TRENDS IN RAINFALL AND ECONOMIC GROWTH IN
A Neglected Cause of the African Growth Tragedy*
European Commission, Directorate General Joint Research Centre
Institute for Prospective Technological Studies
CREA, Université du Luxembourg
Ecole Polytechnique Paris
We examine the role of trends in rainfall in the poor growth performance of sub-Saharan
African nations relative to other developing countries. To do so we use a new cross-
country panel climatic data set in an empirical economic growth framework. Our results
show that rainfall has been a significant determinant of poor economic growth for
Africa, but not for other developing countries. Depending on the benchmark measure
of potential rainfall, we estimate that the direct impact under the scenario of no decline
in rainfall would have resulted in a reduction of between around 15 and 40 per cent of
today’s gap in African GDP per capita relative to the rest of the developing world.
Keywords: development, Africa, climate
JEL classification: O11, O55, Q25, C23
* We are grateful to comments by participants at seminars given at University of Birmingham, University
of Cagliari, CSAE, University of Navarra, University of Cape Town, University of Sassari, University of
Stellenbosh, University of Starsbourg (ULP), University College Dublin, the UNU-WIDER Jubilee
conference in Helsinki and the Eighth Annual Conference on Econometric Modelling for Africa. The
opinions expressed by the authors do not necessarily reflect those of the institutions they are affiliated
with. All errors remain ours alone.
This paper has been awarded the Albert Berry Prize (2008), delivered by the Canadian Development
Economics Study Group (CDESG), and sponsored by the International Development Research Centre,
Section I – Introduction
The poor performance of sub-Saharan Africa during the second half of the last
century has and continues to receive a considerable amount of attention in the
economics literature, see Collier and Gunning (1999a, 1999b) and Artadi and Sala-i-
Martin (2004) for comprehensive reviews.
In the 1960s there was widespread optimism
about its future – African countries' per-capita GDP was higher than that of many Asian
countries and increasing political self-determination seemed to provide further scope for
governments to cater to domestic needs. Indeed, until the early 1970s there was little
difference between the growth performance of African and other developing countries.
By the second half of the 1970s, however, the outlook changed considerably as the
average pace of growth of African economies began to slow down and by the 1980s even
resulted in economic contraction. While Africa’s growth rates have recently begun to
normalise again, the disastrous performance over more than twenty years has now left
standards of living and income levels lagging well behind other developing countries.
A large number of theories have been put forward to explain this relatively poor
economic performance, but the evidence for their importance, although abundant, is
mixed, see Collier and Gunning (1999a, 1999b). In essence the theories can be
categorised into those arising from political and those due to exogenous factors. Political
explanations usually refer to the poor policies or political institutions that are argued to
have hindered growth in Africa, see Elbadawi and Ndulu (1996), Knack and Keefer
(1995), Mauro (1995). These range from poor fiscal, exchange rate, and trade policies,
and badly functioning financial and labour markets, to the lack of sufficient democracy
and good governance; see Collier and Gunning (1999b). Explanations of an ‘exogenous’
1 As is conventional in essentially all of the literature on this topic, we focus on the relative growth
performance of sub-Saharan Africa since the North African countries (i.e., Algeria, Egypt, Lybia, Morocco,
and Tunisia) are considered to be part of the Middle East and thus of a different regional economy with
other distinctive economic issues. In what follows we will interchangeably refer to Africa for sub-Saharan
African countries (SSA), and to non-sub-Saharan (NSSA) countries for all other developing countries.
nature have, in contrast, appealed to features of African economies outside of the
immediate domestic political domain that may have negatively influenced growth. These
include external aid allocation (Burnside and Dollar (1997)), the lack of diversification of
Africa’s exports (Sachs and Warner (1997)), and ethno-linguistic diversity (Easterly and
Levine (1997)), as well as the landlocked geography and tropical climates prominent of
many African nations (Bloom and Sachs (1998)).
One other aspect of Africa that is increasingly more frequently referred to, but
has as of yet not been evaluated empirically as a potential determinant of Africa’s poor
performance, is the distinct change in rainfall trends that has taken place since the 1960s.
In particular, while there is a general awareness of a number of severe droughts over the
period, it has only relatively recently been noted that rainfall in Africa has also in general
been on a decline since its relative peak in the 1960s; see, for instance, Nicholson (1994,
2001). Given the importance of agriculture for African countries and the dependence of
this sector on rainfall, this decline, as suggested by Nicholson (1994), Collier and
Gunning (1999b), O’Connell and Ndulu (2000), and Bloom and Sachs (1998), may have
had potentially severe consequences for economic growth. Additionally, this decline has
also had a detrimental impact on energy supply since Africa is much more reliant than
other developing countries on hydro-power for electricity generation (Magadza, 1996).
In this paper we explicitly investigate for the first time the role these trends in
rainfall have had on Africa’s relative economic performance.
In particular, we use a
newly available climatic data set to construct a comparable rainfall measure across all
developing countries. Trends in this variable confirm that, in contrast to other
2 There are a number of papers that have already suggested the potential importance of rainfall for
economic growth. For example, O’Connell and Ndulu (2000) include a measure of the number of dry
years in a cross-country growth regression of African countries and find this variable to significantly
negatively affect growth rates. Masters and Sachs (2001), in contrast, use IPCC climatic data to show how
income levels have been affected by rainfall for a sample of developing and developed countries. In
somewhat different approach Guillaumont et al (1999) examine how climate, measured as instability in
agricultural value added, has affected African growth rates.
developing countries, precipitation has been on a general decline in Africa since the
1960s. More importantly, in a cross-country panel growth regression framework results
indicate that rainfall has only had a significant impact on growth in the African sample.
Using these results we show that the direct impact of the decline in rainfall has played an
important role in the poor performance of African countries – ceteris paribus, the gap in
GDP per capita between African and non-African developing countries could have been
between around 15 and 40 per cent lower, depending on what level of rainfall is
considered the benchmark.
The paper proceeds as follows. In the next Section we discuss the importance of
rainfall for Africa’s economic performance and the channels through which rainfall
affects it. Section III discusses our main data sources and provides summary statistics of
our main variables. A discussion about the estimated specification is provided in Section
IV. The results of our econometric analysis are given in Section V. Using these results
hypothetical growth scenarios under more benevolent rainfall conditions are explored in
Section VI. The last section provides concluding remarks.
Section II: Rainfall and Economic Growth in Sub-Saharan Africa
Rainfall could potentially have a wide array of economic implications anywhere in
the developing world. Historically, however, shortages in rainfall in Africa seem to have
been associated with particularly damaging consequences. This particular sensitivity to
rainfall seems at least in part to rest on features specific to Africa. We briefly describe the
two main channels, agriculture and hydro-energy supply, through which rainfall is likely
to have affected sub-Saharan Africa’s (SSA) development below.
3 Unless stated otherwise, information from this section is taken from IPCC (2001).
4 Our choice of these two channels does not preclude the prominence of other channels, some of which
have been documented in Christiansen et al. (2002), Rosenzweig and Binswanger (1993), Masters and
McMillan (2001), Barrios et al. (2005).
A. Agricultural Production
The most direct impact of rainfall on Africa is certainly on the agricultural sector,
since water is an important input into agricultural production.
A large part of this is due
to the significance of this sector for Africa’s economy relative to those of most other
developing nations. Table 1 shows, for example, that agriculture has traditionally had a
higher share in GDP in Africa than in other non-sub-Saharan developing countries
(NSSA) – nearly 40 per cent in 1960. Although this share has since been steadily
decreasing, it still represents almost a third of total GDP in 1997, compared to the
average 14.1 per cent in the rest of the developing world.
However, even apart from the importance of agriculture per se, there are other
aspects of the SSA continent that are likely to make the SSA agricultural sector very
susceptible to shortages in rainfall. In considering these it is important to note that the
availability of water in SSA differs widely as a consequence of the large diversity of
geographic conditions across the continent. Parts of both West and the western part of
Central Africa, i.e., mostly the tropics around the equator, are humid throughout the
year. While there is substantial rainfall during the wet season(s) in the sub-humid regions
located to the north and south of the tropics, there is almost no rain during the much
longer dry season(s). Further pole ward from these sub-humid regions are the large
semi-arid climates. These areas receive some water during the wet season, but suffer
from extreme unreliability of rainfall and few permanent water sources, whereas arid
areas receive little direct water. Semi-arid and arid areas turn out to be most vulnerable to
It is also important to point out that while the African continent has several large
water basins and rivers and there is, as just noted, heavy rainfall in some areas, the run-
off from these water sources to the arid and semi-arid areas is particularly low. This is
exacerbated by the year round high temperatures in SSA. Additionally, within the arid
and semi-arid areas there is little water runoff as drier soil absorbs more moisture. As a
matter of fact, the average runoff of about 15% is lower than in any other continent and
very sensitive to changes in rainfall. Reibsame (1989), for example, estimates that in
Southern Africa a reduction of 10 per cent in precipitation would lead to a fall of more
than 50 per cent in runoff. Moreover, compared to other developing areas in the world,
a much smaller proportion of arable land in SSA is irrigated. For instance, figures in
Table 1 show that still less than 10 per cent of arable land in SSA is irrigated, compared
to nearly a fifth in other developing countries, thus increasing the vulnerability of African
agriculture to precipitation shortages.
As becomes apparent, the areas outside the tropics are extremely reliant on
rainfall for moisture.
The availability of water from rainfall depends in turn on the rate of
evapotranspiration, i.e., the share of water that is evaporated and transpired by plants as a
part of their metabolic processes. The rate of this is particularly high in SSA, in part
because high temperatures increase the water-holding capacity of the air. Moreover,
recent trends in desertification may have affected the extent of rainfall in the semi-arid
areas, as a reduction of vegetative cover can also translate into the absence of inter-
annual soil water storage and hence negatively affect agricultural productivity. It has
been estimated, for example, that desertification has reduced the potential vegetative
productivity by 25 per cent for nearly a quarter of Africa’s land area, see UNEP (1997).
The geographical variation of availability of water just described can be in turn
considered in terms of its implications for agricultural production in SSA. More
precisely, despite the abundance of water, the tropical humid regions are generally not
5 Masters and Wiebe (2000) have already shown evidence of the importance of rainfall for agriculture,
using the same climatic data source as we do in this paper, for developing countries in general.
suitable for crop or animal production. For crops, the combination of high temperatures
and abundant rainfall fosters high rates of chemical weathering and the production of
leached clay soils of low inherent fertility. Hence much of crop production is located in
the semi-arid regions, making it susceptible to rainfall shortages. In terms of animal
production, domestic livestock in Africa other than pigs are also generally concentrated
in the arid and semi-arid regions because the relatively more humid areas provide greater
exposure to animal diseases and are characterised by grasses of low digestibility. Since
livestock are directly dependent on grass quantity, rainfall variations in the semi-arid and
arid areas, have, in turn, also direct consequences on livestock production.
B. Energy Production
Rainfall can also significantly affect the energy sector in SSA, and hence affect
other industries indirectly, since energy supply in many of SSA countries now relies
heavily on water as both a direct and an indirect input; see Magadza (1996).
last 50 years, African countries have invested heavily in hydroelectric power. This is
evidenced by the figures provided in Table 1 which show that hydropower energy now
represents about 47 per cent of total power generation in Africa compared to the
relatively stable average of 34 per cent in other developing countries. Additionally, water
also serves as an important secondary input for thermal power generation as a cooling
device and is needed in large quantities for this purpose.
Importantly, hydroelectric and other energy production that uses water as a
secondary input in SSA tend to be heavily reliant on rivers as their source of water. River
flows in African regions are in turn very sensitive to changes in precipitations. One of the
reason for this is that, apart from the Zambezi and Congo Rivers, major African rivers
like the Nile, Niger, Senegal, Senqu/Orange, and Rufiji are located in arid or semi-arid
6 Apart from animal products, domestic livestock often also serve as source of draft power in SSA.
regions. As a matter of fact, there is evidence that shows that the African major rivers’
performance is significantly lower than that of other areas in the world.
these rivers originate in tropical areas where high temperatures increase evaporation
losses. Moreover, lakes and reservoirs, the other sources of water for hydropower, are
also greatly exposed to decreases in rainfall. For example, declines in precipitation led to
a loss of as much as 30% of total hydropower energy from the Kariba dam, which
supplies power to Zambia and Zimbabwe; see Magadza (1996).
Section III – Data and Summary Statistics
The data used in this paper is derived from a number of sources, and we describe
these and the definitions of all our variables in greater detail in the Data Appendix. Our
main variable of interest, the measure of rainfall, is taken from the Inter-Governmental
Panel on Climate Change (IPCC) data set, which provides, amongst other things, time
series data on the annual rainfall for 289 ‘countries’ (comprised of 188 states and 101
islands and territories) from 1901 to 2000; see Mitchell et al (2002) for a complete
description of the data set. The underlying methodology used to derive these country
level series consists essentially of three steps; see New et al (1999). First a high-
resolution 0.5 degree latitude by 0.5 longitude gridded climatology of the world’s land
surface area is constructed. This grid is then, subsequently, used to derive gridded time
series of rainfall over the desired period. Finally, the individual gridded values are
assigned to individual countries to arrive at country-wide time series. Since the spatial
area of each grid box can vary with latitude, a mean measure of rainfall within each
country for each year was calculated by using the cosine of the grid box’s latitude as
7 See Harrison and Whittington (2001, 2002a and 2002b) for studies of climate variability on hydropower
weight. As a country-specific proxy of rainfall we follow the climatology literature and
use ‘anomalies’, defined as the deviations from the country’s long-term mean, divided by
its long-run standard deviation, where the long-run is taken to be the 1901-2000 period.
Using ‘anomalies’ allows one to eliminate possible scale effects and take account of the
likelihood that for the more arid countries variability is large compared to the mean; see
Nicholson (1986) and Munoz-Diaz and Rodrigo (2004)).
One other aspect with regard to our rainfall measure that deserves discussion
since it has plagued many studies examining other potential determinants of Africa’s
poor growth performance is the question of exogeneity. In terms of rainfall we can
argue fairly confidently that it is a strictly exogenous factor given that it measures an
aspect of climate. While one could in theory also hypothesize that perhaps economic
activity itself can affect aspects such as environmental degradation and desertification,
and thereby possibly rainfall, Nicholson (1994) finds no evidence suggesting such, at least
in the case of Africa. Moreover, as just noted, earlier historical data from other studies
suggests that rainfall naturally moves through long cycles of relative troughs and peaks,
and that a cycle similar to the one over the 20
century seems to have also occurred in
Figures 1 and 2 depict the mean long-term trends in our rainfall anomalies for
SSA and NSSA, shown as five-year moving averages.
As can be seen, the mean rainfall
anomalies in SSA were fairly variable in the first part of the century, peaking in the late
1950s. However, since this peak, rainfall has been what appears to be on a downward
trend, at least until the 1990s. More precisely, mean rainfall anomalies experienced a
8 For example, the total runoff as a percentage of precipitation in African rivers is estimated to be around
20% for Africa while it oscillates around 40% in Asia, North America and Europe see IPCC (2001).
9 We also experimented with using the mean and standard deviation of the data prior to our econometric
analysis, i.e., before 1960. This made no qualitative and little quantitative difference to our results.
10 For all graphical depictions and all other tabulations we included more developing countries than we
used for our econometric specification where the use of control variable restricted our sample. This
allowed the graphs to be more representative of the entire population of developing countries. However,
considerable drop in the 1960s and then an even more severe one in the early 1980s,
reaching an all time low for the century. Figure 2 shows, in contrast, that average annual
rainfall anomalies in NSSA are less variable than in SSA. Moreover, there are no
comparatively large downward trends in the latter half of the 20
Given the continent’s diverse geographic and climatic conditions, it is unlikely
that the mean trends over time just described are completely homogenous across all
countries in SSA. In order to investigate this further we follow the general approach in
the climatology literature and take our individual national annual rainfall anamolies over
the 100 available years and perform a principal component analysis to identify climatic
More specifically, this involves running a principal component analysis of all
series, selecting the number of ‘significant’ components, rotating their principal
and then identifying climatic groups according to the rotated
loadings. Important in this regard is determining the significant components and what
cut-off factor to use to identify groups among the rotated loadings of these components.
With regard to the former we implemented Horn’s test and also verified our results
visually via a scree plot. In terms of the latter we chose a cut-off value of 0.2.
The procedure just described resulted in identifying four climatic groups, as
shown in Map 1, where the countries depicted in white did not fall in any group.
Accordingly, using the national series results in groups that are fairly geographically
distinct. The mean long-term anomalies depiction in Figure 3 shows, nevertheless, that
we did restrict this sample to those for which over the years depicted there was a full set of observations,
so as to avoid trends being pushed by sample entry and exit.
11 See, for instance, Singh and Singh (1996) and Munoz-Diaz and Rodrigo (2004). Importantly one should
note that much of the analysis in the climatology literature has dealt with monthly data and hence captured
seasonal co-variability across geographical units. Since we use as a measure of rainfall a moving average of
annual series in our econometric analysis we restricted our regionalization exercise to annual data.
12 The rotation of loadings facilitates interpretation of the principal components; see Richman (1986). We
use an oblimin (oblique) rotation.
13 The choice of cut-off is inevitably subjective as researchers tend to choose a value that results in
reasonable groupings. For instance, in their sub-national study of 90 rainfall stations in Nepal, Singh and
Singh (1996) experiment with cut-off points ranging between 0.2 and 0.5. One should note that using
all four groups have experienced a declining trend in rainfall at least at some stage since
the early 1960s. Applying the same classification methodology to national annual series
of temperature anomalies, also taken from the IPCC data, similarly isolates four distinct
climatic regions, but also left more countries unclassified, and, at least for one group,
resulted in greater geographical scope; see Map 2. As a comparison group we undertook
parallel analyses for Latin American and Caribbean countries (LAC), the results of which
for rainfall and temperature are shown in Maps 3 and 4. As is apparent, for LAC there
are fewer groups using the same cut-off point, and these cover only a small part of the
continent, both in terms of rainfall and temperature anomalies.
Thus far we have implicitly assumed, and as will be necessary for the econometric
analysis, that climate is homogenous within national borders. To examine this in greater
detail we ran a principal component analysis of the 0.5 by 0.5 degree cells separately for
retained the two components with the highest eigenvalues if these were
deemed to be significant according to Horn’s test, rotated their loadings, and then
identified those cells within each country that had a loading greater than 0.05. The result
of this for SSA are shown in Map 5, where coloured cells identify cells that belong in the
first, second, or both components, and white cells were unclassified.
As can be seen,
most countries have a substantial amount of their area following some common annual
movement in rainfall, except notably Sudan and to a lesser extent the Democratic
Republic of Congo, which, unsurprisingly are two of the larger countries on the
lower cut-off points will tend to result in larger not necessarily mutually exclusive groups. At our chosen
value there were no countries that fell into more than one group.
14 For instance, there were a total of 8129 cells for SSA, with an average of 170 cells per country.
15 There were several island economies where there was only one rainfall series observation at the 0.5 by
0.5 degree level and hence the analysis could not be done. These were Cape Verde, Seychelles, Sao Tome,
Comoros, and Mauritius. In the other countries the number of cells ranged from 3 (Gambia) to 833
continent. In contrast, as shown in Map 6, in LAC particularly in large countries there
appears to be somewhat less homogeneity defined according to the same criteria.
B. Real Income
The main purpose of this paper is to link trends in rainfall to growth of real
income. As a measure of real income in a country we use GDP per capita and for this
we take data directly from the 2001 World Penn Tables for all developing countries, as
defined by World Bank criteria according to their 1960 status.
We graph the mean
series of GDP per capita, taking 1960 as the base year for normalization, for sub-Saharan
African and other non-sub-Saharan developing countries in Figure 4.
The picture that
emerges is one that is well known in the literature – the gap remained roughly constant
during the early 1960s and slightly increased up to the early 1970s. It then rose
significantly in the late 1970s and particularly in the 1980s, but appears to have stabilised
in the latter half of the 1990s.
In order to give some graphical indication of how the observed rainfall trends in
SSA may be related to its poor growth performance, we depicted a five year moving
average of real GDP per capita growth rates and rainfall, appropriately rescaled, from
1960 onwards simultaneously in Figure 5 for SSA countries for which we have a
complete series of growth rates over the period.
This reveals that the two series seem
to move remarkably closely together, except during the drop in rainfall in the early 1970s.
A similar pattern is, in contrast, not apparent for other developing countries, as shown in
16 We also undertook a similar exercise with temperature and found similar patterns within countries and
differences across these.
17 See the Data Appendix for further details on the groups as well as the definitional criteria.
18 The mean real GDP per capita in 1996 $US was 1457 and 2611 for Sub-Saharan African and other
developing countries, respectively.
19 This constitutes 22 out of a possible 46 SSA countries for which we have rainfall series, and hence
explains the slight difference in trends compared to Figure 1.
Section IV – Econometric Specification
The graphical trends just depicted seem to suggest that SSA’s relatively poor
growth performance has gone hand in hand with some of the trends in mean
precipitation. In contrast no such relationship is apparent for other developing countries.
In order to investigate this econometrically we follow the standard empirical cross-
country economic growth literature and assume that economies follow the ‘extended
neo-classical’ growth model with conditional convergence first proposed by Barro
(1991). Accordingly, economies have unique steady-state growth rate values to which
they will tend to converge to over time, the rate of which will depend on the current
distance from the steady state value. The steady state of each economy will itself depend
on cross-country differences in postulated factors:
) + β
where GR is the GDP per capita growth rate for country i over the period t-j to t, Y is
the level of GDP per capita of country i at time t-j, X is a vector of hypothesized
determinants of future steady state-income growth rates that will vary over
countries/regions and possibly over time, γ are time specific effects common to all
countries, μ are country specific effects that are unobservable to the econometrician, ε is
an i.i.d. random term, and the β’s are the coefficients to be estimated. The coefficient on
log(Y) determines the speed at which economies converge towards their steady state and
is expected to be negative.
In terms of showing what since Easterly and Levine’s (1997) seminal paper has
become known as the African growth tragedy generally X includes a zero-one type
dummy that takes on the value of one when a country is located in the SSA region, the
20 Also see Barro and Sala-i-Martin (1995) for further details. One should note that this framework
underlies much of the empirical growth literature on what determines differences in growth rates of
countries. See also, just to name a few, Evans (1993), Islam (1995), Lee et al (1997), Barro (1997), Easterly
and Levine (1997), and Masters and McMillan (2001).
coefficient of which has consistently been found to be negative. For purposes of this
paper, we postulate the following specification for the pooled sample of all developing
) + β
where SSA is the sub-Saharan African dummy variable, and RAIN is a country level
measure of precipitation. Our working hypothesis is that the coefficient on the
interaction term SSA*RAIN is positive and significant (and, possibly, the coefficient on
RAIN insignificant) implying that trends in rainfall have affected growth in SSA more
than that of NSSA. Alternatively, we also estimate (2) separately for the SSA and NSSA
sample, excluding the SSA dummy and its interaction with rain, where we then compare
their coefficients on RAIN.
In estimating (2) and other variants of this specification we generally resorted to
estimating the determinants of average GDP per capita growth over five-year intervals.
This was done for a number of reasons. Firstly, the underlying conditional convergence
framework is more concerned with longer term growth patterns than with annual short-
term fluctuations in GDP per capita. In fact, abstracting from annual movements in
GDP per capita is an approach taken essentially by all of the empirical literature using
Moreover, researchers of the African growth tragedy have similarly
been more interested in long-term divergence from the economic growth patterns of
other developing countries.
Within this five-year average economic growth rate
21 See Dobson et al (2006) for a review.
22 We also tried our base specification with ten year intervals with an obviously much reduced sample size.
Reassuringly our main results were qualitatively the same.
empirical framework our proxy of rainfall, RAIN, is then the average anomalies over any
five-year interval t-5 to t.
In terms of choosing other control variables, X, we took into consideration both
what is commonly used in the literature to look at conditional convergence, and what has
in the past been used to investigate the African growth tragedy. A complete list of all
control variables, their definitions, and sources are given in Appendix. One should note
in this regard that, while there have clearly been a sizeable number of other time varying
and time invariant variables that have been used to explain cross-country differences in
growth rates, inclusion of these, where available, would have put severe restrictions on
the number of countries and extent of time span for each in our sample. Use of the ones
listed in the Appendix provided us for the five-year interval growth rate regressions with
a sample of 60 countries, of which 22 where sub-Saharan African, covering the period
One should note that our base sample consists of an unbalanced panel data
set in the sense that not all time periods are available for all countries, although for most
the number of observations across time is complete.
Section V: Econometric Results
23 We also experimented with using the growth rate of rainfall anomalies over the five-year period but this
proved never to be significant. This may not be surprising since rainfall is generally viewed as a flow
variable of the input in the water stock in the hydrology literature; see Dingman (2001).
24 Our time period was limited to 1990 in the base specification because a number of our main and
auxiliary explanatory variables are limited to this period, namely, urbanization growth, education, civil wars,
and hydro-power growth.
25 The mean number of observations for each country (from a possible 6) is 5.87.
A. Main Results
Using standard OLS, we first estimate (2) without any interaction term between
RAIN and SSA or other control variables X, as shown in the first column of Table 2.
Accordingly, the SSA dummy is negative and significant, indicating that SSA countries
had on average lower growth rates, thus supporting the idea of an African growth
tragedy, whereas the coefficient on RAIN is insignificant. In order to determine whether
the lack of significance of the latter may be due to different effects across SSA and NSSA
countries, we included an interaction term of the SSA dummy and rainfall in the second
column. As can be seen, while the coefficient on RAIN remains insignificant, one finds
a positive effect of the interaction term. Put differently, lower rainfall will negatively
affect growth only in SSA countries. As shown in the third and fourth columns, this
result, i.e., a significant positive relationship only in SSA countries but no effect in their
NSSA counterparts, is robust to regressing growth on rainfall for the two samples
We next included our full set of control variables, including time dummies.
Given that our focus here is not on disentangling the effects of the previously mentioned
other theories that have been put forward in the literature trying to explain SSA’s poor
performance, but rather on isolating the impact of rainfall, the full set of results on all
control variables are not discussed, but reported in the Appendix. The estimated
coefficients on our main variable of interest, rainfall, for the full sample and the sub-
samples are provided in the fifth through seventh columns of Table 2. In line with our
simple specification, they similarly indicate that rainfall has only had a significant impact
in SSA countries.
In Table 3, columns 1 to 6, we re-ran the specifications of columns 2 through 7
of Table 2 using a fixed effects estimator, which allows us to purge not only the effect of
our time invariant controls, but all other non-included time invariant factors from the
model. Accordingly, taking account of fixed effects in the specification without (time
varying) controls changes little relative to the OLS results - rainfall influences economic
growth only in SSA nations and gives similar coefficients. The results are also robust to
including our set of time varying explanatory variables, although the coefficient for the
separate SSA sample regression is somewhat higher in the fixed effects specification.
B. Temperature and Main Channel Effects
As indicated in Section II, temperature may also play an important role in SSA.
Feasibly the effect of rainfall on growth in SSA found above may simply be capturing the
effect of temperature trends. For example, previous studies have argued and found
evidence for some industrialised countries that temperature can have a negative impact
on agriculture; see, for instance, Mendelsohn et al (1994). We hence constructed an
‘anomalies’ measure of temperature similar to our rainfall proxy for SSA and NSSA,
which are depicted in Figure 7.
As can be seen, the trend in average temperature
followed a similar pattern in both country groups, first rising until the 1940s, then
embarking on a long decline until the late 1970s, from which onwards they have been on
a steep ascend.
To investigate whether the temperature trends may have affected growth rates
and/or the effect of rainfall on growth in SSA may simply be capturing the effect of
temperature changes, we included temperature in our fixed effects specification for our
two sub-samples in columns 1 and 2 of Table 4. Accordingly, in neither case is
temperature a significant determinant of growth, nor does its conclusion change the
coefficient on rainfall. We subsequently then also experimented interacting temperature
with rainfall, given that the rate of evapotranspiration of water depends on both aspects
26 Given that countries appear many times in the data, we tested for serial correlation within panels with
the test suggested by Wooldridge (2002), but found no evidence of this.
27 The data on temperature was also taken from the IPCC database and constructed in a similar manner.
of climate. As can be seen, the interaction terms also are insignificant, with no apparent
change in the coefficient on rainfall.
Our discussion in Section II also suggested that the two key impact sectors
through which rainfall in SSA affected the economy are agriculture and hydropower.
One would thus expect that countries for which these sectors are more important parts
of the economy to be more vulnerable to trends in rainfall. In order to proxy the
importance of these channels nationally we calculated the average agriculture production
share of GDP, AGR, and the average hydropower production share of GDP,
over the sample period for each country. One should note that we used the average
measure of these over the period rather than their time varying values for two reasons.
Firstly, the data on both proxies is comparatively poor, with many missing values for
most countries in our data set, and thus including these with our control variables would
have further reduced the sample size considerably and made it difficult to compare any
results with our base specification.
Secondly, the use of time invariant averages at least
partially allows us to circumvent likely endogeneity and misspecification issues since both
sectors are components of GDP and are also affected by rainfall itself.
We hence proceeded to interact AGR and
with the rainfall anomalies in
our fixed effects specification. These interaction terms can thus be interpreted to capture
the potentially different effects of rainfall on GDP growth for countries that had on
average greater agricultural and hydropower sectors. Additionally, we also controlled for
different convergence rates in this regard by interacting AGR and
GDP per capita. Our results of this exercise are reported in the final two columns of
Table 4. As can be seen, controlling for either the size of the agricultural sector or for
the importance of hydropower does not change our finding of a non-significant impact
of rainfall on growth in NSSA. In contrast, these interaction terms are positive and
statistically significant for the SSA sample, while rendering rainfall anomalies coefficients
insignificant. This indicates that countries with a larger agricultural sector and/or greater
hydropower production are indeed more vulnerable to trends in rainfall in terms of
economic growth. Moreover, these two channels appear to explain most of the impact
of rainfall trends on growth.
C. Extended SSA Sample and Further Robustness Checks
The inclusion of other control variables has meant that for our SSA sample we
were restricted to less than half (22) of all SSA nations and to a sample period that ended
in 1990. In order to ensure that these data were a representative sample of SSA we thus
also re-ran our base specification with fixed effects without any controls for SSA,
allowing us to include 42 countries over the 1960 to 2000 period, as shown in the first
column of Table 5. Accordingly, despite being able to cover almost all SSA nations and
more than doubling our sample size, the essential result of a positive effect of rainfall on
Moreover, the size of the coefficient is similar.
While rainfall measured in terms of anomalies is the most widely used proxy for
rainfall in the climatology literature we also experimented with other indicators of
precipitation. For instance, in the past the FAO has used the level of rainfall divided by
its long-term mean; see Gommes and Petrassi (1996). As the estimates in the second
column demonstrate, the significant positive effect is robust to employing this alternative
proxy. Arguably one of the advantages of the anomalies relative to the FAO measure of
rainfall, is that the former can better take account of the fact that in countries where
mean rainfall is low the coefficient of variation tends to be high, which would amplify the
effect of droughts. One good example is the Sahel region which tends to have very low
levels of rainfall by SSA standards, but also experiences severe droughts over our sample
period. To investigate this we created zero-one type dummies for the Sahel, SAHEL,
28 For example, the number of SSA countries would have been reduce by nearly a half.
and non-Sahel regions, NON_SAHEL, and interacted these with the two rainfall
measures, the result of including these are given in the third and fourth column of Table
5. Accordingly, while both interaction terms are significant for the anomalies proxy, the
non-Sahel interaction term is now marginally insignificant for the FAO proxy
specification, thus indeed confirming the suspicion that the latter may in some cases be
As noted in Section III, the country-wide measure of rainfall is an average of the
individual cell values, weighted by the latitude in order to control for the area of the cell.
It may, however, be the case that the falling trends in rainfall are mostly occurring in
sparsely populated areas which generate little economic activity, hence attributing too
much weight to these in terms of the impact of rainfall on economic growth. For
example, Masters and McMillan (2001), using the same climatic data set as here, show
that local population density is related to precipitation. To see whether our results are
robust to taking account of this we resorted to information from the African population
database which provides population estimates in a raster format for the years 1960, 1970,
1980, 1990, and 2000. Since these raster grids did not correspond to the format for
which the precipitation data was available, i.e. 0.5 by 0.5 degrees, we had to impose this
latter format on the population data. In order to derive annual population shares from
the decadal data we linearly interpolated the shares between decades and used the 1960
weights for all years prior to 1960. Rather then weighting these by the latitude of the
area as before, we subsequently instead used the population shares of the grids as
weights. The population weighted rainfall anomalies were then averaged at the country
level to obtain annual country level series. Reassuringly, using this measure, RAIN_POP,
produces the same positive and significant effect on growth as our benchmark measure,
as is shown in the fifth column of Table 5.
29 Extending the NSSA sample in a similar manner continued to produce an insignificant effect for rainfall.
D. Oceanic Factors
There is now an extensive literature demonstrating the link between oceanic
factors and precipitation patterns over large distances, commonly known as
teleconnections. For instance, the El Nino-Southern Oscilliation (ENSO) has been
shown to have a diapole association with rainfall anomalies in Africa, where eastern
African rainfall is in phase whereas southern African precipitation moves negatively with
warm ENSO events, see Nicholson and Kim (1997). In contrast, for northern Africa the
North Atlantic Oscillation appears to be the main factor behind climatic interannual
variability, while for the western part of the continent the Atlantic Ocean as well as the
rest of the world oceans appear to play a major role, see IPCC (2001). It thus may be of
interest to determine whether SSA growth itself can ultimately be statistically linked to
movements in such oceanic factors.
The primary approach to modelling the link between oceanic variables and
rainfall has been to employ canonical correlation analysis (CCA) on precipitation and sea
surface temperature (SST) data sets; see, for example, Barnston and Smith (1996) and
Giannini et al (2000).
We similarly follow this approach here to determine whether our
annual SSA rainfall anomalies can be related to movements in SSTs. In this regard we
first use data on measures of SST anomalies of 2 by 2 degree cells covering all the
world’s oceans from the Comprehensive Ocean Atmosphere Data Set (ERSST v.2),
aggregate these to 6 by 6 degree cell measures,
and then reduce the series to a much
smaller set of summary fields using principal component analysis. More specifically, a
Horn’s test indicated the existence of seven components from the 290 SST 6 by 6 degree
cells, and we used the complete set of loadings of these components to generate 7
summary series. For the country level rainfall anomalies data we used the complete set
30 For a given two sets of variables CCA involves finding the linear combinations of these so that the
correlation of these combinations is as high as possible.
of loadings on the previously isolated four components to similarly create summary
series of the data. The reduced set of fields of the predictor (SST) and predictand
(rainfall) were then subjected to a CCA, which produced various sets of linear
combinations of the two input sets that maximized the correlation between these. The
derived linear combinations (modes) of the predictor can then be used to examine how
SST anomalies are linked to ‘local’ rainfall anomalies by calculating the temporal
correlation (also known as the skill) of the CCA predictions of original data with the
observed country level rainfall series in the manner proposed by Barnston (1994).
Map 7 provides a graphical depiction of the skills derived from the first (i.e., the
one with the greatest explanatory power) set of linear combinations. As can be seen,
there is much variation in the manner and the degree to which this first mode can explain
rainfall anomalies in SSA. More precisely, some parts of Africa are positively correlated,
while others are characterized by negative comovements with SST, whereas many other
nations appear to be nearly unrelated to SST movements. Also noteworthy is that, as has
been shown in some of the studies previously cited, correlation patterns are not
necessarily ‘local’ phenomena, but instead that very distant parts of the continent can
move similarly in response to SST changes. The depiction of the skills derived from the
second mode, shown in Map 8, demonstrates how inherently complex the relationship
between SST and rainfall anomalies are in SSA, however, where a largely different pattern
To examine whether the SST anomalies can also be econometrically linked to
growth we used the first two modes of predictors instead of rainfall as explanatory
variables in our base econometric specification of the SSA sample, as shown in the last
31 Barnston and Smith (1996) find in an analysis linking monthly SST anamolies to climatic variables that
results from this more aggregated data produced virtually identically results to the 2 by 2 degree cells.
32 The first set of linear combination of predictand and predictor variables was found to be correlated at
0.95 while the second set had correlation of 0.89. Further linear combinations fell drastically in their
covariability and hence we do not report on these.
column of Table 5. Accordingly, one finds that both components are significantly related
to growth, although with opposing signs. This provides some evidence that growth
patterns in SSA are at least partially linked to movements in SST.
Section VI: Simulations
Our results clearly indicate that trends in rainfall have had a significant impact
only in SSA countries. Given the trends in the growth rates and rainfall outlined in
Section III, this finding suggests that perhaps rainfall may have played a considerable role
in explaining the diverging performance in economic growth of SSA countries relative to
the rest of the developing world as shown in Figure 1. A simple manner of investigating
this is to calculate the trend that GDP per capita in SSA countries would have followed if
rainfall had remained at some previous level using our estimated coefficients.
In considering how rainfall would affect GDP per capita within our conditional
convergence framework, one must realise that it will do so directly through the growth
rate and by influencing the following period’s initial level of GDP per capita and thus the
convergence to the steady state. Consequently, given a benchmark level of rainfall,
, one can construct the hypothetical GDP per capita series at any time T for a
country i by:
where the superscript H indicates the simulated hypothetical series. In essence (3) entails
simulating a ‘hypothetical’ log of GDP per capita series for SSA that has the same initial
value in 1960s as the true series, but differs in terms of the forcing process for RAIN.
33 One should note that we did attempt to control for the importance of the two key sectors, agricultural
and hydropower production, in our simulations since our econometric results in this regard used time
invariant measures of these.
We first calculate such a hypothetical GDP per capita series for SSA holding
rainfall at its maximum mean annual anomaly over the entire 20th century,
coefficient on rainfall from the sixth column and the coefficient on initial GDP per
capita from the fourth column of Table 3.
The resultant hypothetical GDP per capita
series, along with the actual SSA and NSSA series, is depicted in Figure 8. Accordingly,
if rainfall had remained at the high level of the late 1950s, the difference in the mean
growth rates between SSA and NSSA nations, which can be gauged from the relative
slopes of the series, would have been roughly similar until the late 1970s, from which
point onwards SSA countries would have even experienced a temporary slight superiority
in economic growth. Using the underlying figures one finds that if rainfall had remained
at its 1955-1960 level, the gap in GDP per capita between SSA and NSSA would have
been about 40.0 per cent less than what was observed in actuality at the end of our
sample period. Thus the gap would have been reduced by 1418 dollars per capita.
Given the high variability of African rainfall over time, perhaps a more realistic
scenario to examine is the one under which rainfall would have remained at its previous
long-term mean prior to the 1960s (1901-1959). This is shown, also in Figure 7, relative
to the true trends in SSA and NSSA countries. Accordingly, the divergence in growth
rates between SSA and NSSA under this scenario would have actually been slightly
greater in the earlier period due to the fact that the peak in the late 1950s was above the
previous long-term mean. GDP per capita in SSA nations would thus have followed a
roughly similar path to that observed in reality during the late 1970s and early 1980s.
After 1985, however, GDP per capita growth rates in SSA nations would have risen to a
level parallel to their NSSA counterparts. Overall, under this more moderate benchmark
34 This occurred in 1955.
35 We chose the former so as to allow for an estimate from a less restricted error generating process and
the latter to measure convergence relative to all developing countries.
level of rainfall, the gap in GDP per capita between SSA and NSSA would have been
about 15.6 per cent less, reducing it by 550 dollars, than what was observed in actuality.
Finally, it is important to emphasize that our simple simulations should only
serve as a fairly rough demonstration of the potential economic significance that rainfall
trends have played in Africa for several reasons. Firstly, our choice of hypothetical
rainfall series is not based on any criteria of what may be considered ‘normal’
occurrences of precipitation, given that we only observe its values over the period in
which our empirical analysis takes place. Secondly, we are assuming that rainfall only has
a direct effect on economic growth and not through other control variables. Finally, our
estimated impact compared to some hypothetical situation rests on accurately having
measured the rate of convergence. In this regard, the inclusion of control variables other
than the ones we have used here may result in differences in the convergence parameter.
Moreover, one can easily rewrite equation (1) as a standard dynamic panel specification,
so that, as shown by Nickell (1981), our estimate of the convergence rate may be biased.
As a matter of fact Bond et al (2001) argue that the ‘true’ estimate is likely to lie
somewhere between the fixed effects and OLS estimates of it. In our case this would
mean that the absolute value of the coefficient on initial GDP per capita that we use for
our simulations (i.e., the one from the fixed effects specification) is above the ‘true’ one
and hence that we are underestimating the impact of a more favourable precipitation
situation in reducing the GDP per capita gap between SSA and NSSA. To examine this
we re-estimated the pooled sample fixed effects specification with the Kiviet (1995)
correction. As can be seen from the last column in Table 3, however, the estimated
coefficient is not too different from the estimate that we use, hence suggesting that the
bias is in our case likely to be minimal.
Section VII: Concluding Remarks
Using a new cross-country panel climatic data set we provide evidence that
trends in rainfall have affected economic growth rates in sub-Saharan Africa, but that no
such relationship is apparent for other developing countries. This means that the general
decline in rainfall that has been observed in Africa has had adverse effects on its growth
rates, and is likely to explain part of the puzzle of Africa’s relatively poor performance.
As a matter of fact, some simple simulations suggest that if rainfall had remained at
previous levels, the current gap in GDP per capita relative to other developing countries
could have been between 15 and 40 per cent lower.
Our results arguably have important policy implications. In particular, economists
and policy analysts studying African economies should pay closer attention to rainfall as
an explanatory and conditioning factor of economic growth, not only for annual changes
but also in terms of long-term trends. Including an indicator of rainfall anomalies,
directly and potentially as an instrumental variable for other aspects (such as policy
changes), may very well reverse some widely held views about the causes of low growth
in the past and the effects of policy changes in the future.
Artadi, E.V. and X. Sala-i-Martin (2004), The Economic Tragedy of the XXth Century:
Growth in Africa, NBER working Paper 9865.
Barnston, A. G. (1994). “Linear Statistical Short-Term Climate Predictive Skill in the
Northern Hemisphere”, Journal of Climate, 7, pp. 1513-1564.
Barnston, A.G. and Smith, T.M. (1996). “Specification and Prediction of Global Surface
Temperature and Precipitation from Global SST using CCA”, Journal of Climate, 9, pp.
Barrios, S., L.Bertinelli and E. Strobl (2006), Climatic Change and Rural-Urban
Migration: The Case of Sub-Saharan Africa, Journal of Urban Economics, 60, pp. 360-371.
Barro, Robert J, (1991). "Economic Growth in a Cross Section of Countries," The
Quarterly Journal of Economics, 106, pp. 407-43.
Barro,R. and J. Lee (1993), International Comparisons of Educational Attainment, Journal
of Monetary Economics, 32(3): 363-94.
Barro, R. J. and Sala-I-Martin, X. (1995). Economic Growth, McGraw Hill.
Barro, R.J. (1997). The Determinants of Economic Growth, The MIT Press: Cambridge,
Bloom, D.E. and Sachs, J.D. (1998). Geography, Demography, and Economic Growth
in Africa. Brookings Papers on Economic Activity 2: 207-273.
Bond, S., Hoeffle, A., and Temple, J. (2001). “GMM Estimation of Empirical Growth
Models”, CEPR Discussion Paper DP3048.
Burnside, C. and D. Dollar (1997), Aid, Policies and Growth, World Bank Policy
Research, Working Papers on International Economics, Trade and Capital Flows,
Christiansen, L., Demery, L. and Paternostro, S. (2002). Growth, Distribution, and
Poverty in Africa, World Bank: Washington, DC.
Collier, P. and J.W. Gunning (1999a). Why Has Africa Grown Slowly?, Journal of Economic
Perspectives, 13: 3-22.
Collier, P. and J.W. Gunning (1999b). Explaining African Economic Performance, Journal
of Economic Literature, 37: 64-111.
Davis, J. and V. Henderson (2003), Evidence on the political economy of the
urbanization process, Journal of Urban Economics, 53(1) pp.98-125.
Dingman, L. (2001). Phsical Hydrology. Prentice Hall, London
Dobson, S., Ramalogan, C. and Strobl, E. (2006) “Why do Rates of Convergence Differ?
A Meta Regression Analysis”, Scottish Journal of Political Economy, 53, pp. 153-173.
Elbadawi, I. and B. Ndulu (1996), Long Run Development and Sustainable Growth in
Sub-Saharan Africa, in New Directions in Development Economics, M. Lundahl and B.
Ndulu (Eds.). London: Routledge.
Easterly, W. and R. Levine (1997), Africa's Growth Tragedy: Policies and Ethnic
Divisions, Quarterly Journal of Economics, 112(4): 1203-50.
Giannini, A., Kushnir, Y., and Cane, M.A. (2000). “Interannual Variability of Caribbean
Rainfall, ENSO, and the Atlantic Ocean”, Journal of Climate, 13, pp. 297-311.
Gommes, R. and F. Petrassi, (1996), Rainfall variability and drought in sub-Saharan
Africa since 1960. FAO Agrometeorology Series 9, Food and Agriculture Organization
of the United Nations.
Guillaumont, P., Guillaumont Jeanneney, S., and Brun, J.F. (1999). “How Instability
Lowers African Growth”, Journal of African Economies, 8, pp. 87-107.
Harrison, G.P. and H.W. Whittington, (2001). “Impact of climatic change on
hydropower investment”, Proceedings of the 4
International Conference on Hydropower
Development (Hydropower '01), 19-22 June 2001, Bergen, Norway, pp. 257-261
Harrison, G.P. and H.W. Whittington, (2002a), Susceptibility of the Bakota Gorge
Hydroelectric Scheme to Climate change, Journal of Hydrology 264: 230-241.
Harrison, G.P. and H. W. Whittington, (2002b) 'Vulnerability of hydropower projects to
climate change', IEE Proc. Generation, Transmission & Distribution, 149 (3), pp. 249-255.
Intergovernmental Panel on Climate Change, (2001). Climate Change 2001: Impacts,
Adaptation and Vulnerability - Contribution of Working Group II to the Third
Assessment Report of IPCC
Islam, N. (1995). “Growth Empirics: A Panel Data Approach”, Quarterly Journal of
Economics, 110, pp. 1127-1170.
Kiviet, J.F., 1995. “On Bias, Inconsistency and Efficiency of Various Estimators in
Dynamic Panel Data Models”, Journal of Econometrics, 68, pp. 53-78.
Knack, S. and P. Keefer (1995). Institutions and Economic Performance: Cross-Country
Tests using Alternative Institutional Measures, Economics and Politics, 7:207-227.
Lee, K., M. Pesaran and R. Smith (1997). “Growth and Convergence in a Multi-Country
Empirical Stochastic Solow Model”, Journal of Applied Econometrics, 12, pp. 357-392.
Magadza, C.H.D., 1996: Climate change: some likely multiple impacts in southern Africa.
In: Downing, T.E. (ed.), Climate Change and World Food Security Springer-Verlag,
Dordrecht, The Netherlands: 449-483.
Masters, W.A. and M.S. McMillan (2001). “Climate and Scale in Economic Growth”,
Journal of Economic Growth, 6: 167-187.
Masters, W.A. and Wiebe, K.D. (2000). “Climate and Agricultural Productivity”, paper
presented at the Conference on Biophysical Constraints in Tropical Agriculture, Harvard
University, 17 October 2000.
Masters, W.A. and Sachs, J. (2001). “Climate and Development”, paper presented at the
annual meetings of the AEA, New Orleans, LA, January 7, 2001.
Mauro, P. (1995). Corruption and Gowth, Quarterly Journal of Economics, 110: 681-712.
Mendelsohn, R., Nordhaus, W., and Shaw, D. (1994). “The Impact of Global Warming
on Agriculture: A Ricardian Analysis”, American Economic Review, 84, pp. 753-771.
Mitchell, T. D., M. Hulme and M. New (2002). "Climate data for political areas." Area
Munoz-Diaz, D. and Rodrigo, F.S. (2004). “Spatio-Temporal Patterns of Seasonal
Rainfall in Spain (1912-2000) Using Cluster and Principal Component Analysis:
Comparison”, Annales Geophysicae, 22, pp. 1435-1448.
Murdoch, J. and Sandler, J. (2002). Economic Growth, Civil Wars, and Spatial Spillovers,
Journal of Conflict Resolution, Vol. 46(1): 91-110
New, M., M. Hulme and P. Jones (1999), Representing twentieth-century space-time
climate variability. Part I: development of 1961-1990 mean monthly terrestrial
climatology, Journal of Climate, 12: 829-56
New, M., M. Hulme and P. Jones (2000), Representing twentieth-century space-time
climate variability. Part II: development of 1901-96 monthly grids of terrestrial surface
climate, Journal of Climate, 13: 2217-38.
Nicholson, S.E. (1986). “The Spatial Coherence of African Rainfall Anamolies:
Interhemispheric Teleconnections”, Journal of Climate and Applied Meteorology, 25, pp. 1365-
Nicholson, S.E. (1994). Recent rainfall fluctuations in Africa and their relationship to
past conditions over the continent, The Holocene, 4: 121-131.
Nicholson, S.E. and Kim(1997). “The Relationship of the El Nino-Southern Oscillation
to African Rainfall”, International Journal of Climatology, 17, pp. 117-135.
Nicholson, S.E. (2001). Climatic and Environmental Change in Africa During the Last
Two Centuries, Climate Research, 17: 123-144.
Nickell, S. (1981). “Biases in Dynamic Models with Fixed Effects”, Econometrica, 49, pp.
O’Connell, S.A. and B.J. Ndulu (2000). Africa’s Growth Experience, mimeo.
Reibsame, E. (1989). “Assessing the Implications of Climate Fluctuations: A Guide to
Climate Impact Studies”, UNEP, Nairobi.
Richman, M. B. (1986). “Rotation of Principal Components”, Journal of Climatology, 6, 293.
Rosenzweig, M. and Binswanger, H. (1993). “Wealth, Weather Risk and the Composition
and Profitibality of Agricultural Investment”, Economic Journal, 103, pp. 56-78.
Sachs, J. and A. Warner (1997), Sources of Slow Growth in African Economies, Journal of
African Economies, 6(3): 335-76.
Singh, K.K. and Singh, S.V. (1996). “Space-Time Variation and Regionalization of
Seasonal and Monthly Summer Monsoon Rainfall on sub-Himalayan Region and
Gangetic Plains of India”, Climate Research, 6, pp. 251-262.
UNEP, (1997). World Atlas of Desertification, United Nations Environment Program,
Edward Arnold, London.
Wooldridge, J. M. (2002). Econometric Analysis of Cross Section and Panel Data.,
Cambridge, Massachusetts: The MIT Press.
Figure 1: Rainfall in Sub-Saharan African Countries – Long Term Trends
1900 1920 1940 1960 1980 2000
Figure 2: Rainfall in Non Sub-Saharan African Countries – Long Term Trends
-.4 -.3 -.2 -.1 0 .1 .2 .3 .4
1900 1920 1940 1960 1980 2000
Figure 3: Long-Term Trends in Rainfall for SSA Climatic Groups
-1 -.5 0.5 1
1900 1920 1940 1960 1980 2000
Group I Group II
Group III Group IV
Figure 4: GDP per Capita Trends
Normalised GDP per Capita
1960 1970 1980 1990 2000
Figure 5: Trends in real GDP per capita growth rates and Rainfall in Sub-Saharan
-.01 0.01 .02 .03
GDP per C apita Growth Rate
-2 -1 0 1
1960 1970 19 80 1990 2000
Rainfall Anamolies GDP per Capita Growth Rate
Figure 6: Trends in Real GDP per Capita Growth Rates and Rainfall in other
0.01 .02 .03
GDP per Capita Growth Rate
-.2 -.1 0.1 .2 .3
1960 1970 19 80 1990 2000
Rainfall Anamolies GDP per Capita Growth Rate
Figure 7: Long-Term Trends in Temperature – SSA & NSSA Countries
-1 -.5 0.5 11.5
1900 1920 1940 1960 1980 2000
Figure 8: GDP per Capita in Sub-Saharan African Countries – Actual vs.
11.5 22.5 3
1960 1970 19 80 1990 2000
NSSA Actual Series SSA Actual Series
SSA Simulated Series A SSA Simulated Series B
Map 1: Rainfall Climatic Groups – Sub-Saharan Africa
Map 2: Temperature Climatic Groups – Sub-Saharan Africa
Map 3: Rainfall Climatic Groups – Latin American and Caribbean
Map 4: Temperature Climatic Groups – Latin American and Caribbean
Map 5: Rainfall Within Country Homogeneity – Sub-Saharan Africa Map 6: Rainfall Within Country Homogeneity – Latin American and Caribbean
Map 7: Correlation between Sea Surface Temperature and Rainfall:
First Linear Combination – Sub-Saharan Africa
Map 8: Correlation between Sea Surface Temperature and Rainfall:
Second Linear Combination – Sub-Saharan Africa
Table 1: Mean Characteristics for SSA and NSSA
1960 1970 1980 1990 1997
% of Agriculture in GDP:
NSSA 24.4 23.0 18.7 16.3 14.1
39.2 33.9 32.0 29.9 29.7
% of Arable Land Irrigated:
NSSA 14.2 16.3 16.1 17.1 17.2
6.4 7.2 7.7 8.3 8.4
% of Power Generation by Hydro-power:
NSSA 35.0 39.4 37.6 39.6 34.1
27.9 37.3 46.5 42.9 46.6
Notes: (1) Where exact year was not available information from the nearest year was used. (2) The sample
sample of countries may not correspond across the three variables as we only included countries in our
sample for which we had observations for all five periods. Sources: World Development Indicators (World
Bank), FAO and authors’ computations.
Table 2: OLS Results
(1) (2) (3) (4) (5) (6) (7)
RAIN 0.003 -0.004 -0.004 0.011** -0.007 -0.007 0.016**
(0.003) (0.005) (0.004) (0.006) (0.005) (0.005) (0.008)
SSA -0.013*** -0.012*** -0.006
(0.004) (0.004) (0.008)
RAIN*SSA 0.014** 0.018***
log(GDP/CA) -0.006** -0.006** -0.009*** 0.001 -0.014*** -0.016** -0.037***
(0.003) (0.003) (0.003) (0.005) (0.004) (0.007) (0.009)
Constant 0.068*** 0.065*** 0.094*** 0.007 0.137*** 0.155*** -0.215
(0.022) (0.022) (0.026) (0.037) (0.034) (0.051) (0.177)
Sample All All NSSA SSA All NSSA SSA
Controls No No No No Yes Yes Yes
Observations 393 393 254 139 393 254 139
Countries 60 60 38 22 60 38 22
F-Test 3.58*** 3.82*** 4.57*** 2.02*** 4.63*** 5.10*** 3.05***
R-squared 0.03 0.04 0.04 0.03 0.21 0.30 0.39
Notes: (1) Robust standard errors in parentheses. (2) ***, **, and * indicate 1, 5, and 10 per cent
significance levels. (3) Controls include time dummies, openness (OPEN), population size (POP),
schooling (ED), civil war incidence (CIVWAR), civil war incidence in surrounding countries (CIVWAR_S),
investment (INV/GDP), government expenditure (G/GDP), urbanization (URB, landlockedness
(LANDLOCK), ethnic diversity (ETHNIC), tropical area dummy (TROP).
Table 3: Fixed Effects Results
(1) (2) (3) (4) (5) (6) (7)
RAIN -0.004 -0.004 0.011* -0.003 -0.003 0.022*** -0.004
(0.004) (0.004) (0.005) (0.004) (0.004) (0.008) (0.004)
RAIN*SSA 0.015** 0.014* 0.015**
(0.007) (0.007) (0.007)
log(Y) -0.043*** -0.041*** -0.046*** -0.056*** -0.065*** -0.054*** -0.039***
(0.007) (0.008) (0.014) (0.009) (0.012) (0.015) (0.004)
Sample All NSSA SSA NSSA SSA SSA All
Controls No No No Yes Yes Yes No
Obse. 393 254 139 393 254 139 393
Countries 60 38 22 60 38 22 60
F-Test 14.70*** 14.58*** 7.66*** 6.56*** 6.21*** 2.43*** ---
R-squared 0.15 0.15 0.15 0.26 0.33 0.28 ---
Notes: (1) Standard errors in parantheses. (2) ***, **, and * indicate 1, 5, and 10 per cent significance levels.
(3) Controls include time dummies, openness (OPEN), population size (POP), schooling (ED), civil war
incidence (CIVWAR), civil war incidence in surrounding countries (CIVWAR_S), investment
(INV/GDP), government expenditure (G/GDP), urbanization (URB). (4) Coefficients in column (7) are
corrected using the Kiviet (1995), while standard errors are generated via bootrapping.
Table 4: Temperature and Channel Effects
(1) (2) (3) (4) (5) (6)
RAIN -0.005 0.014** -0.004 0.014** -0.009 -0.008
(0.004) (0.006) (0.004) (0.006) (0.007) (0.010)
log(Y) -0.052*** -0.048*** -0.053*** -0.048*** -0.051*** -0.098***
(0.010) (0.014) (0.010) (0.014) (0.018) (0.023)
TEMP -0.002 0.000 -0.003 0.000
(0.003) (0.005) (0.003) (0.005)
RAIN*TEMP -0.008 -0.001
Sample NSSA SSA NSSA SSA NSSA SSA
Controls Yes Yes Yes Yes Yes Yes
Obse. 254 139 254 139 254 124
Countries 38 22 38 22 38 20
F-Test 7.91*** 3.25*** 7.30*** 2.93*** 6.00*** 4.62***
R-squared 0.28 0.23 0.28 0.23 0.24 0.40
Notes: (1) Standard errors in parantheses. (2) ***, **, and * indicate 1, 5, and 10 per cent significance levels.
Table 5: Extended SSA Sample and Further Robustness Checks
(1) (2) (3) (4) (5) (6)
log(Y) -0.047*** -0.046*** -0.047*** -0.047*** -0.049*** -0.045***
(0.008) (0.008) (0.007) (0.008) (0.008) (0.008)
Observations 301 301 301 301 301 301
Number of countries 42 42 42 42 42 42
F-Test 23.68*** 22.22*** 39.06*** 15.12*** 21.30*** 18.63***
R-squared 0.16 0.15 0.16 0.15 0.14 0.18
Notes: (1) Standard errors in parantheses. (2) ***, **, and * indicate 1, 5, and 10 per cent significance levels.
Appendix A: Selected Full Regression Results of Table 2 Columns (5)-(7) and of
Table 3 Columns (4)-(6)
(1) (2) (3) (4) (5) (6)
METHOD OLS OLS OLS FE FE FE
RAIN -0.007 -0.007 0.016** -0.003 -0.003 0.022***
(0.005) (0.005) (0.008) (0.004) (0.004) (0.008)
RAIN*SSA 0.018*** 0.014*
log(Y) -0.014*** -0.016** -0.037*** -0.056*** -0.065*** -0.054***
(0.004) (0.007) (0.009) (0.009) (0.012) (0.015)
URB -0.009 0.007 0.005 -0.042 0.092
(0.019) (0.026) (0.053) (0.068) (0.091)
POP -0.030 -0.054 -0.187* -0.112* -0.065 -0.250**
(0.042) (0.055) (0.104) (0.060) (0.075) (0.115)
OPEN -0.000** -0.000 -0.000 -0.000** -0.000** 0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
ED 0.004** 0.004** 0.004 0.002 0.002 -0.003
(0.002) (0.002) (0.005) (0.004) (0.004) (0.011)
CIVW -0.014** -0.010* -0.013 -0.014** -0.012* -0.027*
(0.005) (0.006) (0.012) (0.006) (0.007) (0.014)
CIVW_S -0.006 0.005 -0.003 -0.005 0.002 -0.048*
(0.007) (0.007) (0.027) (0.008) (0.008) (0.026)
INV/GDP 0.001*** 0.001*** 0.001 0.001*** 0.002*** 0.000
(0.000) (0.000) (0.001) (0.000) (0.000) (0.001)
G/GDP -0.000** -0.000 -0.000 -0.001*** -0.001** -0.001
(0.000) (0.000) (0.000) (0.000) (0.000) (0.001)
LANDLOCK -0.009 0.001 -0.034***
(0.005) (0.008) (0.012)
ETHNIC -0.018** -0.014 -0.014
(0.009) (0.011) (0.023)
TROPICAL 0.000 0.003 0.501***
(0.007) (0.007) (0.171)
AREA -0.000 -0.000 -0.000
(0.000) (0.000) (0.000)
Constant 0.137*** 0.155*** -0.215
(0.034) (0.051) (0.177)
Sample ALL NSSA SSA ALL NSSA SSA
Obs. 393 254 139 393 254 139
F-Test 4.63*** 5.10*** 3.05*** 6.56*** 6.21*** 2.43***
R-squar. 0.21 0.30 0.39 0.26 0.33 0.28
See appendix B for a definitionof the variables.
1. Country Samples
For the purposes of this paper we generally use observations on developing countries,
although as a robustness check we also include developed countries in one of the
specifications. We consider a country to be of developing status if it is either a low,
lower-middle, or upper-middle income nation according to the World Bank definition
which is based on GNP per capita cut-off points that are constant in real values over
time and were first set 1987.
These cut-off points were based on the Bank's operational
lending categories (civil works preferences, IDA eligibility, etc.). In order to avoid
potential sample selection bias where one excludes countries in our sample that at the
beginning of our sample period, 1960, were ‘developing’ but then became ‘developed’ or
vice versa, we used these cut-off points and data from the World Penn Tables to ensure
that countries were classified as ‘developing’ at the beginning of our sample period or at
the earliest date at which data was available.
For those for which there was no
information in the World Penn Tables, but which we did include in our graphical analysis
in the paper we used the 1987 definition of their status. Our classification of countries
included in our analysis is as follows:
Developing: Sub-Saharan Africa:
Angola, Burundi, Benin, Burkina, Botswana, Central Africa, Cote d'Ivoire, Cameroon,
Congo, Comoros, Cape Verde, Ethiopia, Gabon, Ghana, Guinea, Gambia, Guinea-
Bissau, Equatorial Guinea, Kenya, Lesotho, Madagascar, Mali, Mozambique, Mauritania,
Mauritius, Malawi, Namibia, Niger, Nigeria, Rwanda, Senegal, Sierra Leon, Sao Tome,
Seychelles, Chad, Togo, Tanzania, Uganda, South Africa, Zaire, Zambia, Zimbabwe.
Developing: Non Sub-Saharan Africa:
Algeria, Albania, Argentina, Antigua, Bangladesh, Bulgaria, Belize, Bolivia, Brazil,
Barbados, Chile, China, Colombia, Costa Rica, Cyprus, Cuba, Dominica, Dominican
Rep., Ecuador, Egypt, Fiji, Grenada, Guatemala, Guyana, Honduras, Haiti, Hungary,
Indonesia, India, Iran, Is, Israel, Jamaica, Jordan, Cambodia, St. Kitts, Korea, S,
Lebanon, St. Lucia, Sri Lank, Morocco, Mexico, Malta, Malaysia, Nicaragua, Nepal,
Pakistan, Panama, Peru, Philippi, Papua New Guinea, Poland, Puerto R, Portugal,
Paraguay, Romania, Singapore, El Salvador, Syrian A, Thailand, Trinidad, Tunisia,
Turkey, Uruguay, St. Vincent, Venezuela, Vietnam, Yemen.
2. Gridded Climatic Data
The variable used to construct the gridded climatology was each available station’s mean
value of precipitation over the period 1961-1990, where these normals were calculated
from a variety of sources.
In cases were published sources did not provide information
on the chosen normal period, normals outside of this period were substituted. As noted
by New et al (1999), the improvement in accuracy gained by includung additional station
information outweighs any penalty associated with relaxing temporal fidelity. Moreover,
means outside the 1961-1990 were generally assigned a low weighting during the
interpolation. The authors then used a thin-plate spine-fitting technique to interpolate
37 The only countries covered in the World Penn Tables that changed from ‘developing’ to developed
status were Singapore, Cyprus, and Puerto Rico.
38See New et al (1999) for details.
the climate surfaces into the 0.5 by 0.5 degree high-resolution climatology grid. One
should note that this technique is robust even in areas with sparse or irregularly spaced
data points. Moreover, it maximizes the representation of the spatial variability of the
mean climate given the available data.
For deriving the time series for each grid, first each station rainfall series from the
beginning of the 20
century was converted into monthly anomalies calculated as a
percentage of its 1961-1990 mean, since the gridded climatology was calculated from the
same measure. The individual series were then interpolated to obtain overall values for
every grid using the angular distance-weighted method (ADW) on measurements of the
eight nearest stations.
Since measurements from stations far away from the grid point
were unlikely to provide useful information about that grid’s climate, they were forced to
zero if they were beyond the correlation decay distance, thus ‘relaxing’ their value
towards the monthly 1961-1990 mean of that station measurement.
These series were
then converted back into millimeters of precipitation, resulting in time series over the
period 1901-1998. Annual measures are simply the sum of the monthly measures of
3. Other Variables
All other variables used in the analysis are described according to their definition and
source as below:
Variable Definition Nature Source
RAIN Rainfall anamolies Time varying (annual);
RAIN_FAO FAO rainfall measure Time varying (annual);
RAIN_POP Population weight. rainfall
Time varying (annual);
IPCC ; African
TEMP Annual temperature
Time varying (annual);
SST_1, SST_2 CCA Generated Modes Time varying (annual);
SSA 1-0 Dummy Time invariant
SAHEL Sahel Dummy
NON_SAHEL Non-Sahel Dummy
Log(GDP/Cap) Log of initial year GDP
per capita (constant value)
World Penn Tables 6.1
OPEN (exports+imports)/GDP Time varying (annual):
World Penn Tables 6.1
POP Size of population Time varying (annual)
World Penn Tables 6.1
ED Average years of
Barro and Lee (1993)
CIVWAR Number of years of civil
Murdoch and Sandler
CIVWAR_S Number of years of civil
wars in surrounding years
Murdoch and Sandler
39 The ADW essentially “..employs a distance weighting function so that stations closest to the grid point
of interest carry greater weight” (New et al 2000, p. 2221).
40 The correlation decay distance is the distance at which zonally averaged interstation correlation is no
longer significant at the 95 per cent level.
INV/GDP Investment share of eal
GDP per capita
Time varying (annual)
World Penn Tables 6.1
G/GDP Government Spending
share of real GDP per
Time varying (annual)
World Penn Tables 6.1
URB Percentage of population
living in urban areas
Time varying (five year
Davis and Henderson
HYDRO Kilowatts per hour Time varying (annual)
UN Energy Statistics
AGP Aggregate price-weighted
volume of agricultural
with the base period
Time varying (annual) FAOSTAT
LANDLOCK 1-0 Dummy if country is
Time invariant World Bank Global
ETHNIC Index of Ethnic
Time invariant World Bank Global
TROP 1-0 Dummy for tropical
Time invariant World Bank Global
AREA Land Area Time invariant World Bank Global
IRR Percentage of Land
Time Invariant FAO database
6 Regional Dummies Dummies indicating
whether country is in
Asia, Latin America,
Middle East, SSA, South
Asia, and East Asia