. The critical values for both tests are given in Table II(c) of Pesaran (2007).

Pooled mean group estimation of dynamic heterogeneous panels

This paper utilises PMG estimators, provided by Pesaran et al. (1999), to estimate the long- and short-run elasticity of life insurance and banking on economic development. The major characteristic of PMG estimators is that it allows the intercepts, short-term coefficients and error-correction coefficients to be country specific, but restricts the long-run coefficients to be the same.

The error-correction form of an autoregressive distributed lag model, ARDL(p, q), is written as follows:

$$\Delta Y_{i,t} = \phi_{i} \left( {Y_{i,t - 1} - c - \beta^{\prime}X_{i,t - 1} } \right) + \sum\limits_{k = 1}^{p - 1} {\alpha_{ik} \Delta Y_{i,t - k} + \sum\limits_{j = 0}^{q - 1} {\gamma^{\prime}_{ij} } } \Delta X_{i,t - j} + \varepsilon_{i,t} ,$$

(15)

where $X_{i,t} = \left( {LI_{i,t} ,FD_{i,t} } \right)^{\prime }$ in which LI and FD are indicators of life insurance and financial development; p and q are the lag order for dependent and independent variables, respectively; $\beta$ is a vector of long-run coefficients, measuring the possible impacts of life insurance and banking (X_i,t−1) on the real economic sector (Y_i,t−1); $\alpha_{ik}$ and $\gamma_{ij}$ are short-run coefficients, representing the influences of difference development variables ($\Delta Y_{i,t - k}$ and $\Delta X_{i,t - j}$) on growth rate ($\Delta Y_{i,t}$); and $\phi_{i}$ is the speed of adjustment to the long-run equilibrium.

To construct the estimators, Pesaran et al. (1999) suggest jointly estimating the long-run slope coefficients $\left( \beta \right)$ across agents through a maximum likelihood (MLE) approach. Once the pooled MLE of the long-run parameters is successfully computed, the short-run and error-correction coefficients can be consistently estimated by running the individual MLE. Therefore, the mean of error-correction coefficient $\left( {\phi_{MG} } \right)$ and short-run coefficients ($\alpha_{MG}$ or $\lambda_{MG}$) follow asymptotic normality and can be calculated by the equal weighted average of individual coefficients:

$$\phi_{MG} = N^{ - 1} \sum\limits_{i = 1}^{N} {\phi_{i} } ;$$

(16)

$$z_{MG,j} = N^{ - 1} \sum\limits_{i = 1}^{N} {z_{ij} } .$$

(17)

where $z = \alpha \,$ and $\gamma$.

As shown in Pesaran et al. (1999, pp. 625–626), the PMG estimator can be computed using the same algorithm regardless of whether the regressors are integrated of order 1 or order 0, I(1) or I(0). Furthermore, the PMG estimates will be consistent if the regression residuals are independent. Note that by properly choosing lag order p and q, the time-series independence of the disturbances can be satisfied. Next, the common-time period effect is modelled by expressing the regression variables as deviations from their corresponding cross-sectional means in each period, to ensure that the disturbances are independently distributed across countries (Pesaran et al. 1999).