Reducing size bias in epidemic network modelling

Epidemiological models can inform policymaking on disease control strategies, and these models often rely on sampled contact networks. The Random Walk (RW) sampling algorithm, commonly used for network sampling, produces size-biased samples that over-represent highly connected individuals, leading to biased estimates of disease spread. The Metropolis-Hastings Random Walk (MHRW) addresses this by providing samples representative of the underlying network's connectivity distribution. We compare MHRW and RW in reducing size bias across four network types: Erdös-Rényi (ER), Small-world (SW), Negative-binomial (NB), and Scale-free (SF). We simulate disease spread using a stochastic Susceptible-Infected-Recovered (SIR) framework. RW tends to overestimate infections (by 25 % in ER, SW, NB) and secondary infections (by 25 % in ER, SW and 80 % in NB), and underestimate time-to-infection in NB networks. MHRW reduces the size bias, except on SF networks, where both algorithms provide non-representative samples and highly variable estimates. We find that RW is appropriate for fast-spreading, high-mortality epidemics in homogeneous or moderately random networks (ER, SW). In contrast, MHRW is better suited for slower and low-severity epidemics and can be effective in both homogeneous and heterogeneous networks (ER, SW, NB). However, MHRW is computationally expensive and less accurate when duplicate nodes are removed. We also analyse real-world data from cattle movement and human contact networks; MHRW generates disease spread estimates closer to the underlying network than RW. Our findings guide the selection of sampling algorithms based on network structure and epidemic characteristics, enhancing the reliability of disease modelling for policymaking.