An Analysis of the Potential Impact of Climate Change on Dengue Transmission in the Southeastern United States

We downloaded 20 years (1981–2000) of observed daily weather station measurements (temperature and precipitation data) from the Global Historical Climatology Network Database (Menne et al. 2012a, 2012b; National Centers for Environmental Information; https://www.ncdc.noaa.gov/oa/climate/ghcn-daily/) for each of the 23 sites. Future climate change data were accessed from GCM data housed in the statistical weather generator LARS-WG5 (LARS) (Rothamsted Research 2016). Daily-level data from GCMs are not directly comparable to observed daily data; therefore, we used LARS to produce a comparable series. LARS uses historical data to create a new synthetic time series that is statistically similar to the observed data at each site, enabling translation between monthly and daily time scales. We evaluated the observed and synthetic time series to assess the software’s ability to reproduce realistic daily data. The daily probability distributions (Kolmogorov–Smirnov test) and monthly means (Student’s t-test) for precipitation, minimum, and maximum temperatures for each site showed only one statistically significant value (p < 0.01; Jackson, MS, September minimum temperature). The minimal significance differences between the observed and synthetic data sets for these values suggests that the synthetic time series reasonably capture the overall patterns of the observed data sets.

We produced time series for the future climate scenario by calculating GCM ensemble–projected average monthly changes in temperature (absolute) and precipitation (proportion) between the GCM-simulated baseline (1961–1990) and future (2045–2065) periods. We used the closest location for which the GCMs (1.3–3.9° resolution) were run to apply the changes in LARS for each site to the synthetic time series in order to create 21 years of downscaled, daily time series of site-specific future midcentury climate conditions (2046–2056). We applied the same procedure to generate comparable baseline data to use in place of the observed data. We used an ensemble of 15 GCMs (Table 1) to account for models that over- or under-predicted specific variables. The projections used the Intergovernmental Panel on Climate Change (IPCC) SRA1B Special Report on Emissions scenario, which assumes “[a] future world of very rapid economic growth, global population that peaks in mid-century and declines thereafter, and rapid introduction of new and more efficient technologies, with the development balanced across energy sources” (IPCC 2007a). This scenario was selected because it was available for all models and is a midrange emissions and warming scenario (IPCC 2007b).

DyMSiM is a meteorologically driven, process-based model containing entomological and epidemiological components. The model was first parameterized to simulate Culex quinquefasciatus populations (Morin and Comrie 2010) and has now been adjusted to simulate Ae. aegypti and DENV development. Using daily temperature and precipitation data, mosquito populations are simulated by calculating daily rates of development and mortality for mosquito cohorts as they proceed through their life cycle (egg, larva, pupa, and adult). Development rates are calculated using temperature, and mortality rates are dependent on temperature and, in the case of larvae and pupae, water availability. Water habitat is calculated for rain-filled containers (precipitation minus evaporation and spilling) and permanent water sources (levels remain constant). During the adult stage, cohorts of mosquitoes proceed through their gonotrophic cycle, including blood meal questing, ovarian development, and egg laying. Rates of feeding and ovarian development are temperature dependent. Once ovarian development is complete and water is available, the eggs are deposited. During feeding, mosquitoes can become infected with DENV at a probability based on human infection rates (see below) and will then proceed through a temperature-regulated EIP. We input new cohorts of eggs into the model if the mosquito population drops to zero to prevent extinction of the mosquito within the model during inhospitable climate conditions.

The human population is simulated using a compartmental susceptible–exposed–infectious–recovered model. Susceptible humans move to the exposed stage based on the number of infected mosquitoes that have completed the EIP. Exposed individuals remain in the exposed and infectious stages for a period of 5–7 days before recovering from the infection. Mosquito infection rates are calculated using the number of humans in the infectious stage. To prevent extinction of the virus, a minimum human infection rate is used when there are few or no infections in the human population. The governing equations for each stage (mosquito and human) as well as the parameters and their equations can be found in Morin et al. (2015).

San Juan was used for model evaluation because of the availability of multiyear dengue case records, which are unavailable in the mainland United States owing to a lack of long-term transmission. The climate of San Juan is an important driver of local transmission, and DyMSiM can effectively capture inter- and intra-annual variability (Morin et al. 2015). We used the same predictors (7 variables, 96 total parameter values, Table 2) that had been applied in the previous analysis of transmission in San Juan for the present analysis, and although we acknowledge the potential for variation among individual locations, it was not possible to validate these parameters for each location in the present analysis. In addition, standardizing model outputs for each U.S. site against San Juan data while holding all model parameters other than climate constant allowed us to estimate the specific effects of climate and climate change on dengue transmission.

The daily meteorological data sets generated by LARS, described above, were used to drive DyMSiM to simulate current and projected future mosquito populations and human dengue cases. DyMSiM also requires the parameterization of additional environmental variables and the site latitude to determine the duration of sunlight hours (which influences the evaporation of standing water). To account for spatiotemporal variability in parameter values and climate conditions, 96 simulations were performed at each site using different combinations of model parameter values. The parameter values were selected to represent ranges of values reported in the literature (Morin et al. 2015). Six of the seven parameters included in our suite of simulations were represented by only 2 or 3 possible values because their values are well established (Table 2). In contrast, we included 12 possible values for habitat area (determined by the number and surface area of containers available for egg laying) because of greater uncertainty and potential for variation in this parameter (Morin et al. 2013, 2015). Parameter values for the simulation were chosen from the best-fit 1% of simulations for San Juan, Puerto Rico for the period 2010–2013 because of their superior ability to replicate patterns of dengue (Morin et al. 2015) (Table 2). We simulated a total of 21 years of daily total mosquito populations and dengue case data for each site driven by baseline climate, and we repeated this process with the future climate scenarios. After discarding the first year of DyMSiM results to allow for model spin-up time, each of the 23 mainland U.S. sites had four associated 20-year daily time series: a) total mosquitoes, present climate; b) total dengue cases, present climate; c) total mosquitoes, future climate; and d) total dengue cases, future climate.

We compared the outputs of the model runs with output from San Juan for each site. The analysis results are indicative of one of two possibilities for each location: a) transmission in the mainland U.S. site occurs at the same level as that in San Juan, indicating that the climate is suitable for transmission, but socioeconomic factors, lack of importation, or other unaccounted-for factors are limiting transmission; or b) dengue transmission in the mainland U.S. site is significantly reduced or absent compared with that in San Juan, suggesting that the site climate is not suitable for epidemic dengue. By repeating these steps for the future climate scenarios, we also assessed future transmission potential for each location.

Maps were created from the model output to visualize spatio-temporal variance. For visualization purposes, the daily model outputs (mosquito population and dengue cases) were aggregated to seasonal values by dividing the year into four 13-week periods (approximately January–March, April–June, July–September, and October–December). For each mainland U.S. location and parameterization, data values were averaged across the 20-year runs and were then compared with the corresponding data for San Juan, Puerto Rico. The comparison metric (percent of value in San Juan) was then averaged across the parameterizations and mapped (Figures 1 and 2).

In addition to estimating mosquito populations and dengue cases during each season for the 23 locations, we conducted a separate analysis to estimate weekly variation in dengue cases for the Key West, Florida and Brownsville, Texas sites. Specifically, we averaged estimated numbers of weekly dengue cases across the 20-year period to create a single annual time series (at weekly resolution) for each simulation and location. The time series were then standardized against the cumulative annual San Juan dengue cases using the following equation:

StDen = standardized dengue cases, MDen = modeled dengue averaged over the 20 years, SMDen = modeled dengue in San Juan averaged over the 20 years, i = epidemiological week, j = simulation number, x = site. An epidemiological week (epi week) is a standardized way of defining the aggregation of days into weeks so that data are comparable across years. We then averaged across all the simulations to create one annual time series each for Key West and Brownsville under the base and future climate scenarios. This method is similar to the method used to estimate seasonal averages for all 23 locations. Here, however, the standardization process only serves to highlight the seasonality of transmission (i.e., the percent of the total annual dengue cases that occurred each week).

We conducted additional analysis on Key West, Florida, and Brownsville, Texas, because both have had recent dengue cases (see Bouri et al. 2012) but have contrasting climate characteristics. Figure 3 shows dengue cases at these sites as a percentage of total cases in Puerto Rico, along with plotted climate data. Presently, Key West climate conditions are more favorable for transmission than those in San Juan during the summer (weeks 30–44), with the window widening to weeks 27–51 under the future scenario. Our modeled output data showed, on average, low but continual transmission throughout the winter in San Juan. However, this year-round transmission is not maintained in Key West or Brownsville (Figure 1). The climate data in Figure 3 show that San Juan has a smaller annual temperature range than these two sites, with warmer winter temperatures capable of supporting year-round transmission. Conversely, cooler winter conditions are likely to prevent or limit transmission in Key West, even under future climate conditions. Estimates show a similar pattern for Brownsville, but with a shorter time window for transmission that does not currently exceed that of San Juan. However, transmission is projected to increase in future climate scenarios, exceeding current case numbers in San Juan during the summer (weeks 27–35). Nevertheless, temperatures remain substantially cooler than in San Juan during fall and winter, thus limiting the potential for year-round transmission.

A number of important factors known to affect dengue, such as humidity, herd immunity, vector competition, insect resistance, viral mutation, and socioeconomic factors, are not included in the model. Nor do we account for land cover variation, geographic relationships to endemic areas, or potential human adaption strategies. Dengue disease dynamics are complex, and future research should aim to develop more comprehensive models that can better assess the roles of such variables.

In light of the abovementioned factors, our results should be interpreted as climate-based projections of relative differences in estimated risks, not as concrete predictions of future climate change impacts on dengue fever. The generated climate data provide only possible values for daily temperature and precipitation. Synthetic baseline and GCM data sets are inherently subject to inaccuracies; for example, some standard deviation metrics between the observed and synthetic data sets did not perform as well as the evaluation metrics described in “Methods.” We attempted to partially address this concern by using the synthetic baseline data in place of the observed data to standardize such inaccuracies across the baseline and future time series for comparative purposes. The use of larger data sets and improved downscaling techniques may improve the accuracy of future studies.

Case studies have shown that the spatial distribution of Ae. aegypti is heterogeneous within cities and regions (Hayden et al. 2010; Murray et al. 2013). We have attempted to account for this heterogeneity by using the average from 96 different parameterizations that were selected to optimize model performance based on an analysis of data from Puerto Rico. Nevertheless, it remains possible that some locations in our study may have different spatial patterns of breeding sites than those captured in the San Juan validation, and readers should keep in mind that these results were modeled using San Juan parameters. Further field studies are needed to quantify the distribution of Ae. aegypti across U.S. urban landscapes.

Although it would be ideal to have accurate mosquito population and breeding site data for each location, the paucity and accuracy of such records and the length of time needed to generate methodologically sound data sets is an ongoing problem in climate and health research. DyMSiM seeks to fill such a void by generating environmentally driven mosquito populations in the absence of mosquito data. Future work should be focused on evaluating model performance in multiple locations and on obtaining better location-specific information for model parameterization; the model should be periodically re-validated as larger dengue data sets become available.

Finally, more detailed risk assessments are needed to better understand site-specific vulnerabilities. This may include parameterizing the model against local mosquito data, if available, or running the model at finer spatial scales within a city. Given that microclimatic variations within a community can affect mosquito abundance (Hayden et al. 2010), this additional detail may be useful for determining local risk. Inter- and intra-annual future climate variations may also affect the seasonality of dengue transmission in ways that current GCMs cannot account for when downscaled to the local level. As GCM capabilities improve, along with our understanding of shifting weather patterns, additional analysis may be useful for understanding dengue transmission potential.

These limitations are inherent in modeling approaches. Nevertheless, this technique is useful for demonstrating the contributing role of climate in shaping dengue fever transmission risk within the southeastern United States. The impact of climate on Ae. aegypti abundance patterns is also important for the Chikungunya virus and the Zika virus. The current DyMSiM model does not incorporate specific characteristics of these viruses, but the potential for an increase in Ae. aegypti seasonality may have implications for the transmission of these viruses as well. Future studies incorporating specific temperature thresholds for the Zika virus and the Chikunguyna virus would be helpful.