ARTICLES

Potential effect of population and climate changes on global

distribution of dengue fever: an empirical model

Simon Hales, Neil de Wet, John Maindonald, Alistair Woodward

Summary Introduction
There is increasing scientific interest in the potential
Background Existing theoretical models of the potential effects on health of global climate change. One area that
effects of climate change on vector-borne diseases do not has received particular attention is the association
account for social factors such as population increase, or between climate variation and vector-borne diseases.1,2
interactions between climate variables. Our aim was to Dengue fever is the most important viral vector-borne
investigate the potential effects of global climate change on disease in the world.3 The disease affects hundreds of
human health, and in particular, on the transmission of millions of people every year, and is transmitted
vector-borne diseases. predominantly by one species of mosquito, Aedes aegypti,
which has adapted to living near areas of human
Methods We modelled the reported global distribution of habitation. It feeds during the day and prefers human
dengue fever on the basis of vapour pressure, which is a beings to other animals. No effective vaccine or drug
measure of humidity. We assessed changes in the treatment for dengue fever is yet available, so
geographical limits of dengue fever transmission, and in the management of the disease has relied on vector control
number of people at risk of dengue by incorporating future measures, such as reduction of breeding sites and use of
climate change and human population projections into our insecticides. Such measures have succeeded in
model. eradicating mosquitoes in some regions, but have proved
difficult to maintain in the long term.
Findings We showed that the current geographical limits of Mosquito-borne disease transmission is climate
dengue fever transmission can be modelled with 89% sensitive for several reasons; mosquitoes require standing
accuracy on the basis of long-term average vapour pressure. water to breed, and a warm ambient temperature is
In 1990, almost 30% of the world population, 1·5 billion critical to adult feeding behaviour and mortality, the rate
people, lived in regions where the estimated risk of dengue of larval development, and speed of virus replication.4,5 If
transmission was greater than 50%. With population and the climate is too cold, viral development is slow and
climate change projections for 2085, we estimate that about mosquitoes are unlikely to survive long enough to
5–6 billion people (50–60% of the projected global become infectious. Although a suitable climate is
population) would be at risk of dengue transmission, necessary for disease transmission, other factors are
compared with 3·5 billion people, or 35% of the population, if needed for an epidemic to take place, including a source
climate change did not happen. of infection, vector populations, and a susceptible human
population.
Interpretation We conclude that climate change is likely to Mathematical models project substantial increases in
increase the area of land with a climate suitable for dengue transmission of vector-borne diseases in various climate
fever transmission, and that if no other contributing factors change situations.1,6,7 However, these models have been
were to change, a large proportion of the human population criticised on the grounds that they do not adequately
would then be put at risk. account for rainfall,8 interactions between climate
variables,9 or relevant socioeconomic factors.10 An empiri-
Lancet 2002; 360: 830–34. Published online August 6, 2002 cal model of malaria transmission risks, which accounted
http://image.thelancet.com/extras/01art11175web.pdf for interactions between climate variables,9 predicted
little change in the size of the at-risk population by the
year 2050.
The decline of malaria in Europe and the
southern USA in the 20th century was attributable to
social and economic development and improved public
health services, and not to change in climate.10 However,
this finding does not mean that ambient temperature
and rainfall are irrelevant from a health perspective.
Climate is one of the fundamental forces behind
Department of Public Health, Wellington School of Medicine and epidemics, and its effects become evident if adaptive
Health Sciences, PO box 7343, Mein Street, Wellington, New measures falter or cannot be extended to all populations
Zealand (S Hales PhD, Prof A Woodward PhD); International Global at risk.
Change Institute, Hamilton, New Zealand (N de Wet MPHIL); Centre Our aim was to describe the current geographical limits
for Bioinformation Science, John Curtin School of Medical of dengue fever transmission on the basis of climate. In
Research and Mathematical Sciences Institute, Australian an empirical model, we incorporated future projections of
National University, Canberra, Australia (J Maindonald MSc) climate, to estimate changes in the geographical limits of
Correspondence to: Dr Simon Hales dengue fever transmission and the size of populations
(e-mail: shales@wnmeds.ac.nz) at risk.

830 THE LANCET • Vol 360 • September 14, 2002 • www.thelancet.com

For personal use. Only reproduce with permission from The Lancet Publishing Group.
 ARTICLES

Method model was then fitted with these data, and resulted in a
Data modelled baseline risk of dengue fever transmission
Dengue fever was defined as present if any outbreak of between zero and one for each grid cell, on the basis of
the disease had been reported between 1975 and 1996.10 climate data for 1961–90. The accuracy of the model
Every country was a geographic unit, except where more was assessed by comparing our results with the known
detailed information was available, in which case distribution of dengue fever.
subnational administrative regions were used.
Population projections were based on the spatial Future climate predictions
pattern in 1990,11 and region-specific projections for The model results were applied to future climate change
2055 and 2085.12 Monthly averages of vapour pressure, situations to generate projections of dengue fever risk in
rainfall, and temperature recorded between 1961 and the 2050s and 2080s. As recommended by the IPCC,15
1990,13 and global circulation model projections of we estimated future climate by combining recorded
future climate were obtained from the baseline data with projections. In 1992, the IPCC
Intergovernmental Panel on Climate Change (IPCC) defined six alternative scenarios of greenhouse gas
data distribution centre.14 emissions, named IS92a to IS92f. These scenarios use a
Data were entered into a geographic information range of assumptions about how future greenhouse gas
system, converted into ASCII grid format. Global emissions might evolve in the absence of mitigation
circulation model data were interpolated to a resolution policies. All of our analyses used the IS92a scenario,
of 0·5° latitude and longitude, or 720 grid cells east-west which results in a 1% per year increase in greenhouse gas
and 360 cells north-south, with a total of 259 200 cells. concentrations leading to about a three-fold rise of CO2
Of these, the 61 538 cells that represented land were concentration by the year 2100 compared with 1990
included in the analyses. values. This step resulted in a modelled risk of dengue
transmission between zero and one for each grid cell.
Model
We used logistic regression, fitted by the method of Population at risk
maximum likelihood, to model the presence (1) or A crude estimate of the population living within
absence (0) of dengue, on the basis of 1961–90 climate countries or administrative regions where at least one
reports. We used monthly average rainfall, vapour outbreak of dengue fever had been reported was
pressure, and maximum, minimum, and mean generated in the geographic information system by
temperature as exposure variables, singly, and in overlaying the map of dengue fever occurrence with the
combination, with or without statistical interaction terms. gridded 1990 population data. Next, we used output
For model selection, the gridded data were aggregated by from the regression model to estimate the population at
administrative region (n=2000). This approach allowed risk of dengue fever.
us to keep to a minimum the effect of spatial Previously generated baseline data for risk of dengue
autocorrelation in the data, while avoiding over-fitting of fever transmission were reimported into the geographic
the model with 20-fold cross-validation. Leaving out each information system as a text file with a standard ASCII
twentieth part in turn, we compared recorded values for header and one row per grid cell. Following accepted
the omitted part, with predictions from a model that is practice, we defined the population at risk according to a
fitted to the remaining 19/20ths of the data. probability cutoff of 0·5—ie, estimated populations
Data were exported from the geographic information living in an area where the modelled probability of
system as a text file containing columns for presence or transmission was greater than 50% were defined as at
absence of dengue fever, and estimates of climate risk. Transmission probabilities were generated for three
variables, and rows for each grid cell. Our selected situations; the estimated population in 1990 with

Figure 1: The distribution of dengue fever between 1975 and 1996

In yellow areas there was at least one outbreak between 1975 and 1996.3

THE LANCET • Vol 360 • September 14, 2002 • www.thelancet.com 831

For personal use. Only reproduce with permission from The Lancet Publishing Group.
ARTICLES

0·0–0·1 0·1–0·2 0·2–0·3 0·3–0·4 0·4–0·5

0·5–0·6 0·6–0·7 0·7–0·8 0·8–0·9 0·9–1·0

Figure 2: Estimated baseline population at risk in 1990 (A) and estimated population at risk in 2085 (B)
Results of a logistic regression model with vapour pressure (humidity) as the predictor of dengue fever risk, using climate data from 1961 to 1990 (A).
Forecast geographical distribution of dengue transmission based on climate projections for 2080–2100 from a global circulation model (CCGCMA2) (B).
Colours represent probability of dengue fever transmission.

1961–90 (baseline) climate data, the projected not readily available and change fields for specific
population in 2055 and 2085 with baseline climate data, humidity were calculated as ratios for 2080–2100 versus
and projected population in 2055 and 2085 with climate 1975–95 for single global circulation model runs.
change projections for those years. Examination of a map of the residuals from our model
which generated baseline risk of dengue fever
Sensitivity analysis transmission on the basis of baseline climate data,
Risk data for dengue fever transmission with expected suggested that spatial autocorrelation was not
climate change, and for predicted climate change and completely removed by the model. We investigated the
population growth were regenerated with alternative effect of including further humidity variables in the
global circulation models (ECHAM4,16 HADCM2,17 model, which we hoped would account for some, or all,
CCSR/NIES,18 CGCMA2,19 and CGCMA120); and three of this residual autocorrelation. Using the geographic
alternative models (all of which were for the CGCMA2 information system, we estimated the maximum
model, but with slightly different initial conditions). humidity for each grid cell within radii of 5, 10, and 20
Because global circulation model data were not readily cells. These data were included in multivariate logistic
available, slightly different comparison periods were regression models, with or without statistical interaction
used for generating model change fields. For the terms.
HADCM2, CGCMA1 and ECHAM4 models, average
changes in vapour pressure from several models runs Role of the funding source
(2070–99 vs 1961–90) were available from the The funding source did not have a role in the study
intergovernmental panel for climate change. For design, data collection, data analysis, data
CGCMA2 and CCSR/NIES models, these results were interpretation, or writing of the report.

832 THE LANCET • Vol 360 • September 14, 2002 • www.thelancet.com

For personal use. Only reproduce with permission from The Lancet Publishing Group.
 ARTICLES

Results With a suitable climate, other factors, including a

We showed that annual average vapour pressure was the source of infection, a vector, and susceptible human
most important individual predictor of dengue fever populations would also need to be present for an
distribution. Therefore, we chose a cautious final model epidemic to happen. Where and when dengue occurs in
that contained humidity as the only explanatory variable, the future will therefore depend on diverse economic,
which is in keeping with recommendations of the US social, and environmental factors. Although we do not
Committee on Climate, Ecosystems, Infectious Diseases, deny the importance of these factors in determining the
and Human Health.2 Vapour pressure proved a significant actual transmission of vector-borne disease, we do not
predictor of dengue fever risk (OR 1·3; SE 0·003; attempt to account for them in our model.
p<0·0001). When areas at risk of dengue fever were The data for the distribution of dengue fever
defined by a cutoff probability of 0·5, the model fitted epidemics are of uncertain quality. However, data for
with gridded data classified 89% of the grid cells presence or absence of dengue fever, though crude, are
accurately. likely to be more reliable for our purpose than reports of
On the basis of the reported distribution of dengue number of cases. Because we used dengue fever
fever and the estimated population in 1990, we calculated epidemiology data that covered many years, the effect of
that 2 billion people, (about 40% of the 1990 world human interventions is likely to be small. Successful
population) lived in regions where at least one outbreak of vector eradication campaigns would have biased results
dengue was reported between 1975 and 1996 (figure 1).3 towards prediction of the absence of dengue fever.
We refined this estimate using gridded data to obtain However, because of unfavourable climate conditions,
highest accuracy of the estimates of population at risk. vector control measures are difficult to sustain over time,
However, this method was not perfect, and nearly 5% of except in regions where vector survival is precarious.
the population living near, or on boundaries of land and Spatial autocorrelation is likely to have increased the
sea could not be assigned to a risk category by the apparent significance of the results. However, the main
geographic information system. On the basis of this results are likely to be conservative, since when we
regression model, we conservatively estimated the attempted to model residual spatial autocorrelation by
baseline population that was at risk in 1990 to be including additional variables in the model, the
1·5 billion people (about 30%) (figure 2). If humidity population at risk increased. Furthermore, in this study
remains at baseline values, and projected increases in we are mainly interested in applying the model to predict
population are realised, we forecast that 3·2 billion people the effects of an incremental change in humidity at the
(34% of the total population) will be at risk of dengue margins of the present distribution of dengue fever; this
fever in 2055 and 3·5 billion people (or 35% of the total) prediction should be almost unaffected by spatial
in 2085. With these same increases in population, but autocorrelation. A model that predicts the current
with changes in humidity projected by the HADCM2 distribution with high cross-validation accuracy should
global circulation model, we forecast an increase in the continue to perform well when there are incremental
population at risk of dengue fever to 4·1 billion people changes to humidity as a result of climate change.
(44%) by 2055 and 5·2 billion (52%) by 2085. Using four Future social trends might move in a favourable
alternative global circulation models, we estimated the direction, which would reduce the likelihood that disease
population at risk to be in the region of 5–6 billion by potential is translated into disease occurrence. However,
2085 (figure 2). there are other possible situations that might have
Model accuracy was improved, and spatial patterning of negative consequences.1 We have not attempted to
the residuals reduced, by inclusion of a variable indicating account for all possible economic and social futures, nor
the maximum of humidity within a specific radius in the do we claim to be able to predict the distribution of
neighbourhood of each grid cell. The most accurate dengue fever at the end of the 21st century. Instead we
model included local humidity, and the maximum have examined the relations between climate variability
humidity within a radius of ten grid squares, which is and dengue fever, independent of other variables. This is
equivalent to 5° latitude, and a variable representing vital information for policy makers who want to
statistical interaction between these two variables. This understand the potential effects of climate change, and
multivariate model achieved an overall accuracy of 92%, to set priorities for mitigation and adaptation. The
with a sensitivity of 85% and specificity of 93%. We resurgence of dengue fever and global climate change are
repeated risk calculations for population and climate driven by similar issues, which include excessive
projections for 2085 using one global circulation model resource consumption in rich countries, an increase in
(CGCMA2). The population at risk was shown to be social inequality, and population increases in poor
7 billion people (70% of the total). countries. Addressing these issues will help to reduce
dengue fever prevalence and global climate change in the
Discussion future.
Our findings confirm that the geographical limits of
dengue fever transmission are strongly determined by Contributors
climate. On the assumption that other factors affecting S Hales initiated the study and collected and analysed the data.
J Maindonald and A Woodward contributed to the study design. All
dengue fever transmission remain the same, we forecast authors contributed to the interpretation of results and writing the
that climate change will contribute to a substantial paper.
increase in the number of people and proportion of global
population at risk of dengue fever. The finding that the Conflict of interest statement
baseline distribution of dengue is well predicted on the None declared.
basis of reported vapour pressure is biologically credible.
Humidity, defined as vapour pressure or specific Acknowledgments
humidity, is high only where rainfall and temperatures are We thank Matthew Livermore for providing population projections,
Sari Kovats and Diarmid Campbell-Lendrum for helpful suggestions.
high, and these are conditions that are conducive to Simon Hales and Neil de Wet were supported by the Health Research
breeding and survival of vector populations, and rapid Council of New Zealand. The Health Research Council of New Zealand
replication of the virus.4 funded the study.

THE LANCET • Vol 360 • September 14, 2002 • www.thelancet.com 833

For personal use. Only reproduce with permission from The Lancet Publishing Group.
ARTICLES

References 11 Li Y. Global Population Distribution Database.

http://grid2.cr.usgs.gov/globalpop/1-degree/ (accessed March 13,
1 Intergovernmental Panel on Climate Change (IPCC). Climate 2002).
Change 2001: Impacts, adaptation and vulnerability. Contribution of
12 Bos E, Vu M, Massiah E, Bulatao R. World Population Projections
working group II to the third assessment report of the
1994–95: Estimates and projections with related demographic
intergovernmental panel on climate change. Cambridge: Cambridge:
statistics. New York: Johns Hopkins University Press, 1994.
University Press, 2001.
13 New M, Hulme M, Jones P. Representing twentieth-century space-
2 Committee on climate ecosystems infectious diseases and human
time climate variability. Part I: development of a 1961–90 mean
health. Under the weather. climate, ecosystems, and infectious disease.
monthly terrestrial climatology. J Climate 1999; 12: 829–56.
Washington: National Academy Press, 2001.
14 Intergovernmental Panel on Climate Change. Data Distribution
3 World Health Organisation. WHO report on global surveillance of
Centre. http://ipcc-ddc.cru.uea.ac.uk/index.html (accessed March 23,
epidemic-prone infectious diseases. Geneva: WHO, 2000.
2002).
4 Focks D, Haile D, Daniels E, Mount G. Dynamic life table model for
15 Carter T, Hulme M, Lal M. Guidelines on the use of scenario data for
Aedes aegypti (L) (Diptera: Culicidae). Analysis of the literature and
climate impact and adaptation assessment. Geneva: Intergovernmental
model development. J Med Entomol 1993; 30: 1003–17.
Panel on Climate Change, 1999.
5 Patz J, Martens W, Focks D, Jetten T. Dengue fever epidemic
16 Roeckner E, Arpe K, Bengtsson L, et al. The atmospheric general
potential as projected by general circulation models of global climate
circulation model ECHAM-4: model description and simulation of
change. Environ Health Perspect 1998; 106: 147–53.
present-day climate. Hamburg: Max-Planck Institute for Meteorology,
6 Martens W, Jetten T, Focks D. Sensitivity of malaria, 1996.
schistosomiasis and dengue to global warming. Climatic Change 1997;
17 Johns T, Carnell R, Crossley J, et al. The Second Hadley Centre
35: 145–56.
coupled ocean-atmosphere GCM: model description, spinup and
7 Martens P, Kovats R, Nijhof S, et al. Climate change and future validation. Climate Dynamics 1997; 13: 103–34.
populations at risk of malaria. Global Environ Change 1999; 9:
18 Emori S, Nozawa T, Abe-Ouchi A, Numaguti A, Kimoto M,
S89–S107.
Nakajima T Coupled ocean-atmosphere model experiments of future
8 Hales S, Weinstein P, Souares Y, Woodward A. El Nino and the climate change with an explicit representation of sulfate aerosol
dynamics of vector borne disease transmission. Environ Health Perspect scattering. J Meteorol Soc Jpn 1999; 77: 1299–307.
1999; 107: 99–102.
19 Flato G, Boer G. Warming asymmetry in climate change simulations.
9 Rogers D, Randolph S. The global spread of malaria in a future, Geophys Res Lett 2001; 28: 195–98.
warmer world. Science 2000; 289: 1763–66.
20 Flato G, Boer G, Lee W, et al. The Canadian Centre for Climate
10 Reiter P. Climate change and mosquito borne disease. Environ Health Modelling and Analysis global coupled model and its climate. Climate
Perspect 2001; 109 (suppl 1): 141–61. Dynamics 2000; 16: 451–67.

834 THE LANCET • Vol 360 • September 14, 2002 • www.thelancet.com

For personal use. Only reproduce with permission from The Lancet Publishing Group.