[Paper Review] A study of gravity-linked metapopulation models for the spatial spread of dengue fever

Marta Sarzynska, Oyita Udiani+1 more|arXiv (Cornell University)|Aug 20, 2013

COVID-19 epidemiological studies19 references3 citations

TL;DR

This study develops a gravity-linked metapopulation model to simulate spatial dengue fever spread in Peru, using population size and distance to define inter-patch transmission weights. It shows that gravity-based links produce more realistic, desynchronized epidemic peaks across regions compared to uniform link weights, and while fitting the model to real data proved challenging due to irregular incidence and limited climate data, it successfully captured yearly epidemic patterns and highlighted the need for multi-strain and stochastic modeling for improved accuracy.

ABSTRACT

Metapopulation (multipatch) models are widely used to study the patterns of spatial spread of epidemics. In this paper we study the impact of inter-patch connection weights on the predictions of these models. We contrast arbitrary, uniform link weights with link weights predicted using a gravity model based on patch populations and distance. In a synthetic system with one large driver city and many small follower cities, we show that under uniform link weights, epidemics in the follower regions are perfectly synchronized. In contrast, gravity-based links allow a more realistic, less synchronized distribution of epidemic peaks in the follower regions. We then fit a three-patch metapopulation model to regional dengue fever data from Peru -- a country experiencing yearly, spatially defined epidemics. We use data for 2002-2008 (studying the seasonal disease patterns in the country and the yearly reinfection patterns from jungle to the coast) and 2000-2001 (one large epidemic of a new disease strain across the country). We present numerical results.

Motivation & Objective

• To investigate how inter-patch connectivity weights affect spatial disease spread predictions in metapopulation models.
• To compare uniform link weights with gravity-based link weights in simulating dengue fever epidemics across a synthetic network of one central city and 99 surrounding cities.
• To calibrate a deterministic, gravity-based three-patch metapopulation model to real dengue incidence data from Peru (2000–2008), focusing on seasonal and strain-specific transmission dynamics.
• To assess the limitations of the model due to data irregularity and poor fit, and to identify directions for improvement, such as incorporating climate data, stochasticity, and multiple dengue strains.
• To explore the potential of alternative models like the radiation model for more accurate human mobility-based transmission prediction.

Proposed method

• The study uses a deterministic SIR-type compartmental model across multiple patches representing distinct geographic regions.
• Inter-patch transmission weights are modeled using a gravity model: $ P_{ij} = \theta \frac{n_i^\alpha n_j^\beta}{d_{ij}^\gamma} $, where $ n_i $ and $ n_j $ are populations, $ d_{ij} $ is distance, and $ \alpha, \beta, \gamma, \theta $ are fitted parameters.
• Model fitting is performed using the least squares statistic to minimize the sum of squared differences between model predictions and observed dengue incidence data.
• The model is first tested on a synthetic system with one large driver city and 99 smaller follower cities to compare uniform vs. gravity-based connectivity.
• The model is then applied to real dengue data from Peru, initially using 7 patches, but reduced to 3 due to data quality and irregularity.
• The model is further tested on a 49-patch model representing all provinces with dengue cases during the 2000–2001 epidemic, and results are compared to data for shape and magnitude.

Experimental results

Research questions

• RQ1 How do uniform inter-patch transmission weights versus gravity-based weights affect the synchronization of epidemic peaks in a metapopulation model?
• RQ2 Can a gravity-based metapopulation model accurately reproduce the observed spatial and temporal patterns of dengue fever in Peru?
• RQ3 Why does the model overestimate epidemic size despite capturing the general shape of yearly incidence peaks?
• RQ4 How do data limitations, such as irregular incidence and missing climate data, affect model calibration and predictive accuracy?
• RQ5 What improvements—such as multi-strain dynamics, stochasticity, or alternative mobility models—could enhance model performance?

Key findings

• In the synthetic model, uniform link weights led to perfect synchronization of epidemic peaks across all follower cities, while gravity-based links produced a more realistic, desynchronized spread.
• The three-patch gravity-based model captured the general shape of yearly dengue incidence peaks in Peru but predicted higher epidemic sizes than observed.
• The 49-patch model also failed to achieve a good fit, with the best-shaped prediction being outperformed by a lower-amplitude, poorly shaped curve due to the least squares objective function.
• Preliminary fitting of the 49-patch model yielded higher $ \beta $ and $ \gamma $ values, suggesting that the three-patch scale may be too coarse to capture gravity model effects.
• The model’s inability to account for multiple dengue strains—each causing distinct yearly epidemics—limits its accuracy, as immunity to one strain does not protect against others.
• The study concludes that improved models require integration of climate data into transmission parameters, inclusion of stochasticity, and exploration of alternative mobility models like the radiation model.