TSR-Proj
README.md

The pitfalls of incidence-based time series regression for inferring the effects of weather on infectious diseases

This repository hosts all the code needed to replicate the results of the study on the reliability of time series regressions in estimating the effect of weather on infectious diseases from incidence data.

Software versions

This project relies on renv for package management. In order to run the same package versions used originally, it is necessary to install renv and then run:

renv::restore()

Hardware and sofware

Simulations and fitting procedures were executed on the high-performance computing (HPC) SLURM system Viper, of the Max Planck Computing and Data Facility, running Red Hat Enterprise Linux 9.4. Images and tables generation were carried out on a macOS 26.3 workstation equipped with an Apple M2 Pro processor (10 CPU cores) and 32 GB RAM.

Workflow

The entire results of the work can be entirely reproduced using a SLURM cluster or locally, running a subset (or all) of simulations and analyses.

Weather data and effect calibration

First, in m-Extract_weather_Popular.r, the effects of temperature and relative humidity are calibrated based on previous studies (Yaari et al., 2013 and Weber et al., 2001). Then, 10 years of weather data are pulled for selected locations and stored inside their independent folder.

Main analysis via HPC cluster

All combinations of variables are defined in config.sh. Each unique combination results in one job. Mind that main_HPC.sh,copy_rds_parallel.sh and run_aggregate.sh need to be adjusted in the cluster settings (in resources, limits and personal email) specifically.

Job execution

Each cluster array job:

  • determines its variables' combination (1 scenario x 1 location x 1 parameter set)
  • exports variables to the environment
  • runs m-Main.r
  • writes outputs to Rout/,logs/ and temp/
  • results are saved in Results folder

Job batch submission

The master script:

  • computes the total number of simulations
  • submits jobs in batches
  • waits for active jobs to be completed before submitting new ones
  • logs progress
  • supports safe stopping

Results generation

Once on an HPC SLURM cluster, run the line:

nohup ./master_HPC.sh > logs/master_HPC.out 2>&1 &

This will start the entire analysis for all parameter combinations defined in config.sh. Running master_HPC.sh, which iteratively sources main_HPC.sh, embarrassingly parallel jobs are submitted to the cluster until all are completed. Each job will consist of running m-Main.r with a selected parameter set, location and scenario. This latter script will generate simulations and fit the tested models. Finally, performance measures are computed and everything is stored as .RData in a folder named using specific identifiers for that particular analysis. Mind to tune the HPC bash settings according to the rules and resources of the cluster used. For diagnostic purposes, the pipeline outputs runtime and execution logs to the directories logs, temp, and Rout. These record workflow progress and R session output, and can be inspected to monitor execution or troubleshoot failures.

Data aggregation and download

Once all jobs are completed, execute:

batch aggregation_copy.sh

This will, for each parameter set and location, run on the cluster run_aggregate.sh that will generate in each Results subfolder:

  • sims_tot.rds $\rightarrow$ all simulations
  • sims_reg_tot.rds $\rightarrow$ all regressions' results
  • final_pars.rds $\rightarrow$ all combinations of parameters used

Once this is finished, copy_rds_parallel.sh, will automatically start to copy all .rds files in the subfolders of Results in Results_Aggregated. In this way, one can now quickly download the latter folder ($\approx 5$ GB) on a local machine via rsync(more information on this here), or any other preferred software.

Main analysis on local machine

It is also possible to run the entire analysis locally.

Results' subset generation

Editing m-local_master.r, it is possible to select the combination of parameters to use in the full analysis. The entire core pipeline will still find m-Main.r at its core. The running time for the simulation and fit on 100 replicates for one single parameter set, location and scenario is just under 8 seconds locally (see here for harware and sofware) utilizing 5 workers for parallel simulation and 2 for the fitting.

Data aggregation

Once this has been executed, it is possible to aggregate the results via execution of m-Final_Analysis.r, which will generate of each Results subfolder (specific for one location and parameter set):

  • sims_tot.rds $\rightarrow$ all simulations
  • sims_reg_tot.rds $\rightarrow$ all regressions' results
  • final_pars.rds $\rightarrow$ all combinations of parameters used

Paper content

Figures

In order to reproduce the main and supplementary materials, it is sufficient to execute Figures_Manuscript.r, which will use the data and results present in the folder with the complete aggregated results. This will create and save all figures in Figures, as Portable Document Format.

Tables

All the code used to format the tables is present in Table_Maker.r.

Reproducibility of reported statistics

The summary statistics reported in the manuscript are derived from the same underlying data presented and can be reproduced using the script Summ_stats.r.

License

This repository is released under the MIT License.

Contact

For questions about the simulation workflow, contact gemo@mpiib-berlin.mpg.de.

Powered by GitFront