Temperature, rainfall and economic growth in Africa

Abstract

Following a recent line of research, this paper investigates the aggregated effects of temperature and rainfall on economic growth in Africa. Our econometric approach is based on a reduced-form model and takes account explicitly of parameter heterogeneity and cross section dependence, relying on ARDL modelling and panel estimators with multifactor structures. We find clear supportive evidence of short- and long-run relations between temperature and per-capita GDP growth, while the role played by rainfall appears to be less important and the evidence on its statistical significance is less clear-cut. Very similar results are reported when the analysis is carried out by focusing solely on Sub-Saharan African countries or considering GDP growth per worker. This evidence is in sharp contrast to the results obtained via standard MG estimation and this confirms that, by not controlling for cross section dependence, traditional panel estimators are likely to provide misleading inference. The empirical results suggest that, far from adapting quickly to weather shocks, African economies appear to be significantly damaged by them. In the absence of corrective measures, the current trends in climate change may impose a progressively heavier burden on African countries.

Appendices

Appendix A

See Tables 5 and 6.

Table 5 $\Delta$lnP-models, CCEP and AMG estimations

Table 6 $\Delta$lnY-models, CCEP and AMG estimations

 

Appendix B

Bond and Eberhardt (2009) and Eberhardt and Teal (2010) show that the inclusion of the ‘common dynamic process’ $\hat{{\mu }}_t^{\bullet }$ in the country regressions allows the AMG procedure to account for cross section dependence. In order to see the intuition behind this, consider the following empirical model

$$\begin{aligned} y_{it} ={\beta }^{\prime }_i x_{it} +u_{it} \end{aligned}$$

(B1)

$$\begin{aligned} u_{it} =\alpha _i +{\lambda }^{\prime }_i f_t +\varepsilon _{it} \end{aligned}$$

(B2)

where $i=1,2,{\ldots },N$ and $t=1,2,{\ldots },T$, while$x_{it}$ is a $k\times 1$ vector of explanatory variables, $\beta _i$ are the $k\times 1$ coefficient vectors,$\alpha _i$ represents the (group-specific) fixed effect and $f_t$ is a set of unobservable common factors with group-specific factor loading $\lambda _i$. This is a simplified version of the model setup in Eberhardt and Teal (2010), which abstracts from the potential presence of multiple factors in the observable variables. Using (B2) into (B1) and first-differencing (to control for nonstationarity issues) we obtain

$$\begin{aligned} \Delta y_{it} ={\beta }^{\prime }_i \Delta x_{it} +{\lambda }^{\prime }_i \Delta f_t +\Delta \varepsilon _{it} \end{aligned}$$

(B3)

The AMG approach captures the common (or mean) evolution of unobservables across groups (e.g. countries) and over time via the inclusion of $T-1$ year dummies $\left( {\sum _{t=2}^T {\Delta D_t } } \right)$ in a pooled estimation of the model in (B3). Intuitively, if$f_t$ is truly common across groups, in each year $t$ the coefficient on the year dummy variable $D_t$ will provide an average estimate of the common factors across groups in that particular year. Thus, the inclusion of $T-1$ year dummies produces an estimate of how the common factors $f_t$ evolve over time—this is the common dynamic process $\hat{{\mu }}_t^{\bullet }$.

In other words, $\hat{{\mu }}_t^{\bullet }$ is some function $h$ of the common factors $f_t$, i.e. $\hat{{\mu }}_t^{\bullet } =h\left( {\bar{{\lambda }}f_t } \right)$. As such, $\hat{{\mu }}_t^{\bullet }$ can be used as a proxy for $f_t$ in the second stage of the AMG procedure as follows

$$\begin{aligned} y_{it} =\alpha _i +{\beta }^{\prime }_i x_{it} +{\lambda }^{\prime }_i h\left( {\bar{{\lambda }}f_t } \right)_t +\varepsilon _{it} \end{aligned}$$

(B4)

Provided that in (B3) the error term $\Delta \varepsilon _{it}$ is white noise, the estimated common dynamic process $\hat{{\mu }}_t^{\bullet }$ can be included in the second-stage AMG regression (B4) to control for the presence of common factors $f_t$ and their heterogeneous effects on $y_{it}$. The Monte Carlo simulations in Bond and Eberhardt (2009) provide substantial evidence on the performance of the AMG estimator, indicating that the inclusion of $\hat{{\mu }}_t^{\bullet }$ in the second-stage regression allows for the separate identification of $\beta _i$ and $f_t$, and that the small-sample performance of the AMG approach broadly matches that of the CCEMG and CCEP estimators.