superhao-wu/covid19npisecondwave

Code for: Understanding the effectiveness of government interventions in Europe's second wave of COVID-19

This repo contains the code for Sharma et al. (2021), Understanding the effectiveness of government interventions in Europe’s second wave of COVID-19.

Installation

We use Poetry for dependency management and use Python 3.8. Please see Python Poetry for installation instructions for poetry.

Running poetry install, once Poetry has been installed, will setup a virtual environment and install project dependecies. You can then run poetry shell to launch a shell instance where the virtualenv is active.

Repository Structure

The most important files are:

• epimodel/models/models.py contains the main models, implemented in Numpyro, used for this project.
• epimodel/epiparam.py contains the EpidemiologicalParameters class, used to store and discretise delay distributions.
• epimodel/preprocessing.py contains the preprocess_data function used to load data, as well as the PreprocessedData class definition.

Other notes:

• epimodel/models/model_run_utils.py contains a run_model utility function.
• epimodel/models/model_build_utils.py contains a model build utility functions that can be used to build new models.
• epimodel/plotting contains some plotting helpers.
• epimodel/distributions/asymmetric_laplace contains the asymmetric laplace distribution, used for the NPI prior.

Reproducing Results

An example main model run is in notebooks/final_results/main_results.ipynb.

import numpyro
numpyro.set_host_device_count(4)

from epimodel import preprocess_data, run_model, EpidemiologicalParameters, default_model

# load data
data = preprocess_data('../../data/all_merged_data_2021-01-22.csv')
data.featurize() # convert raw NPI data to binary features
data.mask_new_variant(new_variant_fraction_fname='../../data/nuts3_new_variant_fraction.csv') # mask VOC
data.mask_from_date('2021-01-09') #mask end of perior

ep = EpidemiologicalParameters() # load delay dists

samples, warmup_samples, info, mcmc = run_model(default_model, data, ep, num_samples=1250, target_accept=0.75, num_warmup=250, num_chains=4, max_tree_depth=15) # run model

For the sensitivity analyses, sensitivity analysis scripts are in scripts/sensitivity_analysis. Each script define a particular sensitivity analysis type. Many sensitivity analyses can be launched in parallel using sensitivity_dispatcher.py.

For example, you can run the full set of sensitivity analyses on a server as follows:

python scripts/sensitivity_dispatcher.py --max_parallel_runs NRUNs --model_type MODEL --model_config CONFIG --categories [CATEGORY_1 CATEGORY_2 ... ]
# example    
python scripts/sensitivity_dispatcher.py --max_parallel_runs 24 --model_type default --model_config default --categories npi_leaveout output_noise_scale_prior r_walk_noise_scale_prior r_walk_period seeding_days seeding_scaling basic_R_prior_mean basic_R_prior_scale england_ifr_iar cases_delay_mean death_delay_mean gen_int_mean frac_voc infection_noise_scale intervention_prior bootstrap

Note: the parallelisation currently only works on Unix systems, since it requires the use of taskset to correctly parallelise the numpyro code.

• Sensitivity analysis options for a given script are input using Python argparse arguments. e.g., in seeding_days, there is an argparse argument --n_days_seeding N_DAYS which is the number of seeding days.
• model_configs.yaml contains model configs and featurise configs that can be used to overwrite model arguments across many sensitivity analyses. For example, if you wanted to run a full sensitivity analysis suite with a different model input argument, you could use the config.
• sensitivity_analysis.yaml defines the input arguments used for the sensitivity analysis scripts, and sensitivity analysis categories. The arguments in the yaml file are exactly the argparse arguments of the scripts.

All plotting code is in the notebooks/final_results folder. See validation_plotter (Fig. 3), main_result_plotter (Fig. 1, other plots also) and sensitivity_analysis_plotter (Appendix plots).

Questions?

Please email Mrinank Sharma (mrinank at robots dot ac dot uk) for questions regarding the codebase.