Epidemiological data.

In Mexico, the reporting of suspected dengue cases by healthcare centres is mandatory, and dengue cases are recorded by the healthcare system through the National Epidemiological Surveillance System (Sistema Nacional de Vigilancia Epidemiológica; SINAVE). Anonymised line-list data on dengue cases in Mexico for 2020–2023 are available online from the Ministry of Health (Secretaría de Salud) [25], and data for 2016–2019 are available upon request from the Instituto Nacional de Acceso a la Información (INAI), folio 330026924000362. This study is reported as per the Reporting of Studies Conducted Using Observational Routinely-Collected Data (RECORD) guidelines (S1 Checklist).

The combined dataset comprised 1,062,240 entries, with information on the federal entity (entidad federativa; administrative level 1, from here on referred to as “state”, Fig A in S1 Text) and municipality (municipio; administrative level 2) of residence. We only considered probable or confirmed cases (of which there were 833,761) and excluded international cases (those with state of residence listed as either the United States of America [Estados Unidos de América], other Latin-American countries [otros países de Latinoamérica], or other countries [otros países]) and one person with an incorrect age entry from the analysis, leaving 833,629 cases. Case status (estatus caso) was defined as either probable (probable), confirmed (confirmado), or discarded (descartado), where confirmed cases indicate probable cases with DENV confirmed by laboratory techniques, and discarded cases are probable cases with DENV not confirmed by laboratory techniques [26]. The definition of a probable case can be found in the Standardised Procedures Manual for the Epidemiological Surveillance of Vector-Borne Diseases (Manual de Procedimientos Estandarizados para la Vigilancia Epidemiológica de las Enfermedades Transmitidas por Vector) [26].

All patients were categorised as either hospitalised (hospitalizado) or ambulatory (ambulatorio, i.e., non-hospitalised) (variable tipo paciente). Sex (sexo) was defined as either man (hombre) or woman (mujer), and death as either yes (sí) or no (no). The reverse transcription polymerase chain reaction (RT-PCR) test result (resultado PCR) was defined as either DENV-1, DENV-2, DENV-3, DENV-4, or no designated serotype (sin serotipo aislado). Not all probable dengue cases are tested for DENV infection, and not all those positive for DENV have their serotype categorised, with the proportion of probable cases selected for testing dependent on local DENV circulation according to surveillance and on the severity of the case [26]. The case fatality rate (CFR) was calculated as the proportion of confirmed and probable cases who were recorded as having died, and incidence as the total number of probable and confirmed cases per capita per year, with the population of each state each year calculated as below. For the descriptive analysis, uncertainty was calculated using exact binomial confidence intervals (Clopper-Pearson method).

Population data.

Population data at administrative levels 1 and 2 for Mexico were reconstructed from the 2015 intercensal survey and from the 2020 census [6,27]. For each state, we calculated the change in population of each 5-year age group (where A₁ = 0–4 years old, A₂ = 5–9 years old, …, A₁₅ = 70–74 years old and A₁₆ = 75+ years old) between 2015 and 2020 and reconstructed the yearly rates of population change, assuming that those whose age was listed as NA in the 2015 intercensal survey or in the 2020 census (0.07% of people in the 2015 intercensal survey and 0.22% [273,386/126,014,024] in the 2020 census) were distributed according to the age distribution of the rest of the population with known age in that year in each state. The population growth rate of age group , , , was then used to estimate the population size of each age group in each state between 2015 and 2020, and we extrapolated the population size for 2021, 2022 and 2023 assuming the same growth trend, using the formula:

(1)

where is the population of age group in year , .

Serotype-specific case incidence.

We reconstructed the number of serotype-specific cases from the total number of cases reported by state and the respective proportions of cases positive for each serotype among hospitalised and non-hospitalised cases separately. To account for the fact that severity impacts the likelihood a person receives an RT-PCR test [26], we stratified cases with a designated serotype by hospitalisation status and calculated the prevalence of each serotype separately for hospitalised and non-hospitalised cases for each state and year. We then assumed that the prevalence of each serotype among the probable/confirmed cases without a designated serotype was the same as among those with a designated DENV serotype, and reconstructed the number of hospitalised and non-hospitalised cases for each serotype by multiplying the total number of hospitalised and non-hospitalised cases by, respectively, the serotype-specific proportion positive for each serotype among hospitalised and non-hospitalised cases. We then combined these estimates to generate the total reconstructed serotype-specific case incidence for each state and year.

We assumed that the serotype-specific distribution of cases with unknown hospitalisation status (0.006% [53/833,629] of probable and confirmed cases) was the same as was observed among non-hospitalised cases. Building on the available age-distribution of serotype-specific case data (Figs B and C in S1 Text), we assumed that the observed serotype proportions were constant over all ages. We reconstructed serotype-specific case incidence in a given state for the longest sequence of consecutive years for which there were more than 10 RT-PCR positive test results in each year. Furthermore, for a given state, we assumed that a serotype was circulating within the modelled years only if there were more than 10 RT-PCR positive cases for that serotype in at least one of the modelled years.

Non-serotype-specific models.

To estimate the FOI in each of Mexico’s states over the 8 years of data (2016–2023 inclusive), we used both time-constant and time-varying catalytic models, which assume that each serotype confers lifelong immunity against reinfection by the same serotype. The time-constant catalytic model, Model 1, developed by Imai and colleagues [16], assumes that the per-serotype FOI (denoted in the model) is constant throughout the 8 modelled years and historically prior to 2016. We also adapted the time-varying catalytic model developed by O’Driscoll and colleagues [17], denoted Model 2, to allow the per-serotype FOI to vary each modelled year (now denoted ), and used five different historic FOI estimates for the years prior to 2016 ( to reconstruct the initial immunity profile of the population at the start of the time-series data ( is applied to the 20 years directly before the start of the study period, 1995–2015, to the 20 years prior to that [1975–1995], and similarly for , and ). Both Model 1 and Model 2 assume that the four DENV serotypes are equally transmissible and that tertiary and quaternary cases are not symptomatic, and hence not reported (full model equations for Model 2 can be found in S1 Text).

In both models, the reported incidence of dengue cases in age group 1–4-year-olds) in year , , can be described by equation (2) below. Model 1 assumes that the reporting rate of primary infections is smaller or equal to the reporting rate of secondary infections, whereas Model 2 relaxes this assumption. Model 2 also allows for children under 15 years old to report at a different rate to people older than 15 years to reflect potential differences in the diagnosis and reporting of dengue cases through paediatricians.

(2)

In equation (2), is the annual incidence of primary dengue infection in age in year , the annual incidence of secondary dengue in age in year , the population of in year reconstructed as defined in equation (1), the reporting rate of secondary cases, the reporting rate of primary cases (defined as in Model 1), the width of age group (e.g., for , the age group containing 5–9 year olds,), and a multiplier that allows children under 15 years old to report at a different rate than people older than 15 years in Model 2 (which was fixed to 1 in Model 1). We excluded cases in children under one year old from the analysis due to maternally transferred immunity [28], and modified and accordingly (in the case of , using the proportion of 0-year-olds among children aged 0–4 years in the 2020 census [6]).

As a sensitivity analysis, and in line with O’Driscoll and colleagues [17], we tested the assumption that the FOI was constant-in-time for all years prior to 2016 (Model 3), a stricter assumption on the historic FOI than used in Model 2. We assigned Normal(0,1) prior distributions, with parameters constrained to be positive and at most 1 (or at most 0.25 in the case of per-serotype FOI estimates, see Table A in S1 Text), to all parameters in Models 1–3, apart from the scaling factor used in Models 2 and 3, , which was assigned a Normal(1,1) prior distribution (with constrained to be positive and such that reporting rates were never above one). We performed model selection using: the deviance information criterion (DIC), the widely-applicable information criterion (WAIC), and leave-one-out information criterion (LOOIC) [29–31].

Serotype-specific model.

We developed a variation of Model 2 to allow serotype-specific time-varying FOIs, where denotes the proportion of the total FOI attributable to each serotype ( denoting the 4 serotypes, denoting the year, with varying depending on the data available). Equations (3)–(12) describe this new model (Model 4), where denotes the proportion of the population of age susceptible at the end of year , the proportion of the population monotypically infected with DENV serotype at age at the end of year , the proportion of the population multitypically infected at age at the end of year, the annual primary incidence of DENV serotype in age in year , and the annual secondary incidence of DENV serotype in age in year . In equations (3)–(12), denotes the set of DENV serotypes, the set of all serotypes excluding serotype . is a function that defines the historical force of infection of serotype experienced by an individual of age prior to 2016.

(3)

(4)

(5)

where ,

(6)

For ,

(7)

(8)

(9)

For ,

(10)

(11)

The serotype-specific number of cases reported in age group during year for serotype , , was modelled as:

(12)

where represents the reporting rate of secondary cases, the reporting rate of primary cases, the population of age group in year (as defined above), the width of age group , and where is, as above, a multiplier that allows children under 15 years old to report at a different rate than people older than 15 years ( , which we estimate, and otherwise).

We assigned a Dirichlet prior distribution to at each timestep and included a scaling factor ( for the historic ’s) to allow the total FOI to be below 1, which was assigned a Uniform(0,1) prior distribution. Prior distributions for the other parameters were the same as for Model 2.

Model fitting.

We fitted our models to the observed age-stratified case-notification data (Models 1–3) and to the age-stratified serotype-specific case incidence data reconstructed from the observed case-notification data (Model 4). We assume that the total number of cases (over the study period for Model 1 and over each year for Models 2–3) was Poisson distributed and that the age-specific number of cases was multinomially distributed. Similarly, for Model 4, we assume that the yearly total number of cases (over all age groups) for each serotype was Poisson distributed, and that for each serotype the yearly age-specific distribution of cases was multinomially distributed. We report the median of the posterior distribution of the estimated parameters and the 95% credible interval (CrI) to show the uncertainty.

All models were fitted to the data using the Hamiltonian Monte Carlo method implemented using Stan (RStan version 2.32.7) [32,33]. For each state, models were run using 3 chains, each with 6,000 iterations, of which the first 1,000 were considered warm-up. The chains were considered to have successfully converged if the diagnostic value Rhat was below 1.01 and the bulk effective sample size (ESS) was over 300 for all parameters, as per Stan guidelines [33]. We excluded 5 states from our modelling analyses: Baja California, Ciudad de México, Chihuahua, Tlaxcala and Zacatecas, as there was no evidence of sustained transmission over the study period in these states.

Time-constant model.

Model 1 converged for all 27 states modelled. Fig 3c shows the per-serotype FOI () estimates from Model 1 by state, showing large variations from 0.0092 (95% CrI [0.0086, 0.0097] in Durango to 0.0182 (95% CrI [0.0179, 0.0184] in Yucatán. Overall, our results suggest a higher FOI in the south of the country, particularly in the Yucatán peninsula. Estimated secondary reporting rates () varied from 0.0051 (95% CrI [0.0046, 0.0058]) in Querétaro to 0.3245 (95% CrI [0.3207, 0.3284] in Morelos, with low estimated secondary reporting rates often accompanying low estimates of per-serotype FOI. We estimate that the sensitivity of surveillance in Mexico during 2016–2023, defined as the total number of reported probable and confirmed cases divided by the total model-derived number of infections, varied from between 0.210% (95% CrI [0.196%, 0.229%]) of infections detected in Aguascalientes to 15.3% (95% CrI [15.2%, 15.4%]) of infections detected in Morelos (Fig 3d).

Time-varying models.

The model which allowed the historical FOI to vary over time (Model 2) outperformed the model with a constant historical FOI (Model 3) by DIC, WAIC and LOOIC in, respectively, 27, 23 and 25 states out of a total of 27 (Table C in S1 Text). As a result, we present the results of Model 2 below (which converged for all states according to the tests set out in Methods). We estimate lower reporting rates for primary infections compared to secondary infections in all 27 states, and the reporting rate of children under 15 to be higher than that of people older than 15 in all states but four (Coahuila, Guanajuato, Durango and Querétaro), with ranging from 0.795 (95% CrI [0.574, 1.084]) in Querétaro to 1.364 (95% CrI [1.284, 1.447]) in Puebla (Fig F in S1 Text).

Fig 4 shows the median per-serotype FOI estimate, (by definition assumed to be the same for all four serotypes) from Model 2, which is highly heterogeneous both in space and time. For instance, in Morelos, which is the state with the highest per-capita number of detected dengue cases over the reporting period, the median estimated varied from 0.0206 (95% CrI [0.0187, 0.0226]) in 2016 to 0.0051 (95% CrI [0.0046, 0.0056]) in 2022, before rising again to 0.0804 (95% CrI [0.0734, 0.0879]) in 2023. During 2023, the most recent year modelled, estimates of ranged from a median of 0.0003 (95% CrI [0.0002, 0.0004]) in Nuevo León in the north-east of the country and 0.00030 (95% CrI [0.00029, 0.00032]) in Jalisco in the west, to 0.1155 (95% CrI [0.1077, 0.1231]) in Guerrero and to as high as 0.2492 (95% CrI [0.2459, 0.2500]) in Yucatán and 0.2468 (95% CrI [0.2350, 0.2499]) in Campeche in the south-east.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Spatiotemporal heterogeneity in annual state-level per-serotype FOI.

Map of the median estimated time-varying per-serotype FOI () for each of the 27 modelled states in Mexico between 2016 and 2023, obtained using a non-serotype-specific model (Model 2), which allows for a variable historic FOI and different reporting rates for children under 15 years, fitted to line-list case data of probable and confirmed dengue cases. The shapefiles used to generate the maps in the figure above were obtained from GADM (https://gadm.org/) and can be accessed as described above.

https://doi.org/10.1371/journal.pmed.1004874.g004

Our model estimates that 10 out of 27 states had their highest DENV FOI during 2023. The states that saw the highest estimates of over the 8-year period were not, however, geographically clustered: Yucatán, Campeche and Guerrero in the south; Aguascalientes, Querétaro and Morelos in the centre; Sonora in the north; and Colima in the west. In contrast, lower FOI estimates were clustered in the centre of the country, especially in Guanajuato, San Luis Potosí and Durango.