Insurance–growth nexus and macroeconomic determinants: evidence from middle-income countries

Abstract

This paper examines the causal relationships between insurance market activities, economic growth, financial depth, and government consumption expenditure. We utilize a panel vector autoregressive model to test Granger causality for 18 middle-income countries over 1980–2012—a group that has not been previously studied in this literature. The results show a robust long-run economic relationship between insurance market activities, economic growth, financial depth, and government consumption expenditure. Moreover, in the short run, we find bidirectional causality between financial depth and economic growth, between financial depth and government consumption expenditure, and between insurance market activities and government consumption expenditure. Unidirectional causality exists from insurance market activities to economic growth, from financial depth to insurance market activities, and from government consumption expenditure to economic growth.

Appendix: Panel data unit root test

The aim of the unit root test is to ascertain the order of integration of the variables, which is where they attain stationarity. This is a primary requirement for conducting cointegration and Granger causality test. This study uses two different panel unit root tests for determining the order of integration where the time series variable attains stationarity. These include: LLC (Levine-Lin-Chu: Levine et al. 2002) and IPS (Im-Pesaran-Shin: Im et al. 2003). Together LLC and IPS have been deployed on the principles of the conventional augmented Dickey-Fuller (ADF) test. While LLC allows for heterogeneity of the intercepts across members of the panel, IPS allows for heterogeneity in intercepts as well as in the slope coefficients. Both LLC and IPS are applied by averaging individual ADF t-statistics across cross-sectional units.

The test follows the estimation of the following equation:

$$\begin{aligned} \Delta Y_t =\mu _i +\gamma _i Y_{it-1} +\sum _{j=1}^{p_i } {\beta _{ij} } \Delta Y_{it-j} +\lambda _i t+\varepsilon _{{ it}} \end{aligned}$$

(8)

where $i = 1, 2{\ldots }.{N}; {t} = 1, 2{\ldots }. {T}; {Y}_{{it}}$ is the series for country i in the panel over period t; ${p}_{{i}}$ is the number of lags selected for the ADF regression; $\Delta$ is the first difference filter (I–L); and $\varepsilon _{{ it}}$ are independently and normally distributed random variables for all i and t with zero means and finite heterogeneous variances $(\sigma _{{i}}^{2})$.

LLC considers the coefficients of the autoregressive term as homogenous across all individuals, i.e., $\gamma _{\mathrm{i}}=\gamma \forall i$. The LLC tests the null hypothesis that each individual in the panel has integrated time series, i.e., $H_{0}$: $\gamma _{{i}}=\gamma = 0 \forall i$ against an alternative $H_{A}: \gamma _{\mathrm{i}}=\gamma < 0 \forall i$. LLC considers pooling the cross-sectional time series data, and the test is based on the following t-statistics:

(9)

where $\gamma$ is restricted by being kept identical across regions under both the null and alternative hypotheses.

It is clear that the null hypothesis of the LLC test is very restrictive and the IPS test relaxes this assumption by allowing $\gamma$ to vary across i under the alternative hypothesis. Hence, the null hypothesis of the IPS test is ${H}_{0}{:} \gamma _{i} = 0 \forall i$, while the alternative hypothesis is that at least one of the individual series in the panel is stationary, i.e., alternative ${H}_{A}: \gamma _{i}< 0 \forall i$. The alternative hypothesis simply implies that $\gamma _{i}$ differs across countries.

Due to heterogeneity, each equation is estimated separately by the ordinary least squares technique and the test statistics are obtained as averages of the test statistics for each equation.

The IPS t-bar statistic is simply defined as the average of the individual Dickey-Fuller $\tau$ statistics. This is as follows:

(10)

Assuming that the cross sections are independent, IPS test uses the mean-group approach and obtains $\tau _{\mathrm{i}}$ and then proposes the use of the standardized t-bar statistic as shown below.

$$\begin{aligned} \bar{{Z}}=\sqrt{N}\left( {\bar{{t}}-E(\bar{{t}})} \right) \big /\sqrt{\mathrm{var}\left( {\bar{{t}}} \right) } \end{aligned}$$

(11)

where $E\left( {\bar{{t}}} \right)$ and $var\left( {\bar{{t}}} \right)$ represents the mean and variance of each $\tau$ statistic, respectively. They are generated by simulations and are tabulated in IPS (1997). The statistic $\bar{{Z}}$ converges to a standard normal distribution as N and T stand to infinitely large and we can compute the significance level in a simple way (see, for instance, Im et al. 2003).