Research ArticleDengue

Mapping global variation in dengue transmission intensity

See allHide authors and affiliations

Science Translational Medicine  29 Jan 2020:
Vol. 12, Issue 528, eaax4144
DOI: 10.1126/scitranslmed.aax4144

At last, an atlas of dengue spread

Transmitted by Aedes mosquitos worldwide, dengue virus causes flu-like symptoms and, in severe cases, hemorrhaging and potentially death. In this Research Resource, Cattarino et al. present a downloadable high-resolution map of the global variation in dengue transmission intensity. The authors developed this map by fitting geospatial serology and surveillance data to force of infection estimates derived from environmental and demographic predictors. Using this data to evaluate potential dengue control strategies, they predicted that Wolbachia-based intervention may be more likely than the Sanofi-Pasteur vaccine to reduce the global disease burden.


Intervention planning for dengue requires reliable estimates of dengue transmission intensity. However, current maps of dengue risk provide estimates of disease burden or the boundaries of endemicity rather than transmission intensity. We therefore developed a global high-resolution map of dengue transmission intensity by fitting environmentally driven geospatial models to geolocated force of infection estimates derived from cross-sectional serological surveys and routine case surveillance data. We assessed the impact of interventions on dengue transmission and disease using Wolbachia-infected mosquitoes and the Sanofi-Pasteur vaccine as specific examples. We predicted high transmission intensity in all continents straddling the tropics, with hot spots in South America (Colombia, Venezuela, and Brazil), Africa (western and central African countries), and Southeast Asia (Thailand, Indonesia, and the Philippines). We estimated that 105 [95% confidence interval (CI), 95 to 114] million dengue infections occur each year with 51 (95% CI, 32 to 66) million febrile disease cases. Our analysis suggests that transmission-blocking interventions such as Wolbachia, even at intermediate efficacy (50% transmission reduction), might reduce global annual disease incidence by up to 90%. The Sanofi-Pasteur vaccine, targeting only seropositive recipients, might reduce global annual disease incidence by 20 to 30%, with the greatest impact in high-transmission settings. The transmission intensity map presented here, and made available for download, may help further assessment of the impact of dengue control interventions and prioritization of global public health efforts.


Dengue is an acute viral infection transmitted between humans by Aedes mosquitos. The virus is responsible for a substantial burden of disease across the tropics and subtropics (1). Secondary infections with one of the four dengue serotypes (DENV1–4) are on average more severe than primary infections, an observation thought to be explained by heterotypic antibody-dependent enhancement (2). Dengue prevention options have historically been limited and largely restricted to insecticide-based mosquito population suppression (3), but new opportunities are now offered by the first licensed dengue vaccine (4, 5) and vector control measures such as Wolbachia (6). However, these new interventions have imperfect efficacy, meaning their impact will depend on local dengue transmission intensity (7, 8) as routinely quantified by the force of infection (FOI; per capita rate at which susceptible individuals become infected with a pathogen) or the reproduction number (average number of secondary cases resulting from the introduction of a single infectious individual into a susceptible population). Given that dengue transmission intensity is highly spatiotemporally variable and this variation is poorly characterized for much of the world, optimal targeting of interventions against dengue is a major public health challenge.

Modern geostatistical methods provide powerful tools for characterizing the relationships between geolocated outcome variables of interest and possible environmental, social, or demographic drivers. Their application to vector-borne diseases has grown in popularity, but for dengue has hitherto been restricted to modeling the local probability of reported cases (or occurrence) of dengue rather than local transmission intensity (9). Probability of occurrence is well suited to delineating the boundaries of endemicity of dengue, but not for distinguishing between different degrees of transmission intensity within endemic regions. However, predictive modeling of dengue transmission intensity is particularly challenging as individuals develop lifelong homotypic immunity upon infection with any of the serotypes, which limits the maximum number of dengue infections an individual can have to four. This causes incidence (and disease burden) to saturate at high transmission intensities, limiting our ability to infer the latter from the former (10, 11).

Lack of a fine-scale geographical characterization of dengue transmission intensity has hindered assessment of the impact of candidate control strategies such as Wolbachia and vaccination. To date, analyses of intervention impact have been restricted to single geographic settings or have provided impact estimates for a range of transmission intensities without mapping those onto specific geographies (5, 8). Here, we quantify dengue transmission intensity globally at high spatial resolution and project the likely impact of control measures on dengue transmission and disease burden, drawing on examples from a recent vector control strategy, namely, the release of Wolbachia-infected mosquitoes and application of the first licensed dengue vaccine.


Global predictions of dengue FOI

Under the hypothesis that climate drives vector carrying capacity and competence, we fitted a random forest model to a dataset of first administrative unit-level geolocated estimates of the average FOI of dengue (Fig. 1), derived from age-stratified seroprevalence or case notification data (tables S1 and S2), using a set of environmental explanatory variables (figs. S1 to S8 and table S3) (1114). The resulting FOI map predicted areas of high transmission in all continents straddling the tropics: South America (Colombia, Venezuela, and Northeastern Brazil), Africa (western and central African countries), Southeast Asia (Thailand, Indonesia, and the Philippines), and Australasia (Papua New Guinea) (Fig. 2A). Our model had high in-sample predictive performance as indicated by the strong correlation between data and predictions (coefficient of determination, R2 = 0.99) (tables S4 and S5 and figs. S9 and S10; raw data from these and other figures are in data file S1). Out-of-sample predictive performance for a complete random validation set was good (R2 = 0.75). As expected, out-of-sample accuracy decreased (R2 = 0.69) when making predictions for locations distant (about 500 km) from data points included in the training set. Our predictions were robust to the number of explanatory variables used, as demonstrated by the maps generated with 16 and 25 predictors, which showed similar patterns (figs. S11 to S14).

Fig. 1 Geographic location of FOI data.

A total of 382 data points are shown comprising the national administrative division of all countries and the second national administrative division (admin 2) of Mexico, Colombia, Venezuela, Brazil, India, and Australia. Color scale shows FOI value (average per-serotype).

Fig. 2 Predicted global dengue risk.

Means (A) and SDs (B) of FOI estimates in dengue endemic countries across 200 geographically stratified bootstrap samples.

Mean nocturnal and diurnal temperature and their annual seasonal variations were among the most highly explanatory variables (figs. S15 and S16). Biannual seasonal variations in diurnal temperature and annual and biannual seasonal variations in precipitation, middle infrared reflectance (mean, annual, and biannual seasonal variations), birth rate, and population density also contributed to the model fit. After accounting for the effect of other variables, elevation had a minor effect.

Model fit was poorest for the highest transmission intensity locations, likely due to the limited numbers of data points for these settings. Uncertainty around our mean estimates was spatially heterogeneous, with southeast Brazil and East Africa among the areas with the highest variability in predictions (Fig. 2B), reflecting the sparseness of data in these regions. Uncertainty was also high in northern Thailand due to relatively high variation in estimated FOI in the underlying data points used for model fitting in that area.

Predictions of global dengue burden

The global FOI map produced by our model also allowed us to estimate mean annual dengue infection and disease incidence. We estimate that there are an average of 105 [95% confidence interval (CI), 95 to 114] million dengue infections globally per year, 51 (95% CI, 32 to 66) million febrile disease cases, and 4 (95% CI, 2 to 5) million symptomatic infections, which might require hospitalization (Table 1). Most burden of disease (58%) was concentrated in South and Southeast Asia (India, Bangladesh, Indonesia, the Philippines, Thailand, and Vietnam), with half of the burden in the region occurring in India (data file S2). Sub-Saharan Africa carried almost 26% of the global dengue burden, with hot spots in Central and Eastern Africa (Nigeria, the Democratic Republic of the Congo, and Ethiopia). Latin America had 16% of the global burden, mostly occurring in Brazil, Mexico, Colombia, and Venezuela.

Table 1 Estimated number of dengue infections, febrile dengue cases, and potentially hospitalized cases by continent.

Values are quoted to the nearest 10,000.

View this table:

Global estimates of the reproduction number of dengue

Under the simplifying assumption that dengue dynamics have equilibrated in each region, we translated our FOI predictions into estimates of the average basic reproduction number (R0) of dengue (figs. S17 and S18). Given the lack of definitive data on infectiousness differences between serotypes, we estimated R0 under two alternative assumptions about the relative infectiousness of sequential number of dengue infections (with any serotype): that primary to quaternary infections have the same infectiousness or that only primary and secondary infections contribute to transmission (table S6). The former assumption gives optimistic (lower) estimates of R0, and the latter, pessimistic (higher) estimates. Geographic heterogeneity in R0 predictions largely mirrored the variation seen in FOI, albeit modified by country-to-country variation in demography.

Projected impact of interventions

We used the resulting R0 maps to predict the potential impact of interventions on dengue transmission and disease burden. We first considered an intervention that reduces transmission by a fixed proportion (modeled as a multiplicative reduction of R0). Interventions such as the release of Wolbachia-infected mosquitoes have been predicted to act on transmission in this manner (7).

The predicted long-term “best-case” impact of such a transmission-reducing control measure varied with the infectiousness scenario assumed, but even with the scenario producing the highest R0 estimates (namely, that only primary and secondary infections contribute to transmission), we predicted that an intervention capable of reducing R0 by 80% in theory could come close to effectively eliminating dengue globally (Fig. 3 and figs. S19 to S21), although note the caveats to this result highlighted in Discussion. For the other infectiousness scenario, a 60% reduction in R0 was predicted to achieve a similar impact (Fig. 3). Even an intervention with intermediate efficacy (50% reduction) was predicted to reduce global case incidence by at least 70% (95% CI, 68 to 77) in the more pessimistic infectiousness scenario. To put these results into context, successful release of Aedes aegypti mosquitos carrying the wMel strain of Wolbachia has been predicted on the basis of laboratory studies to achieve a 70% reduction in R0 (7).

Fig. 3 Predicted impact of a transmission-reduction intervention.

Mean (solid lines) and 95% CI (envelopes) of percentage reduction relative to no-intervention scenario of global number of dengue mild febrile cases (A), absolute number of mild febrile cases (B), and number of countries where average R0 is reduced below 1 (C), for different percentage amounts of R0 reduction and serotype infectiousness assumptions. 4S, all infections are infectious; 2S, primary and secondary infections are infectious. The vertical dotted lines indicate 30 and 70% reductions in transmission intensity.

For lower levels of intervention effectiveness, impacts were predicted to be highly spatially heterogeneous, with the least impact seen in the highest transmission intensity areas (Fig. 4). This highlights one of the potential caveats to these predictions of potential intervention impact: local hot spots of dengue transmission intensity not able to be resolved at the spatial resolution of our analysis may see persistence of transmission even at higher intervention efficacies.

Fig. 4 Geographic variation in the predicted impact of a transmission-reduction intervention.

Maps show the mean percentage reduction in incidence of febrile dengue cases at level 1 administrative units, across 200 bootstrap samples, for 30% transmission reduction and 4S serotype infectiousness assumption (A), 70% transmission reduction and 4S serotype infectiousness assumption (B), 30% transmission reduction and 2S serotype infectiousness assumption (C), and 70% transmission reduction and 2S serotype infectiousness assumption (D).

We also explored the impact of introducing childhood vaccination with the Sanofi-Pasteur dengue vaccine (CYD-TDV), a recombinant chimeric live-attenuated vaccine that has been licensed for use in 20 countries. The vaccine, which has a complex efficacy profile (8, 15), acts on reducing the risk of symptomatic disease rather than transmission. On the basis of the results of a published transmission dynamic model previously used to analyze the Sanofi vaccine trial data (8, 16), we examined the impact of a single round of screen-and-vaccinate [that is, only vaccinating seropositive individuals, as currently recommended by World Health Organization (WHO) (17)] per birth cohort targeting 9-year-olds. We assumed 80% policy coverage and that the diagnostic test for seropositivity has 90% sensitivity and 95% specificity, meaning that 72% of eligible seropositive individuals and 4% of seronegative individuals who falsely tested positive were assumed to receive the vaccine. We predicted that this policy would lead to reductions of 11 to 31% in symptomatic dengue cases (Fig. 5) and 15 to 39% in hospitalized dengue (figs. S22 to S24) over the first 30 years, depending on the infectiousness scenario assumed. Targeting 16-year-olds was predicted to give a slightly larger effect, as individuals from this age group are more likely to be seropositive in low to moderate transmission intensity settings and, thus, more likely to receive the vaccine. Tuning the age of vaccination to be optimal for the intensity of transmission seen at the local level 1 administrative unit is expected to have a slightly greater effect (Fig. 5 and fig. S24). Given the Sanofi vaccine can increase the risk of hospitalized dengue in seronegative recipients and that no test for seropositivity is 100% specific, minimizing individual harm requires vaccinating at an age where seroprevalence is expected to be high. Maps of the proportion of 9- and 16-year-olds expected to be seronegative (and thus at risk of a false-positive screening test) show the potential magnitude of this issue (figs. S25 and S26). For each infectiousness scenario, geographic variation in the predicted impact of vaccination was lower than seen for a pure transmission-reducing intervention (Fig. 6 versus Fig. 4). Unlike previous predictions for the blanket use of the vaccine without serological testing (16), our results (Fig. 6) suggest that vaccine use, in conjunction with a serological test, would have a slightly larger relative impact on dengue incidence in low-transmission settings than in higher transmission settings.

Fig. 5 Predicted impact of the Sanofi dengue vaccine.

Effect of child vaccination with the Sanofi dengue vaccine on the global incidence of febrile dengue cases over the first 30 years after vaccine introduction, for different vaccine screening ages and serotype infectiousness assumptions (4S, all infections are infectious; 2S, primary and secondary infections are infectious). Bar represents mean percentage reduction relative to no-intervention scenario, with 95% CI shown.

Fig. 6 Geographic variation in the predicted impact of the Sanofi dengue vaccine.

Maps show the mean percentage reduction in incidence of febrile dengue cases in the first 30 years after vaccine introduction, for vaccination of 9-year-olds and 4S serotype infectiousness assumption (A), vaccination of 16-year-olds and 4S serotype infectiousness assumption (B), optimal choice of vaccination age and 4S serotype infectiousness assumption (C), vaccination of 9-year-olds and 2S serotype infectiousness assumption (D), vaccination of 16-year-olds and 2S serotype infectiousness assumption (E), and optimal choice of vaccination age and 2S serotype infectiousness assumption (F).

It is also notable that the impact of vaccination was substantially larger in the higher-R0 infectiousness scenario (where only primary and secondary infections contribute to transmission). This is because vaccination had a much larger impact on transmission in this scenario, as all breakthrough infections in seropositive recipients were assumed not to be infectious (given the assumed effect of the vaccine is to make secondary natural infections “tertiary-like”).


This study characterized geographic variation in dengue transmission intensity by using environmental and demographic routine predictors to fit a machine-learning model to FOI estimates derived from serological surveys and routine surveillance data. Our finding that temperature and precipitation, as well as their annual and multiannual variation, are key predictors of dengue transmission variations is in agreement with previous evidence on the environmental determinants of both A. aegypti ecology and dengue transmission (1820). We also found that human population density contributed to the model predictive capacity, explaining why dengue transmission seems to be higher in urban environments.

By mapping annual average FOI, we were able to generate “bottom-up” estimates of dengue disease burden. Our estimates of the burden of symptomatic disease are relatively consistent with recent studies (1), but lower than a notable earlier study (9). However, our estimates of the number of infections are over threefold lower than the estimates from the latter study (9). This is because our approach explicitly takes into account that dengue is an immunizing disease, meaning that incidence is constrained by a biologically realistic cap of a maximum of four lifetime infections per individual. In agreement with past work (1, 9), we found that sub-Saharan Africa accounts for a disproportionally high proportion (27%) of the global disease burden due to the high birth rate of this region. However, dengue surveillance is traditionally weak in this region and few serological surveys have been undertaken, meaning that our current estimates are based on very limited data. More data would therefore be valuable in allowing current uncertainty in our transmission intensity and disease burden estimates to be reduced.

Only a minority of clinically apparent dengue cases seek health care in many settings (21, 22). However, since the estimates of FOI we use here are derived from the age-specific trends in serology and case notification data, not absolute case incidence, they are robust to underascertainment of incidence so long as this does not vary substantially with age.

A unique benefit of our approach to understanding global variation in dengue transmission and disease burden is that it allows the potential public health impact of interventions to be assessed. Like malaria, the vector-borne nature of dengue leads to much larger geographic variation in R0 than is typical of most directly transmitted infections (10). Hence, results from a randomized control study conducted in one setting cannot be directly and easily extrapolated to other settings. A mechanistic, model-based understanding of the impact of an intervention on transmission is required, together with reliable estimates of transmission intensity across the range of endemic settings where that intervention might be considered. It is the latter gap this current study fills. We take existing models and associated effect size estimates for generic transmission-reduction interventions and project the potential impact of these—and two existing interventions (the Sanofi-Pasteur vaccine and the release of Wolbachia-infected mosquitoes) (7, 8, 16)—across the entire region of dengue endemicity.

We find that there is geographic variation in the predicted impact of both transmission-reduction interventions and vaccination, driven by the spatial variation in local transmission intensity and by the mechanism of impact of these interventions. Both classes of interventions show larger effects in lower transmission intensity settings, although it should be noted that in the case of the vaccine, this is a result of modeling a policy of prevaccination testing of potential recipients for seropositivity with a very high-specificity diagnostic test (required when used in settings where a higher proportion of potential recipients are expected to be seronegative). Previous predictions for the blanket use of the vaccine without serological testing (16) showed the opposite trend, namely, larger impact in high-transmission settings. It should also be noted that the predictions of vaccine impact presented here are of overall population impact; vaccination may still enhance the risk of hospitalized dengue in the small proportion of seronegative individuals who receive the vaccine due to falsely testing positive as a result of imperfect test specificity (15, 17).

Interventions that reduce R0 by a fixed proportion [as Wolbachia has been predicted to do (7)] can achieve disease elimination in low- to moderate-transmission settings (by reducing R0 to below the threshold of 1 for self-sustaining transmission), but have a smaller impact in the highest transmission settings. It should be noted that even in very high transmission intensity settings where an intervention such as Wolbachia may fail to reduce R0 to below 1, high baseline herd immunity means that introduction of such an intervention is predicted to stop dengue transmission for a decade or more until that herd immunity declines (due to new births into the population) sufficiently to allow low-level transmission to resume (6).

Caution is needed in evaluating our estimates of the impact of transmission-reduction interventions as predictive of what release of Wolbachia-infected mosquitoes might achieve. Field-based evaluation of Wolbachia is ongoing (6, 23) and in the absence of results from randomized control trials, we have relied on effect size estimates derived from laboratory studies. In addition, we have assumed that the effect size will be uniform across settings, that estimates assume that the efficacy of Wolbachia on reducing dengue transmission shown in laboratory experiments will be maintained at the population level in the field and will have a uniform effect across settings. Neither of these assumptions have yet been tested. Once data from intervention studies become available, more precise estimates of the potential global public health impact of Wolbachia will be able to be derived.

An important additional limitation of our analysis is that we do not consider temporal variation in dengue incidence but rather focus on characterizing long-term average transmission intensity and disease burden. Dengue incidence exhibits high interannual variability driven by climate (for example, the El Niño–Southern Oscillation) (20), immune-mediated serotype dynamics (24), and virus genotype–specific phenotypic variation (25). Of particular importance in assessing the impact of control measures is that seasonal peak R0 values may be substantially larger than the annual average values presented here. In highly seasonal settings, our transmission-reduction intervention impact predictions are approximations, as they underestimate intervention impact in seasonal troughs but overestimate it during transmission intensity peaks.

A further limitation affecting our estimates of R0 and the impact of interventions is that we have had to assume dengue transmission dynamics are at quasi-equilibrium in the sense of a balance having been achieved between herd immunity, infection incidence, and new births. In areas only relatively recently invaded by one or more dengue serotypes [for example, Latin America (26)], this assumption may only hold to a first approximation. However, accurately characterizing long-term trends in dengue transmission requires incidence or seroprevalence data spanning decades—data that are available in relatively few settings (2729).

In addition, our analysis makes use of FOI estimates available at a relatively coarse spatial scale (administrative level 1). Although we used a spatial disaggregation expectation-maximization algorithm to address this, finer-scale raw input data would permit fuller characterization of local spatial heterogeneity in dengue transmission intensity and might reveal higher transmission intensity hot spots where interventions might have a more limited impact than predicted here.

Last, we have used machine learning–based regression to derive predicted maps of transmission intensity. Although this gives some insight into the relationships between dengue transmission and environmental predictors such as temperature, rainfall, and vector abundance and competence, a more mechanistic approach would offer deeper scientific insight (23, 30).

Although these limitations are priorities for ongoing research, the results presented here represent the most detailed characterization of geographic heterogeneity in dengue transmission intensity yet undertaken. By using age-stratified data sources and accounting for the immunizing effect of dengue, our transmission intensity map and burden estimates represent a methodological advance, most notably in allowing the effect of interventions on disease burden to be explored. Our analysis has critical relevance for future evidence-based prioritization, planning, and implementation of vaccine- and vector-control–based control policies.


Study design

We collated a dataset of geolocated estimates of per-serotype dengue FOI, derived from the analysis of age-stratified seroprevalence and case notification data. FOI is the per capita rate at which susceptible individuals acquire infection and is a key measure of infectious disease transmission. The resulting 382 georeferenced point estimates are publicly available (31). Additional method details are available in the Supplementary Materials.

Serology-based FOI estimates. We sourced 33 FOI estimates derived from serology studies (13). The surveys tested age-stratified seroprevalence by immunoglobulin G (IgG) enzyme-linked immunosorbent assays or plaque reduction neutralization assays (PRNTs). Average (per single serotype) FOI estimates were obtained by fitting IgG and PRNT data to a catalytic model, which assumed that all serotypes had the same constant FOI over the time span of the ages of the individuals surveyed. As additional seroprevalence studies became publicly available, more FOI estimates were generated in the same manner and included into the dataset. In total, we identified 34 datasets (table S1) from which we estimated 116 geolocated FOI estimates.

Case notification–based FOI estimates. We sourced 233 FOI estimates derived from age-specific routine case surveillance data (11) at the level 1 administrative unit scale in Thailand, Colombia, Brazil, and Mexico and, separately (but using identical methods), Venezuela and the Philippines (31) (table S2). For Thailand and Brazil, age-specific reports on suspected cases (according to case definition) of dengue hemorrhagic fever (DHF) were used. For Colombia and Mexico, data on all reported suspected dengue cases were used because numbers of DHF were insufficient to estimate FOI. For each administrative unit, the average per-serotype FOI over the last 20 years was estimated (11), assuming, in the absence of serotype-specific data, that all serotypes have the same FOI.

Explanatory variables

A range of environmental, socioeconomic, and demographic factors drive the distribution and spread of dengue, with rainfall, temperature, and humidity being the key determinants of vector efficacy in virus transmission (9, 32). In addition, FOI is strongly affected by human demography (13, 28, 33). We selected a set of environmental and socioeconomic variables that facilitate dengue transmission and for which data were available at the global scale. We considered eight environmental and demographic variables, which were available at high resolution (1 to 10 km), measuring (i) precipitation, (ii) diurnal temperature, (iii) nocturnal temperature, (iv) enhanced vegetation index, (v) middle infrared reflectance, (vi) altitude, (vii) population density, and (viii) per capita human birth rate. The environmental variables were available as mean and harmonic amplitudes derived from a fast Fourier transform of the original time series data. This resulted in a suite of 28 potential explanatory variables. See the Supplementary Materials for further discussion on choice and source of explanatory variables.

Predicting dengue burden

For each 1/6-degree pixel, we used a simple hazard model to estimate total annual number of dengue infections, febrile dengue cases, and cases requiring hospitalization (“hospitalized cases”) from the prediction of FOI for that pixel. We first calculated the total number of secondary, tertiary, and quaternary infections at endemic equilibrium (14). We then calculated the number of mild febrile and hospitalized cases using data on the proportions of primary, secondary, and tertiary infections that are symptomatic, derived from a published analysis of the phase 3 trials of the Sanofi-Pasteur vaccine (8).

Generating R0 estimates and predicting the impact of control measures

We explored how the impact of interventions varied under two different infectiousness scenarios: (i) all four infections (with any of the four serotypes) have the same infectiousness, and (ii) only primary and secondary infections contribute to transmission. We converted the 1/6-degree resolution FOI predictions into R0 estimates using a susceptible-infected-recovered model of dengue transmission at endemic equilibrium. We modeled transmission-reducing interventions (such as Wolbachia) by multiplying our R0 estimates by a scaling factor (α), representing the assumed reduction in transmission intensity. We then calculated the FOI corresponding to the modified reproduction number R0'. The resulting reduced FOI was used to estimate total number of infections, febrile dengue cases, and hospitalized cases in the presence of the intervention using the same approach used for burden estimation.

To model the impact of the Sanofi-Pasteur vaccine, we assumed 80% coverage and that potential recipients were screened for seropositivity using a test with 90% sensitivity and 95% specificity, with only those testing positive being vaccinated. We used the transmission dynamic model of Ferguson et al. (8) to estimate the proportion of infections, febrile dengue cases, and hospitalized cases averted within 30 years after the introduction of such a vaccination program, for a grid of 20 values of transmission intensity, for ages at vaccination between 9 and 18 years, and for both of the infectiousness scenarios outlined above. We used linear interpolation to generate pixel-level predictions from this grid of impacts. We also considered a scenario where vaccination age was optimally chosen (at admin level 1) in each setting to maximize the reduction in infections, mild febrile cases, and hospitalized cases, as commented in the revised guidance issued by WHO (34). See the Supplementary Materials for more details.

Statistical analysis

Full details are given in the Supplementary Materials. In brief, we used random forests to model the statistical relationship between FOI and the explanatory variables (35). To calibrate the model and better discriminate between predicted areas of endemicity, we generated pseudo-absence data. We randomly sampled pseudo-absence points from areas that are known to be free of dengue based on a past study of national-level dengue endemicity status (36). To generate predictions at finer spatial scales than the original training data, we used an expectation-maximization algorithm to spatially disaggregate the original FOI dataset. A spatial block bootstrap approach was used to select predictor variables based on out-of-sample predictive accuracy and to assess prediction uncertainty (37). We conducted a sensitivity analysis of the effect of the random forest parameters on model out-of-sample accuracy (figs. S27 to S31). Further outputs from the variable selection and expectation-maximization fitting routine are provided in table S7 and figs. S32 to S34. We assessed model predictive performance by calculating the coefficient of determination (R2), which represents the proportion of variance in the data explained by the model. Prediction 95% CIs were calculated using spatial block bootstrapping at a scale depending on the distance of the point to be predicted from the nearest data point.


Materials and Methods

Fig. S1. NOAA RFE2 precipitation.

Fig. S2. MODIS diurnal temperature.

Fig. S3. MODIS nocturnal temperature.

Fig. S4. MODIS enhanced vegetation index.

Fig. S5. MODIS middle infrared reflectance.

Fig. S6. WorldClim altitude.

Fig. S7. Landscan 2015 population density.

Fig. S8. United Nations 2015 per capita human birth rate.

Fig. S9. FOI observations against predictions with pseudo-absence data.

Fig. S10. FOI observations against predictions without pseudo-absence data.

Fig. S11. Predicted FOI from the model with 16 top predictors.

Fig. S12. Predicted FOI from the model with 25 predictors.

Fig. S13. SD of FOI predictions from the model with 16 predictors.

Fig. S14. SD of FOI predictions from the model with 25 predictors.

Fig. S15. Partial dependence plots for the model with 16 predictors.

Fig. S16. Partial dependence plots for the model with 25 predictors.

Fig. S17. Predicted R0 assuming only primary and secondary infections are infectious.

Fig. S18. Predicted R0 assuming all infections are equally infectious.

Fig. S19. Estimated impact of Wolbachia on global dengue infections.

Fig. S20. Estimated impact of Wolbachia on global febrile dengue cases.

Fig. S21. Estimated impact of Wolbachia on hospitalized dengue cases.

Fig. S22. Estimated impact of the Sanofi-Pasteur vaccine on global dengue infections.

Fig. S23. Estimated impact of the Sanofi-Pasteur vaccine on global febrile dengue cases.

Fig. S24. Estimated impact of the Sanofi-Pasteur vaccine on global hospitalized dengue cases.

Fig. S25. Estimated proportion (%) of 9-year-olds expected to be seronegative.

Fig. S26. Estimated proportion (%) of 16-year-olds expected to be seronegative.

Fig. S27. Sensitivity analysis of block bootstrapping grid size.

Fig. S28. Saturating function for setting pseudo-absences case weights.

Fig. S29. Sensitivity analysis of number of trees.

Fig. S30. Sensitivity analysis of minimum node size.

Fig. S31. Sensitivity analysis of pseudo-absence value.

Fig. S32. Root mean square error during forward selection algorithm.

Fig. S33. Frequency distribution of numbers of selected predictors.

Fig. S34. Expectation maximization algorithm convergence diagnostics.

Table S1. Summary of seroprevalence datasets identified and associated demographics.

Table S2. Summary of case notification datasets.

Table S3. Potential explanatory variables.

Table S4. Model R2 calculated with pseudo-absences.

Table S5. Model R2 calculated without pseudo-absences.

Table S6. Weights of primary to quaternary infections for different infectiousness scenarios.

Table S7. Rank of explanatory variables selected by the forward selection algorithm.

Data file S1. Primary data.

Data file S2. Burden estimates.

Data file S3. Raster FOI map.

Data file S4. Raster R0 map, assumption 1.

Data file S5. Raster R0 map, assumption 2.

References (3980)


Acknowledgments: We thank L. Mease and R. Coldren for sharing the age-specific seroprevalence data from Kenya. Funding: This study was primarily funded by the National Institute of General Medical Sciences “MIDAS” program (5U01GM110721 to NMF) and received additional support from the U.K. National Institute of Health Research (PR-OD-1017-20002 to NMF), Centre (MR/R015600/1) funding from the U.K. Medical Research Council (MRC) and the U.K. Department for International Development (DFID) under the MRC/DFID Concordat agreement, Institute funding from the Abdul Latif Jameel Institute for Disease and Emergency Analytics (J-IDEA), U.S. CDC Southeastern Center of Excellence in Vector-borne Diseases (CDC Cooperative Agreement U01CK000510 to D.A.T.C.), U.S. National Institute of General Medical Sciences (5U54GM08849 to D.A.T.C.), and the Bill and Melinda Gates Foundation (OPP1092240 to N.M.F.). Author contributions: N.M.F. conceived the analysis. L.C. and N.M.F. conducted the formal analysis. N.M.F. acquired the funding. L.C. conducted the investigation. L.C. and N.M.F. developed the methodology. L.C. prepared a first draft of the manuscript. L.C., I.R.-B., N.I., D.A.T.C., and N.M.F. contributed to the manuscript review and editing. Competing interests: N.F. advises the World Health Organization on aspects of dengue control and receives expense payments for this activity. N.F. additionally sits on an advisory board for Takeda Pharmaceuticals in relation to their dengue vaccine candidate, but receives no payment of any kind (honorarium, expenses, or research funding) for this activity. Data and materials availability: All data associated with this study are present in the paper or the Supplementary Materials. Code to reproduce study findings is available at (38).

Stay Connected to Science Translational Medicine

Navigate This Article