The West Nile Virus Prediction Model
This project forecasts the risk that any subsequent mosquitos will test positive for West Nile Virus (WNV) after any mosquitos have already been found to have WNV at a particular trap. The model allows the Chicago Department of Public Health to proactively react to the spread of WNV. A multilevel (hierarchical) regression to predict the likelihood that WNV will be present if it was present last week. A cross-validation study has shown that this model correctly predicts consecutive results for WNV virus about 80 percent of the time.
This repository provides all of the source code used develop the model and generate predictions, however this project can not be executed in its entirety outside of the City of Chicago computer network. Though it does provide the statistical methodology to be reviewed, evaluated, and replicated elsewhere.
Data and Model
Most of the work in this project is manipulation to get the data into a format where each row represents the results for the test results of a week / trap location collection, and the interval between the dates is regular (i.e. a week apart).
The model relies on the presence of past West Nile Virus, weather data, and trap locations with a Generalized Linear Mixed Model to develop predictions of whether a trap will have WNV present on during a particular collection week. The models used rely on R’s
arm package written by (by Yu-Sung Su, Daniel Lee, and Andrew Gelman).
The overall performance of the model depends highly upon the choice of the cutoff for what is considered a positive prediction. We carefully chose a cutoff to balance capturing positive results without projecting too many positive results, which would result in unnecessary treatment. In machine learning terms, we sought to balance the precision and recall of the model.
Based on previous results we chose a cutoff of 39%, which accurately predicts the positive results 78% of the time in the test case (94 / 120), and these predictions were correct 65% of the time (94 / 144).
How run the model
The R programming language was used to develop the model, which is free and available for download at https://cran.r-project.org. The preferred IDE for this project is R Studio, which is also free and available for download at https://www.rstudio.com/.
You can open the project file in R Studio
WNV_R_Model.Rproj, and step through the code in the
.\R directory. The files are organized sequentially. Data will be downloaded to a
./data directory, which will be created if it does not already exist.
We recommend using R Studio as the IDE along with R the programming language. The user can then open the
.Rproj file and explore / run the code. However, it is not possible to run the code in its entirety from end to end without access to our databases. The West Nile Virus results on the data portal contain inexact locations for the traps, but exact locations are used in the production model. Because of the threat of vandalism / etc. we do not make exact trap locations available.
Within R you will need to install several libraries, which are declared at the beginning of each script. The libraries are loaded using a function called
loadinstall_libraries from the
geneorama package, which is a convenience function that attempts to install libraries if they are not installed.
geneorama can be installed by using the devtools package (available on CRAN) and running
You will also need to store a token to access the NOAA api. This token is easy to obtain by registering at the NOAA website: https://www.ncdc.noaa.gov/cdo-web/token. Store the token in a plain text file:
untracked\weather_noaa_token.txt with no other text.