Table 1.
A summary of literature that includes Connecticut or New York as part of the study area.
Studies varied in their choice of dependent variable (De). Independent variables were classified as Surveillance (Su), climate (Cl), land cover (La), Sociological (So), host-related (Ho), or Other (Ot).
Fig 1.
Observed mosquito infection rate (MLE) vs. predicted MLE from the WNV model using the entire data set.
Background colors correspond to a classification of model predictions based on MLE of 5 [22]. Green corresponds to a correct prediction of high WNV MLE (27 records, 12.4%), blue corresponds to a correct prediction of low WNV MLE (157 records, 72.0%). Yellow corresponds to an error where the model predicts MLE to be high, but it is not (14 records, 6.4%), whereas orange corresponds to an error where the model predicts MLE to be low, but MLE was high (20 records, 9.2%). Future models should aim to improve the model’s sensitivity (0.57), although the specificity (0.92) is also of concern. Note that some predictions can be quite accurate, and still result in misclassification if they are near the classification threshold.
Fig 2.
Observed number of human cases of WNV across all of New York and Connecticut vs. predicted number of human cases of WNV from the model using the entire data set.
Background colors correspond to a classification of model predictions based on a threshold of 1 human case. Green corresponds to a correct prediction of one or more human cases (65 records, 7.4%), blue corresponds to a correct prediction of no human cases (704 records, 79.8%). Yellow corresponds to an error where the model predicts at least one human case, but none were observed (38 records, 4.3%), whereas orange corresponds to an error where the model predicts no human cases, but at least one was observed (75 records, 8.5%). Sensitivity (0.46) and specificity (0.95) were similar to the estimates for county-scale mosquito infection rates.
Fig 3.
Predicted and observed WNV mosquito infection rates (MLE, a, c) and human cases (b, d) for 2012, a particularly widespread WNV year.
MLE thresholds from Little et al. [22]: blue corresponds to MLE < 1 mosquito per 1000, yellow corresponds to MLE 1–5 per 1000, and red to MLE > 5 per 1000. White indicates excluded counties for which we did not have mosquito surveillance data. For human cases (b, d), blue indicates no human cases, yellow indicates 1–5 cases, and red indicates more than 5 cases.
Fig 4.
Predicted (unfilled ≤ 5, filled > 5) and observed (black ≤ 5, red > 5) infected mosquitoes per 1000 for each county and year for WNV.
Missing points correspond to missing years for those counties. Point sizes are scaled relative to the observed infection rate.
Fig 5.
Predicted (open < 1, filled ≥ 1) and observed (black < 1, red ≥ 1) number of human WNV cases for each county and year.
Data were not available for New York for 2000–2002, hence the missing points.
Table 2.
Model fit results for the calculated mosquito infection rates (per 1000).
Climate indicates whether climate variables were included, N indicates sample size, while WNV+ N indicates the number of samples estimated to have WNV present. RMSE, Median RMSE, Max Error, Scaled RMSE, R2, rp, and rs are defined in Methods: model fit statistics.
Table 3.
Model fits for the human data at the county-scale.
The All Counties analysis was based on 882 county × year records, while the subset contained 206 county × year records for which surveillance data were available. RMSE, Max Error, Median RMSE, Scaled RMSE, R2, rp, and rs are defined in Methods: model fit statistics.
Fig 6.
Predicted mosquito infection rates (MLE, contours) increase non-linearly with 2nd quarter soil moisture anomaly and 3rd quarter temperature.
Cool years with normal soil moisture were associated with the lowest MLE. Warm years showed high MLE regardless of soil moisture and dry years often (but not always) had high MLE. Observations (red circles, size is proportional to MLE) broadly support these predictions. Contour lines correspond to predictions made for a regular grid of 100 points covering the range of both variables. Predictions were made for mean values for all other covariates (see Tables 4 and 5 for included variables, see S1 File for mean values), while observed values correspond to the exact variable combinations and therefore may not exactly correspond to the predictions. Observations are plotted as a general guide to identify major patterns and highlight particular exceptions.
Fig 7.
Warm winter temperatures and dry summers were associated with the highest risk of mosquito infection with WNV.
Observations (red circles, size is proportional to infection rate) broadly support these predictions. Contour lines correspond to predictions made for a regular grid of 100 points covering the range of both variables. Predictions were made for mean values for all other covariates (see Tables 4 and 5 for included variables, see S1 File for mean values), while observed values correspond to the exact variable combinations and therefore may not exactly correspond to the predictions. Observations are plotted as a general guide to identify major patterns and highlight particular exceptions.
Fig 8.
For individual trap sites, the risk of WNV increased with increasing mosquito abundance, especially when the mean minimum temperature in the 3rd quarter was high.
Contour lines correspond to predictions from a regular grid of 100 points, (with values from other covariates fixed at a mean value). Observed infection rates (red circles, size is proportional to infection rate) are plotted for comparison, but note that they use exact parameter combinations and not the mean conditions used for making the predictions.
Fig 9.
Risk of human cases of West Nile were highest for locations with high total populations, especially in years with a warm summer.
Data correspond to the human subset analysis. Contour lines correspond to predictions from a regular grid of 100 points, (with values from other covariates fixed at a mean value). Observed infection rates (red circles, size is proportional to infection rate) are plotted for comparison, but note that they use exact parameter combinations and not the mean conditions used for making the predictions.
Table 4.
Climate variables identified as important by the random forest model when the model with all covariates was run, and when a model with only climate covariates was run (only C).
Model results are presented for human cases in those counties where mosquito surveillance data were collected, and for mosquito infection rates (MLE) at both the county and trap scales. Values in the table indicate the amount of unique variation explained by the variable using variance partitioning, while a blank indicates that the variable was not included in the final predictive model.
Table 5.
Non-climatic variables identified as important by the random forest model when the model with all covariates was run, and when a model without climate covariates was run (-C).
Model results are presented for human cases in those counties where mosquito surveillance data were collected, and for mosquito infection rates (MLE) at both the county and trap scales. Values in the table indicate the amount of unique variation explained by the variable using variance partitioning.
Fig 10.
Mean minimum temperature (a), soil moisture anomaly (b), mosquito infection rate (c), and human case counts (d) by year for five example counties.
Fig 11.
Total human population of the study region.
Note that the five counties of New York City have been merged into a single entity. Data taken from the US Census [100,111].