Infectious disease modeling for public health practice: projections, scenarios, and uncertainty in three phases of outbreak response

This paper presents a practical toolkit and conceptual guide developed by the Insight Net Modeling Guidance for Public Health Working Group, which uses early COVID-19 data from Michigan to demonstrate how to generate transparent, evidence-backed scenario projections with quantified uncertainty across three critical phases of outbreak response to better support public health decision-making.

The Gist

Imagine you are the captain of a ship (a public health department) sailing into a storm you've never seen before. You need to know: How big will the waves get? Will we run out of lifeboats (hospital beds)? Should we drop anchor now or wait?

This paper is essentially a user manual for building a "storm simulator" to help captains make those decisions. The authors, a team of modelers and public health experts, explain how to create simple, useful predictions without getting lost in overly complicated math. They argue that you don't need a supercomputer to be helpful; sometimes a "back-of-the-envelope" calculation is exactly what's needed.

Here is how they break it down, using the three stages of an outbreak as their guide:

The Three Stages of the Storm

The paper says the tools you need change depending on how far along the storm is. They divide the response into three phases:

Phase 1: The Fog Before the Storm (Before Local Spread)

• The Situation: You know a storm is coming, but it hasn't hit your town yet. You have no local data, only stories from other places.
• The Tool: A simple "exponential growth" calculator. Think of this like a snowball rolling down a hill. You know how fast snowballs grow in general, so you can guess how big yours might get in six weeks if you do nothing.
• The Goal: To help leaders plan ahead. "If we don't act, we might need 500 extra beds by next month." It's not a precise forecast; it's a warning to get the lifeboats ready.

Phase 2: The First Big Waves (Local Exponential Growth)

• The Situation: The storm has hit. Cases are rising fast, and you have local data for the last few weeks.
• The Tool: A slightly more tuned version of the snowball calculator. Now, instead of guessing based on other towns, you measure the speed of the snowball right here.
• The Goal: To give a short-term look (about two weeks) at what happens if things stay exactly as they are. It answers: "If the wind keeps blowing this hard, how many people will get sick by next Tuesday?"

Phase 3: Navigating the Eye of the Storm (Established Transmission)

• The Situation: The storm is fully here. It's complex. People are changing their behavior, and you might want to try different strategies (like asking people to stay home).
• The Tool: A "Mechanical Engine" (a compartmental model). Imagine a giant, complex water tank with pipes. You can open and close valves to see how the water level changes.
  • Valve A: How fast the virus spreads.
  • Valve B: How many people get sick enough to go to the hospital.
  • Valve C: How many people recover.
• The Goal: To run "what-if" experiments. "What if we close the valve and reduce spread by 25%? What if we do nothing?" This helps leaders compare different strategies.

The Secret Sauce: Uncertainty (The "Maybe" Zone)

The authors emphasize a crucial point: Never trust a single number.

If a model says "We will have 1,000 cases," that's dangerous because it sounds too certain. Instead, the paper suggests showing a range, like a weather forecast that says "There's a 50% chance of rain, but it could be a drizzle or a downpour."

• The Analogy: Imagine throwing a dart at a board. You might hit the bullseye (the best guess), but you should also show the area where the dart could land if your hand shook a little (the uncertainty).
• Why it matters: If you only show the bullseye and the dart lands in the outer ring, people lose trust. If you show the whole ring of possibility, they understand that the future is fuzzy and they need to prepare for the worst-case scenario just in case.

The Golden Rules of the Paper

1. It's a Scenario, Not a Crystal Ball: The authors are very clear: These models are not predicting the future. They are showing what happens if nothing changes. If the public decides to wear masks or stay home, the "status quo" prediction will be wrong, and that's okay. It just means the situation changed.
2. Keep it Simple: You don't need a PhD in physics to help. Simple math that is fast and easy to explain is often better than a complex model that no one understands.
3. Talk to Each Other: The modelers (the people building the math) and the public health partners (the people making the decisions) need to sit down and agree on what questions they are trying to answer. It's like a pilot and a co-pilot agreeing on the destination before takeoff.

Summary

This paper is a guidebook for how to use math to help communities prepare for disease outbreaks. It teaches you how to build simple "what-if" machines that show the range of possible futures, helping leaders decide when to sound the alarm, when to open the lifeboats, and when to try new strategies to calm the storm. The key message is: Don't promise a perfect prediction; promise a clear picture of the possibilities.

Technical Summary

Technical Summary: Infectious Disease Modeling for Public Health Practice

Problem Statement
Public health departments require evidence-backed scenario projections to support decision-making during infectious disease outbreaks. However, traditional infectious disease models are often not readily deployable or responsive to the urgent, evolving priorities of local health departments. Furthermore, uncertainty in model outputs is frequently inadequately assessed or communicated, which can undermine trust among practitioners and the public. A critical distinction exists between forecasting (predicting specific future outcomes) and scenario modeling (projecting outcomes under status quo or alternative conditions), yet this distinction is often blurred, leading to misplaced expectations. The paper addresses the need for a framework that balances the provision of actionable numbers with the accurate communication of uncertainty, specifically tailored to the varying data availability and modeling needs across different phases of an outbreak.

Methodology
The authors propose a tiered modeling framework divided into three phases of outbreak response, utilizing data streams from reported cases, hospitalizations, and deaths. The approach integrates specific data parameters: infection–case ratio, case–hospitalization ratio, case–fatality ratio, and reporting lags for cases, hospitalizations, and deaths.

1. Phase 1: Preliminary Projections (Pre-local Introduction)

   • Context: Before local spread is detected.
   • Model: Exponential growth model.
   • Method: Projections are generated based on historical or ongoing outbreak data. Since no local data exists, uncertainty is quantified using Sobol' sampling to generate a large number of parameter combinations from realistic ranges. Credible intervals are calculated for each day of the projection.
   • Goal: To estimate the potential speed and magnitude of local spread to support advanced planning and resource requests.

2. Phase 2: Local Exponential Growth

   • Context: Local transmission is sustained, and the epidemic is growing exponentially.
   • Model: Exponential growth model with parameters informed by recent local data.
   • Method: Models are fitted to local data (typically a 2-week window) using maximum likelihood estimation (Minimizing negative log-likelihood assuming Poisson distributions). Parameters such as initial case counts, doubling time, and outcome ratios are estimated. Unidentifiable parameters (e.g., infection–case ratio) are handled via sensitivity analysis using ranges from Phase 1. Uncertainty is captured by sampling parameter combinations within ranges (e.g., ±25% of best-fit values) to generate confidence intervals.
   • Goal: To provide short-term (2-week) status quo projections to guide immediate planning.

3. Phase 3: Established Transmission with Potential Interventions

   • Context: Established transmission where interventions are being considered.
   • Model: Mechanistic compartmental model (Susceptible–Latent–Infectious–Recovered, or SLIR) simulated via differential equations.
   • Method: The model accounts for effective transmission rates and delays between compartments. A two-step uncertainty analysis is employed:
     1. Sampling a large number of combinations for unidentifiable parameters (e.g., initial fraction infectious, latent/recovery rates).
     2. For each sampled set, estimating identifiable parameters (e.g., effective reproduction number $R_e$, reporting ratios) via maximum likelihood.
     3. Discarding parameter sets with negative log-likelihoods exceeding a threshold above the minimum.
     4. Generating projections for remaining sets to plot 50% and 95% credible intervals.
   • Scenario Analysis: The framework allows for "in silico" experiments, such as simulating ±25% changes in the reproduction number to illustrate potential trajectories under different intervention scenarios or behavioral changes.

Key Contributions

• Phased Modeling Framework: The paper delineates specific modeling strategies appropriate for three distinct outbreak phases, acknowledging that data availability and modeling needs evolve over time.
• Distinction Between Forecasting and Scenario Modeling: It explicitly frames models as tools for scenario planning (answering "what if" questions) rather than precise forecasting, emphasizing that deviations from projections often reflect changes in underlying parameters (e.g., transmission rates) rather than model failure.
• Uncertainty Quantification: The authors provide a structured approach to quantifying uncertainty, moving beyond single-point estimates to provide best-case, worst-case, and credible intervals. This includes handling unidentifiable parameters through sampling and sensitivity analysis.
• Practical Toolkit: The work offers open-source code and examples (using Michigan COVID-19 data from 2020–2021) to serve as a starting point for public health partners and modelers, aiming to reduce the need to "reinvent the wheel" for each new outbreak.

Results
Using Michigan COVID-19 data, the authors demonstrated the application of their framework:

• Phase 1: Projected 6-week outcomes starting from a single introduction, yielding wide ranges (e.g., 1,500–15,100 reported cases at 95% CI) to illustrate potential magnitude.
• Phase 2: Generated 2-week status quo projections based on data from November 2020, estimating 6,600 reported cases (95% CI: 4,100–12,000) and 400 hospital admissions, with a best-fit doubling time of 17.2 days.
• Phase 3: Modeled established transmission (April 2021) using an SLIR model, estimating an $R_e$ of 1.71. The model projected a median of 12,000 cases four weeks out under status quo. Crucially, the study showed how alternative scenarios (e.g., a 25% increase or decrease in $R_e$) produced vastly different trajectories, and how a retrospective scenario with a 45% reduction in $R_e$ aligned with actual reported data, validating the utility of scenario modeling in hindsight.

Significance and Claims
The paper claims to provide a "conceptual guide and practical toolkit" to support more transparent, timely, and appropriate use of models in outbreak response. Its primary significance lies in:

• Accessibility: Demonstrating that relatively simple, "back-of-the-envelope" analytic approaches (like exponential growth models) can be highly effective for public health practitioners if applied thoughtfully, rather than requiring complex mechanistic models in all instances.
• Trust Building: By rigorously communicating uncertainty and distinguishing between status quo projections and forecasts, the framework aims to prevent the loss of trust that occurs when actual outcomes deviate from overly confident predictions.
• Actionability: The models are designed to inform planning (e.g., hospital capacity, PPE needs) and motivate action rather than predict precise future outcomes.
• Collaboration: It emphasizes the necessity of a shared understanding between modelers and public health partners regarding the goals, limitations, and expectations of modeling efforts.

The authors explicitly state that these models are starting points intended to be modified for specific outbreaks and are not suitable for all transmission dynamics (e.g., highly clustered networks or zoonotic point-sources). The work serves as a foundation for the Insight Net Modeling Guidance for Public Health Working Group to improve preparedness for future outbreaks.