The authors have declared that no competing interests exist.
Over the last 5 years (2013–2017), the Caribbean region has faced an unprecedented crisis of co-occurring epidemics of febrile illness due to arboviruses transmitted by the
Distributed lag nonlinear models (DLNMs) coupled with a hierarchal mixed-model framework were used to understand the exposure–lag–response association between dengue relative risk and key climate indicators, including the standardised precipitation index (SPI) and minimum temperature (Tmin). The model parameters were estimated in a Bayesian framework to produce probabilistic predictions of exceeding an island-specific outbreak threshold. The ability of the model to successfully detect outbreaks was assessed and compared to a baseline model, representative of standard dengue surveillance practice. Drought conditions were found to positively influence dengue relative risk at long lead times of up to 5 months, while excess rainfall increased the risk at shorter lead times between 1 and 2 months. The SPI averaged over a 6-month period (SPI-6), designed to monitor drought and extreme rainfall, better explained variations in dengue risk than monthly precipitation data measured in millimetres. Tmin was found to be a better predictor than mean and maximum temperature. Furthermore, including bidimensional exposure–lag–response functions of these indicators—rather than linear effects for individual lags—more appropriately described the climate–disease associations than traditional modelling approaches. In prediction mode, the model was successfully able to distinguish outbreaks from nonoutbreaks for most years, with an overall proportion of correct predictions (hits and correct rejections) of 86% (81%:91%) compared with 64% (58%:71%) for the baseline model. The ability of the model to predict dengue outbreaks in recent years was complicated by the lack of data on the emergence of new arboviruses, including chikungunya and Zika.
We present a modelling approach to infer the risk of dengue outbreaks given the cumulative effect of climate variations in the months leading up to an outbreak. By combining the dengue prediction model with climate indicators, which are routinely monitored and forecasted by the Regional Climate Centre (RCC) at the Caribbean Institute for Meteorology and Hydrology (CIMH), probabilistic dengue outlooks could be included in the Caribbean Health-Climatic Bulletin, issued on a quarterly basis to provide climate-smart decision-making guidance for Caribbean health practitioners. This flexible modelling approach could be extended to model the risk of dengue and other arboviruses in the Caribbean region.
Rachel Lowe and colleagues model for the delayed effects of climate change on outbreaks of Dengue in Barbados, an approach that can be extended to model risk for other arboviruses in the Caribbean
Changes in local climate conditions (i.e., rainfall, temperature) can affect the risk of outbreaks of diseases transmitted by mosquitoes, such as dengue fever, chikungunya, and Zika.
Climate information can be used to develop forecasts of disease outbreaks.
In this study, we devised a statistical model to test whether dengue outbreaks in the Caribbean island of Barbados could be predicted using weather station data for temperature and a precipitation index—used to monitor drought and extreme rainfall—as model inputs from June 1999 to May 2016.
The model was able to successfully predict months with dengue outbreaks versus nonoutbreaks in most years.
The risk of dengue outbreaks increased with increasing minimum temperature (Tmin; up to 25°C). Disease outbreaks were more likely to occur 4 to 5 months after periods of drought and 1 month after periods of excess rainfall.
The modelling results suggest that a drought period followed by intense rainfall 4 to 5 months later could provide optimum conditions for an imminent dengue outbreak.
In practice, the Regional Climate Centre (RCC) could use this dengue model to generate disease forecasts using their seasonal climate forecast products.
Public health decision-makers can use the dengue forecasts as an early warning tool to plan interventions to reduce the risk of dengue and other mosquito-borne diseases.
The ability of the model to predict dengue outbreaks is complicated by the emergence of new diseases with similar symptoms that are transmitted by the same mosquito vector, including chikungunya and Zika.
Small Island Developing States (SIDS) in the Caribbean are among the most vulnerable countries to extreme climate events (e.g., droughts and tropical storms), which are becoming more frequent and severe due to climate change [
In recent years, the Caribbean region has experienced an unprecedented crisis of co-occurring epidemics of febrile illness due to dengue, chikungunya, and Zika viruses. These diseases are transmitted principally by the female
Prior studies have shown that climate variability influences dengue transmission and
Prior studies in Barbados documented the seasonal linkages between climate and dengue (from 1980 to 2000), finding that epidemics occurred in the latter part of the year [
The El Niño Southern Oscillation (ENSO) is a major driver of regional year-to-year climate variability in the Caribbean [
Before an effective disease forecast can be developed, it is important to identify the key climate drivers, lag periods, and appropriate model formulation that reflects the local disease transmission ecology and climatology. It takes time for anomalies in the climate to manifest and contribute to disease risk. The various components of the time lag include the period for mosquito larval habitat to increase, the development period of the mosquito, the time before the first blood meal in which the mosquito transmits the virus to a human host, and the time before the appearance of clinical manifestations of dengue [
This study aims to quantify the nonlinear and delayed effects of climate impacts, such as drought, extreme rainfall, and temperature variations, on dengue risk in the eastern Caribbean island of Barbados by coupling DLNMs with a Bayesian model estimation framework. The model is then used to produce out-of-sample predicted probabilities of exceeding island-specific outbreak thresholds. Barbados is an ideal case study due to high-quality historical climate and health data, a relatively high burden of disease, and prior experience in building projects on climate and health, paving the way for a sustainable collaboration to develop integrated early warning information to predict the risk of dengue and other mosquito-transmitted diseases in the Caribbean.
This study was carried out for the Caribbean island of Barbados (13° N 59° W) in the Lesser Antilles. Barbados is relatively small in land area (425 km2), with a resident population of over 277,000 people [
The surveillance unit of the Environmental Health Department in the Ministry of Health of Barbados is responsible for maintaining records on the incidence of dengue and other arboviral disease cases. As of January 2018, the surveillance unit liaises with the Leptospirosis lab (national reference lab) to record laboratory-confirmed cases of dengue virus, chikungunya virus, and Zika virus. Specimens are tested for the dengue virus nonstructural protein 1 (NS1) antigen, dengue virus IgM/IgG, and chikungunya virus IgM/IgG using commercial ELISA kits. Since September 2016, the lab has also conducted real-time PCR using the CDC Trioplex assay for dengue virus, chikungunya virus, and Zika virus, and they conduct serotyping of positive dengue virus samples. National population estimates, obtained from the Barbados Population and Housing Census (1990, 2000, and 2010) [
The climatology of Barbados fits well within that of the Lesser Antilles. The wet season lasts from June until November and coincides with the Atlantic Hurricane Season [
Given its low topography, Barbados records significantly less rainfall, on average, than neighbouring mountainous islands, from less than 1,200-mm rainfall per annum in some low-lying areas to around 2,000-mm per annum above 300-m elevation [
Monthly rainfall totals and average minimum, mean, and maximum temperatures were extracted from the 2 synoptic weather stations on the island—at the Grantley Adams International Airport (GAIA), at an elevation of 56 m in the southern parish of Christ Church, and at the Caribbean Institute for Meteorology and Hydrology (CIMH), at an elevation of 112 m in the western parish of St. James (see
The Standardised Precipitation Index (SPI) is a drought index first developed by McKee et al. [
SPI at different time scales are representative of different types of drought. Generally, the longer the meteorological drought persists and the more negative the SPI value, the greater the societal impact. The 1-month SPI (SPI-1) allows for early detection of meteorological droughts. Operationally, with the aim to enable drought early warning systems, the Caribbean Drought and Precipitation Monitoring Network (CDPMN)—coordinated by the CIMH—tracks drought using SPIs at 1 month, 3 months (SPI-3), 6 months (SPI-6), 12 months (SPI-12), and 24 months (SPI-24) [
With drought being a slow-onset climate hazard, early warning (e.g., based on SPIs) with ample lead time is a realistic and effective option. This is especially the case when monitoring is combined with forecasting. Operationally, the Caribbean Climate Outlook Forum (CariCOF) provides drought forecasts using predictions of SPI-6 with a lead time of 3 months (i.e., 3 months in advance) and SPI-12 with lead times up to 6 months [
For this study, all SPI values were calculated based on the records of daily rainfall totals at the CIMH and the GAIA stations in Barbados, which start in January of 1981 and 1971, respectively. The historical climatological reference was calculated using the period 1981–2010 to conform with World Meteorological Organization regulations.
Annual cycle of (a) dengue incidence rates (per 100,000 population) in Barbados, (b) SPI-6, and (c) Tmin (°C) averaged over CIMH and GAIA weather stations at the monthly time scale from June 1999 to May 2016. CIMH, Caribbean Institute for Meteorology and Hydrology; GAIA, Grantley Adams International Airport; SPI-6, 6-month Standardised Precipitation Index.
As dengue incidence tends to peak between September and February (see
First, a baseline model was formulated by including a first-order random walk latent model for month βt'(t), where t'(t) = 1, …, 12 and β1 represents the parameter estimate for the month of June. This term helps to capture the seasonality in dengue, which is assumed to be stationary each year (monthly random effect). The first-order random walk prior allows dengue incidence rates in 1 month to depend on the previous month, to reflect both seasonality and the infectious nature of the disease. Next, exchangeable random effects for each year γT'(t) (where T'(t) = 1, …, 17 and γ1 represents the dengue year June 1999–May 2000) was included in the model to account for interannual variation in dengue over time (yearly random effect). This term potentially allows for changes in population immunity between outbreak years, lapses in vector control, and other slowly changing factors, such as changes in mosquito-control intensity and the introduction of new dengue virus serotypes or other viruses, which could be mistaken for dengue. Note that chikungunya and Zika viruses were introduced to Barbados in 2014 and at the end of 2015, respectively [
DLNMs [
Model parameters were estimated in a Bayesian framework using Integrated Nested Laplace Approximation (INLA;
To test the predictive ability of the model, we refitted the model 17 times, removing 1 year at a time to produce out-of-sample predictions (i.e., 17-fold cross-validation). Note that the annual random-effect term was included in the model-fitting stage to better quantify the association between climate factors and variation in dengue incidence rates. However, when producing out-of-sample predictions, no effect was estimated for the year in which the prediction was valid, hence the term does not contribute to incidence rate estimates in prediction mode. Therefore, model predictions are based solely on exposure–lag–response functions of key climate variables and the seasonality term, meaning the model can be used to predict any year not included in the model-fitting process.
To evaluate the output from the selected Bayesian hierarchical model, posterior predictive distributions of the response variable were simulated using samples from the posterior distribution of the parameters and hyperparameters in the model [
Dengue control programmes often monitor new cases against historical case data, which define a typical dengue transmission season [
As part of the model selection procedure, a range of climate variables were tested, first as linear terms for individual lags from 0 to 5 months and second as exposure–lag functions for individual climate indicators (see
The CV mean logarithmic score, the DIC, and the likelihood ratio RLR2 statistic for models of increasing complexity.
Model | log(ρt) | CV log score | DIC | RLR2 |
---|---|---|---|---|
1 | α + βt'(t) |
4.46 | 1,801.36 | 0.23 |
2 | α + βt'(t) + γT'(t) |
4.23 | 1,719.46 | 0.54 |
3 | α + βt'(t) + f.w(x1t, l) |
4.32 | 1,759.07 | 0.41 |
4 | α + βt'(t) + f.w(x2t, l) |
4.34 | 1,770.52 | 0.34 |
5 | α + βt'(t) + f.w(x1t, l) + f.w(x2t, l) |
4.28 | 1,742.98 | 0.47 |
6 | α + βt'(t) + γT'(t) + f.w(x1t, l) + f.w(x2t, l) |
4.09 | 1,664.94 | 0.68 |
Abbreviations: CV, cross-validated; DIC, deviance information criterion; SPI-6, 6-month Standardised Precipitation Index; Tmin, minimum temperature.
Associations between the selected climate variables (SPI-6 and Tmin) and dengue relative risk are presented as three-dimensional graphs and two-dimensional contour plots in
Three-dimensional (upper panel) and contour (lower panel) plots of the exposure–lag–response association between (left) SPI-6 relative to normal conditions (SPI-6 = 0) and (right) mean temperature anomalies relative to Tmin of 20°C, at lags between 0 and 5 months. SPI-6, 6-month Standardised Precipitation Index; Tmin, minimum temperature.
Lag–response association for scenarios of (a) SPI-6: exceptionally dry (SPI-6 = −2.5) and exceptionally wet (SPI-6 = 2.5) conditions relative to the baseline (SPI-6 = 0) and (b) Tmin: Tmin = 21.5°C and Tmin = 25.5°C relative to the baseline (Tmin = 20°C), at lags between 0 and 5 months. SPI-6, 6-month Standardised Precipitation Index; Tmin, minimum temperature.
Moving outbreak thresholds were calculated as the 75th percentile of the distribution of dengue cases per month between June 1999 and May 2016, excluding the ‘dengue’ year for which the prediction was valid. We applied this moving threshold to produce out-of-sample probabilistic predictions of exceeding the outbreak threshold for all 17 years (June 1999–May 2016) as a demonstration.
Posterior predicted mean (dashed purple curve) and 95% prediction interval (shaded area) for dengue incidence rates (per 100,000 population) in Barbados from June 1999 to May 2016, simulated from the final model (refitted 17 times, leaving out 1 year at a time). Observed values (solid orange curve) and moving outbreak threshold (blue dotted curve) are included. Year labels are included at the start of each calendar year (e.g., in January).
Probability of exceeding the moving outbreak threshold (75th percentile of observed dengue cases per month, excluding the year for which the prediction is valid) from June 1999 to May 2016, using out-of-sample predictive distributions simulated from the final model. The graduated colour bar represents the predicted probability of observing an outbreak (ranging from 0, pale colours, to 1, deep colours). Months in which the moving outbreak threshold was exceeded are marked with a cross. Note: the ‘dengue season’ year runs from June to May.
The outbreak detection performance of the model is summarised in
Summary of ROC and contingency table analysis for observed dengue incidence rates exceeding the moving outbreak threshold (75th percentile of observed dengue cases per month, excluding the year for which the prediction is valid) using out-of-sample probabilistic predictive distributions from the final and baseline models (95% CIs included in parentheses).
Performance measures | Final model | Baseline model |
---|---|---|
AUC | 0.9 (0.85–0.94) | 0.75 (0.71–0.79) |
Probability trigger threshold | 0.3 | 0.27 |
Hit rate | 0.9 (0.82–0.97) | 0.79 (0.69–0.9) |
False alarm rate | 0.31 (0.22–0.39) | 0.57 (0.51–0.63) |
Proportion correct | 0.86 (0.81–0.91) | 0.64 (0.58–0.71) |
Abbreviations: AUC, area under the ROC curve; ROC, relative (receiver) operating characteristic.
Drought conditions were found to positively influence dengue incidence rates at longer lead times up to 5 months, while excess rainfall increased the risk at shorter lead times between 1 and 2 months. Therefore, the modelling results suggest that a drought period followed by intense rainfall 4 to 5 months later could provide optimum conditions for an imminent dengue outbreak. The use of the SPI-6, designed to monitor drought and extreme rainfall, better explained variations in dengue risk than summary statistics of measured precipitation. Tmin explained more variation in dengue incidence rates than mean or maximum temperature. Furthermore, including bidimensional exposure–lag–response functions of these indicators—rather than linear effects for individual lags—more appropriately described the climate–disease associations than traditional modelling approaches. To our knowledge, DLNM methodology has not previously been combined with a Bayesian hierarchical mixed-modelling framework to produce probabilistic predictions of exceeding dengue outbreak thresholds. To demonstrate the added value of our climate-driven dengue model, we formulated a baseline model to represent current practice (i.e., monitoring dengue cases throughout the year against historic seasonal averages). A probability trigger threshold was calculated by using the ROC curve to select an optimal cut-off value that maximised sensitivity and specificity for the moving outbreak threshold of the upper quartile of the observed dengue distribution per month. The model successfully distinguished outbreaks from nonoutbreaks, except in the last 2 dengue seasons (2014–2015, 2015–2016), with an overall proportion of correct predictions (hits and correct rejections) of 86% compared with 64% for the baseline model.
The observed effect of Tmin on dengue transmission is consistent with findings from prior studies in the Americas, which also found that Tmin was a better predictor of dengue transmission than mean or maximum temperature [
We hypothesise that the effect of excess rainfall and drought on dengue risk operated at different time scales due to different mechanisms associated with the availability of larval habitat and water storage in the urban environment. Following a rainfall event, the availability of larval habitat increases (e.g., rain-filled abandoned containers, rubbish), and within a relatively short period of time, eggs hatch and adult mosquito densities increase (i.e., approximately 2–3 weeks after rainfall, depending on ambient temperatures). Subsequently, the risk of arbovirus infections increases several weeks later, a lag associated with the intrinsic and extrinsic viral incubation periods.
In contrast, we observed that drought conditions, as determined by the SPI-6, were associated with dengue risk over longer periods of time (5-month lag). Barbados is among the 10 most water-scarce countries in the world [
To our knowledge, there is little known about the effects of prolonged drought on dengue transmission. Some studies have shown that rainfall shortages can increase dengue risk in regions where people store water [
In recent decades, the Caribbean has experienced several drought events. Notable events during our study period include 2002–2003, 2009–2010, and 2015–2016. Drought and water scarcity are projected to increase in the future due to climate change [
In Barbados, over the last 5 years, there have been 3 outbreaks of dengue fever (2013–2014, 2014–2015, and 2015–2016), 1 outbreak of chikungunya virus (2014–2015), and 1 outbreak of Zika virus (2015–2016). Better management of arboviruses is a high priority for Caribbean SIDS given the burden of
Climate anomalies in the Caribbean are associated with tropical Atlantic and Pacific sea surface temperatures, which are usually well predictable at seasonal time scales [
In practice, the CIMH could run this dengue model using their seasonal climate forecast products. Results could be updated on a monthly basis as the target month is approached. This should reduce forecast uncertainty as one more month of climate observations becomes available and one fewer month of climate forecasts is needed. CariCOF’s SPI-6-based drought outlooks have a prediction lead time of 3 months. For example, CariCOF’s SPI-6-based drought outlook issued in July, which estimates the drought situation by the end of October, could be used to predict the dengue risk in October (target month). Observed SPI-6 data for May (i.e., calculated from the observed rainfall totals from December of the previous year to May of the current year) and June (i.e., 4 and 5 months before the target month, to capture the long-lag impacts on dengue risk) would be used along with forecast SPI-6 conditions from July to October (i.e., lags of 0–3 months with respect to the target month). This would produce a 3-month lead dengue forecast (see
Schematic to show the type (e.g., observed or forecast) of climate information (e.g., SPI-6 and Tmin variables, lags 0 to 5) used to produce a dengue forecast for the target month of October. The first forecast would be issued 3 months ahead of the target (i.e., 3-month lead time), in July, using observed climate data (for SPI-6 and Tmin) for May and June and forecast climate data for July, August, September, and October. The forecast would then be updated in August (i.e., 2-month lead time) using observed climate for July and updated climate forecasts for August, September, and October. In September, 1 month before the target (i.e., 1-month lead time), the dengue forecast would be updated again, using observed climate from May to August and updated climate forecasts for September and October. SPI-6, 6-month Standardised Precipitation Index; Tmin, minimum temperature.
Seasonal forecast models can contribute to reducing the risk of disease outbreaks by increasing preparation time to better allocate scarce resources. To date, there are few localised studies in this region that quantify the associations between subseasonal to seasonal forecasts of climate and dengue risk outcomes. In the absence of a robust evidence base, the operational use of climate information for predicting increased dengue risk remains an untapped opportunity to support climate-driven dengue early warning information systems. This is an area of further investigation, along with the transferability of this framework to model and predict the risk of dengue and other arboviruses in other Caribbean countries.
There are several limitations of the study. In Barbados, the absolute number of dengue cases across the island is relatively small, even in outbreak years (i.e., 1,140 confirmed cases in 2013, 488 confirmed cases in 2014, and 587 confirmed cases in 2016). The time series is relatively long compared with previously studied dengue records in other countries. However, it is still limited for statistical modelling purposes because our knowledge of outbreak dynamics is based on the behaviour of only a few outbreaks. Ideally, data on vector control activities and other interventions or policy changes would be included in the model. In the absence of such data, monthly and yearly random effects are included in the model to try and ensure that the variance in the predictions includes uncertainty that could be generated by these unknown features of the disease system. However, while yearly effects are useful to identify missing information and better quantify the contribution of explanatory variables to the relative risk, they cannot contribute to out-of-sample predictions unless prior assumptions about the year ahead can be made.
The model failed to predict the peak and magnitude of the last 2 dengue transmission seasons, which coincided with the first epidemics of chikungunya and Zika. The emergence of these viruses, which are transmitted by the same mosquito vector, likely altered diagnostic, reporting, and healthcare-seeking practices, and we suspect that this might have affected the model performance for the last 2 dengue seasons in 2015 and 2016. This may also be true in other countries and territories across Latin America and the Caribbean during this time period. Dengue virus, chikungunya virus, and Zika virus have a similar clinical presentation (e.g., mild febrile illness, rash, joint pain), and differential diagnosis is challenging without molecular diagnostics (e.g., PCR) due to cross-reactivity of the serological assays. Only after September 2016 was Barbados able to test suspected arbovirus infections using a triplex PCR; prior to then, a subset of samples was sent for confirmation to CARPHA’s diagnostic laboratory in Trinidad and Tobago. The emergence of Zika virus may have also changed healthcare-seeking behaviour, with more women seeking diagnostics for mild febrile illness due to concerns about Zika congenital syndrome [
Overall, the final model had a lower false alarm rate and a lower miss rate than the baseline model. However, the model did result in false alarms 31% of the time. Repeatedly issuing false alarms can result in a lack of trust from the general public. However, in practice, false alarms (i.e., issuing a high probability alert of a dengue outbreak) can, in fact, result from effective interventions in response to the alert. Furthermore, when dealing with detrimental health events, such as disease outbreaks, false alarms are preferable to missed events.
The objective of this study was to select a model to infer delayed and nonlinear impacts of climate on dengue risk and predict the probability of exceeding an outbreak threshold relevant to the Ministry of Health in Barbados and other islands in the Caribbean. Forecasts of the probability of exceeding outbreak thresholds allow decision-makers to quantify the level of certainty of the model predictions. However, when using predefined outbreak and alarm trigger thresholds, the translation of probabilities into discrete warnings might not always reflect the predictive power of the model [
We are working towards implementing a climate-based early warning system for arboviral diseases in all Caribbean islands and territories. Not only does this study contribute to enriching the knowledge base, it represents a considerable opportunity to translate investment in health–climate research into practice to improve national and regional health outcomes. The goal is to enhance the health warnings issued in the quarterly Caribbean Health-Climatic Bulletin with quantitative probabilistic forecasts of disease risk rather than expert statements on probable health outcomes. Integration of an early warning information product into national and sectoral planning and practice has the potential to reverse the upward trend of new infections of arboviral diseases, which currently undermines the productivity and sustainable development of SIDS in the Caribbean.
Map of Caribbean (left) and Barbados showing population (middle) and elevation (right) and the locations of the 2 main meteorological stations (CIMH and GAIA). This figure was created in ArcGIS version 10.3.1 [
(TIF)
The chance of a wet day (i.e., a calendar day with >0.85 mm; left panel) and the average rainfall intensity on a wet day for each Julian day of the year (right panel). Source: modified from Trotman and colleagues [
(TIF)
Annual cycle of (a) dengue incidence rate (per 100,000 population), (b) precipitation (mm/month), and (c) Tmin (°C) averaged over CIMH and GAIA weather stations. CIMH, Caribbean Institute for Meteorology and Hydrology; GAIA, Grantley Adams International Airport; Tmin, minimum temperature.
(TIF)
Annual cycle of (a) SPI-6 and (b) Oceanic Niño Index, defined as the 3-month running-mean sea surface temperature departures from average in the Niño 3.4 region (120–170° W, 5° S-5° N), from June 1999 to May 2016. SPI-6, 6-month Standardised Precipitation Index.
(TIF)
Annual cycle of dengue cases given tercile categories of the SPI-6: drier than normal (solid curve), normal (dashed curve), and wetter than normal (dotted curve). During wetter than average years, dengue cases tended to peak earlier in the season, in September, whereas dry years coincided with late-season peaks. SPI-6, 6-month Standardised Precipitation Index.
(TIF)
Exposure–response association across all lags (0 to 5 months) for (a) the SPI-6 relative to the baseline SPI-6 = 0 and (b) Tmin relative to the baseline Tmin = 20°C. SPI-6, 6-month standardised precipitation index; Tmin, minimum temperature.
(TIF)
Observed versus posterior predicted mean dengue incidence rates (per 100,000 population) from the final model (purple circles) and the baseline model (red squares). Note: logarithmic scale.
(TIF)
ROC curve for binary event of dengue incidence rates exceeding the moving outbreak threshold (75th percentile of observed dengue incidence rates per month, excluding the year for which the prediction is valid) for (a) final model (without year effect contribution) and (b) baseline model. Numbers indicate values of probability thresholds along the curve, and the purple circle indicates the position of an ‘optimal’ ROC cut-off (alarm trigger threshold), defined as the point on the curve closest to the point of perfect discrimination (0, 1). Note: false alarms are a desired outcome when the predicted probability of threshold exceedance is very low. ROC, relative (receiver) operating characteristic.
(TIF)
SPI, Standardised Precipitation Index.
(DOCX)
SPI, Standardised Precipitation Index.
(DOCX)
Summary statistics for a model including monthly and yearly random effects and (a) SPI-6 and (b) Tmin at individual lags of 0 to 5 months: posterior mean, 95% CIs, the CV mean logarithmic score, the DIC, and the likelihood ratio RLR2 statistic. CI, credible interval; CV, cross-validated; DIC, deviance information criterion; SPI-6, 6-month Standardised Precipitation Index; Tmin, minimum temperature.
(DOCX)
The CV mean logarithmic score, the DIC, and the likelihood ratio RLR2 statistic for models including monthly and yearly random effects and an exposure–lag–response function with time lags from 0 to 5 months for the SPI averaged over 1 month, 3 months, 6 months, and 12 months). The best indicator is shaded in grey. CV, cross-validated; DIC, deviance information criterion; SPI, Standardised Precipitation Index.
(DOCX)
The CV mean logarithmic score, the DIC, and the likelihood ratio RLR2 statistic for models including monthly and yearly random effects and an exposure–lag–response function with time lags from 0 to 5 months for individual climate variables: precipitation, SPI (1-month, 3-month, 6-month, and 12-month), Tmin, mean temperature, and maximum temperature. CV, cross-validated; DIC, deviance information criterion; SPI, Standardised Precipitation Index; Tmin, minimum temperature.
(DOCX)
(DOCX)
The authors are grateful to Trevor Bailey, Hårvard Rue, Ben Armstrong, Charmaine Blades, Shermaine Clauzel, Steve Daniel, Sally Edwards, Lyndon Forbes Robertson, Adrianus Vlugman, Marquita Gittens-St. Hilaire, and Karen Polson-Edwards for valuable expert advice during this study.
area under the ROC curve
Caribbean Climate Outlook Forum
Caribbean Public Health Agency
Caribbean Drought and Precipitation Monitoring Network
Caribbean Institute for Meteorology and Hydrology
cross-validated
dengue virus
deviance information criterion
distributed lag nonlinear model
El Niño Southern Oscillation
Early Warning Information Systems Across Climate Timescales
Grantley Adams International Airport
Integrated Nested Laplace Approximation
nonstructural protein 1
Regional Climate Centre
relative (receiver) operating characteristic
Small Island Developing States
Standardised Precipitation Index
SPI averaged over a 6-month period
minimum temperature