Arboviral diseases are major global public health threats. Yet, our understanding of infection risk factors is, with a few exceptions, considerably limited. A crucial shortcoming is the widespread use of analytical methods generally not suited for observational data – particularly null hypothesis-testing (NHT) and step-wise regression (SWR). Using Mayaro virus (MAYV) as a case study, here we compare information theory-based multimodel inference (MMI) with conventional analyses for arboviral infection risk factor assessment.
A cross-sectional survey of anti-MAYV antibodies revealed 44% prevalence (n = 270 subjects) in a central Amazon rural settlement. NHT suggested that residents of village-like household clusters and those using closed toilet/latrines were at higher risk, while living in non-village-like areas, using bednets, and owning fowl, pigs or dogs were protective. The “minimum adequate” SWR model retained only residence area and bednet use. Using MMI, we identified relevant covariates, quantified their relative importance, and estimated effect-sizes (β±SE) on which to base inference. Residence area (βVillage = 2.93±0.41; βUpland = −0.56±0.33, βRiverbanks = −2.37±0.55) and bednet use (β = −0.95±0.28) were the most important factors, followed by crop-plot ownership (β = 0.39±0.22) and regular use of a closed toilet/latrine (β = 0.19±0.13); domestic animals had insignificant protective effects and were relatively unimportant. The SWR model ranked fifth among the 128 models in the final MMI set.
Our analyses illustrate how MMI can enhance inference on infection risk factors when compared with NHT or SWR. MMI indicates that forest crop-plot workers are likely exposed to typical MAYV cycles maintained by diurnal, forest dwelling vectors; however, MAYV might also be circulating in nocturnal, domestic-peridomestic cycles in village-like areas. This suggests either a vector shift (synanthropic mosquitoes vectoring MAYV) or a habitat/habits shift (classical MAYV vectors adapting to densely populated landscapes and nocturnal biting); any such ecological/adaptive novelty could increase the likelihood of MAYV emergence in Amazonia.
Citation:Abad-Franch F, Grimmer GH, de Paula VS, Figueiredo LTM, Braga WSM, et al. (2012) Mayaro Virus Infection in Amazonia: A Multimodel Inference Approach to Risk Factor Assessment. PLoS Negl Trop Dis 6(10): e1846. doi:10.1371/journal.pntd.0001846
Editor: Scott C. Weaver, University of Texas Medical Branch, United States of America
Received: March 7, 2012; Accepted: August 20, 2012; Published: October 11, 2012
Copyright: © Abad-Franch et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding:This research was funded by the International Development Research Center (IDRC), Canada, and the Fiocruz-Fapeam agreement, Brazil. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Arboviral infections are a major global public health concern; dengue is the most widespread, but other viruses in the families Flaviviridae, Togaviridae or Bunyaviridae are also emerging worldwide –. A solid understanding of the epidemiology of emerging arboviral diseases is crucial for the development and operation of functional control/surveillance systems , . However, except for dengue virus (e.g., –) and a few other arboviruses (e.g., –), risk factors for infection remain poorly understood.
Apart from overall neglect resulting in fewer epidemiological studies than would be needed , poor data analysis and presentation in published reports also hinder our understanding of arboviral infection risk factors. On the one hand, most reports are merely descriptive, thus precluding formal inference; on the other, infection survey data are often analyzed with inadequate statistical techniques. In particular, null hypothesis-testing (NHT) statistics and step-wise regression (SWR) analysis have been repeatedly criticized for their many drawbacks in the analysis of observational data (e.g., –).
Among the major practical shortcomings of NHT is the fact that p-values provide no information on the size, direction, or precision of effect estimates; such estimates, in the form of, for instance, regression slope parameters or odds ratios (with their associated standard errors and/or confidence intervals), are central to inference –. In addition, NHT p-values represent the probability of the observed (or more extreme) data, given the (presumably false) null hypothesis , . As Jacob Cohen put it, this is not “what we want to know”; rather, we want to know, at least, how likely the null hypothesis is, given the data (ref. , p. 997). Taking this argument a little further, we aim to examine the likelihood of (or strength of evidence for) several alternative, plausible hypotheses by confronting them with empirical data –. In epidemiology, this is often accomplished with the aid of statistical models. Since several candidate covariates (putative risk factors and confounders) are usually examined in different combinations, model selection procedures are used to ‘retain’ only those that appear as important in a final, ‘minimum adequate model’ on which inference is then based. The most widely used of these procedures apply step-wise algorithms in which NHT-derived p-values are used to decide whether a particular covariate should be retained or dropped from the model . Apart from relying on a mechanical application of p-values from multiple null hypothesis tests, step-wise procedures can lead to biased parameter estimates and disregard the variance component due to model selection uncertainty , , –. A framework for inference based on likelihood and information theories has been developed that avoids many of the pitfalls of NHT and SWR; it relies on comparing multiple models, representing alternative a priori hypotheses, based on both their fit to the data and their complexity –. Multimodel inference (MMI) then proceeds by examining model-averaged effect-size estimates for all the covariates of interest , , . These approaches are being increasingly applied in infectious disease epidemiology (e.g., –), but have seldom been used for assessing emerging arboviral disease risk –, –.
Here, we analyze data from a cross-sectional serological survey on Mayaro virus (MAYV) infection as a case-study to illustrate how MMI can enhance arbovirus infection risk factor analyses. MAYV is an alphavirus transmitted primarily by diurnal, canopy-dwelling mosquitoes of the genus Haemagogus , . It causes an acute, dengue-like febrile illness accompanied by rash and severe arthralgia that is often highly incapacitating , –. MAYV infection is a candidate for emergence as a major public health problem, much in the way it recently happened with the closely-related chikungunya virus when it adapted to urban Aedes mosquitoes , , . However, available epidemiological evidence suggests that MAYV transmission is largely restricted to sylvatic cycles involving non-human vertebrates, with limited spillover to human hosts who make frequent use of forest habitats in tropical South America , , –, –. Such a scenario implies that MAYV infection risk must be higher among human groups living or working regularly in well-preserved, forested landscapes than among those living in degraded landscapes or rarely entering forest habitats (e.g., children). Here we use MAYV serology (IgG) data to test this prediction. Furthermore, we compare the performance of NHT, SWR, and MMI at identifying and quantifying risk factors for MAYV infection in a typical central Amazon rural setting.
This study was approved by the Fiocruz Institutional Review Board, Brazil (Protocol 384/07); all participants gave written informed consent. Laboratory procedures involving mice followed the guidelines of the Brazilian National Council for the Control of Animal Experimentation (CONCEA) and were approved by the Institutional Review Board for Animal Research of the School of Medicine, University of São Paulo at Ribeirão Preto, Brazil (Protocol 115/2008).
In the context of a study on infectious disease ecology in the central Brazilian Amazon, we conducted a cross-sectional serological survey (see Text S1) in a rural settlement of the agricultural frontier. The settlement (N = 583 inhabitants in 158 households within a terra firme rain forest matrix) is located ~150 km north of Manaus (~1°48′S; 60°19′W) and consists of two village-like household clusters (30×50 m plots for a house with courtyard) plus extensive upland and riverbank areas with scattered households (in ~250×2000 m farm plots) and hence lower population density; old-growth forests comprise most of the ~280 km2 of the settlement. Typical upland houses are located in large forest clearings used for farming, whereas most riverbank houses, which have only boat access, are located in smaller clearings within better-preserved forest. To the north and west, the settlement shares boundaries with a large indigenous reserve composed of pristine forests. Agriculture is the most important economic activity in the settlement; the main crops are banana, manioc, papaya, black beans, rice, maize, and the native cupuaçu (Theobroma grandiflorum) and pupunha (Bactris gasipaes). Most of the harvest is consumed within the community. Only a few settlers raise cattle for commercial purposes, but many families breed other domestic animals (mainly fowl and pigs) for their own use; dogs and cats are common. Forest extractivist products include timber, Brazil nuts (Bertholletia excelsa), and a variety of medicinal herbs. Non-commercial hunting and fishing are also relatively common.
The allochthonous dengue vectors, Aedes aegypti and Ae. albopictus, have never been collected in longitudinal mosquito studies conducted in this remote settlement. At the time of our survey, Culex (Culex) and Cx. (Melanoconion) were the dominant mosquito subgenera in the village-like clusters, although Psorophora, Anopheles, and Coquillettidia were also present. Haemagogus, Sabethes, Ochlerotatus, Wyeomyia, and Trichoprosopon were only recorded in forest sites, whereas Aedeomyia, Mansonia, and Uranotaenia occurred only in crop-plots; Anopheles, Culex, Psorophora, and Coquillettidia were present in all three environments (SLBL, unpublished data).
Finger-prick bloodspots from 270 randomly selected subjects (83 households: 42 in upland, 26 in village-like, and 15 in riverbank areas) were collected onto filter paper in 2007; the sample was representative of the community and ensured 0.05 precision for an expected prevalence of 50% (required sample size, finite population correction: n = 232). Sera were tested for anti-MAYV IgG antibodies using an enzyme-linked immunoassay with infected cultured cells as antigenic matrix (EIA-ICC), following procedures described elsewhere , . Briefly, Aedes albopictus C6/36 cells were cultured in Leibovitz L-15 medium with bovine fetal serum (Invitrogen), penicillin, and streptomycin. Cells were infected with lyophilized baby mouse infected brain tissue resuspended in PBS; cells and virus suspension were incubated for 3–4 days at 28°C. Infected and uninfected (negative control) cells were transferred to TPP tissue culture plates (Sigma-Aldrich), which were incubated at 28°C for 24 h. A formalin solution was then added (18 h at 4°C) and wells washed with PBS. Non-specific binding sites were blocked for 1 h at 37°C with skim milk (5% plus 0.05% Tween-20 in PBS). Eluted serum samples (100 µL) were then added to each well pair (one with and one without viral antigen); mouse immune ascitic fluid was used as a positive control. Plates were then incubated (1 h, 37°C) and washed; 100 µL of 1:200 peroxidase-labeled anti-human (test wells) or anti-murine (positive control wells) IgG antibody (KPL Inc.) were added to each well. Plates were again incubated at 37°C for 1 h, washed, and 50 µL of ABTS peroxidase substrate (KPL Inc.) were added to each well; after 15 min at room temperature, enzyme activity was blocked with 50 µL H2SO4, and plates read in a spectrophotometer at 450 nm. In each plate, absorbance values (test well minus paired negative control well) were averaged (â) and the standard deviation (SD) calculated; the plate cut-off value for positivity was â+3SD.
Participants were interviewed for information on three groups of putative risk factors and/or confounders:
- individual-level traits: age (years), gender (male/female), and regular bednet use (yes/no);
- household-level traits: residence area within the settlement (village-like clusters, riverbanks, upland); basic sanitation (since there was no sewage system in the settlement, this covariate described whether or not there was a closed toilet/latrine in the household; yes/no); adequate solid waste disposal (yes/no; here, ‘adequate’ means that waste was eliminated from the household's surroundings, mainly by burying/burning it at a sufficiently large distance); and ownership of a crop-plot – locally known as roça and usually located in forest clearings (yes/no); and
- whether or not domestic fowl, dogs, cats, or pigs, which may represent bloodmeal sources for female mosquito vectors, were reared or kept near the household (yes/no for each one).
We first screened the dataset for associations between anti-MAYV seropositivity and putative risk factors with NHT statistics, using either Fisher's exact tests or likelihood-ratio (LR) χ2 tests from bivariate logistic regression. At this stage, we also checked for correlation between potential predictor variables. If any pair of covariates proved to be highly correlated, we planned to retain only that for which a clear hypothetical relationship with MAYV transmission could be specified; however, all correlation coefficients were <0.30 (details not shown).
We then adopted the SWR approach  that has become the conventional standard in risk factor analysis (e.g., for arboviruses, , , –). Starting with a saturated, additive logistic model including all covariates for which NHT suggested a correlation (defined a priori as those with bivariate p≤0.10), we removed, at each step, the covariate with the largest p-value (from LR tests) until all covariates in the final, ‘minimum adequate’ model  had p-values<0.05.
- fitting three subsets of models with only individual-level, only household-level, and only domestic animal covariates;
- selecting unequivocally important covariates, as quantitatively assessed by their relative importance (see below) for predicting seropositivity status in each subset;
- specifying and fitting the complete set of additive logistic regression models for the selected covariates (i.e., an all-subsets approach);
- back-checking that none of the covariates excluded after step (ii) improved the performance of the models with substantial support from the data identified in step (iii); and
- estimating weighted mean effect-sizes (see below) and the relative importance of each covariate based on the final model set.
Logistic regression models were of the simple general formwhere α is the intercept and βi represents the effect of covariate i (covi) on the (logit-scale) probability that a subject is MAYV-seropositive. Models were fit in JMP 9.0.0 (SAS Institute), with parameters estimated via maximum likelihood, and compared using Akaike's Information Criterion corrected for small sample size (AICc); AICc combines likelihood and information theories to identify, within a given set of models, those with a better compromise between fit and complexity, providing an estimate of Kullback-Leibler information loss. AICc is given bywhere Lm is the likelihood of the data given each fitted model, K is the number of estimable parameters in each model, and n is sample size.
For each model i, we calculated the variation in AICc relative to the best-ranking (lowest AICc) model (ΔAICc = AICci−AICcmin); models with ΔAICc<2 are generally taken to be substantially supported by the data. The likelihood of each model given the data was estimated as L (model | data) = exp(−ΔAICc/2); these values were then used to compute Akaike weights (denoted wi), which are normalized model likelihoods, as:The relative importance of each covariate (denoted w) was estimated, within each model set, as the sum of Akaike weights over all models in which the covariate was present; covariates with w≤0.35 were considered unimportant. Weighted mean effect-sizes (βs) were estimated, for each covariate within the final model set, as the sum of model-specific effect sizes times model-specific Akaike weights. Finally, model fit was assessed using goodness-of-fit χ2 tests and scaled generalized R2 values , withwhere L0 is the likelihood of the data given the intercept-only model, Lm is the likelihood of the data given the fitted model, and n is sample size.
Descriptive results and null hypothesis-testing
Anti-MAYV antibodies were detected in 119 serum samples (44.1%). NHT suggested no departure from a random distribution of seropositivity in relation to sex or age; 36.8% of 19 toddlers fewer than three years old were seropositive, and there was very little variation across age classes (Table 1). Surprisingly for a virus transmitted by forest-dwelling vectors, seropositivity was strongly, positively associated with living in household-like village clusters and negatively associated with living in the better-preserved riverbank areas. Using a closed toilet/latrine also increased risk, whereas regularly sleeping under a bednet and owning/rearing fowl, pigs, or dogs were apparently protective. No association was detected between seropositivity and owning cats, owning crop-plots, or whether solid waste disposal at the household level was or was not adequate (Table 1).
A saturated, additive logistic model including all covariates for which bivariate NHT suggested association with seropositivity (bold typeface in Table 1) was then built as the starting point of backward SWR. The pre-established selection criterion/procedure resulted in the sequential exclusion of the following covariates: dog (LR χ2 = 0.006, p = 0.939), fowl (LR χ2 = 0.36, p = 0.851), closed toilet/latrine (LR χ2 = 2.04, p = 0.153), and pig (LR χ2 = 1.83, p = 0.177). The SWR ‘minimum adequate’ model therefore retained just two covariates: residence area within the settlement (LR χ2 = 131.31, p<0.0001), and regular bednet use (LR χ2 = 17.36, p<0.0001); this same model was selected by standard forward SWR (details not shown). This model (Table 2) suggests that village-like cluster residents were at much higher risk of MAYV infection, independent of bednet use, and that, conversely, regularly sleeping under a bednet protected from infection irrespective of residence area. The remaining covariates were considered irrelevant when adjusted for one another.
All model sets included a null model (estimating only the intercept), which represents the hypothesis of no predictable variation in seropositivity; as expected, this model always proved to be relatively very poor at explaining the data (Tables 3,4, and 5 and Table S1). Note that this null model is the same used in LR tests and in generalized R2 calculations.
The individual-covariate model set included seven models (Table 3); both the best-performing model and model-averaged effect-size estimates (details not shown) suggested a strong protective effect of regular bednet use; it was relatively much more important (w = 0.945) than age (w = 0.280) or gender (w = 0.270), whose effects were indistinguishable from zero. Therefore, only bednet use was retained for further analysis.
Next, we fitted all possible additive household-level models; this model set can be envisaged as representing hypotheses stating that in our study setting MAYV seropositivity simply varies among households with different traits, and comprises 15 models with all combinations of four candidate covariates (Table 4). All these covariates were retained for further assessment because they had relatively high w values: residence area (w = 1.000), crop-plot ownership (w = 0.965), closed toilet/latrine (w = 0.498), and, to a lesser extent, solid waste disposal (w = 0.385).
Third, we assessed all 15 possible model specifications including the four domestic animal covariates in our dataset (Table 5). On account of their relative importance, pigs (w = 0.999) and dogs (w = 0.735) were retained for testing, whereas cats (w = 0.267) and fowl (w = 0.337) were not considered any further.
Based on these results, we finally specified and compared the 128 models with all possible combinations of important individual-level (bednet use), household-level (residence area, crop-plot, toilet/latrine, waste disposal), and domestic animal covariates (pigs, dogs); the complete model set is provided in Table S1, and the subset of models with highest support from the data (ΔAICc<2) is presented in Table 6. Figure 1 illustrates variation in ΔAICc values across the 128 models in this set.
Models were ranked according to variation in the Akaike's information criterion value of each model with respect to the best-performing model in the set (i.e., ranked by ΔAICc). Arrows highlight ΔAICc ‘leaps’ associated with the removal of the two most important covariates, residence area and bednet use. The position and ΔAICc value of the saturated model (Full) and the intercept-only model (Null) are also indicated. Note that the y-axis is in log10 scale.
The best-performing model within this set had a generalized R2 = 0.55 and a goodness-of-fit test χ2 = 10.21 (6 d.f., p = 0.116); the overall misclassification rate (model-predicted seropositivity status different from observed status) was just 0.19. These satisfactory fit metrics were similar for the rest of models in Table 6 (details not shown), and suggest that the models succeed in capturing important processes governing the relationships between covariates and the dependent variable. Adding interaction terms did not improve model performance or resulted in failure to reach convergence. Moreover, a posteriori addition of individual-level and domestic animal covariates removed in previous steps did neither improve the performance of any of the models with ΔAIC<2 nor change their relative positions (details not shown); therefore, we based inference on this 128-model set.
When ranked according to their relative importance, the covariates considered in the final model set performed as follows: residence area, w = 1.000; regular bednet use, w = 0.997; crop-plot ownership, w = 0.627; closed toilet/latrine, w = 0.501; solid waste disposal, w = 0.364; pig, w = 0.351; and dog, w = 0.308. Table 7 presents the weighted average effect size over all models in the set (βs) for each of these covariates; exp(β) values, which estimate adjusted odds ratios, are presented in Figure 2. While MMI agreed with SWR in identifying residence area and no regular bednet use as strong independent predictors of MAYV seropositivity, it also showed that further covariates, and particularly crop-plot ownership, were important factors with non-negligible effects on infection risk. In sharp contrast with NHT results, domestic animals had little or no influence on MAYV seropositivity when adjusted for other covariates in our study setting.
Covariates describe: residence area (Village: village-like household clusters; Upland: upland areas; Riverbanks: better-preserved riverbank areas); bednet use (Bednet); crop-plot ownership (Crop-plot); use of a closed toilet/latrine (Toilet/latrine); adequate solid waste disposal (Waste); and the keeping/rearing of pigs (Pig) or dogs (Dog). Estimates are presented as odds ratios (solid circles) and 95% confidence intervals. The dotted line at odds ratio = 1 represents no effect; values >1 indicate a positive effect (increased risk of infection), and values <1 a negative (protective) effect.
Arboviral infections are increasingly recognized as major public health threats. Geographic range expansions by dengue virus, West Nile virus, Japanese encephalitis virus, or, more recently, chikungunya virus have focused attention on their importance , . Other arboviruses, however, have so far remained endemic to their putative areas of origin within developing countries, and this has perhaps contributed to their neglect . Among the many viruses one could mention as examples, Rift Valley fever virus and O'nyong-nyong fever virus are of particular concern in Africa, and Venezuelan equine encephalitis virus and Mayaro virus in Latin America . With effective vaccines unavailable (except for yellow fever), preventing infection heavily relies on vector control and personal protection measures, but both perform poorly. One key weakness of efforts in this direction is that robust risk factor analyses are lacking for most of these viruses; this limits our understanding of determinants of infection and, therefore, our ability to (i) identify risk situations/areas and (ii) design improved prevention strategies.
Here, we address this gap by presenting a case-study on MAYV; in addition to providing the first MMI-based assessment of risk factors for infection with this virus, we aimed at illustrating the caveats of conventional approaches often used to analyze observational data from cross-sectional surveys – and how MMI provides a powerful alternative. We discuss our results within the hypothetical framework outlined in the Introduction: if MAYV transmission is largely sylvatic, then spillover should increase infection risk mainly among people making frequent use of forested landscapes; those living in more degraded landscapes and those rarely entering forests, such as young children, should be relatively protected.
Prior to more detailed discussion, we consider several limitations of this study that must be kept in mind when interpreting our conclusions. First, even if EIA-ICC has been shown to be specific at detecting different anti-Alphavirus antibodies , some of our positive EIA-ICC results might be due to cross-reactions. We nonetheless tested all sera for anti-Venezuelan equine encephalitis virus antibodies and found just three putatively reactive samples, one of which was MAYV-negative. Thus, we feel confident that our serological results do reflect past MAYV infections, although we cannot completely exclude the possibility of a few cross-reactions with antibodies to closely-related but rarer viruses, particularly Una virus. In addition, we are unaware of any reliable data on the typical duration of anti-MAYV IgG; if high titers last long, this might weaken the relationships between seropositivity and (current) covariate values. Finally, we treated all subjects as independent, random samples of the settlement population; however, since several members of single families typically participated in the survey, the data are expected to present some degree of non-independence. This possibility, which has not been considered in previous MAYV risk factor analyses, could result in parameter variance underestimation , especially for household-level covariates. To check for possible effects of seropositivity clustering within households, we re-ran our ‘best’ model adding a term that indexed, for each subject, whether there were other seropositive individuals living in the same house; we found evidence of moderately increased risk (βOther seropositive = 0.52±0.18; odds ratio 1.68, 95% confidence interval 1.18–2.39), but adjusted covariate effect estimates were very similar to those derived from MMI (cf. Table 7): βVillage = 2.76±0.42; βUpland = −0.57±0.32; βRiverbank = −2.18±0.53; βBednet = −0.85±0.29; and βCrop-plot = 0.53±0.34. In addition, the overall biological plausibility of MMI results and the satisfactory model fit metrics both give us confidence that our results are a fair approximation to the data-generating processes .
We report an antibody prevalence in the middle-upper range of previous cross-sectional surveys , , , –. While, as throughout rural Amazonia, malaria is considered the main vector-borne disease in our study settlement, 75% of ~1400 blood-smears from febrile patients seeking malaria diagnosis in 2004–2007 were Plasmodium-negative (SLBL and FAF, unpublished data). Even if some of these results are false-negatives, this suggests that pathogens other than Plasmodium are a major cause of acute febrile illness in the settlement. Our data, including the frequent mention by local residents of ‘joint pain’ as a typical feature of non-malarial febrile illness (unpublished observations from informal interviews), suggest that MAYV is likely involved in generating this epidemiological scenario.
But what factors modulate MAYV infection risk? Previous work is strongly suggestive of a pattern overtly dominated by forest transmission, with outbreaks sporadically recorded in rural communities embedded within rainforest environments , , –, –, –. Antibody prevalence is also typically higher, and clinical illness more frequent, among post-pubertal men, suggesting labor-related exposure involving forest activities , , , ; sex bias may nonetheless be absent in traditional communities in which both men and women make regular use of forested habitats , , . All these conclusions are however based on merely descriptive accounts or on bivariate, NHT-based data treatments, rarely with age adjustment (e.g., ); they must therefore be interpreted with caution.
Our results are in partial contrast with these findings – and, hence, partly at odds with the prevailing hypothesis of forest transmission and sporadic spillover. Notably, age- and sex-specific seroprevalence suggest that people of both sexes and all age classes were similarly exposed to MAYV in our study population of non-indigenous settlers. The youngest seropositive subject was a girl just under two years of age residing in one of the village-like clusters; seropositive <3-year-olds were living in the two village-like clusters and in two distinct upland sites. This suggests, at least for these cases, that transmission was relatively recent and geographically widespread at the scale we consider, and casts doubts on the forest-transmission-only scenario.
In addition, our data strongly suggest that residing in the more densely populated village-like clusters greatly increased MAYV infection risk, whereas living in the better-preserved, sparsely populated riverbanks was protective. This sharply contrasts with previous reports explicitly showing the opposite pattern , ; again, it also contradicts the predictions of the forest-transmission hypothesis. However, our MMI approach revealed a role for crop-plot ownership as a relatively important risk factor, which would partially reconcile both results. Note that, had we based inference on NHT-SWR only, this relationship with forest crop-plots would have gone undetected and unreported (Table 1): the ‘best’ SWR model implicitly estimates  a zero effect for this covariate, even after adjustment, whereas MMI estimates a positive, marginally non-significant effect on the risk of MAYV seropositivity (Table 7, Figure 2). This crop-plot effect was in fact larger when young children (<8-years-old, n = 56) were excluded from the ‘best’ model in Table 6 (β = 0.74±0.40 vs. β = 0.54±0.33), suggesting that crop-plot work did increase risk-exposure.
The strong protective effect of regularly sleeping under a mosquito bednet (Figure 2) is also at odds with the assumption that MAYV is transmitted only by diurnal, canopy-dwelling mosquitoes. In addition, both NHT and MMI suggested an intriguing (albeit weak after adjustment) relationship between owning a closed toilet/latrine and an increased risk of seropositivity (Figure 2). Finally, and in plain disagreement with bivariate NHT results, domestic animals had very small effects (effectively not distinguishable from zero) on MAYV infection risk (Figure 2).
Thus, when considered as a whole, which MMI allows us to do, the effects of residence area and bednet use, and the clear lack of effect of age and gender, all suggest the possibility that MAYV cycles other than the classical ones (maintained by diurnal, forest-dwelling vectors) might occur in our study settlement. Ongoing research examines two main hypotheses: (i) that an alternative, nocturnal, endophilic vector species is involved in transmission, and (ii) that some local Haemagogus populations have shifted habitat and habits, adapting to densely populated landscapes and nocturnal biting. That MAYV can naturally infect Psorophora and Mansonia – and be transmitted by Culex and Aedes – lends more support to the first scenario. Culex and Psorophora were the most abundant vectors in the village-like clusters of our study site, where non-native Aedes spp. have never been recorded. Haemagogus and Mansonia seem to prefer forest and crop-plot habitats; both could therefore be involved in more typical MAYV transmission cycles in forested landscapes, which would help explain why owning a crop-plot increases risk, particularly among >8-year-olds.
We have presented the first MMI-based assessment of risk factors for MAYV infection. The results suggest that two different, possibly overlapping MAYV transmission cycles might co-occur in a typical settlement of the Amazon agricultural frontier. If transmission by synanthropic vectors is confirmed, it could ultimately increase the risk of MAYV emergence in non-forest settings – perhaps even urban or periurban. This might have serious public health consequences , as the chikungunya example has shown , and calls for a tighter surveillance of arboviruses and their vectors in the Amazon. Finally, our analyses show how NHT and SWR result in the loss of valuable information when used to analyze observational data – a pervasive problem that is by no means particular to arbovirus epidemiology –. By revealing subtle associations that conventional risk factor analyses miss, MMI can foster our understanding of emerging infectious disease epidemiology and thus enhance disease control/surveillance systems.
The complete set of 128 models; covariates, AICc and related metrics, and number of parameters are given for each model.
We thank Ricardo Agum Ribeiro, Daniel Forsin Buss, and the settlement neighbors for their help at different stages of this project. This is contribution number 14 of the Research Program on Infectious Disease Ecology in the Amazon (RP-IDEA) of the Instituto Leônidas e Maria Deane – Fiocruz Amazônia.
Conceived and designed the experiments: FAF SLBL. Performed the experiments: GHG VSP FAF. Analyzed the data: FAF GHG. Contributed reagents/materials/analysis tools: GHG LTMF WSMB SLBL. Wrote the paper: FAF. Commented on and approved the final version of the manuscript: FAF GHG VSP LTMF WSMB SLBL.
- 1. Gubler DJ (1998) Resurgent vector-borne diseases as a global health problem. Emerg Infect Dis 4 (3) 442–450. doi: 10.3201/eid0403.980326
- 2. Institute of Medicine (2008) Vector-borne Diseases: Understanding the Environmental, Human Health, and Ecological Connections. Washington, DC: The National Academies Press.
- 3. Weaver SC, Reisen WK (2010) Present and future arboviral threats. Antiviral Res 85 (2) 328–345. doi: 10.1016/j.antiviral.2009.10.008
- 4. Forshey BM, Guevara C, Laguna-Torres VA, Céspedes M, Vargas J, et al. (2010) Arboviral etiologies of acute febrile illnesses in western South America, 2000–2007. PLoS Negl Trop Dis 4 (8) e787. doi: 10.1371/journal.pntd.0000787
- 5. McBride WJ, Mullner H, Muller R, Labrooy J, Wronski I (1998) Determinants of dengue 2 infection among residents of Charters Towers, Queensland, Australia. Am J Epidemiol 148 (11) 1111–1116. doi: 10.1093/oxfordjournals.aje.a009589
- 6. Cobelens FG, Groen J, Osterhaus AD, Leentvaar-Kuipers A, Wertheim-van Dillen PM, et al. (2002) Incidence and risk factors of probable dengue virus infection among Dutch travellers to Asia. Trop Med Int Health 7 (4) 331–338. doi: 10.1046/j.1365-3156.2002.00864.x
- 7. van Benthem BH, Vanwambeke SO, Khantikul N, Burghoorn-Maas C, Panart K, et al. (2005) Spatial patterns of and risk factors for seropositivity for dengue infection. Am J Trop Med Hyg 72 (2) 201–208.
- 8. Naish S, Hu W, Nicholls N, Mackenzie JS, Dale P, et al. (2009) Socio-environmental predictors of Barmah forest virus transmission in coastal areas, Queensland, Australia. Trop Med Int Health 14 (2) 247–256. doi: 10.1111/j.1365-3156.2008.02217.x
- 9. Hu W, Clements A, Williams G, Tong S, Mengersen K (2010) Bayesian spatiotemporal analysis of socio-ecologic drivers of Ross River virus transmission in Queensland, Australia. Am J Trop Med Hyg 83 (3) 722–728. doi: 10.4269/ajtmh.2010.09-0551
- 10. Vazquez-Prokopec GM, Vanden Eng JL, Kelly R, Mead DG, Kolhe P, et al. (2010) The risk of West Nile virus infection is associated with combined sewer overflow streams in urban Atlanta, Georgia, USA. Environ Health Perspect 118 (10) 1382–1388. doi: 10.1289/ehp.1001939
- 11. LaBeaud AD (2008) Why arboviruses can be neglected tropical diseases. PLoS Negl Trop Dis 2 (6) e247. doi: 10.1371/journal.pntd.0000247
- 12. Gardner MJ, Altman DG (1986) Confidence intervals rather than P values: estimation rather than hypothesis testing. Br Med J (Clin Res Ed) 292 (6522) 746–750. doi: 10.1136/bmj.292.6522.746
- 13. Cohen J (1994) The earth is round (p<.05). Am Psychol 49 (12) 997–1003. doi: 10.1037//0003-066x.49.12.997
- 14. Greenland S (2006) Bayesian perspectives for epidemiological research: I. Foundations and basic methods. Int J Epidemiol 35 (3) 765–775. doi: 10.1093/ije/dyi312
- 15. Hobbs NT, Hilborn R (2006) Alternatives to statistical hypothesis testing in ecology: a guide to self teaching. Ecol Appl 16 (1) 5–19. doi: 10.1890/04-0645
- 16. Whittingham MJ, Stephens PA, Bradbury RB, Freckleton RP (2006) Why do we still use stepwise modelling in ecology and behaviour? J Anim Ecol 75 (5) 1182–1189. doi: 10.1111/j.1365-2656.2006.01141.x
- 17. Anderson DR, Burnham KP, Thompson WL (2000) Null hypothesis testing: problems, prevalence, and an alternative. J Wildl Manage 64 (4) 912–923. doi: 10.2307/3803199
- 18. Burnham KP, Anderson DR (2001) Kullback-Leibler information as a basis for strong inference in ecological studies. Wildl Res 28 (2) 111–119.
- 19. Burnham KP, Anderson DR (2002) Model Selection and Multimodel Inference: A Practical Information-Theoretic Approach. New York: Springer.
- 20. Burnham KP, Anderson DR (2004) Multimodel inference – Understanding AIC and BIC in model selection. Sociol Methods Res 33 (4) 261–304. doi: 10.1177/0049124104268644
- 21. Lloyd-Smith JO, Schreiber SJ, Kopp PE, Getz WM (2005) Superspreading and the effect of individual variation on disease emergence. Nature 438 (7066) 355–359. doi: 10.1038/nature04153
- 22. van Boven M, Koopmans M, Du Ry van Beest Holle M, Meijer A, Klinkenberg D, et al. (2007) Detecting emerging transmissibility of avian influenza virus in human households. PLoS Comput Biol 3 (7) e145. doi: 10.1371/journal.pcbi.0030145.eor
- 23. Carneiro I, Roca-Feltrer A, Griffin JT, Smith L, Tanner M, et al. (2010) Age-patterns of malaria vary with severity, transmission intensity and seasonality in sub-Saharan Africa: a systematic review and pooled analysis. PLoS ONE 5 (2) e8988. doi: 10.1371/journal.pone.0008988
- 24. Ezenwa VO, Milheim LE, Coffey MF, Godsey MS, King RJ, et al. (2007) Land cover variation and West Nile virus prevalence: patterns, processes, and implications for disease control. Vector Borne Zoonotic Dis 7 (2) 173–180. doi: 10.1089/vbz.2006.0584
- 25. Winters AM, Eisen RJ, Lozano-Fuentes S, Moore CG, Pape WJ, et al. (2008) Predictive spatial models for risk of West Nile virus exposure in eastern and western Colorado. Am J Trop Med Hyg 79 (4) 581–590.
- 26. Rizzoli A, Hauffe HC, Tagliapietra V, Neteler M, Rosà R (2009) Forest structure and roe deer abundance predict tick-borne encephalitis risk in Italy. PLoS ONE 4 (2) e4336. doi: 10.1371/journal.pone.0004336
- 27. Tesh RB, Watts DM, Russell KL, Damodaran C, Calampa C, et al. (1999) Mayaro virus disease: an emerging mosquito-borne zoonosis in tropical South America. Clin Infect Dis 28 (1) 67–73. doi: 10.1086/515070
- 28. Pinheiro FP, Freitas RB, Travassos da Rosa JF, Gabbay YB, et al. (1981) An outbreak of Mayaro virus disease in Belterra, Brazil. I. Clinical and virological findings. Am J Trop Med Hyg 30 (3) 674–681.
- 29. Azevedo RSS, Silva EVP, Carvalho VL, Rodrigues SG, Joaquim P, et al. (2009) Mayaro fever virus, Brazilian Amazon. Emerg Infect Dis 15 (11) 1830–1832. doi: 10.3201/eid1511.090461
- 30. Pialoux G, Gaüzère BA, Jauréguiberry S, Strobel M (2007) Chikungunya, an epidemic arbovirosis. Lancet Infect Dis 7 (5) 319–327. doi: 10.1016/s1473-3099(07)70107-x
- 31. Tsetsarkin KA, Vanlandingham DL, McGee CE, Higgs S (2007) A single mutation in chikungunya virus affects vector specificity and epidemic potential. PLoS Pathog 3 (12) e201. doi: 10.1371/journal.ppat.0030201
- 32. Lloyd-Smith JO, George D, Pepin KM, Pitzer VE, Pulliam JR, et al. (2009) Epidemic dynamics at the human-animal interface. Science 326 (5958) 1362–1367. doi: 10.1126/science.1177345
- 33. LeDuc JW, Pinheiro FP, Travassos da Rosa AP (1981) An outbreak of Mayaro virus disease in Belterra, Brazil. II. Epidemiology. Am J Trop Med Hyg 30 (3) 682–688.
- 34. Hoch AL, Peterson NE, LeDuc JW, Pinheiro FP (1981) An outbreak of Mayaro virus disease in Belterra, Brazil. III. Entomological and ecological studies. Am J Trop Med Hyg 30 (3) 689–698.
- 35. Talarmin A, Chandler LJ, Kazanji M, de TB, Debon P, Lelarge J, et al. (1998) Mayaro virus fever in French Guiana: isolation, identification, and seroprevalence. Am J Trop Med Hyg 59 (3) 452–456.
- 36. Torres JR, Russell KL, Vasquez C, Barrera R, Tesh RB, et al. (2004) Family cluster of Mayaro fever, Venezuela. Emerg Infect Dis 10 (7) 1304–1306. doi: 10.3201/eid1007.030860
- 37. de Thoisy B, Gardon J, Salas RA, Morvan J, Kazanji M (2003) Mayaro virus in wild mammals, French Guiana. Emerg Infect Dis 9: 1326–1329. doi: 10.3201/eid0910.030161
- 38. Figueiredo LTM, Shope RE (1987) An enzyme immunoassay for dengue antibody using infected cultured cells as antigen. J Virol Methods 17 (3–4) 191–198. doi: 10.1016/0166-0934(87)90129-7
- 39. Figueiredo LT, Simões MC, Cavalcante SM (1989) Enzyme immunoassay for the detection of dengue IgG and IgM antibodies using infected mosquito cells as antigen. Trans R Soc Trop Med Hyg 83 (5) 702–707. doi: 10.1016/0035-9203(89)90403-3
- 40. Bartley LM, Carabin H, Vinh Chau N, Ho V, Luxemburger C, et al. (2002) Assessment of the factors associated with flavivirus seroprevalence in a population in Southern Vietnam. Epidemiol Infect 128 (2) 213–220. doi: 10.1017/s0950268801006495
- 41. Wilder-Smith A, Foo W, Earnest A, Sremulanathan S, Paton NI (2004) Seroepidemiology of dengue in the adult population of Singapore. Trop Med Int Health 9 (2) 305–308. doi: 10.1046/j.1365-3156.2003.01177.x
- 42. Tuboi SH, Costa ZG, Vasconcelos PFC, Hatch D (2007) Clinical and epidemiological characteristics of yellow fever in Brazil: analysis of reported cases 1998–2002. Trans R Soc Trop Med Hyg 101 (2) 169–175. doi: 10.1016/j.trstmh.2006.04.001
- 43. da Silva-Nunes M, de Souza VA, Pannuti CS, Sperança MA, Terzian AC, et al. (2008) Risk factors for dengue virus infection in rural Amazonia: population-based cross-sectional surveys. Am J Trop Med Hyg 79 (4) 485–494.
- 44. DeMaris A (2002) Explained variance in logistic regression. A Monte Carlo study of proposed measures. Sociol Methods Res 31 (1) 27–74. doi: 10.1177/0049124102031001002
- 45. Ze-Shuai X, Li-Li J, Niao-Su Q, Yong-He Z (1986) Application of enzyme immunoassay on infected cells (EIA-IC) for arboviruses. Acta Virol 30 (6) 487–493.
- 46. Legendre P (1993) Spatial autocorrelation: trouble or new paradigm? Ecology 74 (6) 1659–1673. doi: 10.2307/1939924
- 47. Hilborn R, Mangel M (1997). The Ecological Detective. Confronting Models with Data. Princeton: Princeton University Press.
- 48. Spence L, Downs WG (1968) Virological investigations in Guyana, 1956–1966. West Indian Med J 17 (2) 83–89.
- 49. Neel JV, Andrade AH, Brown GE, Eveland WE, Goobar J, et al. (1968) Further studies of the Xavante Indians. IX. Immunologic status with respect to various diseases and organisms. Am J Trop Med Hyg 17 (3) 486–498.
- 50. Madalengoitia J, Flores W, Casals J (1973) Arbovirus antibody survey of sera from residents of eastern Peru. PAHO Bull 7 (4) 25–34.
- 51. Tavares-Neto J, Freitas-Carvalho J, Nunes MR, Rocha G, Rodrigues SG, et al. (2004) Pesquisa de anticorpos contra arbovírus e o vírus vacinal da febre amarela em uma amostra da população de Rio Branco, antes e três meses após a vacina 17D. Rev Soc Bras Med Trop 37 (1) 1–6. doi: 10.1590/s0037-86822004000100001
- 52. Silva-Nunes M, Malafronte RS, Luz BA, Souza EA, Martins LC, et al. (2006) The Acre Project: the epidemiology of malaria and arthropod-borne virus infections in a rural Amazonian population. Cad Saude Publica 22 (6) 1325–1334. doi: 10.1590/s0102-311x2006000600021
- 53. Cruz AC, Prazeres AS, Gama EC, Lima MF, Azevedo RS, et al. (2009) Vigilância sorológica para arbovírus em Juruti, Pará, Brasil. Cad Saude Publica 25 (11) 2517–2523. doi: 10.1590/s0102-311x2009001100021
- 54. Aitken TH, Downs WG, Anderson CR, Spence L, Casals J (1960) Mayaro virus isolated from a Trinidadian mosquito, Mansonia venezuelensis. Science 131 (3405) 986. doi: 10.1126/science.131.3405.986
- 55. Aitken TH, Spence L, Jonkers AH, Downs WG (1969) A 10-year survey of Trinidadian arthropods for natural virus infections (1953–1963). J Med Entomol 6 (2) 207–215.
- 56. Galindo P, Srihongse S, De Rodaniche E, Grayson MA (1966) An ecological survey for arboviruses in Almirante, Panama, 1959–1962. Am J Trop Med Hyg 15 (3) 385–400.
- 57. Galindo P, Srihongse S (1967) Transmission of arboviruses to hamsters by the bite of naturally infected Culex (Melanoconion) mosquitoes. Am J Trop Med Hyg 16 (4) 525–530.
- 58. Smith GC, Francy DB (1991) Laboratory studies of a Brazilian strain of Aedes albopictus as a potential vector of Mayaro and Oropouche viruses. J Am Mosq Control Assoc 7 (1) 89–93.
- 59. Mitchell CJ (1991) Vector competence of North and South American strains of Aedes albopictus for certain arboviruses: a review. J Am Mosq Control Assoc 7 (3) 446–451.