Immune-mediated protection and enhancement of dengue drives patterns of infant cases in Brazil

Abstract

Infants represent a vulnerable population for severe outcomes from infectious diseases. For dengue virus, maternal antibodies can cause severe disease in infants aged 5-12 months through antibody-dependent enhancement, but the role of maternal immunity in neonates aged 0-30 days has not been quantified. Our study leverages the re-emerging setting of Brazil, where the proportion of mothers with anti-dengue antibodies varies across space and time. Infant dengue burden has risen 11-fold from 2000-2024, in line with expansion in the general population. We fit mechanistic models to infant dengue and severe dengue cases reported through surveillance to disentangle the competing roles of maternal immunity, transmission, and age on the risk of dengue and severe dengue in infants across the first year of life. We observed peaks in severe dengue risk in neonates and at 7-8 months of age. These two peaks were explained by competing mechanisms. Infants born to seropositive mothers had lower overall risk of dengue, but increased risk of severe disease from 5-12 months, than those born to seronegative mothers. As dengue incidence increases, and vaccination becomes available, maternal antibodies may decrease infant burden but increase the risk of severe disease.

Introduction

Across pathogens, infants are particularly susceptible to severe disease due to the underdeveloped neonatal innate immune system^1,2,3,4. Transfer of maternal antibodies during pregnancy and breastfeeding, therefore, plays a key role in protecting newborns from infection during this period⁵. However, for dengue virus (DENV), a mosquito-borne arbovirus that is endemic in over 100 countries, the waning of maternally derived antibodies can increase the severity of disease through antibody-dependent enhancement (ADE)^{6,7,8,9,10,11}. The increase of anti-DENV antibodies in mothers following rapid global expansion of dengue in the last two decades¹² may explain the rise in severe infant dengue in many countries, including Brazil, that had previously reported a low burden of severe disease in this age group. Discerning the underlying risks to infants in the Americas and other emerging settings is a key step towards understanding how DENV burden will evolve in the future, and towards designing and implementing effective interventions, especially vaccines.

Our ability to quantify the role of maternal antibodies in the rise in severe infant dengue in the Americas and other emerging settings is limited by four factors. First, transmission has increased across all ages due to the invasion and increasing abundance of the Aedes mosquitoes that transmit arboviruses^13,14,15. Second, increased burden of reported cases could also arise from changes in diagnostic capacity, surveillance practices, and awareness in the population¹⁶. Third, maternal antibodies may reduce the risk of dengue infection^17,18, meaning that maternal immunity has two conflicting effects on infant risk (Fig. 1A). Because the bulk of evidence for the immunological drivers of infant risk comes from highly endemic settings where most mothers are seropositive^{6,8,9,10,18,19,20,21}, the risk profile of infants born to seronegative mothers is not well understood. This knowledge gap limits our ability to quantify these conflicting effects of maternal antibodies on severe dengue in infants in settings with lower maternal seroprevalence. Fourth, the profile of risk across the first year of life likely varies independently of maternal antibodies due to the inherent fragility of neonates and due to behavioural factors including reduced exposure to mosquitoes for the youngest infants¹⁰, introducing separate mechanisms that can shape the age distribution of infant cases (Fig. 1A). As a result, modelling approaches considering multiple interacting mechanisms are required to disentangle the contribution of maternal immunity to the evolving burden of dengue in infants.

Fig. 1: Proposed hypotheses regarding infant dengue risk profile, and schematic of mechanistic model structure.

A The risk profile of infant dengue and severe dengue may change over the first year of life due to protection against infection when antibodies are high, or antibody-dependent enhancement of severe disease. Additionally, the inherent susceptibility to infection, disease, and propensity for reporting may decrease over the first year of life, while exposure to mosquitoes may also increase due to behavioural changes. B The population dengue model infers dengue force of infection ($\lambda$) from the age distribution of reported cases in individuals aged >1 year old. This estimated $\lambda$ is a key input to the infant dengue model. C Infants born to seronegative (blue) and seropositive (red) mothers can experience different age-specific hazards of becoming a reported dengue case (left) and odds of severe disease upon infection (middle), which together lead to different age distributions of reported severe cases. As maternal seroprevalence varies from 0% (blue), to 50% (purple), and 100% (red), the age distribution of reported cases (left) and severe cases (right) changes. These age distributions are compared to the observed age distributions to infer the age profile of infection, reporting, and severe disease among children with and without maternal anti-dengue antibodies. Illustration of antibody in C from NIAID NIH BioArt Source (bioart.niaih.nih.gov/bioart/250).

Here we make use of the unique setting of Brazil, where dengue reemerged in 1986 following 20 years without widespread transmission^13,22. Spatial heterogeneity in re-emergence and transmission has led to variable accumulation of maternal immunity over time across the Brazilian states^{23,24,25,26,27}, affording us an opportunity to discern the impact of changing maternal seroprevalence on neonate and infant dengue risk.

Results

Burden of infant dengue has increased in Brazil, with two age peaks in severe disease incidence

We obtained surveillance data, including 186,837 infant DENV cases and 3100 severe DENV cases, from all 27 states of Brazil from 2000 to 2024²⁸. Across Brazil, we found an 11-fold increase in infant dengue incidence from 2000-2004 to 2020-2024. The increase was greater in the South (dark blue), Southeast (orange), and Central-West (pink) regions, consistent with the spatiotemporal pattern of re-emergence in the population (Fig. 2A, Fig. S1). Incidence of severe infant DENV exhibited substantial annual heterogeneity with no clear trend (Fig. 2A); burden was highest in the Central-West and Southeast regions. Among children, infants represent a high-risk group for severe disease: infants made up a greater share of severe dengue cases among under-18-year-olds than of all dengue cases, and this share rose overall from 2000 to 2024 (Fig. S2).

Fig. 2: Spatiotemporal patterns of infant dengue cases and severe dengue cases in Brazil.

A total reported incidence rate of infant cases and severe cases over time by region. The solid black line represents cases aggregated across Brazil, while transparent coloured lines represent regions, with the legend key shown in the inset map; B age distribution of reported infant dengue cases and severe dengue cases, aggregated across Brazil and across the years 2000-2024; C average age of all infant cases and severe infant cases, by region and time. Lines represent average age across regions and time periods, coloured according to the legend in the panel A inset. A map of Brazil was obtained from GADM.

Aggregated over the 25 years, the probability of severe disease in this population exhibited a bimodal pattern: one peak in the neonates (0–30 days in age) and another in 7–8-month-olds (Fig. 2B, Fig. S3). The second peak aligns with the peak age of 8 months among infant hospitalisations in Thailand⁹, which has been shown to be driven by antibody-dependent enhancement (ADE) from maternal antibodies. The bimodal age distribution was observed in some states and time periods (Fig. S4–8), but the pattern was often unclear and affected by small numbers, and where observed, the relative magnitude of cases among 5-12-month-olds and neonates differed. The average age of severe infant cases increased from 2000 to 2024 (Fig. 2C) in the South by four months and in the Central-West by two weeks, while in the North (grey), the average age decreased by 5 weeks from 2000 to 2024. An increase in the age of severe infant cases is consistent with a shift from children born without anti-DENV antibodies, who can acquire severe dengue as neonates, to those born with anti-DENV antibodies, who can acquire severe dengue through ADE. Meanwhile, the average age of all infant cases has been similar over time, with small decreases in the Northeast (green) and small increases in the South (Fig. 2C).

Estimation of spatiotemporal dengue risk and maternal seroprevalence using case data

Under our current understanding of infant dengue, two primary factors affecting infant case numbers in a given state and year are the dengue risk and the proportion of infants born with maternal antibodies. In general, the age distribution of cases is informative of historical transmission, with locations with very young or very old age of reported cases likely having higher seroprevalence²². If maternal seroprevalence prevented neonates from infection, we would expect lower and higher case age to be associated with higher age of infant cases (Fig. S9). However, case age is associated with incidence and reporting, and each of these quantities displays spatial variation and temporal trends (Fig. S9), limiting the ability of a statistical analysis to identify specific drivers of risk.

We developed a mechanistic model of dengue and severe dengue incidence in infants by age to explicitly account for the observed bimodal age distribution in infants as well as spatiotemporal differences in reporting and the non-linearity between age of reported cases and infection history^29,30. The model is divided into two independent modules: first, a model that estimates population-level dengue hazard (or annual, cumulative force of infection, FOI) and maternal seroprevalence from the age distribution of reported non-infant cases in Brazil (Fig. 1B); and second, a model of the hazard and severe disease risk among infants over the first year of life. We restricted this model and the data to parameterise it to 2000-2014 due to uncertainty in estimates of dengue transmission after the Zika epidemic, and the following results pertain to this time period. Table S1 shows a description of all data sources used to parameterise both models.

The estimated FOI based on the age distribution of cases in the general population was positively correlated with incidence (Fig. S10) and showed spatial variation with lower estimated FOI in the South region and the highest FOIs in the Northeast and Southeast regions, consistent with case burden prior to and during this period. Maternal seroprevalence also reflects historical dengue transmission, with lowest seroprevalence in the South region and highest in the Northeast and Southeast regions (Fig. 3A), with specific states exhibiting seroprevalence over 90% (Rio Grande do Norte/RN, Pernambuco/PE, Amazonas/AM, and Rio de Janeiro/RJ), and most states exhibiting only small rises in seroprevalence from 2000 to 2014 (Fig. 3B). Across Brazil, estimated maternal seroprevalence increased modestly from 56% to 64%. Five studies have been conducted in Brazil estimating DENV seropositivity in pregnant or post-partum women: three studies found seroprevalence of 91-94% in Salvador, BA^31,32,33 (vs. model-estimated 66% in Bahia State); one study found seroprevalence of 95% in Recife, PE³⁴ (vs. model-estimated 95% in Pernambuco State); and one study found seroprevalence of 54% in Goiânia, Goiás/GO²³ (vs. model-estimated 55% in Goiás State). Published serosurveys conducted among adults in Brazil, extracted from the reviews in Salgado et al.²⁶ and Vicco et al.³⁵ similarly show varying consistency between published and model-estimated seroprevalence (Fig. S11). While there is agreement in high seroprevalence in Rio de Janeiro/RJ, Amazonas/AM, Rio Grande do Norte/RN, and Pernambuco/PE, and moderate seroprevalence in São Paulo/SP, Ceará/CE, and Goiás/GO, the model may underestimate seroprevalence in Bahia/BA and overestimate seroprevalence in Minas Gerais/MG.

Fig. 3: Spatiotemporal variation in maternal dengue seroprevalence.

Model-estimated maternal dengue seroprevalence by state averaged across 2000-2014 (A) and change in estimated maternal dengue seroprevalence from 2000 to 2014 (B). Maternal seroprevalence is estimated from the age distribution of case data. In both panels, darker colours represent higher values. The mechanistic model was fit to dengue case data from 2000 to 2014. Map of Brazil was obtained from GADM.

States with very low estimated maternal seroprevalence generally had low infant incidence (Fig. S12), suggestive of minimal current or historical transmission. Among other state-years, there was a trend of lower dengue incidence in infants in states with higher estimated maternal seroprevalence (Fig. S12), consistent with maternal protection.

Mechanistic model of age-dependent infant risk uncovers protective and harmful effects of maternal antibodies

We developed a model of infant dengue risk to test the consistency of proposed immunological mechanisms with the spatiotemporal pattern of infant case age distribution. A model schematic showing hypothetical age profiles of dengue case hazard and severe disease stratified by maternal serostatus is shown in Fig. 1C. Briefly, we derive expressions for the reported dengue and severe dengue incidence in infants by month of age, based on FOI estimates from the population dengue model (Fig. 1B). Model-derived age profiles of reported dengue disease and severe disease were compared to the observed age distribution of cases and severe cases in infants, and we used Bayesian inference to estimate the underlying mechanistic model parameters representing the association between age, maternal antibodies, and dengue outcomes.

We incorporated age-specific infection (representing increased frailty of neonates aged 0-30 days; Fig. 1A and C), age-specific reporting, age-specific severe disease risk (Fig. 1C), differential severe disease risk among infants born to seropositive mothers (Fig. 1C), and reporting rate varying across states and time. Finally, we allowed protection at birth that decays with age. We estimate separate protection for infants born to seronegative mothers (through behavioural protection, i.e. reduced risk in all neonates due to reduced exposure to mosquitoes, that has been hypothesized to lead observed low incidence of infections in neonates¹⁰; Fig. 1A) and to seropositive mothers (i.e. immune-mediated in addition to behaviour-mediated; Fig. 1A). In both seronegative and seropositive children, this protection leads to reduced hazard of infection at one month relative to the general population, and the protection decays exponentially as the infant ages. A lower hazard for infants born to seropositive vs. seronegative mothers is consistent with immune-mediated protection. We fit 24 primary models to the data (Table S2): a model with and without each independent mechanism (severe disease risk varying by age and mother’s serostatus, and one or both of age-specific infection hazard and age-specific reporting), and with two separate models for spatiotemporal variation in reporting probability.

There was evidence for increased hazard of infection among neonates born to seronegative mothers (hazard ratio (HR) 2.96, 95% credible interval (CrI) 2.79 to 3.13 relative to the population), and a small decrease in hazard among one-month-olds born to seronegative mothers (HR 0.96, 95% CrI 0.88 to 1.00). Severe disease risk exhibited a rise and fall among infants born to seropositive mothers, reaching a peak OR of 2.63 (95% CrI 2.30 to 3.02 relative to a child born to a seronegative mother) at 5.8 months (95% CrI 4.52 to 6.72 months), consistent with ADE. Finally, infants born to seropositive mothers had a lower hazard of infection than infants born to seronegative mothers among neonates (HR 0.37, 95% CrI 0.34 to 0.41) and among one-month-olds (HR 0.38, 95% CrI 0.35 to 0.43) (Fig. 4, Table S3). The best-fit model (Table S2) fit the overall age profile of infant cases and severe cases well (Fig. 4) and displayed good convergence according to trace plots (Fig. S13). Some parameters displayed correlation in the posterior (Fig. S14), but identifiability was demonstrated through a simulation study (Figs. S15, 16). The model captured some spatial and temporal trends in age of cases (Fig. S17), including the lower age of infant cases in the South, and higher age of severe cases in the Southeast and Northeast.

Fig. 4: Mechanistic model fit to case and severe case age distribution in Brazil and estimated infant dengue risk profiles.

(Top row) Best-fit model-estimated median (solid line) and observed (points) reported dengue cases by age (left), best-fit model-estimated and observed severe dengue cases by age (right). Dengue case and severe case numbers are aggregated across Brazil and the years 2000–2014. (Bottom row) Median model-estimated age profile of infection hazard ratio (left) and severe disease odds ratio given infection (right) relative to a 12-month-old born to a seronegative mother, stratified by maternal serostatus (blue line: seronegative mother, red line: seropositive mother). 95% credible intervals of the aggregated monthly case numbers, hazard ratio, and odds ratio are represented as ribbons but are not always visible due to the low width of the intervals relative to the y-axis scale.

In a sensitivity analysis incorporating uncertainty in FOI estimates into the model, parameter estimates from the model and model fit were similar (Fig. S18 and S19). Reporting probabilities of infant cases did not show a clear temporal trend and had high estimated annual variation (Fig. S20). In addition, although the fit to age distribution of infant cases by state is good (Fig. S21), the fit to severe infant cases is not good in all states (Fig. S22) – in particular, underestimation of severe infant cases in Rio de Janeiro, Espírito Santo, and Maranhão, overestimation of severe infant cases in São Paulo, and failure to capture aggregate temporal trends in severe infant case burden (Fig. S23).

Finally, we used our model to predict changes in infant risk with increasing maternal seroprevalence. Across infants of all ages, increasing maternal seroprevalence is associated with lower incidence of DENV but higher incidence of severe DENV, and increasing age of infant cases and severe cases (Fig. S24), although the absolute differences are small.

Discussion

Arbovirus surveillance data from Brazil reveals that the burden of dengue virus in infants has increased exponentially over the past 25 years, with faster growth in the Central-West, Southeast, and South regions in which dengue has recently expanded. We used mechanistic models to understand how maternal immunity has contributed to the spatiotemporal pattern of the age distribution of infant dengue, and in particular how risk is distributed between neonates and older infants. Our study produced two major findings. First, we demonstrated that spatiotemporal patterns in infant dengue cases are consistent with immune-mediated protection to children born to seropositive mothers. Second, the age distribution of severe infant cases has two peaks: one in neonates (0–30 days) and one at 7–8 months of age. We showed that the second peak is consistent with the proposed model of antibody-dependent enhancement (ADE), increasing the risk of severe disease in children born to seropositive mothers, while the first peak is explained by increased susceptibility to infection among neonates.

ADE has been proposed as a major mechanism causing severe dengue, both in individuals experiencing a second infection and in infants. Epidemiological studies of cases in Southeast Asia and paediatric cohorts in Thailand³⁶ and Nicaragua³⁷ have demonstrated the clinical effect of ADE. O’Driscoll et al. ⁹ used immunological evidence together with the age distribution of hospitalised infant cases to derive the age profile of severe disease risk in Thailand from 1974-2019, where the mean age of hospitalised cases was 7.5 months, the peak was 8 months, and there were no hospitalisations among neonates. In Brazil, this peak is observed at 7-8 months, but the peak in severe disease among neonates afforded us an opportunity to further disentangle underlying frailty from the role of maternal antibodies in this re-emerging setting. Although this study relied on the rich surveillance data from Brazil, we expect the findings to be applicable to other settings in Latin America that have undergone re-emergence and expansion of dengue serotypes over the last few decades. Data on the age of infant cases in months is necessary to apply this model, but reproduction of our findings in contexts in which that data is available would strengthen our understanding of infant dengue risk in the Americas.

There are three outstanding mechanisms that we did not address that warrant further research. Dengue antibody kinetics may vary between children born to dengue-monotypic and dengue-multitypic mothers. A birth cohort enroled in the Brazilian city of Recife in 2011-12 found that infants born to DENV-3-positive mothers lost neutralising antibodies by 2-4 months³⁸, suggesting that the window of ADE risk is at younger ages for children born to monotypic compared to multitypic mothers. If the observed ADE peak is only experienced by children born to multitypic mothers, the changing burden of infant severe dengue will be determined by the distribution of monotypic and multitypic mothers in the population. Second, given the immune interactions between ZIKV and DENV antibodies^39,40, it is plausible that maternal ZIKV serostatus may influence infant dengue risk and vice versa. Finally, we did not consider vertical transmission of dengue; given the available evidence, the risk of vertical transmission is likely significant^41,42,43.

The importance of answering questions regarding the aetiology of severe infant dengue should be viewed in the context of the coming availability of dengue vaccines to the general population. Mass vaccination of adolescents will have a significant impact on their immune profile, which may alter the infant dengue profile born to mothers who have previously been vaccinated in the next decade. If there are differences between the experience of a child born to a seronegative, monotypic seropositive, and multitypic seropositive mother, vaccination could have positive or negative effects on infant risk. Given the urgency of this question, it is imperative to design studies to address these questions using multiple approaches, including measurement of correlates of protection and risk in infants and mothers experiencing vaccination and/or infection in cohorts using both active and passive surveillance, and long-term follow-up of participants in Phase II and III vaccine trials.

The model fit to severe cases was lacking for individual states and time periods. Such lack of fit may be driven by errors in the estimated maternal seroprevalence, particularly in the North region which has had limited incidence from 2000–2014 to inform historical transmission and in the Southeast which underwent rapid changes during this period. In addition, the relative reporting rate of severe vs. non-severe cases was assumed to be one, but severe cases are likely reported more frequently than non-severe cases. If relative reporting of non-severe cases increased over time, this would contribute to our underestimation of severe cases in later years. Another mechanism not considered in this model is the effect of serotype interactions which could drive up the burden of severity. For example, the re-introduction of DENV-2 caused an epidemic with high numbers of severe cases in Rio de Janeiro in 2008⁴⁴, which was not captured in our model (Fig. S22).

This study has several limitations. As an ecological study, comparing population-level infant dengue rates to maternal seroprevalence, it is unable to infer associations at the individual level. Due to a lack of a nationally representative serosurvey, together with high heterogeneity within states in dengue transmission, we used data on reported cases to derive estimates of maternal seroprevalence, and therefore these estimates are likely subject to measurement error. We found that estimates in Bahia and Minas Gerais were inconsistent with published serosurveys in mothers; however, a serosurvey conducted in rural Bahia state found seroprevalence to DENV-1, 2, and 4 of 53-60%, and to DENV-3 of 2%⁴⁵, emphasising that within-state heterogeneity complicates comparison of location-specific seroprevalence and statewide estimates. Development of methods for catalytic models to incorporate serosurveys conducted at specific times and locations and allow for fine-scale spatial variation in FOI will improve our understanding of the immune landscape in the Americas. In addition, we did not explore models with variation in risk between infants born to monotypic and multitypic mothers. There is no strong evidence in the literature for differences in infant antibody kinetics between these groups, and identification of specific mechanisms driven by maternal mono- vs. multitypic status would be challenging without strong priors. A large proportion of cases were excluded from all eligible cases, primarily due to missing reliable data on age at infection. We assumed that the age distribution of these cases was similar to included cases, but changes in the pattern of missingness across space and time may explain some lack of model fit in specific states and time periods. As we excluded cases with age recorded as 0-1 days, the analysis may have underrepresented the size of the peak among neonates. Further work should be conducted using chart review or active hospital-based surveillance to improve our understanding of the age distribution of infant dengue in Brazil. Finally, our models did not account for immunity to circulation of specific serotypes; the sequential re-introduction of serotypes into Brazil may account for large outbreaks in specific state-years that are not well explained by our model⁴⁶.

As dengue expands and becomes endemic in new territories, maternal antibodies may have protective and harmful effects on infants. The situation of dengue in Brazil presents a unique opportunity to disentangle the mechanisms shaping risk in infants and inform our estimates of the evolving burden in the future. These insights will also be critical for informing surveillance in infants and the development of vaccination strategies to mitigate the impact on infant cases.

Methods

Data source and preparation

We accessed publicly available data on dengue cases reported to Brazil’s “Sistema de Informação de Agravos de Notificação” (SINAN) surveillance system (https://portalsinan.saude.gov.br/dados-epidemiologicos-sinan) from 2000 to 2024 using the R package microdatasus⁴⁷. SINAN routinely collects and collates case notification data for nationally notifiable diseases, including dengue²⁸. Analysis of this publicly available data, including identifiable information, was approved as exempt by the University of Florida Institutional Review Board (IRB# IRB202401051). Data entry is the responsibility of each municipality, and the data is checked and stored by the Secretary of Health of each of the 27 states in Brazil before being transferred to the national system. Case notification data for dengue has been collected since 2000, with improvements in reliability and comprehensiveness of data collection occurring over time²⁸. We separated cases into infant cases (age ≤365 days) and non-infant cases.

For infant cases, eligibility criteria included: reporting a suspected or confirmed (laboratory or clinically confirmed) DENV case between 2000 and 2024 in Brazil with a test not classified as “Discarded”; and age <365 days. Exclusion criteria included: having a test classified as “Inconclusive” or “Missing”; having >1 year between symptom onset and notification date (to remove cases with likely erroneous sample date or age, such as those who had date of birth entered as symptom onset date); having age in days recorded as 0-1 or >30, age in months recorded as 0 or >12, or age in years recorded as 0 (each of these reported ages were likely to be misclassified age) (a flow chart is shown in Fig. S25). The exclusion of infant cases with age recorded as 0 or 1 days was made after inspection of the data, as an unrealistically large number of cases were recorded with this age. These cases likely represent missing age. Missing data on age was assumed to be missing at random with respect to age, and other sources of missing data were assumed to be missing at random with respect to age. Figure S26 shows a trend towards lower exclusion over time, possibly due to better data recording quality, but no clear trend by region. Severe cases were defined as being classified as “Dengue with complications”, “Dengue haemorrhagic fever”, “Dengue shock syndrome”, or “Severe dengue”. The classification of “severe dengue” represents the updated WHO classification system, which was adopted by the Brazilian Ministry of Health in 2014⁴⁸. For non-infant cases, eligibility criteria were the same, except that age was required to be ≥365 days.

Estimating dengue hazard in the population

To account for the interaction between infection, severe disease, and reporting, we developed a mechanistic model for infant dengue surveillance data. This model relies on estimates of dengue FOI in the general population, along with estimates of maternal seroprevalence, to provide estimates of infant infection rates in the absence of other mechanisms. We first provide an overview of the FOI estimation model, then of the mechanistic model of infant risk. These models were fit only to surveillance data from 2000 to 2014, as the Zika epidemic of 2015-2016, together with possible immune interactions in subsequent years⁴⁹, invalidate certain model assumptions.

We adapted an existing discrete-time model for estimating dengue FOI from case data in settings where the four dengue virus serotypes co-circulate endemically³⁰ to better suit Brazilian data. The key parameter of interest in this model is annual, cumulative, per-serotype dengue FOI, which we refer to as FOI. Specifically, we extended the model allowing FOI of each of the four serotypes at any time point to differ to reflect their asynchronous (re-) emergence in Brazil and the dominance or absence of specific serotypes over long stretches of time^44,50. Mathematically, this translates to differing FOI of each serotype k at time t and state s, ${\lambda }_{{stk}}$. We express ${\lambda }_{{stk}}$ as $\lambda {{\rm B}}_{{stk}}$ where $\lambda$ is the maximal per-serotype infection hazard and ${{\rm B}}_{{stk}}$ is the proportion of this maximal hazard faced at in state s at time t by serotype k. As each state is fit separately in this model, we suppress the state subscript for the following equations. Assuming long-lived protection against the infection serotype k and no cross protection against other serotypes, the probability that an individual of birth cohort h has escaped serotype k up to age a is ${p}_{{esc}}^{k}={e}^{-{\sum }_{t=h}^{h+a}{\lambda }_{{tk}}}$. Letting $K=\{A,B,C,D\}$ be the set of four possible serotypes in circulation, the probability that the individual has acquired one infection is simply the probability of acquiring one serotype and escaping the other three. We sum over the four serotypes, i.e.

$$P\left(S=1\right)={\sum }_{k\,{{{\rm{in}}}}\,K}\left(1-{p}_{{esc}}^{k}\right){\prod }_{j\,{{{\rm{in}}}}\,K-\{k\}}{p}_{{esc}}^{\,\,j}$$

(1)

Note that the mapping between serotypes A-D and dengue serotypes 1-4 cannot be discerned from the observed datasets which are non-serotype-specific. The probability of not having acquired any infection is obtained by multiplying the escape probability over all four serotypes, i.e. $P\left(S=0\right)={\prod }_{{j}\,{{{\rm{in}}}}\,{K}}{p}_{{esc}}^{\;j}$ and the probability of having acquired more than one infection is $P(S > 1)=1-P(S=0)-P(S=1)$. The increase in proportion of individuals in birth cohort h that have acquired at least i infections between time t and $t+\delta t$ indicates occurrence of the i^th infection during the time interval, ${I}_{\delta t}(i,h,t)$.

The rest of the formulation follows the existing model. In brief, the tendency of an i^th infection being reported is ${p}_{{\rm{clinical}}}\left(i\right){\phi }_{a}\left(a\right){\phi }_{t}(t)$ where ${p}_{{{{\rm{clinical}}}}}(i)$ is the clinical tendency of an i^th infection, ${\phi }_{a}\left(a\right)$ is the piecewise constant age-specific reporting rate and ${\phi }_{t}\left(t\right)$ is the piecewise constant time-specific reporting rate. Considering ${{{\rm{Pop}}}}(h,t)$ the population size of birth cohort h at time t, we would expect the number of reported cases from this birth cohort in this time interval to be

$$\sum {I}_{\delta t}\left(i,h,t\right){p}_{{{{\rm{clinical}}}}}\left(i\right){\phi }_{a}\left(a\right){\phi }_{t}\left(t\right){{{\rm{Pop}}}}\left(h,t\right)$$

(2)

We use this expression as the expectation in a negative binomial likelihood, and estimate parameters using Bayesian Markov Chain Monte Carlo (MCMC) with priors as defined in Table S4. Parameters were estimated from the data using cmdstan v2.34.1⁵¹ with 12 independent chains, each of length 2000 (400 discarded as warm-up). Posteriors of all chains combined were considered converged when the Gelman-Rubin R-hat<1.1^52,53 and effective sample size >300 for all parameters.

We use population FOI estimates to estimate seroprevalence among mothers in each state and year. Age of mothers was derived from the Brazilian census, which provides the number of live births by age of mother in each state and year⁵⁴. Figures S27, S28 show the Gelman-Rubin R-hat statistics and effective sample sizes for annual FOI and maternal seroprevalence by state, highlighting that estimates for Acre, Espírito Santo, and Mato Grosso do Sul exhibited lower convergence than other states.

Mechanistic model of infant dengue

We derive expressions for two observed quantities: the expected number of reported incident infant cases occurring at age m months in each state-year, based on a hazard of infection and reporting probability; and the number of reported incident severe cases occurring at age m months in each state-year, based on the modelled number of infections and a probability of severe disease. For the infection, reporting, and severe disease model, we incorporate age-specific changes in hazard or risk as well as differences in children born to seronegative and seropositive mothers. For the infection model, we estimate hazard ratios of infection relative to the population-level FOI estimated from the model described in the section Estimating dengue hazard in the population and in Eqs. (1) and (2). We estimate separate parameters for (i) hazard ratio at birth for infants born to seropositive and to seronegative mothers, (ii) hazard ratio at one month for infants born to seropositive and to seronegative mothers, and (iii) decay of the hazard ratio for a one-month-old towards one (the null). These parameters allow us to estimate the age profile of infection hazard separately for infants born to seropositive and seronegative mothers, with increased frailty at birth counteracted by lower infection hazard caused by maternal antibodies (immune-mediated) and/or reduced exposure to mosquitoes (behavioural protection). A lower hazard ratio for infants born to seropositive vs. seronegative mothers is consistent with immune-mediated protection. A list of model parameters and interpretations is given in Table S5.

Incident reported infection model

We assume that the probability of a reported infant dengue case Y_stm (i.e. reported in surveillance data, and categorized as non-severe or severe) in state s in year t of age m months, with maternal seroprevalence m_st and annual population FOI λ_st, is

$$P\left[{Y}_{{stm}}=1\right]=P\left[{{{\rm{reported}}}}|{{{\rm{infection}}}}\right]P\left[{{{\rm{infection}}}}\right].$$

(3)

We assume that m takes values from 0 to 11, meaning that children are infected on the first day of the month. We model the hazard function by age as a step function, constant within a month. The hazard of infection at age m (in months) is determined by (a) the population hazard, (b) a hazard ratio for neonates aged 0–30 days that represents frailty of the youngest children and varies by maternal serostatus, and (c) maternal protection that decays with age and varies by maternal serostatus. In the absence of these mechanisms, the monthly FOI is given by $\frac{{\lambda }_{{st}}}{12}$. Accounting for these mechanisms, the hazard of infection for a child in year t, in month m of their life, born in state s is

$${\lambda }_{{stm}}^{{sn}}=\left\{\begin{array}{c}\frac{{\lambda }_{{st}}}{12}{h}_{0}^{{sn}},m=0\\ \frac{{\lambda }_{{st}}}{12}{h}_{1}^{{sn}}{e}^{-d\left(m-1\right)},m\ge 1\end{array}\right.$$

(4)

$${\lambda }_{{stm}}^{{sp}}=\left\{\begin{array}{c}\frac{{\lambda }_{{st}}}{12}{h}_{0}^{{sp}},m=0\\ \frac{{\lambda }_{{st}}}{12}{h}_{1}^{{sp}}{e}^{-d\left(m-1\right)},m\ge 1\end{array}\right.$$

(5)

for a child born to a seronegative and seropositive mother respectively. Note that this population includes children born in year t and t−1 (see Accounting for year and month of birth below). In these equations, ${h}_{0}^{{sn}}$ and ${h}_{0}^{{sp}}$ represent the hazard ratio of infection for a neonate relative to a 12-month-old born to a seronegative mother, and ${h}_{1}^{{sn}}$ and ${h}_{1}^{{sp}}$ represent the hazard ratio of infection for a one-month-old born to a seronegative and seropositive mother relative to a 12-month-old born to a seronegative mother. $d$ represents the decay of each hazard ratio towards 1 over age since birth. This model structure allows us to assess evidence for behaviour-mediated and immune-mediated protection: behaviour-mediated protection would be suggested if ${h}_{1}^{{sn}} < 1$, so that one-month-olds born to seronegative mothers have reduced hazard of infection compared to older children; if there is immune-mediated protection, ${h}_{0}^{{sp}}$ and ${h}_{1}^{{sp}}$ should be less than ${h}_{0}^{{sn}}$ and ${h}_{1}^{{sn}}$ respectively, so that children born to seropositive mothers experience reduced risk of infection compared to infants born to seronegative mothers.

To derive the proportion of infants who become infected at age m, we use the estimated maternal seroprevalence ${m}_{{st}}$. The probability of an incident infection at age m for a child born in state s in year t is given by

$${{{{\rm{pinf}}}}}_{{stm}}={{m}_{{st}}}{e}^{{-\varLambda }_{{stm}}^{{sp}}}\left(1-{e}^{-{\lambda }_{{stm}}^{{sp}}\,}\right)+(1-{m}_{{st}}){e}^{-{\varLambda }_{{stm}}^{{sn}}}\left(1-{e}^{-{\lambda }_{{stm}}^{{sn}}}\right)$$

(6)

where ${\Lambda }_{{stm}}^{{sn}}$ is the cumulative hazard up to age m experienced by children in year t, born in state s to seronegative mothers, and ${\Lambda }_{{stm}}^{{sp}}$ having an equivalent interpretation for children born to seropositive mothers. Specifically,

$${\Lambda }_{{stm}}^{{sn}}=\left\{\begin{array}{c}0,m=0\\ {\sum }_{m=0}^{m-1}{\lambda }_{{stm}}^{{sn}},m\ge 1\end{array}\right.$$

(7)

$${\Lambda }_{{stm}}^{{sp}}=\left\{\begin{array}{c}0,m=0\\ {\sum }_{m=0}^{m-1}{\lambda }_{{stm}}^{{sp}},m\ge 1\end{array}\right.$$

(8)

We model the probability of reporting as age-dependent, specifically

$${{{\rm{logit}}}}({\alpha }_{{stm}})=\left\{\begin{array}{c}{{{\rm{logit}}}}\left({\alpha }_{{st}}\right)+{\beta }_{r0},m=0\\ {{{\rm{logit}}}}\left({\alpha }_{{st}}\right),m\ge 1\end{array}\right.$$

(9)

where ${\alpha }_{{st}}$ is the population-level reporting rate in state s in year t and ${\beta }_{r0}$ is the log-odds ratio of reporting for a neonate relative to a 12-month-old.

We fit models in which reporting log-odds ${\alpha }_{{st}}$ is constant across time (i.e., one parameter ${\alpha }_{s}$ per state), and reporting log-odds is independent in each state-year (i.e., one parameter per state-year).

Accounting for year and month of birth

Because population-level FOI varies annually, children of the same age may have experienced different cumulative hazard of infection depending on when they were born (e.g., a 9-month-old case reported in January would have experienced a different cumulative hazard than a 9-month-old case reported in December). Therefore, to calculate the cumulative hazard experienced by a child in year t at age m, we need to account for the month of birth. We only have data on the month of age for cases, so we calculate the cumulative hazard for a child m months of age in state s and year t as a weighted average over children born in calendar month c (taking values 1 to 12 for January to December). The reasoning behind these equations is as follows: (i) if a child of age m was born in a calendar month c > m, their cumulative hazard will accrue from the calendar year of their birth; (ii) if a child of age m was born in January (c = 1), their cumulative hazard will accrue from the previous calendar year; and (iii) otherwise, the cumulative hazard will consist of the first c months of hazard in the year of their birth, and the remaining m-c + 1 in the year preceding their birth. Specifically, ${\varLambda }_{{stmc}}^{{sn}}=0$ for m = 0, and

$${\varLambda }_{{stmc}}^{{sn}}={\sum }_{i=1}^{m}{\lambda }_{s,t,m-i}^{{sn}}\,{for}\,c > m > 0$$

(10)

$${\varLambda }_{{stmc}}^{{sn}}={\sum }_{i=c}^{m}{\lambda }_{s,t-1,m-i}^{{sn}}\,{for}\,m > 0,c=1$$

(11)

$${\varLambda }_{{stmc}}^{{sn}}={\sum }_{i=1}^{c-1}{\lambda }_{s,t,m-i}^{{sn}}+{\sum }_{i=c}^{m}{\lambda }_{s,t-1,m-i}^{{sn}}\,{for}\,1 < c\le m$$

(12)

Finally, if ${n}_{{stc}}$ is the number of children born in state s, in calendar month c of year t, overall cumulative hazard is given by

$${\varLambda }_{{stm}}^{{sn}}=\frac{{\sum }_{c=1}^{12}{n}_{{stc}}{\varLambda }_{{stmc}}^{{sn}}}{\mathop{\sum }_{c=1}^{12}{n}_{{stc}}}$$

(13)

${\varLambda }_{{stm}}^{{sp}}$ is given by an equivalent expression.

Severe disease model

We assume that the probability of a severe infant dengue case Sev_stm in state s in year t of age m months is

$$P[{{{{\rm{Sev}}}}}_{{stm}}=1]=P[{{{\rm{reported}}}}|{{{\rm{severe}}}}\; {{{\rm{case}}}}]P[{{{\rm{severe}}}}\; {{{\rm{case}}}}|{{{\rm{infection}}}}]P[{{{\rm{infection}}}}]$$

(14)

The probability of severe disease upon infection at age m (in months) is determined by (a) the population probability of severe disease, (b) an age-specific term, representing frailty of the youngest children, and (c) antibody-dependent enhancement, applied only to children born to seropositive mothers (Fig. 1). Specifically, the probability of severe disease for a child of age m born in state s in year t is

$${{{\rm{logit}}}}({{{{\rm{psev}}}}}_{m}^{{sn}})=\left\{\begin{array}{c}{{{\rm{logit}}}}\left({{{\rm{psev}}}}\right)+{\beta }_{{{{\rm{s}}}}0,}m=0\\ {{{\rm{logit}}}}\left({{{\rm{psev}}}}\right),m\ge 1\end{array}\right.$$

(15)

$${logit}\left({{psev}}_{m}^{\,{sp}}\right)=\left\{\begin{array}{c}{{{\rm{logit}}}}\left({{{\rm{psev}}}}\right)+{\beta }_{s0},m=0\\ {{{\rm{logit}}}}\left({{{\rm{psev}}}}\right)+f\left({m;}\mu,\kappa,\rho \right),m\ge 1\end{array}\right.$$

(16)

where psev is the population probability of severe disease, ${\beta }_{s0}$ is the log-odds ratio of severe disease for a child <1 month relative to the general population, and $f({m;}\mu,\kappa,\rho )$ is a function that describes the rise and fall of log-odds of severe disease for children born to seropositive vs. seronegative mothers due to ADE. For this function, we chose a form similar to the four-parameter beta distribution, with bounds at $m={m}_{0}$ and $m={m}_{1}$:

$$f\left({m;}\mu,\kappa,\rho \right)=\rho {\left[\frac{\left(m-{m}_{0}\right)}{\mu \left({m}_{1}-{m}_{0}\right)}\right]}^{\mu \kappa }{\left[\frac{\left({m}_{1}-m\right)}{\left(1-\mu \right)\left({m}_{1}-{m}_{0}\right)}\right]}^{\left(1-\mu \right)\kappa }$$

(17)

This function is unimodal and takes the value 0 at the bounds of the interval $[{m}_{0},{m}_{1}]$. The parameter $\mu$ represents the fraction of the interval $[{m}_{0},{m}_{1}]$ where the mode occurs, so that the mode is at ${m}_{0}+\mu ({m}_{1}-{m}_{0})$ and the function takes the value $\rho$ at the mode. Therefore, ${e}^{\rho }$ represents the odds ratio of severe disease risk among children born to seropositive vs. seronegative mothers at the peak of ADE. The concentration $\kappa$ determines the shape of the distribution, with higher $\kappa$ translating to faster decay away from the mode. This parameterisation of the function was chosen to improve the model's identifiability, since the mode and concentration are not correlated. We choose the bounds to be 1 and 11 months, to avoid correlation with the population severe disease probability psev and the odds ratio among neonates ${\beta }_{s0}$. Finally, we assume that reporting probability is the same for severe and non-severe cases.

The probability of an incident severe infection at age m for a child born in state s at time t is therefore

$${{{{\rm{psevinf}}}}}_{{stm}}={m}_{{st}}{{{{\rm{psev}}}}}_{m}^{{sp}}{{{{\rm{pinf}}}}}_{{stm}}^{{sp}}+(1-{m}_{{st}}){{{{\rm{psev}}}}}_{m}^{{sn}}{{{{\rm{pinf}}}}}_{{stm}}^{{sn}}$$

(18)

Finally, to estimate the number of infants at risk of having an infection at age m in state s and year t, ${n}_{{stm}}$, we account for the calendar month and year of birth as above. ${n}_{{stm}}$ is the sum of all children born (${b}_{s,t,c}$) in months c = 1 to 12-m of year t and in months c = 12-m + 1 to 12 of year t−1, i.e. ${n}_{{stm}}=\mathop{\sum }_{c=1}^{12-m}{b}_{s,t,c}+{\sum }_{c=12-m+1}^{12}{b}_{s,t-1,c}$.

Likelihood of the model

The likelihood of observing y_stm reported cases is given by a Poisson distribution with location parameter ${n}_{{stm}}{{{{\rm{pinf}}}}}_{{stm}}{\alpha }_{{stm}}$, and the likelihood of observing sev_stm severe cases is given by a Poisson distribution with location parameter ${n}_{{stm}}{{{{\rm{psevinf}}}}}_{{stm}}{\alpha }_{{stm}}$.

Model evaluation and selection

To verify identifiability of model parameters, we simulated data from the model using the input FOI and maternal seroprevalence estimates from the population-level model, and fit the infant model to this simulated data with a given set of model parameters. We performed 100 simulations with all mechanisms present (age- and maternal serostatus-dependent hazard of infection, age-dependent reporting odds and severe disease odds, and ADE) and constant population-level reporting odds over time. We summarised the simulation results in two ways: the median estimated parameter value compared to the input true value; and normality of the percentile-based residuals⁵⁵.

Convergence was checked by calculating that Gelman-Rubin statistic^52,53 was less than 1.1 for all parameters for all parameters, and we visually inspected trace plots and pairs plots to examine correlation between parameters in the posterior. We compare the model-estimated age distribution of cases and severe cases by age across all states and the time to the observed age distribution. We compared models to each other using expected log-predictive density (ELPD) and leave-one-out information criteria (LOOIC), and we assess estimated parameters and credible intervals to assess specific hypotheses:

• 1. To understand whether maternal seroprevalence is associated with risk of severe disease in children aged 5-12 months, as in previous studies, we estimate ρ, the log-odds ratio for children born to seropositive mothers at the peak of ADE. If the 95% CrI for ρ includes 0, this provides evidence against the ADE hypothesis.

• 2. To understand whether the spatiotemporal pattern in age distribution of infant cases is related to maternal seroprevalence and thus support a behaviour-mediated or immune-mediated mode of protection, we estimate $\frac{{h}_{0}^{{sp}}}{{h}_{0}^{{sn}}}$ and $\frac{{h}_{1}^{{sp}}}{{h}_{1}^{{sn}}}$ which correspond to relative reduction in hazard in infants at birth and one month for infants born to seropositive vs. seronegative mothers. If the 95% CrIs for these ratios exclude 1, this provides evidence for the immune-mediated protection hypothesis, as it suggests that children born to seropositive mothers experience stronger protection at birth and one month than children born to seronegative mothers. In contrast, if these CrIs include 1 this suggests that maternal seroprevalence has no association with age distribution of infant cases, and would provide evidence for behaviour-mediated protection.

Finally, there is uncertainty around the population-level FOI estimates. We fit two versions of our model. In one, we ignore uncertainty around these FOI and maternal seroprevalence estimates by including the posterior means as fixed values in our inference, thus underestimating uncertainty in our model parameters. In a second model, we accounted for uncertainty in FOI and maternal seroprevalence by fitting a two-component mixture normal model to the posterior for each FOI and maternal seroprevalence estimate, and using this distribution as a prior. For all other model parameters, we used weakly or non-informative uniform priors. All analyses were conducted in R version 4.4.1⁵⁶. Parameter estimation for the infant model was performed using rstan version 2.32.7⁵⁷. In addition, we used the R packages dplyr (version 1.1.4)⁵⁸, tidyverse (version 2.2.0)⁵⁹, ggplot2 (version 4.0.1)⁶⁰, grid (version 4.4.3)⁶¹, gridExtra (version 2.3)⁶², lubridate (version 1.9.3)⁶³, arrow (version 17.0.0.1)⁶⁴, geodata (version 0.6-2)⁶⁵, sf (version 1.0−17)⁶⁶, scales (version 1.4.0)⁶⁷, ggh4x (version 0.3.1)⁶⁸, and cowplot (version 1.1.3)⁶⁹ for data processing and visualisation. All code and data needed to reproduce these analyses are available on Github⁷⁰.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.