Ethics statement

All data on dengue cases were part of the information routinely collected for surveillance purposes by the local health authorities. All data were anonymized prior to analyses. No ethical approval was required.

Study area

The Autonomous City of Buenos Aires (hereon, Buenos Aires) has a population of ~2.9 million in an area of 204 km² and is part of the largest urban conglomerate of South America which encompasses ~13 million people (29% of Argentina population). The city has vast areas dominated by apartment buildings and some industrial areas; ~24% of the dwellings are single-family houses (hereon, houses) and harbor ~30% of the inhabitants, the remainder living in apartments. A large proportion of houses are located in middle-class, low-rise residential areas usually with small gardens or courtyards and many with roof terraces. A portion of the most underprivileged people in the city live in slums (population, ~186,000; area, 9.3 km²), characterized by very precarious housing, high population density, poor or no urban planning (i.e., limited access to running water and proper sewage), poor garbage collection services, and high violence, among other characteristics. The temperate climate of Buenos Aires has hot humid summers (average low, 19.3°C; average high, 29.1°C; mean rainfall, 384.4 mm) and rather cold winters (average low, 8.1°C; average high, 16.2°C). These temperatures allow Ae. aegypti development approximately from October to May [45]. DENV transmission in Argentina was first recorded in 1997, with minor seasonal transmission, until 2009 when the first major outbreak took place and the first autochthonous cases were registered in Buenos Aires (240 confirmed cases, 600 suspected cases) [46]. In 2016 the hitherto largest outbreak in Argentina took place, with 41,207 confirmed autochthonous cases, mostly concentrated in Northeastern Argentina [47]. Since DENV introduction, local health authorities have implemented control measures intermittently, focusing on breeding sites year-round and focal control during outbreaks. In 2016, although no effectiveness was measured, focal control was very scarce with little chance of having any relevant effect on transmission.

Data

Data on dengue cases during the 2016 season (end of 2015 to autumn 2016) were obtained from cases reported by healthcare facilities to the SNVS (National System of Health Vigilance, Ministry of Health of Buenos Aires City) and from medical records. These data included all cases for which dengue was considered a possible diagnosis. Following laboratory studies (PCR, NS1 antigen detection, and/or serology) or epidemiological assessment, cases were classified as confirmed (positive PCR or symptomatic with epidemiological link), probable (symptomatic positive to NS1 or IgM without epidemiological link), suspected (compatible symptoms without lab or epidemiological confirmation), or discarded (IgM negative or alternative diagnosis confirmed) [48]. Some cases included serotyping of DENV, particularly for initial cases. Health personnel registered the date of onset of symptoms, date of first medical visit, and travel history in the 15 days prior to onset of symptoms. A case was considered imported if the person had recently travelled to a region (either another country or province within Argentina) with an ongoing dengue outbreak at that time. Cases were georeferenced to their reported residential address, either street and number, intersection or slum name and block number. For cases with an address indicated by the name of a slum but with no street or block indication (14.4% of cases in slums) a random location within the slum was assigned to avoid concentrating those cases in a single point. Cases with incomplete or incongruent address (1.9% of confirmed and probable cases) could not be georeferenced and were considered in non-spatial analyses but not in spatial and spatiotemporal analyses. Confirmed cases represented 62.1% of cases assessed, while 5.3% were probable and 32.6% suspected ones. Only confirmed and probable cases were considered for the current analyses.

Data on daily mean temperature and daily rainfall were obtained from the National Meteorological Service for the weather station in Aeroparque, the local airport, which is located in the city and is representative of temperature and rainfall within the city. Demographic data at census tract level were obtained from the 2010 national survey [49]. Census tracts, the highest resolution available for census data, had a median area of 3.55 ha (90% interquantile range, 1.24–12.10 ha) and a median population of 786 (440–1,229). For the hierarchical modeling, census tracts were grouped into census fractions which had a median size of 39.4 ha (13.2–110.9 ha) and a median population of 7,994 (5,433–11,334). Large uninhabited areas (totaling 34.3 km²) were manually identified based on satellite images [50] and knowledge of the terrain and excluded from analyses. Demographic and geographical information on slums was obtained from the City Housing Institute (https://vivienda.buenosaires.gob.ar/integracion/). The distinctive characteristics for defining slums are that they have no urban planning, building regulations are not followed, and most ownership is irregular [51].

Data on monthly inbound border crossings were obtained from the migration office of Argentina (Dirección Nacional de Migraciones). Data had a monthly resolution for each individual checkpoint with a bordering country with registered dengue transmission. Since it was not possible to get information on the origin of each person entering Argentina, only overland checkpoints were considered, assuming that incoming people had stayed in the neighboring country corresponding to each checkpoint. These checkpoints were 500–2000 km away from Buenos Aires. Nevertheless, this traffic is informative of the international movement of inhabitants of Buenos Aires since the vast majority of people traveling to neighboring countries with DENV transmission tend to go overland through these checkpoints.

Temporal variations: Effects of temperature and rainfall

In order to understand the overall temporal variations in DENV transmission, the weekly incidence rate ratio (IRR), calculated as the number of new cases during week t divided by the number of new cases during week t - 1, was assessed in relation to temperature and rainfall, regardless of the spatial dimension. Temperature governs Ae. aegypti and DENV development and its effects are a result of the specific time series of temperature. Therefore, we summarized the cumulative and delayed effect of temperature along time using a temperature-sensitive model of the basic reproduction number, R₀(T) [52]. This model accounts for the effects of temperature on development, behavior, and mortality of Ae. aegypti, and on the EIP. By calculating the daily effects of temperature on these processes, we estimated the developmental history of the mosquitoes transmitting DENV on each given day. This enabled to account for the delay between hatching and adult biting and, therefore, to calculate the rainfall at the approximate time of hatching of each cohort and estimate R₀(T) for each day accounting for the cumulative effects of recent temperatures. In this we considered the incubation period in humans [12,53]. The absolute value of R₀ indicates when transmission is sustainable but depends on additional elements (e.g., density of people, frequency of human-mosquito contact, housing, and vector control) not captured in this model. Therefore, since our goal was only to estimate temperature cumulative effects, R₀(T) was rescaled from zero to one–which keeps R₀(T) = 0 (when no transmission is possible) unaltered–. Further details in S1 Methods.

Spatial variations: Hierarchical modeling of demographic variables

Assessing which variables are–statistically–related with the spatial distribution of autochthonous cases requires to account for the history of cases in each part of the city, even when the assessed variables may not change temporally. For instance, a very favorable area for transmission (plenty of Ae. aegypti and opportunities for human-mosquito contact) may show no cases just because no infected people introduced DENV in that specific area. This is particularly relevant when a low proportion of all inhabitants becomes infected as in the 2016 outbreak. Additionally, an unknown–but estimable–fraction of cases goes unregistered [54].

A hierarchical spatiotemporal model was built to assess the effects of demographic covariates on DENV transmission outside of slums while accounting for the temporal autocorrelation of cases. Slums were excluded as their specific social and environmental characteristics are distinctively associated with overall poor health. Excluding slums allowed better understanding of which factors were related to DENV transmission in the rest of the city. The spatial unit, i, was the census fraction; the time unit, t, was a week. Due to the likely possibility of overdispersion in data, the number of cases was assumed to follow a negative binomial distribution with mean μ_it and success probability ϕ. The mean, μ_it, was modeled as the linear combination of: i) the joint effects of the demographic covariates v_j given the number of previous cases, y_i,t-1, occurring in census fraction i on the previous week, t - 1 ii) the contribution of neighboring cases, z_i,t-1, on the previous week in fractions immediately adjacent to i, iii) the contribution of overall cases, Σ_i y_i,t-1, in all census fractions of the city on the previous week, and iv) an overall constant term γ₀. Overall cases were considered because it was hypothesized that a higher overall incidence in the city implied higher chances that any inhabitant got infected somewhere in the city regardless of its neighborhood suitability.

The number of autochthonous DENV cases resulting from the negative binomial distribution became the number of trials of a binomial distribution. This was a way of accounting for missing observations of cases. Thus, α would be the probability of observing/detecting a DENV case in a week given that it was present in that week (complete formulation of the model in S1 Methods).

For each census fraction i the considered covariates were the fraction of dwellings that were houses (as opposed to apartments), the fraction of dwellings with poor (as opposed to good) building quality, the fraction of population having no further than high school studies, the fraction of unemployed population, and the fraction of homes with ≥2 people per room (as indicative of overcrowding) (Fig 1). These covariates were centered and standardized to facilitate comparison between their effects. The chosen prior distributions were weakly informative and ensured that μ_it ≥ 0 and 0 < α ≤ 1 (S1 Methods).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Distribution of inhabitant density, demographic covariates, and dengue autochthonous cases in Buenos Aires.

Covariates are quantified as the fraction (or proportion) of dwellings or people–depending on the covariate–in each census fraction that correspond to the considered category. These five covariates were included in the hierarchical model. All autochthonous cases occurring during the outbreak are displayed as red points. Grey shaded areas are uninhabited. Green polygons show the limits of slums. The spatial unit displayed is the census fraction. Map base layer downloaded from the open-data repository of the Government of the City of Buenos Aires (https://data.buenosaires.gob.ar/dataset/informacion-censal-por-radio).

https://doi.org/10.1371/journal.pntd.0009465.g001

The model was fitted to observed data using Markov Chain Monte Carlo sampling using the Nimble package [55,56] in R [57]. Briefly, four chains were run, each with 90,000 total iterations, 9,000 burn-in iterations, and a thinning interval of 3. The effect of each covariate v_j was evaluated according to which fraction f of the posterior samples of its coefficient β_j had the same sign as the mean of the posterior. Further details can be found in the S1 Methods.

Spatial structure and foci detection: Spatiotemporal analysis

To describe the overall spatial distribution of cases, global spatial analyses were performed using the mark-connection function [58]. The mark-connection function evaluates the spatial distribution of qualitatively marked points within a ring of given radius and width. This contrasts with more popular statistics (such as Ripley’s K- and L-functions) that evaluate distribution within a disc of a given radius. The disc implies that effects at shorter distances obscure effects at larger distances [59]. Considering a ring avoids this issue; thus, the mark-connection function enables characterization of patterns by assessing independently for each distance the contribution to the observed spatial pattern.

In turn, local spatiotemporal analyses were performed to identify areas of statistically significant spatiotemporal clustering. This, applied in real-time, would serve as a tool to prioritize and target control efforts during an outbreak. These analyses were performed weekly using the local K-function calculated for each point on a uniform grid of 100 m cells [60]. The K-function enabled assessment of the whole area surrounding each reference point within a given radius, instead of only a ring as with the mark-connection function. Likewise, the time window considered for the local analysis spanned from week t–τ to week t, where τ is analogous to the spatial radius [61]. Such cumulative function was chosen, instead of a ring function, because the global analyses showed an almost monotonical decrease of clustering with distance and the interest was in detecting clusters rather than characterizing them. To define what radius and time window to use, the average distance and time of clustering due to local transmission at a weekly base was characterized performing a global spatiotemporal analysis using the mark-connection function. The temporal dimension was incorporated by considering all cases within a given space-time ring analogous to the spatial ring. The initial phase of the outbreak would provide the best estimates of typical clustering distance since at that time most clusters were still small, more likely reflecting the transmission pattern from a few cases.

For both global and local analyses, the observed patterns were compared to random patterns generated by random labeling (labels being “infected” and “not infected”) among the fixed location of the homes of all inhabitants in the city [59]. Analyses only evaluated the location of infected individuals (either observed or simulated). The random labeling ensured that the simulated random patterns accounted for the underlying population distribution, thus, distinguishing the observed clustering process from artifacts arising from the heterogeneous population distribution. The locations of the inhabitant homes were roughly estimated by uniformly distributing the number of inhabitants of each census tract among the area of the census tract. Excluding large uninhabited areas helped reducing distortions in population spatial distribution. For each analysis 500 simulations were performed and used for building 95% confidence envelopes. In the case of the local analysis, cells with a K-function value above the confidence envelope for that cell was considered as having clustered cases within the radius and time window specified. All spatial analyses were performed using Spatstat package version 1.64–1 [62] for R.

Overall dynamics

Out of a total of 5,202 reported dengue cases in Buenos Aires, 422 (8.1%) were allegedly imported cases whereas the remaining 4,780 cases were considered autochthonous. The outbreak began approximately at the turn of the year with an increasing number of reported imported cases (Fig 2). Concomitantly, autochthonous cases increased, exhibiting a sharp growth from mid-February to mid-March, reaching steady numbers for ~4 weeks, until the beginning of April. By mid-May, the outbreak was over (Fig 2). Out of the 5,104 georeferenced cases, 1,682 (33.0%) were located in slums, with 1,538 (32.8%) autochthonous cases (Fig A in S1 Results). The dynamics of new cases was similar between slums and the rest of the city (Fig B in S1 Results). The proportion of imported cases in slums and outside slums did not differ significantly (Fisher test, P = 0.32). DENV per-capita incidence in slums was higher than outside slums (overall 7.3 times greater), even after stratifying by the different demographic covariates considered for the hierarchical model (Fig C in S1 Results).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Overall dynamics of dengue transmission, temperature, and rainfall in Buenos Aires during the 2016 outbreak.

A) Weekly cases of dengue according to date of onset of symptoms and origin of infection. B) Daily mean temperature (black line) and weekly rainfall (grey bars).

https://doi.org/10.1371/journal.pntd.0009465.g002

Onset of transmission: Imported cases

The major origin registered for imported cases was Paraguay (60.4%), followed by Northeastern Argentina (21.1%), Brazil (7.6%), northern South America and The Caribbean (3.1%), and Northwestern Argentina (2.6%). Most imported cases in slums came from Paraguay (83.3%) and Northeastern Argentina (9.0%), whereas outside slums these origins represented 47.2 and 27.9% of importations, respectively (Table 1). People incoming overland from Paraguay represented 64.3% (1.2 million people in January 2016) of all border crossings with neighboring countries registering dengue transmission, followed by people incoming from Brazil (19.2%, 353,662 people) (Fig D in S1 Results).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Number of imported dengue cases in Buenos Aires according to likely origin of infection and residential address located in or outside slums.

https://doi.org/10.1371/journal.pntd.0009465.t001

A total of 319 (6.1%) of the confirmed cases was positively serotyped. DENV-1 was detected in 97.8% of the autochthonous and in 94.7% of the imported serotyped cases (Table 2). The remaining serotyped cases were DENV-2 and DENV-4. Two of these cases were located in different slums, whereas the others were located elsewhere, away from other cases registered with their serotype. Serotyping was relatively more frequent for imported than for autochthonous cases, as most serotyping was concentrated during the beginning (January and February) of the outbreak.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Number of dengue cases in Buenos Aires according to serotype and likely origin of infection.

https://doi.org/10.1371/journal.pntd.0009465.t002

Temporal variations: Effects of temperature and rainfall

Taking the normalized R₀(T) as a summary of the cumulative effects of temperature, the weekly IRR followed closely R₀(T) dynamics (Kendall correlation coefficients: median = 0.68, 95% CI = 0.61–0.75, P < 0.05), evidencing the strong effects of temperature on DENV transmission (Fig 3). In fact, R₀(T) exhibited very low values until mid-December, and reached relatively high values at the beginning of January (Fig 3B), precisely when the outbreak began. At the end of February, after the sharpest increase in autochthonous cases (i.e., the maximum IRR), R₀(T) fell steadily, concomitantly with the approximate plateau (IRR ≈ 1) and later decrease (IRR < 1) displayed by autochthonous cases (Figs 2A, 3A and 3B). In contrast, the weekly rainfall at the estimated time of egg hatching of the Ae. aegypti supposedly involved in transmission did not show a clear correspondence with transmission dynamics (Kendall correlation coefficients: median = -0.18, 95% CI = -0.41 –-0.04, P > 0.05) (Fig 3A and 3C).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Dengue transmission follows closely R₀(T) but not rainfall in Buenos Aires during the 2016 outbreak.

A) Weekly incidence rate ratio (IRR) of autochthonous cases; the dashed line indicates the IRR value of steady weekly cases (IRR = 1). B) Normalized temperature-dependent basic reproduction number, R₀(T), as a summary of the cumulative effects of temperature along time on transmission. C) Weekly rainfall at the estimated time of egg hatching of Ae. aegypti that were putatively involved in DENV transmission on each given date. Black dots indicate the value corresponding to the beginning of the epidemiological week. Shaded areas indicate the 50, 75, and 95% confidence intervals.

https://doi.org/10.1371/journal.pntd.0009465.g003

Spatial variations: Hierarchical modeling of demographic variables

The hierarchical modeling identified the fraction of houses per census fraction and the fraction of inhabitants with no formal studies beyond high school as the most influential covariates on DENV transmission outside slums, both with a positive effect on the number of cases. Unemployment showed a weaker effect with 95% of its posterior being negative. The remaining covariates–building quality and overcrowding–showed effects not clearly different from zero, with f ≤ 0.87 (Table 3). The total number of autochthonous cases in the city on the previous week had a small contribution (γ₁ = 1.1 x 10⁻³ per case), whereas the effect of adjacent neighbors was negligible (γ₂ = 5.1 x 10⁻⁴ per case). The hierarchical model estimated that the probability α of observing an autochthonous case in a week given that it was present averaged 0.16 (95% CI: 0.11–0.24). The fitted value of the success probability ϕ of the negative binomial distribution showed that a certain degree of overdispersion remained in the model data. Model residuals distributed randomly in space, evidencing that the model captured the spatial structure present in the observed number of cases (Fig E in S1 Results). Additionally, 98.3% of the observed values fell within the 80% interquantile range of the distribution of the predictive posterior. All these results held when considering only confirmed cases as well as a Poisson instead of a negative binomial distribution (Tables A-C in S1 Results).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Hierarchical spatiotemporal modeling of dengue transmission during the 2016 outbreak in Buenos Aires.

https://doi.org/10.1371/journal.pntd.0009465.t003

Spatial structure: Global spatial analyses

The overall spatial distribution of imported cases was clustered at <600 m (Fig 4A). Excluding the 265 imported cases in slums showed that imported cases in the rest of the city (144 cases) were not clustered between them. In turn, autochthonous cases were clustered in and outside slums (and overall) up to 600 m and 900 m, respectively (Fig 4D and 4F). The spatial association between all imported and all autochthonous showed clustering up to ~800 m (Fig 4G) but this owed to the contribution of the clustering in slums (Fig 4I). When excluding cases in slums, autochthonous cases were not clustered around imported cases (Fig 4H).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Imported dengue cases clustered only in slums, whereas autochthonous cases clustered citywide.

The mark-connection function, p, assessed the spatial distribution between the groups of selected cases in Buenos Aires. A, B and C) Imported cases. D, E and F) Autochthonous cases. G, H and I) Autochthonous cases according to imported cases. A, D and G) Cases in all the city. B, E and H) Only cases outside slums. C, F and I) Only cases in slums. The black line shows the values for the observed pattern. The red line shows the mean value for the random patterns. The shading represents the 95% confidence interval for the random patterns. Random patterns were generated by random labeling among the inhabitants of the city. Observed values above 95% confidence interval denote clustered distribution.

https://doi.org/10.1371/journal.pntd.0009465.g004

The global spatiotemporal analysis performed weekly for different space-time rings showed that cases outside slums were not statistically clustered until the end of February, precisely when transmission rose sharply (Fig F in S1 Results). Until then, cases were clustered mostly at 250–400 m within the current week as well as with respect to cases one week earlier; clustering was less often observed for time lags of two or three weeks. During the following weeks (March and first half of April), clustering distance increased with respect to cases in the 1–3 previous weeks (as clusters consolidated and increased in size) ranging 500–800 m, while clustering within the same week became rather negligible.

Foci detection: Local spatiotemporal analysis

Based on the global spatiotemporal analysis (Fig F in S1 Results) we considered a radius of 300 m and a time window encompassing only the current week for the local spatiotemporal analysis. This identified many of the intuitively recognized patterns, particularly those in less populated areas (Fig 5). Overall, 53.1% of cases belonged to clusters. Within slums, 91.3% of cases and 80.0% of the population were in clusters; outside slums, 34.9% of cases and 20.4% of the population were in clusters. Exposition to DENV transmission was also longer in slums than outside slums: 41.9% of the population in slums against 0.3% outside slums were in clusters for ≥10 weeks, whereas for 1–9 weeks 38.1% in slums and 20.2% outside slums were in clusters (Fig G in S1 Results). Increasing the radius and/or time window assessed translated in only slightly more cases but much more population included in clusters (Fig H in S1 Results). Increasing the time window slightly increased the size and duration of clusters (Fig H in S1 Results), mainly because of the “memory” that widening the time window implied, as areas on a given week could be identified as clusters based on cases on the previous weeks, even without cases on the current week (Fig I in S1 Results). This became particularly evident and eventually misleading during the end of the outbreak.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Weekly local spatiotemporal analysis identified multiple dengue transmission foci in Buenos Aires.

Foci are indicated by orange areas. Autochthonous cases are displayed as black points. The displayed dates indicate the beginning of the corresponding week. The time window considered encompassed only the current week (i.e., τ = 0), with a radius of 300 m. Map base layer downloaded from the open-data repository of the Government of the City of Buenos Aires (https://data.buenosaires.gob.ar/dataset/informacion-censal-por-radio).

https://doi.org/10.1371/journal.pntd.0009465.g005