The Global Spread of Malaria in a Future, Warmer World

David J. Rogers1* Sarah E. Randolph2

The frequent warnings that global climate change will allow falciparum malaria to spread into northern latitudes, including Europe and large parts of the United States, are based on biological transmission models driven principally by temperature. These models were assessed for their value in predicting present, and therefore future, malaria distribution. In an alternative statistical approach, the recorded present-day global distribution of falciparum malaria was used to establish the current multivariate climatic constraints. These results were applied to future climate scenarios to predict future distributions, which showed remarkably few changes, even under the most extreme scenarios.

1 Trypanosomiasis and Land-use in Africa Research Group,
2 Oxford Tick Research Group, Department of Zoology, University of Oxford, South Parks Road, Oxford OX1 3PS, UK.
*   To whom correspondence should be addressed. E-mail: david.rogers@zoology.ox.ac.uk


Predictions of global climate change have stimulated forecasts that vector-borne diseases will spread into regions that are at present too cool for their persistence (1-5). For example, life-threatening cerebral malaria, caused by Plasmodium falciparum transmitted by anopheline mosquitoes, is predicted to reach the central or northern regions of Europe and large parts of North America (2, 4). falciparum malaria is the most severe form of the human disease, causing most of the ~1 million deaths worldwide among the ~273 million cases in 1998 (6). Despite these figures, the epidemiology of malaria, like many other vector-borne tropical diseases, remains inadequately understood. Only the most general of maps for its worldwide distribution are available (7), and its global transmission patterns cannot be modeled satisfactorily because crucial parameters and their relations with environmental factors have not yet been quantified. Most importantly, absolute mosquito abundance has not yet been related to multivariate climate.

Nevertheless, the problem of malaria has led to its being included in most predictions about the impact of climate change on the future distribution of vector-borne diseases (8). These studies, which draw on the forecasts of future climate from various global circulation models (GCMs) (9, 10), generally use only one or at most two climatic variables to make their predictions. Biological models for malaria distribution are based principally on the temperature dependence of mosquito blood-feeding intervals and longevity and the development period of the malaria parasite within the mosquito, each of which affects the rate of transmission (4, 11). Those models based on threshold values include a lower temperature threshold, below which all development of the malaria parasite ceases, and an upper limit of mosquito lethality (2). In addition, the suitability (or unsuitability) of habitats for these vectors, which require a minimum atmospheric moisture, is defined by the ratio of rainfall to potential evapotranspiration (2). The output of such models, therefore, represents predicted areas where parasite development within the vector is fast enough to be completed before the vector dies, bounded by limits imposed by habitat suitability (2). The fit of these predictions to the current global malaria situation shows noticeable mismatches in certain places (12); false predictions of presence (e.g., over the eastern half of the United States) are accounted for by past control measures or by "peculiar vector biogeography," whereas false predictions of absence are dismissed as model errors (2).

Refinements of these biological models (3-5) are based on modifications of an equation describing transmission potential, expressed as the basic reproduction number R0, which must equal at least 1 for disease persistence (13, 14). For an estimation of the correct value of R0 from which to predict malaria distribution, absolute, not relative, estimates of all quantities in the equation are needed. Instead, by omitting certain unquantified but important parameters and rearranging the equation (15), a relative measure of "epidemic potential" (EP) [now "transmission potential" (5)] has been derived as the reciprocal of the vector/host ratio required for disease persistence. This predicts a more extensive present-day distribution of malaria than is currently observed (12). The ratio of future EP to present EP is then presented as indicating the relative degree of the future risk of malaria, but this is an inappropriate measure of changing risk because a high ratio may still leave R0 < 1.

Until such biological approaches can give accurate descriptions of the current situation of global malaria, they cannot be used to give reliable predictions about the future. Instead, an alternative two-step statistical approach to mapping vector-borne diseases gave a better description of the present global distribution of falciparum malaria and predicted remarkably few future changes, even under the most extreme scenarios of climate change. First, the present-day distribution was used to establish the climatic constraints currently operating on malaria. Then, the results were applied to future GCM scenarios to predict future distribution. Simple maximum likelihood methods were used (16) (Fig. 1A), based on the mean, maximum, and minimum of three climatic variables: temperature, precipitation, and saturation vapor pressure. The match between prediction and reality was significantly closer than that achieved by previous models (12). Some false-positive areas, in eastern South America and Iran, were recorded as of "limited risk" on earlier maps (17), whereas others, in the southern United States and northern Australia, coincided with successful vector control campaigns. Because these predictions were based on present-day malaria maps, the disappearance of malaria in historical times from the edges of its global distribution has effectively been incorporated (18). This itself is a reflection of climatic conditions. In cooler regions, where mosquito life-spans barely exceed extrinsic incubation periods, transmission cycles are inherently more fragile. Not only the range of each climatic variable, but also the covariation between variables proved to be important in setting distributional limits in this model. Biologically, this implies that organisms can cope with extremes of some variables (e.g., temperature) only if others (e.g., humidity) are at certain levels.


Fig. 1. (A) Current global map of malaria caused by P. falciparum [yellow hatching, data from (7)] and distribution predicted with maximum likelihood methods (red through green posterior probability scale in key; light blue areas indicate no prediction, i.e., conditions very different from those in any of the sites used to train the analysis). These methods give predictions that are 78% correct, with 14% false-positives and 8% false-negatives (12). (B) Discriminating criteria derived from the current situation were then applied to the equivalent climate surfaces from the high scenario from the HadCM2 experiment (19) which predicts mean global land surface changes of +3.45, +3.63, and +3.29C in mean, minimum, and maximum temperatures, respectively; +1.87 hPa for SVP; and +0.127 mm/day for precipitation by the year 2050. The yellow hatching and the probability scale are the same as in (A). (C) The difference between the predicted distributions given in (A) and (B), showing areas where malaria is predicted to disappear (i.e., probability of occurrence decreases from >0.5 to <0.5) (in red) or invade (i.e., probability of occurrence increases from <0.5 to >0.5) (in green) by the 2050s in relation to the present situation. The gray hatching is the current global malaria map shown in yellow hatching in (A). [View Larger Version of this Image (67K GIF file)]


The results from this first step were applied to the most widely used GCM scenario of the future, which envisages a 1% annual compound increase in overall greenhouse gas concentrations (9, 19) and a range of climate sensitivity to this increase. Predictions show the future distribution of habitats similar to those where falciparum malaria occurs today (Fig. 1B). If introduced, by travel or trading activities, for example, both vectors and parasites could survive in such places. Only a relatively small extension was predicted as compared to the present-day situation: northward into the southern United States and into Turkey, Turkmenistan, and Uzbekistan; southward in Brazil; and westward in China. In other areas, malaria was predicted to diminish (Fig. 1, B and C). The net effect of this on the potential exposure of humans to malaria by the year 2050, compared with the present as modeled in Fig. 1A (20), varied with climate sensitivity to greenhouse gases; for example, there was an increase in exposure of 23 million people (+0.84%) under the HadCM2 "medium-high" scenario (19) or a decrease in exposure of 25 million people (-0.92%) in the HadCM2 "high" scenario (i.e., higher mean temperatures) (Fig. 1C). These changes are modest because covariates limit potential expansion along certain dimensions of environmental space. For example, in the present exercise, a univariate model driven by the minimum of the mean temperature (the single most important predictor in the multivariate fit) would predict more extensive malaria than at present along the southern fringes of the Sahara Desert and an expansion northward into the Sahara, as global warming lifts the cold minimum (nighttime) temperature constraint on mosquito or malaria development. Multivariate models gave more accurate predictions of the present situation and do not predict this expansion, because of the limitations imposed by the covarying rainfall or moisture variables.

Whereas others have raised qualitative doubts about the predicted impact of climate change on malaria (18), the quantitative model presented here contradicts prevailing forecasts of global malaria expansion. It highlights the use of multivariate rather than univariate constraints in such applications and the advantage of statistical rather than biological approaches in situations where biological knowledge is incomplete. Whatever the method adopted, the usefulness of GCMs as a basis for making predictions about the future of biological systems needs further clarification. The current low spatial resolution of such models hides considerable local variation and represents mean conditions across large geographical areas that may not occur in many places within them. Furthermore, the accuracy of GCMs in predicting the covariation of climatic variables, to which biological systems are very sensitive, is unknown.

REFERENCES AND NOTES

  1. Y. Matsuoka and K. Kai, J. Global Environ. Eng. 1, 1 (1994) ; W. C. Reeves, J. L. Hardy, W. K. Reisen, M. M. Milby, J. Med. Entomol. 31, 323 (1994) [ISI] [Medline]; W. J. M. Martens, J. Rotmans, L. W. Niessen, "Climate change and malaria risk: An integrated modelling approach," Rep. 461502003 (Dutch National Institute of Public Health and the Environment, Bilthoven, Netherlands, 1994); S. W. Lindsay and W. J. M. Martens, Bull. WHO 76, 33 (1998) [ISI] [Medline]; J. A. Patz, W. J. M. Martens, D. A. Focks, T. H. Jetten, Environ. Health Perspect. 106, 147 (1998) [ISI] [Medline]; J. A. Patz and S. W. Lindsay, Curr. Opin. Microbiol. 2, 445 (1999) [ISI].
  2. M. H. Martin and M. G. Lefebvre, Ambio 24, 200 (1995) .
  3. W. J. M. Martens, L. W. Niessen, J. Rotmans, T. H. Jetten, A. J. McMichael, Environ. Health Perspect. 103, 458 (1995) [ISI] [Medline]; W. J. M. Martens, T. H. Jetten, J. Rotmans, L. W. Niessen, Global Environ. Change 5, 195 (1995) [ISI]; W. J. M. Martens, T. H. Jetten, D. A. Focks, Clim. Change 35, 145 (1997) [ISI].
  4. P. Martens, Health and Climate Change: Modelling the Impacts of Global Warming and Ozone Depletion, (Earthscan, London, ed. 1, 1998).
  5. W. J. M. Martens, et al., Global Environ. Change 9, S89 (1999) .
  6. B. Schwartlander, Lancet 350, 141 (1997) [ISI] [Medline]; R. W. Snow, M. H. Craig, U. Deichmann, D. le Sueur, Parasitol. Today 15, 99 (1999) [ISI]; The World Health Report 1999: Making a Difference [World Health Organization (WHO), Geneva, 1999].
  7. WHO, Wkly Epidemiol. Rec. 72, 285 (1997).
  8. A. J. McMichael, J. Patz, R. S. Kovats, Br. Med. Bull. 54, 475 (1998) [ISI] [Medline]; A. Haines and A. J. McMichael, Eds., Climate Change and Human Health (Royal Society, London, 1999).
  9. T. C. Johns, et al., Clim. Dyn. 13, 103 (1997) [ISI].
  10. M. New, M. Hulme, P. D. Jones, J. Clim. 12, 829 (1998) .
  11. C. Garrett-Jones, Bull. WHO 30, 241 (1964) .
  12. The recorded global limits of falciparum malaria (7) were digitized and turned into a raster grid with a longitude and latitude resolution of 0.5. Predictive maps of seasonal or perennial falciparum malaria distribution (2, 4) were scanned at the same spatial resolution. Five hundred points of presence and 500 points of absence within 10 of longitude and latitude of the nearest presence areas were selected at random from the WHO map; 20 independent sets of these points were used to assess the accuracy of the map predictions. For (2), mean accuracies were 75.79% (95% confidence limits  0.656) correct, with 13.16% (0.420) false-positives (i.e., false predictions of presence) and 11.04% (0.368) false-negatives (i.e., false predictions of absence); the index of agreement (21), kappa , was equal to 0.516 (0.013). For (4), mean accuracies were 67.26% (0.607) correct, with 18.63% (0.553) false-positives and 14.11% (0.352) false-negatives; kappa  = 0.345 (0.012). Comparable results for the predicted map shown in Fig. 1A are 77.71% (0.673) correct, with 13.98% (0.418) false-positives and 8.31% (0.368) false-negatives; kappa  = 0.554 (0.013). These results are significantly better than those for (2) and (4) [Student's t test for the difference between the mean percent correct = 4.11, P < 0.01; and 23.26, P < 0.001 for the comparisons with (2) and (4), respectively].
  13. R. M. Anderson and R. M. May, Infectious Diseases of Humans: Dynamics and Control (Oxford Univ. Press, Oxford, 1991).
  14. For the malaria parasite, R0 may be calculated from (13)
    R<SUB>0</SUB>&thinsp;=&thinsp;<FR><NU>a<SUP>2</SUP>bcmp<SUP>T</SUP></NU><DE>(&minus;<UP>ln<SUB>e</SUB></UP>&thinsp;p)r</DE></FR> (1)

    where a is the mosquito daily biting rate on humans; b and c are the probabilities of transmission of parasites from infected vertebrate to uninfected vector and from infected vector to uninfected vertebrate, respectively; m is the ratio of vectors to vertebrates; p is the vector daily survival rate; T is the incubation period (in days) of parasites in vectors (thus, pT is the proportion of infected vectors that survive to become infectious); and r is the daily rate of recovery of the vertebrate from infectiousness.

  15. Spatial variation in the relative risk of infection has been modeled (3-5) by rearranging the R0 equation to predict the vector/host ratio, m, required for disease persistence
    m&thinsp;=&thinsp;<FR><NU>(&minus;<UP>ln<SUB>e</SUB></UP>&thinsp;p)r</NU><DE>a<SUP>2</SUP>bcp<SUP>T</SUP></DE></FR><UP>&ensp;for&ensp;</UP>R<SUB>0</SUB><UP>&thinsp;=&thinsp;1</UP> (2)

    The reciprocal of m, the EP [now the transmission potential (5)], is taken as a direct measure of the favorability of conditions for disease persistence. In applying this equation to climate change scenarios, parameters r, b, and c, which have not been adequately quantified, are effectively omitted by setting them equal to 1.0, thereby turning absolute threshold conditions for transmission into a relative measure. The omitted quantities are then canceled out by deriving a ratio of future EP to present EP, ignoring areas where the present EP is less than 0.01 to avoid excessively high ratios.

  16. The analysis randomly selected a training set of 1500 points within the mapped limits of falciparum malaria (7) and 1500 points outside the limits, but within 10 of longitude and latitude. Data for each point were derived from 30-year (1960-90) average monthly climate surfaces (10) for the mean (TM), maximum (TX), and minimum (TN) temperature; rainfall (R); and saturation vapor pressure (SVP) variables. These surfaces were preprocessed by temporal Fourier analysis (22) of the monthly data (essentially smoothing the data), from which the mean, maximum, and minimum for each variable were extracted for each training set location. Data were first clustered by using the "k-means cluster" option of SPSSv.9 (SPSS, Chicago, IL), producing three clusters each for presence (p) and absence (a) sites. The means and covariances of the six resulting clusters are available at www.sciencemag.org/feature/data/1050940.shl. Experience has shown that clustering improves the accuracy of predictions because different parts of a global distribution often have different covariances between critical variables; clustering essentially allows for nonlinear responses of biological systems to gradual changes in climatic variables. Stepwise discriminant analysis (22) of the resulting six clusters, using the criterion of maximizing the Mahalanobis distance between all pairs of "dissimilar" (i.e., p to a and a to p) clusters, chose minimum TM, minimum R, minimum SVP, mean SVP, and mean TX as the five most important variables for distinguishing areas of malaria presence and absence. These variables were used to generate the maximum likelihood predictions in terms of posterior probabilities (Fig. 1, A and B). For comparisons with other malaria maps, a threshold prediction probability of 0.5 was taken as distinguishing absence and presence.
  17. "Geographical distribution of arthropod-borne diseases and their principal vectors," Rep. WHO/VBC/89.967 (Vector Biology and Control Division, WHO, Geneva, 1989).
  18. G. Taubes, Science 278, 1004 (1997) [ISI] [Full Text]; P. Reiter, Emerg. Infect. Dis. 6, 1 (2000) [ISI] [Medline] (available at www.cdc.gov/ncidod/eid/vol6no1/reiter.htm).
  19. The chosen future scenario comes from the UK Hadley Centre for Climate Prediction and Research, which is available from the Intergovernmental Panel on Climate Change (IPCC) Data Distribution Centre (http://ipcc-ddc.cru.uea.ac.uk). In 1992, the IPCC defined six alternative scenarios, named IS92a to IS92f. IS92a, which involves a 1% per year compound increase in overall greenhouse gas concentrations, is widely used in impact studies and predicts an increase in total atmospheric CO2 from 7.4 Gt of C in 1990 to 14.52 Gt of C in 2050 and 20.28 Gt of C in 2100. The HadCM2 medium-high and high scenarios represent different climate sensitivities to this level of gas emission (available at www.met-office.gov.uk/sec5/CR_div/Brochure97/).Although some uncertainties exist about these climatic responses (23), the medium-high scenario commonly forms the basis of current attempts to predict the impact of climate change on human health. Outputs of the medium-high scenario are the average of four separate GCM runs and are given as differences between the modeled present and modeled future conditions; the high-scenario outputs are scaled versions of the medium-high outputs (23). Following usual practice, the GCM differences were added to the observed 30-year climatic means (after cubic-spline interpolation to the same spatial resolution), to generate the predicted future climate surfaces that were used in the present analysis.
  20. The "Gridded Population of the World" unsmoothed population density data file created by the Socioeconomic Data and Applications Center at Columbia University (Palisades, NY) was obtained from the Center for International Earth Science Information Network at ftp://ftp.ciesin.org/pub/data/Grid_Pop_World. This record of the 1994 human population density per square kilometer was turned into a raster image at 1/12 spatial resolution and was subsequently used to estimate the total human population within the malarious areas shown in Fig. 1, A through C, allowing for the different land areas corresponding to pixels at different latitudes. Land pixels in the malaria map imagery were mapped onto their equivalent 6 by 6 grid in the population density imagery, from which population totals were extracted and summed. This method estimated a total global population of 5611 million people in 1994, of which 2727 million lived within the predicted malarious areas of Fig. 1A. Under the medium-high scenario, 357 million people live within areas that are currently malaria-free but are predicted to become malarious by 2050, and 334 million live within currently malarious areas that are predicted to become unsuitable by 2050, a net increase of 23 million, or +0.84% on the 1994 baseline population data. For the high scenario, the corresponding figures are 389 million, 414 million, and a net decrease of 25 million or -0.92%, respectively (Fig. 1C).
  21. Z. K. Ma and R. L. Redmond, Photogramm. Eng. Remote Sensing 61, 435 (1995) [ISI].
  22. D. J. Rogers, S. I. Hay, M. J. Packer, Ann. Trop. Med. Parasitol. 90, 225 (1996) [ISI] [Medline].
  23. M. Hulme and G. J. Jenkins, "Climate change scenarios for the UK: Scientific report" (Climatic Research Unit, Norwich, UK, 1998).
  24. We thank the Department for International Development (grant R6626 to D.J.R.) and the Wellcome Trust (S.E.R.) for financial support and G. B. White and S. I. Hay for helpful comments.


Volume 289, Number 5485, Issue of 8 Sep 2000, pp. 1763-1766.
Copyright 2000 by The American Association for the Advancement of Science.