Spatiotemporal relationships between extreme weather events and arbovirus transmission across Brazil

Reviewer #1 (Public review):

Summary:

The authors used fine-level resolution epidemiological data to describe the spatiotemporal patterns of dengue, chikungunya and Zika. They assessed which factors best captured the historic transmission dynamics in Brazil. It was used epidemiological data from 2013 to 2020. They tested the association between arbovirus incidence and environment, human connectivity and socioeconomic, and climate variables, including extreme weather conditions.

Strengths:

The authors used granular epidemiological data at the subnational level and weekly case notification time series. Furthermore, they considered more than one hundred variables. Among the variables, it is highlighted that they also considered human connectivity and extreme weather events.

The authors used appropriate statistical methods accounting for the spatiotemporal structure and used the negative binomial to handle overdispersion; They applied a systematic covariate screening, using WAIC and performed sensitivity analysis. Their results suggest an important role of climate variables such as El Niño South Oscillation Anomalies, and that extremes in wetness and drought may drive infections outside regular patterns; it also suggests that temperature variations and extremes may be more associated with the incidence than the mean temperature; in addition, human connectivity networks are also pointed out as a key driver factor at fine level scale.

Weaknesses:

The authors have not accounted for the correlation between diseases. They have not considered the co-occurrence of diseases by applying a joint modelling approach, nor have they discussed this as a possibility for future work. Still, regarding the methods, they used a simplified lag treatment. They could have included into the discussion, examples of methods like Distributed Lag Models. This can be used in contexts when analysing meteorological covariates and extreme weather events.

They also have not considered the population's immunity to the different serotypes of dengue, which can reflect in peaks of incidence when a new serotype starts to circulate in a certain region. It is important to bring this into the discussion section.

Whether the authors achieved their aims, and whether the results support their conclusions:

The authors assess variables which may be associated with different vector-borne disease incidence and the magnitude of these associations. Conducting a fine-scale resolution analysis (spatial and temporal), they emphasised the role of environmental and extreme weather conditions. Their findings are coherent with their analysis and corroborate some of the existing literature.

Discussion of the likely impact of the work on the field, and the utility of the methods and data to the community:

Their work shows how the different vector-borne diseases are influenced by environmental and climatic factors and that human connectivity may play an important role at the fine level spatial and temporal scale. This work brings a picture of the spatial and temporal distributions of dengue, chikungunya and Zika, at the municipal level in Brazil (2013-2020). The material and methods are well described, and the source is made available, allowing reproducibility by other researchers and academics.

Reviewer #2 (Public review):

Summary:

This manuscript looks at a wide variety of likely important drivers of arbovirus transmission across municipalities in Brazil. The results are intriguing due to their relevance and breadth, but the approach also brings challenges, which make the results hard to interpret.

Strengths:

Important and complex problem, excellent spatiotemporal resolution, collection of important covariates, and holistic analysis.

Weaknesses:

There are two key weaknesses. First, it is difficult to understand the actual contributions of each included covariate. The principal fit metric is WAIC, and importance is characterized by rank based on univariate fit. WAIC is a valuable comparison metric, but does not indicate how well the best model (or any other) fits the data. Figures 5B and S2-S4 show what look like good fits, but it also seems possible that most of this fit could be coming from the random effects rather than the covariates. It would be helpful to show the RE-only model as a comparator in these figures and also to consider other metrics that could help show overall fit (e.g., R^2). How much variance is actually being explained by the covariates?

Relatedly, the mean absolute errors reported are approximately 2-8 across the viruses, which sounds good on the surface. But many of the actual counts are zeros, so it's hard to tell if this is really good. Comparison to the mean and median observed case counts would be helpful.

Second, some of the results/discussion on specific variables and covariates were confusing. For example, the relationships between relative humidity and temperature vary substantially between pathogens and minimum or maximum temperature values. However, as transmission of three viruses relies on the same mosquito and minimum and maximum temperatures are highly correlated, we would expect these relationships to be very similar. One concern is clarity, and another is that some of the findings may be spurious - potentially related to how much of the variance is accounted for by the random effects alone (see above) and the wide range of covariates assessed (thus increasing the chance of something improving fit).

Underlying much of this are likely nonlinear relationships. The authors comment on this as a likely reason for some of the specific relationships, but it is not a very strong argument because the variable selection process is completely based on (generalized) linear univariate regressions.

Lastly, the mischaracterization of arboviral disease is a big challenge, as noted in the discussion. Only a subset of cases in Brazil are laboratory confirmed, but I couldn't find any statement about whether the cases used here were laboratory confirmed or not. I suspect that they are a combination of confirmed and suspect cases. A sensitivity analysis with only confirmed cases would increase confidence in the results.

Author response:

We thank the reviewers for their time and work assessing our manuscript, and for their constructive suggestions for improvements. Based on the reviews, our plan is to adapt the work as follows:

(1) Perform a sensitivity analysis considering only confirmed dengue, Zika, and chikungunya cases,

(2) Explore and discuss the potential correlation between diseases,

(3) Compare the baseline and final models,

(4) Assess model fit using a wider variety of metrics.

We would like to emphasise that our research question was to explore drivers of arbovirus incidence outside of seasonal trends. We therefore designed our models with flexible spatiotemporal random effects to capture baseline patterns, and as the reviewers have highlighted, much of the variance is explained by these random effects. To expand on point 3 above, we will perform a comparison of the baseline random effect models and the final multivariable models to show the differences between the models and quantify the additional impact of the meteorological variables in the final models.