delayscompare

A research project comparing the impact of misspecified delay distributions on epidemic forecasting using EpiNow2. This repository contains simulation studies and real-world case studies examining how incorrect assumptions about generation intervals, incubation periods, and reporting delays affect real-time reproduction number (Rt) estimates and case forecasts.

Overview

This project investigates how misspecification of epidemiological delay distributions affects forecast accuracy across different epidemic scenarios. The analysis uses the EpiNow2 R package to:

• Simulate epidemic data with known delay distributions
• Estimate infections and Rt under various delay misspecification scenarios
• Compare forecast performance across different diseases and epidemic trajectories
• Evaluate the impact of prior weighting on delay estimation

Project Structure

delayscompare/
├── R/                      # Core functions
│   ├── funcs_data.R       # Data processing functions
│   ├── funcs_plots.R      # Plotting functions
│   ├── funcs_rtraj.R      # R trajectory simulation functions
│   ├── generate_scores_func.R  # Scoring metrics
│   └── scenario_loop.R    # Main scenario simulation wrapper
├── scripts/               # Analysis scripts
│   ├── 01_packages.R      # Package dependencies
│   ├── *_simulatedata.R   # Data simulation scripts
│   ├── *_scenariorun.R    # Scenario execution scripts
│   └── 07_resultsprocessing_*.R  # Results processing
├── data/                  # Input data files
├── results/               # Analysis outputs
├── scenario_loop.R        # Main scenario functions
├── plots_baseline.R       # Baseline plotting script
└── plot_runtimes.R        # Runtime analysis

Case Studies

The analysis includes three epidemic case studies:

1. COVID-19 (England, 2021 Delta wave)

2. Ebola (Sierra Leone, 2014)

3. Cholera (Yemen, 2016)

Scenarios

Each case study examines delay misspecification scenarios:

• No delay: All delays set to zero (Fixed(0))
• Very low: 0.25× correct delay mean
• Low: 0.8× correct delay mean
• Correct: 1× correct delay mean (ground truth)
• High: 1.25× correct delay mean
• Very high: 2× correct delay mean

Additional scenarios tested:

• Different Rt projection methods (rt_opts: "latest" vs "project")
• Under-reporting (observation scale: 0.3 vs 1.0)
• Prior weight specifications for delay distributions

Requirements

R Packages

• EpiNow2
• incidence2
• readxl
• dplyr, tidyr, purrr
• here
• ggplot2
• scoringutils
• viridis, RColorBrewer, cowplot

Install packages with:

pkgs <- c("EpiNow2", "incidence2", "readxl", "dplyr", "tidyr",
          "purrr", "here", "ggplot2", "scoringutils", "viridis",
          "RColorBrewer", "cowplot")
install.packages(setdiff(pkgs, rownames(installed.packages())),
                 repos = c("https://epiforecasts.r-universe.dev",
                          getOption("repos")))

Usage

Running Scenario Analysis

Basic scenario execution:

source("scripts/01_packages.R")
source("R/scenario_loop.R")

# Run COVID-19 scenario (example)
res <- sim_scenarios(
  case_data = covid_data,
  var = 1,            # Generation time scenario
  inc = 4,            # Incubation period scenario (1-6)
  gen_mean = 3.6,
  gen_sd = 3.1,
  gen_max = 15,
  inc_mean = 5.2,
  inc_sd = 1.52,
  inc_max = 21,
  rep_mean = 4.4,
  rep_sd = 5.6,
  rep_max = 18,
  freq_fc = 4,        # Forecast frequency (every 4 weeks)
  weeks_inc = 12,     # Use 12 weeks of data
  rt_opts_choice = "latest",
  obs_scale = 1
)

Running with Prior Weights

For uncertainty in delay distributions:

res <- sim_weightprior(
  case_data = data,
  var = 1,
  gen_mean_mean = 3.6,
  gen_mean_sd = 0.5,
  gen_sd_mean = 3.1,
  gen_sd_sd = 0.3,
  gen_max = 30,
  # ... other parameters
  weight_prior = TRUE 
)

Command Line Execution

Scripts accept command line arguments for parallel execution:

Rscript scripts/casestudy_covid_scenariorun.R 1

Output

Results are saved as RDS files in the results/ directory:

• *_samples*.rds: Posterior samples for case predictions
• *_R*.rds: Posterior samples for Rt estimates
• *_id*.rds: Scenario identifiers
• *_summary*.rds: Summary statistics
• *_warnings*.rds: Model warnings

Plots are generated showing:

• Impact of delay misspecification on CRPS for Rt and case forecasts
• Timeseries of best- and worst-performing forecasts
• Runtime comparisons

Configuration

The analysis uses 8 timepoints per case study to balance computational efficiency with temporal coverage. Timepoints are spaced every 4 weeks throughout the epidemic period.

MCMC settings:

• Samples: 3000
• Adapt delta: 0.99
• Max treedepth: 20
• Forecast horizon: 14 days

License

This project is licensed under the GNU General Public License v3.0 - see the LICENSE file for details.