Geography and macroeconomics: New data
and new findings
William D. Nordhaus*
Yale University, 28 Hillhouse Avenue, New Haven, CT 06520-8268
This contribution is part of the special series of Inaugural Articles by members of the National Academy of Sciences elected on May 1, 2001.
Contributed by William D. Nordhaus, December 2, 2005
The linkage between economic activity and geography is obvious:
Populations cluster mainly on coasts and rarely on ice sheets. Past
studies of the relationships between economic activity and geog-
raphy have been hampered by limited spatial data on economic
activity. The present study introduces data on global economic
activity, the G-Econ database, which measures economic activity
for all large countries, measured at a 1° latitude by 1° longitude
scale. The methodologies for the study are described. Three appli-
cations of the data are investigated. First, the puzzling ‘‘climate-
output reversal’’ is detected, whereby the relationship between
temperature and output is negative when measured on a per
capita basis and strongly positive on a per area basis. Second, the
database allows better resolution of the impact of geographic
attributes on African poverty, finding geography is an important
source of income differences relative to high-income regions.
Finally, we use the G-Econ data to provide estimates of the
economic impact of greenhouse warming, with larger estimates of
warming damages than past studies.
economic growth ? development ? climate change
generally ignore geographic factors such as climate, proximity to
water, soils, tropical pests, and permafrost. This inaugural essay
examines this intellectual division, presents data on geographically
based economic activity, and examines some of the major relation-
web site (http:??gecon.yale.edu).
Why has macroeconomics generally ignored geography? As will
be discussed in subsequent sections, three factors have prevented a
thorough integration of geographic factors into macroeconomic
analysis. First, economic growth theory has emphasized the role of
endogenous and policy factors, such as capital formation, educa-
tion, and technology, rather than exogenous factors such as geog-
raphy or even population. Although natural resources (particularly
land and minerals) have been featured in some studies, climate,
soils, tropical diseases, and similar ‘‘unchanging’’ factors have
typically been omitted from modern economic growth analysis.
Second, studies of the impact of geography on economic activity
this focus is sensible for a discipline like economics, which focuses
on national economic policies and living standards, it is difficult to
capture time-invariant geographic factors in such studies. To sep-
arate out geographic factors, this study examines areal density of
output and per capita output. We will see that shifting the measure
dramatically changes the estimated effects of geography on eco-
Third, most measures of economic activity have been time series
or panels measured at the level of the country, which provide ?100
observations at enormously different geographic scales. The data
set presented here (GECON 1.1), which measures output with a
resolution of 1° latitude by 1° longitude, covers 25,572 terrestrial
grid cells. Such an increase in resolution is analogous to pictures
to most people: populations cluster mainly on coasts and rarely
from the Hubble telescope, which provide clear and crisp answers
to many previously difficult and fuzzily answered questions.
The change in emphasis proposed here has an enormous effect
on the estimated impact of geographic attributes on economic
activity. The G-Econ database (described in detail in the second
part of this article) can be useful not only for economists interested
in spatial economics but equally for environmental scientists look-
ing to link their satellite and other geographically based data with
I begin with a brief survey of the role of geographic factors in
economic analysis and empirical work. In this survey, I will discuss
mainly macroeconomics, and it must be emphasized that these
remarks present a highly condensed view of studies that relate to
global economic processes. The vast and impressive literature in
geography and regional economics is largely outside the scope of
It will be useful to state what I mean by ‘‘geographical’’ factors
(or, better, geophysical factors studied in ‘‘physical geography’’).
These physical attributes are tied to specific locations. They may be
nonstochastic on the relevant time scale (such as latitude, distance
from coastlines, or elevation) or they may be stochastic with slowly
moving means and variability (such as climate or soils). One of the
critical features of the present approach is that geographic factors
are statistically exogenous in the sense that they cause, but to a first
approximation are not caused by, economic and other social
variables. For our purposes, we omit most environmental and
endogenous geographic variables, such as pollution, land use, and
the natural-resource content of trade or output. Although these
factors are of critical importance for many purposes, the focus here
is on exogenous and large-scale factors that are largely unaffected
by human activities on decadal time scales.
In reflecting on the wealth of nations, early economists assumed
that climate was one of the prime determinants of national differ-
presumption was probably correct. Earlier civilizations, such as
those investigated in Landes’s history of economic growth (1) or
Diamond’s analysis of societal collapses (2), were highly dependent
and trade than most economies today.
However, one of the major factors in economic development has
been the movement from climatic-sensitive farming and into cli-
mate-insensitive manufacturing and services. In 1820, 72% of U.S.
employment was on farms, whereas by 2004, the share was down to
1.2%. Many studies suggest that the market economy in the
developed world is relatively insensitive to moderate and gradual
changes in climate or similar geographic conditions (see below).
give short shrift to climate as the basis for the differences in the
Conflict of interest statement: No conflicts declared.
Freely available online through the PNAS open access option.
Abbreviation: GCP, gross cell product.
© 2006 by The National Academy of Sciences of the USA
March 7, 2006 ?
vol. 103 ?
wealth of nations. A review of a handful of textbooks on economic
development shows that climate is confined to a few lines in
hundreds of pages. (Exceptions are ref. 3 and more recent work
discussed later.) The modern view of economic growth presents
development as an engine fueled by capital, labor, and technology;
sometimes, mineral resources are included, but only with a major
stretch of interpretation would we equate resources with geo-
graphic attributes. The recent wave of studies investigating inter-
a determining variable.
Over the last decade, economists have begun to introduce
geography into studies of economic growth and development. One
early set of studies by Hall and Jones (4) investigated the reasons
for the enormous diversity of per capita incomes across nations.
Their main hypothesis was that average output differences across
nations are primarily determined by institutions and government
policies. In examining statistically exogenous instruments, they
found that geography (measured as distance from the equator) was
among the most significant variables behind differences in per
capita output by country. They speculated that location affects
economic success because of patterns of human settlements, which
The study of economic geography has been revitalized by the
work of Sachs and his colleagues (5, 6). The major thrust of this
work is to understand the economic problems of tropical Africa.
They examine geographic factors such as the percent of the land
area in the tropics as influencing growth. Their surprising conclu-
approximately two-thirds of the weight of Africa’s growth shortfall
to the ‘noneconomic’ conditions, and only one-third to economic
policy and institutions’’ (5). Other studies examine the role of
‘‘landlockedness,’’ coastal settlements, and tropical diseases on
economic activity (6).
The geographically based studies on Africa have come under
issues concerns the statistical ‘‘endogeneity’’ of the independent
variables. A second and more far-reaching criticism concerns the
relative importance of institutions. Several studies argue (along the
best understood as determined by historical conditions in which
geography led to settlement patterns that were favorable to good
that good institutions led to high incomes (7). Although these
West Germany, North and South Korea, and Baja and Alta
California surely suggests the importance of institutions in eco-
Existing studies serve many useful purposes, but they have three
distinct shortcomings for determining the impact of geographic
attributes on economic activity, all of which are remedied by the
present data set. First, virtually all studies focus on national data. If
institutions are indeed a key ingredient in economic growth, then
it would be very difficult to sort out geographic from national
influences without disaggregating below the national level. The
20,000 terrestrial observations (hence, many per nation) as com-
studies just reviewed.
Second, the analysis here is primarily concerned with the geo-
graphic intensity of economic activity rather than the personal
intensity of economic activity. In other words, it focuses on the
intensity of economic activity per unit area rather than per capita
or per hour worked. Although geographic intensity may be less
interesting for many policy purposes than the determinants of per
capita income, the present approach places the emphasis clearly on
Third, virtually all prior studies have focused on proxies for
geographic variables rather than those that are intrinsically impor-
tant. Distance from the equator and percent area in the tropics, for
example, have no intrinsic economic significance. One of the
advantages of using gridded data rather than national data is that
of the important geographic data (including climate, location,
distance from markets or seacoasts, and soils) are collated on a
geographic basis rather than based on political boundaries. There
is also an important interaction between the finer resolution of the
economic data and the use of geographic data because, for many
countries, averages of most geographic variables (such as temper-
ature or distance from seacoast) cover such a huge area that they
are virtually meaningless, whereas for most grid cells the averages
cover a reasonably small area.
Methodology for Estimating Gross Cell Product
The Concept of Gross Cell Product (GCP). The major statistical
contribution of the present research program has been the devel-
opment of ‘‘gridded output’’ data, or GCP. In this work, the ‘‘cell’’
is the surface bounded by 1-degree latitude by 1-degree longitude
contours. A full description of the data and methods can be found
at the project web site (http:??gecon.yale.edu).
The globe contains 64,800 such grid cells; we provide output
cells are outside Antarctica, 17,433 have complete and minimum-
nonzero population and output.
The grid cell is selected because it is the unit for which data,
particularly on population, are most plentiful. It also is the most
convenient for integrating with global environmental data. Addi-
tionally, this coordinate system is (to a first approximation) statis-
tically independent of economic data (which obviously is not the
case for political boundaries), and the elements are of (almost)
used in the paper.
product and gross regional product as developed in the national
income and product accounts of major countries, except that the
geographic unit is the latitude-longitude grid cell. GCP is gross
value added in a specific geographic region; gross value added is
equal to total production of market goods and services in a region
less purchases from other businesses. GCP aggregates across all
cells in a country to gross domestic product. We measure output in
purchasing-power-corrected 1995 U.S. dollars by using national
aggregates estimated by the World Bank. We do not generally
The exception to this rule is that we make purchasing-power
adjustments for oil and mineral production in countries with a high
proportion of output coming from these sources.
The general methodology for calculating GCP is the following:
GCP by grid cell ? (population by grid cell)
? ?per capita GCP by grid cell).
The approach in Eq. 1 is particularly attractive because a team
of geographers and demographers has recently constructed a
detailed set of population estimates by grid cell, the first term on
the right-hand side of Eq. 1.†Estimates of GCP, therefore,
primarily require new estimates of per capita output by grid cell.
Methodologies for Estimating Per Capita GCP. The detail and accu-
racy of economic and demographic data vary widely among coun-
tries, and we have developed alternative methodologies depending
on the data availability and quality. The methodologies are de-
†The gridded population data are available online at http:??sedac.ciesin.columbia.edu?
plue?gpw with full documentation in ref. 8 and updated in ref. 9.
March 7, 2006 ?
vol. 103 ?
no. 10 ?
scribed (http:??gecon.yale.edu; W.N., Q. Azam, D. Corderi, N. M.
Victor, M. Mohammed, and A. Miltner, unpublished data), and
data for each country are also available upon request.
attributes are central: the level of spatial disaggregation and the
source data used to construct the estimates of gross cell product. In
terms of spatial disaggregation, there are usually three political
subdivisions: (i) national data, (ii) ‘‘state data’’ from the first
political subdivision, and (iii) ‘‘province data’’ from the second
political subdivision. We use the lowest political subdivision for
which data are available, although different levels are sometimes
There are four major sources of the economic data: (i) gross
(ii) regional income by industry (such as labor income by industry
and counties or provinces for the United States and Canada), (iii)
regional employment by industry (such as detailed employment by
industry and region for Egypt), and (iv) regional urban and rural
population or employment along with aggregate sectoral data on
agricultural and nonagricultural incomes (used for African coun-
tries such as Niger). For each country, we combine one or more of
the four data sets at one or more regional levels.
Specific Methodologies. Some examples illustrate the variety of
available for gross state product for 50 states. We use detailed data
gross county product. We then apply spatial rescaling described
below to convert the county data to the 1,369 terrestrial grid cells
for the United States. We would judge these estimates to be highly
reliable. A similar approach was used for Canada, the European
Union, and Brazil. (ii) For most other high-income countries, we
use gross regional product by first political subdivision (such as
oblasts for the Russian Federation). For small- or medium-sized
countries (Argentina), this approach will be relatively reliable,
large and the spatial resolution is consequently poor. (iii) For many
middle-income countries, such as Egypt, we have data from recent
We then use these data along with national accounts data on
national output by industry to estimate output by region and
industry and then aggregate these data across industries to obtain
lowest-income countries, we have no regional economic data. In
these cases, we combine population censuses on rural and urban
populations with national employment and output data to estimate
output per capita by region. For these countries, because of the
sparse economic data and limited regional data, estimates of GCP
are less accurate than those for high-income countries.
Spatial Rescaling. The data on output and per capita output are
estimated by political boundaries. To create gridded data, we need
to transform the data to geographic boundaries. I call this process
‘‘spatial rescaling,’’ although it goes by many names in quantitative
aggregation,’’ or ‘‘areal interpolation’’ (10–12). Spatial rescaling
arises in a number of different contexts and requires inferring the
distribution in another set of spatial aggregates, where neither is a
subset of the other. The scaling problem arises here because all
economic data are published by using political boundaries, and
these data need to be converted to geographic boundaries.
Having reviewed alternative approaches and done some simu-
lations with economic data, we settled on the ‘‘proportional allo-
cation’’ rule (details available upon request). The first step is to
divide each grid cell into ‘‘subgrid cells,’’ each of which belongs
uniquely to the smallest available political unit (call them ‘‘prov-
inces’’). The next step is to collect or estimate per capita output for
each province. Third, the proportional allocation rule assumes that
per capita output is uniformly distributed in each province and that
assumptions, we can calculate a tentative estimate of output for
each subgrid cell as the product of the subgrid cell area times the
province. We next calculate the GCP as the sum of the outputs of
each subgrid cell. The final step is to adjust the GCPs to conform
to the totals for the province and the country.
This approach is data-intensive and computationally burden-
some because it requires estimating the fraction of each grid cell
belonging to each province and estimating the economic data for
each of the provinces. Calculations indicate that there are signifi-
cant gains in accuracy from disaggregating. For the United States,
using actual county data, we estimate that disaggregating from the
of the cell average by a factor of 5.
Impact of Geography, Climate, and Other Geographic
Activities on Economic Activity
This final section presents some results of analyzing the patterns of
economic activity by using the new G-Econ data set. This study is
not meant to be a comprehensive analysis, which must await
integrating the data with a further geographic attributes and time
series on spatial economic data. Moreover, at this point, we are
primarily examining basic patterns and reduced-form estimates;
future work should focus on structural estimates of the major
An Economic Map of Europe. Fig. 1 shows an economic contour map
of Europe, with some important mountains and lowlands marked.
Unlike familiar contour maps, this one has height proportional to
the output density (output per square kilometer) in different
from southern England through northern Italy, whereas the pe-
ripheral areas, particularly arctic Europe, are the economic low-
lands. Maps for other countries are available upon request.
Fig. 2 shows fractile kernel plots of five major geographic
variables. A fractile plot first orders the variable from lowest to
highest observation. It then estimates a kernel density function or
smoothed nonlinear relationship between the fractile and the
logarithm (log10) of output density. For example, output density
differs by a factor of 105from fractile trough to peak for mean
temperature, whereas the difference for mean precipitation is 102
from trough to peak. Clearly, all geographic variables have major
systematic impacts on economic activity. The relationships are
difficult to capture in a simple fashion because the impacts are
The Climate-Output Reversal. The first set of tests examines the
relationship between economic activity and a limited set of geo-
graphic activities, focusing primarily on climate. Many economic
studies have examined the relationship between geography and
economic activity. One of the major findings is that output per
the unit of observation is refined to grid cells within countries?
Fig. 3 shows a ‘‘box plot’’ of the relationship between mean
temperature in each grid cell and the output per capita in that grid
cell. A box plot groups the observations in each bin and then
estimates several statistics for those observations; the different
statistics are explained in the Fig. 3 legend. For this purpose, bins
have a 2°C width (every second bin is shown on the bottom axis).
between temperature and per capita output.
unit area is the key variable from a geographic and ecological point
of view. Fig. 4 shows a box plot of the relationship between mean
For this purpose, we have assumed that the data are censored and
km2, so log10(truncated observations) ? 0. The estimates are
relatively unreliable for log10densities ?1.
The striking finding is the very sharp positive gradient between
output density and temperature from the lowest observations to
?5°C; the difference between the peak and the lowest temperature
(polar) regions is a factor of at least 105. The temperature output
varies modestly above 0°C, peaking between 7°C and 14°C. Output
density falls by a factor of ?100 from the peak to the high-
The striking paradox shown in Figs. 3 and 4 can be labeled the
climate-output reversal. This reversal indicates opposite relation-
output per person or output per area. (This relationship is similar
if the geographic variable is latitude.)
fractile plots for key variables (mean temperature, mean precipitation, mean
distance from coast, mean elevation, and absolute value of latitude). Fractiles
the log10(output density) and the geographic variable. Zero values of output
are included as equal 0 (n ? 17,796).
Fractile plot for key geographic variables. The figure shows the
Note how economic activity clusters in the core, whereas the periphery has much lower economic elevations. The observations measure economic activity in
millions of 1995 U.S. dollars per km2at a 1° latitude by 1° longitude scale.
Economic map of Europe. This figure shows an economic topographical map of Europe with heights proportional to gross domestic product per area.
that high-latitude countries have higher output per capita than those in low
latitudes. This relationship is verified by using mean temperature as the
geographic variable for grid cells. Coldest regions have an output per capita
?12 times that of warmest regions. In boxplots, the means are the red circles,
the medians are the heavy red horizontal line, the one-sigma ranges of the
median are the blue shaded regions, and the interquartile ranges are the
boxes. The width of the box is proportional to the square root of the number
of observations in each bin. The bin is shown on the horizontal axis, but only
is a factor of 10. Zero observations are omitted (n ? 15,755).
March 7, 2006 ?
vol. 103 ?
no. 10 ?
approximation, the reason is that people are mobile, whereas land
is not. Under economic conditions that have existed for recorded
history, areal productivity is low in ice-covered and very cold
regions. This point is obvious for agriculture, but with few excep-
tions (such as skiing and glaciology), it is also true for other sectors
of the economy. We can use different regions of high-income
countries to illustrate. The output density in northern Greenland is
$500 per km2, the non-oil output density in Alaska averages $6,000
km2. Unless the global economy becomes devoted substantially to
the extraction of ice, it seems likely that the low areal productivity
of cold regions will prevail.
The reasons for high per capita productivity of low-temperature
regions are not so obvious. To see what can be explained by human
behavior, take the case of perfect economic mobility over space. In
other words, assume that people migrate until average outputs are
equalized in all regions. Under this assumption, the temperature?
output-per-capita gradient of Fig. 3 would be horizontal. Although
the assumption of perfect mobility does not hold for recent years,
particularly across national boundaries, human mobility is surely at
the heart of the difference between Figs. 3 and 4.
Three other factors might give the temperature?output-per-
capita gradient its negative slope. First, some output in low-
temperature regions is highly capital-intensive (such as oil produc-
tion) and, therefore, tends to have high output per capita. This
factor can be ruled out given the relative unimportance of mineral
production in high-latitude regions. Second, there may be ‘‘com-
pensating differentials,’’ whereby people require higher real wages
to live in unpleasant frigid conditions. Here again, although cold
regions are unattractive, evidence on compensating differentials
cannot explain the large differences. Moreover, tropical regions
have their own perils and fail to show similarly large compensating
that countries in temperate and colder regions have higher per
As described in the first section, a major debate in the economic-
growth literature is whether these differences involve primarily
geography or national institutions.
We can use the G-Econ data set to separate out institutional and
equation of the natural logarithm of output per capita on mean
temperature can be run with and without country fixed effects for
all large and medium countries (n ? 15,229). If national effects
involving institutions and political systems are most important, the
country effects would capture such impacts. The country effects
decrease the sensitivity of log per capita output with respect to
mean temperature by approximately one-third, from ?0.058
(?0.00068) to ?0.040 (?0.00092). This result indicates that ap-
proximately two-thirds of the gradient shown in Fig. 3 cannot be
explained by country-specific factors such as institutional differ-
ences, history, major locational advantages, and such national
factors. At this point, the remaining sources of the climate-output
reversal are still an open question.
Are Output Differences in Output Explained by Pure Geography? One
of the central questions in economic geography is how much of the
dispersion of output is explained by geographic variables. The
G-Econ data provides an ideal laboratory to answer this question.
For this purpose, I estimated a multivariate regression with the
logarithm of output per km2as the dependent variables, with
independent variables being temperature, precipitation, and other
geographic variables. More precisely, the equation is
ln (yij) ? ?0jCountj??
?kgk?Geoijk? ? ?ij,
Geographic variables, Geoijk,are mean annual temperature, mean
annual precipitation, mean elevation, ‘‘roughness’’ measured as
standard deviation of elevation in grid cell, soil categories, and
distance from coastline. The gkrepresent polynomial functions of
geographic variables. The Greek variables ?0jare coefficients on
regions, whereas the ?kare regression coefficients on geographic
variables. It should be noted that we omit all clearly endogenous
variables (such as coastal density, proximity to markets, and health
This test uses a dense set of exogenous variables to capture all
interactions.‡The equation explains 91% of the variance of output
density for all 17,409 minimum-quality observations. The geo-
graphic variables are all highly significant (as is clear for temper-
ature in Figs. 3 and 4).
The equation has some interesting features. It indicates that
the ‘‘optimal’’ temperature (which maximizes output density) is
?12°C. Moreover, it suggests that some countries do particularly
well or badly given their climates. Countries that are big negative
outliers are Australia, Mozambique, Madagascar, and Angola.
Those with positive country effects are Denmark, Japan, France,
and Italy. The low density of output in Greenland, Canada,
Russia, and Alaska are consistent with the economically inclem-
ent climates in those regions.
It should be recognized that much of the dispersion of economic
activity is unexplained; the standard error of the multivariate
regression is 1.97, which is the equivalent of an average error of a
factor of exp(1.97) ? 7.2. Geographical variables will probably
never explain the high densities of economic activity in Madrid,
Paris, or Moscow –nor the relatively low levels in temperate South
America and South Africa. Geography is important, but much
Africa: Geography, Economics, and Destiny. Africa is widely recog-
nized to be the globe’s troubled continent. In terms of economic
statistics, although gross domestic product per capita in 2004 was
‡The precise specification in Eq. 2 contains 72 country effects plus nine polynomial terms in
temperature and precipitation, six statistics on extremes and higher moments in temper-
ature and precipitation, the first and second moments of elevation, three variables for
17,305 degrees of freedom, although that is probably overstated because of spatial
correlation. Undertaking further analysis of these data by using the techniques of spatial
statistics is an important area of research. All results are described in detail in the
background documentation available upon request.
distribution of output density by temperature. Output density varies by at least
five orders of magnitude from cold to temperate region. For the explanation of
the boxplots, see Fig. 3. Zero observations are set at log10(x) ? 0 (n ? 18,995).
Boxplot of output density and temperature. This boxplot shows the