📊 statistics

A varying-coefficient model for characterizing duration-driven heterogeneity in flood-related health impacts

This paper proposes a Bayesian exposure duration varying-coefficient modeling (EDVCM) framework that leverages a two-dimensional Gaussian process prior to quantify duration-driven heterogeneity in flood-related health impacts, demonstrating superior performance over conventional methods when applied to nationwide Medicare data on musculoskeletal hospitalizations.

Original authors: Sarika Aggarwal, Phillip B. Nicol, Brent A. Coull, Rachel C. Nethery

Published 2026-01-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sarika Aggarwal, Phillip B. Nicol, Brent A. Coull, Rachel C. Nethery

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Why "How Long" Matters

Imagine you are studying how rain affects a garden. Most previous studies just asked, "Did it rain?" and looked at the damage. But in reality, a sudden, 10-minute downpour (a flash flood) is very different from a slow, soaking rain that lasts for a week.

This paper argues that how long a flood lasts changes how it hurts people's health. A flood that lasts two days might affect people differently than a flood that lasts ten days. However, until now, scientists didn't have a good tool to measure these differences. They usually treated all floods the same, or they just lumped all the days of a flood together into one big "messy" bucket.

The New Tool: The "Duration-Specific" Recipe

The authors created a new statistical recipe called the Exposure Duration Varying-Coefficient Model (EDVCM).

Think of it like a smart thermostat versus an old dial.

Old Method: The old way was like an old dial that just says "It's hot" or "It's cold." It doesn't care if the heat has been on for 10 minutes or 10 hours.
The New Method (EDVCM): This is like a smart thermostat that knows exactly how long the heat has been on and what time of day it is. It can say, "The first hour of a 2-day flood is fine, but the last hour of a 10-day flood is dangerous."

This tool allows researchers to look at every single day of a flood and ask: "Is the health risk different today compared to yesterday? And is the risk different for a 3-day flood compared to a 7-day flood?"

How They Did It: The "Self-Matching" Game

To make sure they were measuring the flood and not just other things (like the weather or pollution), the researchers used a clever trick called self-matching.

Imagine you are trying to see if a specific day of rain makes people sick.

You look at a town that got flooded on a Tuesday in 2010.
Instead of comparing it to a totally different town, you compare that Tuesday to a Tuesday in 2011 or 2012 when it didn't flood in that same town.
Because it's the same town on the same day of the year, the population, the local habits, and the general health of the area are the same. The only big difference is the flood.

This is like comparing yourself to yourself yesterday to see if a new diet worked, rather than comparing yourself to your neighbor.

The Secret Sauce: Borrowing Clues

One problem with studying floods is that some floods are very rare. Maybe there are plenty of 2-day floods, but very few 14-day floods. If you try to study a 14-day flood on its own, you might not have enough data to get a clear answer.

The authors solved this with a Gaussian Process, which is a fancy way of saying "borrowing clues from neighbors."

Imagine you are trying to guess the temperature in a town you've never visited. You don't just guess randomly; you look at the temperature in the town next door and the town across the river.
The model assumes that a 7-day flood is probably similar to a 6-day or 8-day flood. So, if the data for a 7-day flood is thin, the model "borrows" information from the 6-day and 8-day floods to make a smart guess. This keeps the results stable and prevents the model from getting confused by missing data.

What They Found: The "Muscle" Mystery

The team tested this new tool on real data from the US, looking at hospital visits for musculoskeletal diseases (problems with muscles, bones, and joints) in older adults during floods from 2000 to 2016.

Here is what the "smart thermostat" revealed that the old methods missed:

Short Floods = Less Hospital Visits: For short floods (1, 4, or 5 days), people actually went to the hospital less often.
- The Analogy: Think of it like a "stay home" signal. When a short rainstorm hits, people might just wait it out. If they have a sore back, they might decide to wait a few days to see if it gets better before driving to the doctor, especially if roads are wet.
Long Floods = More Hospital Visits (Later On): For longer floods (6 days or more), hospital visits went up, but not right away.
- The Analogy: Imagine a long, slow leak in a boat. At first, you can bail it out. But after day 6 or 7, the water is too high, and you have to call for help.
- The data showed that the biggest spike in hospital visits happened on the last day of long floods. It seems people can wait a few days, but eventually, the pain or injury becomes too much to ignore, or the roads become too dangerous to delay care any longer.

The Takeaway

The paper concludes that duration matters.

If you just look at "floods" as a whole, you might miss the fact that short floods make people stay home (lowering hospital numbers), while long, dragging floods eventually force people to seek help (raising hospital numbers).
The new model (EDVCM) successfully untangled these two effects, showing that the "critical window" of danger for older adults with muscle and bone issues is usually at the end of a long flood event.

This helps public health officials understand that the risk isn't just about the water being there; it's about how long the water stays there and how that changes people's behavior and health over time.

Technical Summary: Exposure Duration Varying-Coefficient Model (EDVCM) for Flood-Related Health Impacts

Problem Statement
Floods are highly heterogeneous events, ranging from short-duration "flash floods" to slow-moving events lasting weeks. While previous research has established associations between flood exposure and adverse health outcomes (e.g., musculoskeletal diseases, injuries, mortality), existing epidemiological methods fail to adequately incorporate flood duration into exposure modeling. Current approaches often treat multi-day flood events as simple sequences of exposure days or collapse them into a single exposure period, thereby ignoring duration-driven effect heterogeneity. This limitation prevents researchers from answering critical questions: Do health effects vary with flood length? Do different flood durations create different critical windows of vulnerability? There is a methodological gap in time-series and case-crossover frameworks used in climate epidemiology to handle consecutive-day exposures of variable lengths while accounting for lagged effects.

Methodology: The EDVCM Framework
To address this gap, the authors propose the Exposure Duration Varying-Coefficient Model (EDVCM). The framework integrates a self-matched study design with a Bayesian conditional Poisson regression model.

Study Design: The model utilizes an area-level self-matched design. Flood-exposed county-days are matched to non-flooded days within the same county on the same day-of-year in preceding or subsequent years. This design eliminates time-invariant confounders. Time-varying confounders (temperature, humidity, wind, PM2.5, O3, NO2) are adjusted for via covariates.
Model Formulation: The core of the EDVCM is a conditional Poisson regression where the log-rate of hospitalization is modeled as a function of binary flood exposure. Unlike standard models, the exposure coefficient ( $\beta$ $β$ ) is not a single scalar but varies jointly across two dimensions:
- Duration ( $d$ ): The total length of the flood event.
- Exposure Day ( $t$ ): The specific day within the flood event ( $1 \le t \le d$ ).
- The model also includes coefficients for lagged days ( $l$ ) following the flood event.
Bayesian Smoothing via Gaussian Processes: To estimate the high-dimensional matrix of coefficients ( $\beta_{d,t}$ $β_{d, t}$ ) where data may be sparse (especially for long-duration floods), the authors assign a two-dimensional Gaussian Process (GP) prior to the coefficients.
- The prior assumes a separable product kernel structure: $Cov(\beta_{d,t}, \beta_{d',t'}) = \sigma^2_\beta \exp(-\frac{|d-d'|}{\phi}) \exp(-\frac{|t-t'|}{\tau})$ .
- This allows for information sharing across similar flood durations and consecutive exposure days, stabilizing estimation without requiring arbitrary discretization of duration.
- A similar GP structure is applied to lagged effect coefficients ( $\theta_{d,l}$ ).
Estimation: The model conditions out stratum-specific intercepts to yield a multinomial likelihood, reducing computational burden. Parameters are estimated using Hamiltonian Monte Carlo (HMC) sampling.

Key Contributions

Novel Framework: The EDVCM is the first framework to explicitly model exposure effects as a function of both the duration of a climate event and the specific day within that event, allowing for the investigation of duration-driven heterogeneity.
Methodological Innovation: It extends varying-coefficient models (VCM) and distributed lag models (DLM) by parameterizing effects across a constrained two-dimensional space (duration and day) using a separable Gaussian process prior.
Simulation Performance: In simulation studies, the EDVCM demonstrated superior performance compared to conventional approaches (Bayesian models with independent normal priors and frequentist GLMs). Specifically, it achieved lower Root Mean Square Error (RMSE) and better bias reduction, particularly in "noisy" data scenarios and when estimating coefficients for durations with sparse data. It also maintained nominal coverage probabilities where independent prior models under-covered.

Empirical Results
The model was applied to nationwide Medicare claims data (2000–2016) for adults aged 65+, linked with high-resolution satellite-derived flood data from the Global Flood Database. The outcome was county-level hospitalization rates for musculoskeletal system diseases.

Duration-Dependent Heterogeneity: The study found that the direction and magnitude of flood effects on musculoskeletal hospitalizations depend heavily on flood duration.
- Shorter Floods (1, 4, 5 days): Associated with decreased hospitalization rates (protective effect), with the largest reduction observed on day 5 of a 5-day flood (Rate Ratio [RR] = 0.79).
- Longer Floods (2–3 days, 6+ days): Associated with increased hospitalization rates (adverse effect).
Critical Windows: For adverse effects (longer floods), the risk was concentrated in the latter quarter of the flood event. The maximum adverse effect was observed on the final day of a 6-day flood (RR = 1.28).
Cumulative Effects: While cumulative rate ratios for most durations crossed the null (1.0), 7-day floods showed the largest cumulative adverse impact, while 5-day floods showed the largest cumulative protective effect.
Interpretation: The authors suggest that shorter floods may lead to delayed care (postponing non-emergency visits), resulting in lower observed rates. Conversely, longer floods may eventually force individuals to seek care due to worsening conditions or system disruptions, leading to spikes in hospitalizations in the later days.

Significance and Limitations
The paper claims that the EDVCM provides "highly-resolved insights" into duration-driven effect heterogeneity, moving beyond the limitations of collapsed exposure periods which can mask critical windows of vulnerability. By identifying that adverse effects are concentrated in the later days of longer floods, the model offers actionable intelligence for public health response planning and targeted adaptation strategies.

The authors acknowledge several limitations:

Spatial Resolution: Exposure was aggregated to the county level, potentially leading to misclassification for individuals not directly flooded.
Residual Confounding: As with all observational studies, unmeasured time-varying confounders may exist.
Computational Scalability: The model's dimensionality grows quadratically with maximum flood duration, which could be burdensome for extremely long events.
Distributional Assumptions: The use of Poisson regression assumes mean-variance equality, though overdispersion in hospitalization data is plausible.

The study concludes that the EDVCM is a robust tool for characterizing the complex, duration-dependent health impacts of climate events, with potential applicability to other variable-duration hazards like heat waves and wildfires.