Detailed information about how we collect, process, and analyze dengue case data to provide near real-time estimates of the current global dengue situation.

Data Collection

Real-Time Data Collection

We collect near real-time dengue case data for each country from several WHO global and regional dashboards: the WHO Global Dengue Surveillance Dashboard, the WHO South-East Asia Region Dengue Dashboard, and the Pan American Health Organization PLISA database. Automated data scraping tools visit each dashboard daily to check for new updates since the last download and store the data in a structured format within the relevant data repository. The scraping tools for each dashboard, along with all downloaded data, are publicly available here:

WHO Global Dengue Surveillance Dashboard:
https://github.com/DengueGlobalObservatory/WHOGlobal-crawler

WHO South-East Asia Region Dengue Dashboard:
https://github.com/DengueGlobalObservatory/SEARO-crawler

PAHO PLISA Pan American National Dengue Data:
https://github.com/DengueGlobalObservatory/PAHO-crawler

We combine data from these dashboards into a single monthly view for each country. Where the same country appears in more than one source, we choose the most complete and reliable record, generally favouring regional dashboards over the global WHO dashboard.

Historical Data

Our predictions are based on the OpenDengue project, which compiles dengue case data reported by national health authorities, international organisations such as the World Health Organization, and literature. This global database provides a long-term view of dengue patterns across many regions and years. To ensure full coverage and consistency, we use a version of the OpenDengue data that fills in missing months and adjusts gaps in reported data using statistical models that estimate monthly dengue cases for 143 countries from 1990 to 2024 based on typical seasonal patterns https://github.com/ahyoung-lim/OD_gap_filling_public. These harmonised data form the foundation of our real-time nowcasting system and for our evaluation of the current season compared to historical years.

Backfilling

It takes time for dengue cases to move through laboratories and reporting systems before they appear on public dashboards. Recent weeks and months therefore often show fewer cases than will eventually be reported. Backfilling adjusts for this reporting delay so that current totals better reflect the true burden.

We estimate, for each country, how incomplete reporting tends to be at different time lags. This is done by comparing many past dashboard downloads: early reports for a given month are compared with later, more complete reports for the same month. From these comparisons we derive reporting factors that scale up under-reported counts.

As of June 2026, this adjustment is applied to data from the PAHO dashboard (Americas) and the WHO Global dashboard. Data from the SEARO dashboard are included using the reported figures directly, without this adjustment, until enough historical data are available to estimate reliable factors for that region.

Correction is applied only where we have sufficient historical evidence and the estimated adjustment is within reasonable bounds. Where correction is not appropriate — for example, because reporting patterns are too irregular — we use the original reported count rather than omitting the country or month. Belize is excluded from PAHO correction because of unstable reporting-delay patterns.

For PAHO, weekly cumulative counts are converted to monthly cases after adjustment. The corrected and uncorrected series are then combined with SEARO and WHO data to produce the monthly country totals shown in the Observatory.

Defining the Dengue Season

Due to the seasonal nature of Dengue, which is highly affected by temperature and precipitation, countries in different regions will experience the peak dengue season at different times. This affects critical dynamics of the disease when considering nowcasting or forecasting. In our work, we share the data and our results in calendar years but rely on an understanding of the dengue season for nowcasting.

For each country, the dengue season starts in the month with the lowest average case counts and ends 12 months later. For example, if April has the lowest average case load, it marks the beginning of the season, and March of the following year marks the end. Using this alternative window helps evaluate the average dengue season by addressing cases where peak months span across the new year. In such cases, analysing by calendar year could cause small shifts in the timing of peak months, leading to significant redistribution of cases between years, which may distort the results. Aligning data with the dengue season instead of the calendar year mitigates this by defining a time frame that prevents peak months from overlapping consecutive seasons.

Nowcasting

Using the season-aligned data, we define an average seasonal profile. First, monthly data is normalised by dividing the number of cases observed in each month by the total across the entire season. This normalisation scales the data to exist between 0 and 1.

\[\text{Monthly proportion of cases} = \frac{\text{Monthly cases}}{\text{Total season cases}}\]

Once normalised, the typical season is characterised by calculating the average proportion of cases observed in each month. This typical season helps to fill data gaps. The total number of cases expected for that season can be estimated by combining the cases observed so far with the expected proportion.

\[\text{Expected total seasonal cases} = \frac{\text{Cases observed to date}}{\text{Expected proportion to date}}\]

Once the total number of cases expected for that season has been calculated, the monthly proportions of the average seasonal profile can be used to estimate expected cases for months without data so far.

\[\text{Predicted monthly cases} = \text{Expected total seasonal cases} \times \text{Monthly proportion}\]

This method is here used here for nowcasting, the process of filling missing data gaps to date.

Uncertainty around nowcasted months

We quantified uncertainty for the proportion-based nowcast using empirically calibrated prediction intervals derived from a retrospective leave-one-season-out validation. For each country with at least three complete seasons, we iteratively withheld one season, estimated the mean monthly and cumulative seasonal proportions from the remaining seasons, and applied the operational nowcasting algorithm at each information cutoff k (last observed season month, k = 1, …, 11). For each withheld season and cutoff we then compared predicted and observed monthly case counts for all subsequent months.

Forecast errors were summarised as signed relative residuals, defined as the difference between the predicted and observed count divided by the observed count, and computed only when the observed count was greater than zero. We estimated the 2.5th, 25th, 75th, and 97.5th percentiles of these relative residuals separately for each country, cutoff month, and prediction month. Strata with fewer than five non-missing residuals were excluded from the operational lookup table to avoid unstable quantile estimates. Region- and global-level residual quantiles were computed in the same way for diagnostic use.

At deployment, intervals around a point forecast Ĉ are constructed multiplicatively from the stored quantiles: the lower limit is max(0, Ĉ(1 + q_α)) and the upper limit is max(0, Ĉ(1 + q_1−α)), yielding 50% intervals from the 25th and 75th residual percentiles and 95% intervals from the 2.5th and 97.5th percentiles. This approach anchors uncertainty to historical forecast error for the same country and timing within the season rather than assuming a parametric error model.

On country pages we display the 95% interval as whiskers on estimated months of the current-year time series. When a country–cutoff–prediction-month cell has fewer than five validation residuals, the dashboard falls back to the corresponding region-level quantiles, and then to global quantiles, so that all estimated months still carry an interval. Coverage statistics reported in the validation summary are computed from the country-only lookup.