Study design

This ecological study used data collected during routine dengue surveillance in Rio Grande do Sul, Brazil between 2007 and 2023. The dynamics of this disease in the state were analyzed using temporal, spatial, and spatiotemporal statistical analyses.

Study location

Rio Grande do Sul is located at latitude 27˚03,042@ and 33˚45,009@ South and longitude 49˚42,041@ and 57˚40,057@ West and is the southernmost state in Brazil (Fig. 1), bordering Argentina and Uruguay. The state covers an area of 281,707.151 km², and is divided into 497 municipalities. In 2021, the population was 11,466,630, representing a density of 39.79 inhabitants/km2, of which 85.1% lived in urban areas, and the HDI was 0.746¹³. In 2015, 93.7% of the population was served directly or indirectly through garbage collection, and 40% was served through a sewage collection network¹³. The climate of Rio Grande do Sul is subtropical temperate, classified as humid mesothermal (Köppen classification). Because of its geographical position, it presents significant differences from the rest of Brazil. Temperatures show great seasonal variation, with hot summers and rigorous winters, including frost and occasional snowfall. Average temperatures range between 15 and 18 °C, with minimum temperatures dropping to -10°C and maximum temperatures reaching 40 °C¹⁴.

Division of the macro-regions of Rio Grande do Sul, Brazil, according to SIRGAS 2000, highlighting the regions Centro Ocidental, Centro Oriental, Metropolitan Porto Alegre, Northeast, Northwest, Southeast, and Southwest Rio-grandense.

There are three major economic regions in Rio Grande do Sul: (1) the southern region, with a large concentration of land, large livestock farms, and mechanized planting of rice, soybeans, and wheat. This area also has a greater income inequality. (2) The northeastern region, which includes the state capital, has more industries and smaller properties. (3) The northern region, which European immigrants mostly colonize, has forest cover, valleys, and plains with small agricultural lands (Fig. 1)¹⁵. In addition, the state is divided into seven mesoregions: Metropolitan Porto Alegre, Central Westem of Rio Grande, Central Eastem of Rio Grande, Northeast of Rio Grande, Northwest of Rio Grande, Southeast of Rio Grande, and Southwest of Rio Grande (Fig. 1).

Data collection

Since 1996, the reporting of dengue cases has been mandatory in Brazil through the Notifiable Diseases Information System (SINAN), coordinated by the Ministry of Health using a passive surveillance system¹⁶. In 2007, the system was updated and fully implemented nationwide. That same year, the state of Rio Grande do Sul reported its first autochthonous dengue case⁹. Data on the number of confirmed cases, deaths, and municipalities with dengue occurrences in Rio Grande do Sul between 2007 and September 2023 were obtained from SINAN, provided by the Brazilian Ministry of Health^17,18. Cases included in this study were confirmed based on clinical-epidemiological or laboratory criteria, as outlined in the Ministry of Health’s Epidemiological Surveillance Guide¹⁹.

Laboratory confirmation of dengue involves the detection of viral RNA using molecular techniques such as reverse transcription-polymerase chain reaction (RT-PCR), the identification of the NS1 antigen through immunochromatographic testing or enzyme-linked immunosorbent assay (ELISA), or serological testing to detect specific IgM antibodies. Serological testing is recommended from the sixth day after symptom onset to minimize the risk of false-negative results. Additionally, cases can be confirmed using clinical-epidemiological criteria, which require identifying a suspected case in an individual residing in or having traveled to an area with confirmed dengue transmission and presenting symptoms consistent with the disease¹⁹. Once confirmed, whether by laboratory or clinical-epidemiological criteria, cases are recorded in SINAN, ensuring comprehensive monitoring and surveillance of dengue cases across the country.

Statistical analysis

Using data collected over a 17-year observation period, the incidence and mortality rates (per 100,000 inhabitants) and lethality (%) were calculated based on population projections from the Brazilian Institute of Geography and Statistics (IBGE). The 17-year period was chosen to encompass long-term trends and minimize the influence of short-term fluctuations. Data were analyzed on an annual basis to preserve the temporal resolution necessary for capturing year-to-year variations and trends, which are critical for understanding the dynamics of dengue transmission and the impact of interannual factors. Trends in incidence rates were analyzed using generalized additive models (GAM). Furthermore, the incidence rates were mapped to confirm the existence of spatial and spatiotemporal clusters, thereby identifying priority municipalities for dengue surveillance and control in the state^20,21.

Spatial descriptive analyses

For spatial description, we constructed annual maps of incidence rates (cases per 100,000 inhabitants) by municipality. In all descriptive maps created in the present study, the Jenks natural break classification method was used to determine the class intervals used for visual representation²².

Spatial autocorrelation

For each 3-year period, a spatial pattern analysis of dengue cases was performed using Moran’s Global Index and Local Indicators of Spatial Association (LISA) to analyze the association and clustering in the areas²³. The triennial interval was chosen to reduce the influence of short-term variability, ensuring sufficient data aggregation to detect stable spatial patterns and trends. To perform Moran’s Global and LISA, we created a first-order neighborhood matrix (Queen) that considers neighbors as shared border areas or vertices.

The Global Moran’s Index was calculated to analyze the spatial autocorrelation for the entire dataset (the entire study region), which provides a single measure for the set of all municipalities characterizing the entire study region. LISA was used to analyze the spatial distribution pattern and intensity of clusters at the municipality level, with a significance level of p < 0.05. The LISA index must meet the following criteria: (i) have, for each municipality, an indication of significant spatial clusters of similar values around the municipality and (ii) the sum of the LISAs for all municipalities should be proportional to the Global Moran’s I²³.

The results of the LISA test can be interpreted as follows: positive values (between 0 and + 1) indicate positive autocorrelation, suggesting that the values in municipalities tend to be similar to the values of neighboring areas (High-High or Low-Low). Negative values (between 0 and -1) suggest an inversely proportional correlation, meaning that if a particular municipality shows values above average, the surrounding municipalities tend to be below average (High-Low and Low–High). The results found for the LISA are presented in the form of thematic maps, where: (i) high–high correlations identify municipalities with high dengue incidence surrounded by other municipalities with high dengue incidence; (ii) low–low correlations identify municipalities with low dengue incidence surrounded by municipalities with low dengue incidence; (iii) high–low correlations represent municipalities with high dengue incidence surrounded by municipalities with low dengue incidence; and (iv) low–high correlations identify municipalities with low dengue incidence surrounded by municipalities with high dengue incidence. The municipalities identified as high-high in the results of LISA were considered high-priority areas²⁴.

In this study, Global Moran and LISA were created using GeoDa software, version 1.22.

Spatiotemporal analysis

A space–time scan cluster analysis was performed to identify high- and low-risk spatiotemporal clusters of dengue occurrence in Rio Grande do Sul²⁵. This technique is used to identify clusters of events in time and space, thus allowing recording of the number of expected and estimated cases at each location²¹. Spatiotemporal clusters for the dengue incidence were created using the following information: population, number of cases, expected cases, annual cases, observed/expected cases, and location. To perform the tests, information regarding each municipality was inserted into the software: (1) number of cases, (2) year of infection, (3) population average of the 3 years that make up the 3-year period studied, and 4) geocode of each municipality. This information was entered for the entire period (2007-september/2023).

The discrete Poisson model was used to identify spatiotemporal clusters and to perform the scan statistic, with the following settings: 2007–2023 study period, no geographic overlap of clusters, clusters of maximum size equal to 50% of the exposed population, circular sets, 999 repetitions, and time accuracy standardized at 1^26,27. This model considers the space and time in which cases occur^26,27,28. The significance test of the identified clusters was based on a comparison of the null distribution obtained by Monte Carlo simulation. To compare different areas, the program presents the relative risk (RR) and likelihood ratio of each cluster, which represents the relationship between the risk of occurrence of the injury within the cluster and that outside it²⁶. The analyses were carried out using SatCan software version 10.2.1.

Generalized additive models

Generalized Additive Models (GAM) represent an extension of the generalized linear model and are alternatives for modeling nonlinear relationships that do not have a defined shape²⁹. They are based on non-parametric functions called smoothing curves, in which the data defines the association shape^29,30.

The general formula for the GAM with Poisson likelihood is as follows:

where \({y}_{i}\) is the observation I and i = 1,…, T (last observed time period), \({\text{\rm Z}}_{i}\) is the mean number of cases observed at the time i, \({b}_{0}\) is the intercept term, \({s}_{j}\) is the spline function applied to the predictor \({x}_{ij}\), and \(k\) is the number of knots, a parameter that defines the number of windows into which the dataset will be broken, defined by the users³¹.

The response variable was the number of confirmed dengue cases in the year i with a Poisson distribution, with the population as the offset term and a spline function as the continuous time variable³². A Poisson distribution with a log link was used^32,33,34. Smooth terms were used to select a GAM regression model. We considered the following smoothing terms to select the best model fit: (i) thin-plate regression splines, (ii) uchon splines, (iii) cubic regression splines, (iv) B-splines, and (v) P-splines³⁴. The knot-based penalized cubic regression splines exhibited the best performance. The unbiased risk estimator (UBRE) was scaled according to Akaike’s information criterion (AIC) (generalized case) (R, 2010). The UBRE and percentage of deviance explained were used to identify the appropriate smoothness and select the best model fit³⁴ (S1 Table and S1 Figure).

Analyses were performed using the R packages mgcv and ggplot2³⁶.