Investigating poultry trade patterns to guide avian influenza surveillance and control: a case study in Vietnam

Live bird markets are often the focus of surveillance activities monitoring avian influenza viruses (AIV) circulating in poultry. However, in order to ensure a high sensitivity of virus detection and effectiveness of management actions, poultry management practices features influencing AIV dynamics need to be accounted for in the design of surveillance programmes. In order to address this knowledge gap, a cross-sectional survey was conducted through interviews with 791 traders in 18 Vietnamese live bird markets. Markets greatly differed according to the sources from which poultry was obtained, and their connections to other markets through the movements of their traders. These features, which could be informed based on indicators that are easy to measure, suggest that markets could be used as sentinels for monitoring virus strains circulating in specific segments of the poultry production sector. AIV spread within markets was modelled. Due to the high turn-over of poultry, viral amplification was likely to be minimal in most of the largest markets. However, due to the large number of birds being introduced each day, and challenges related to cleaning and disinfection, environmental accumulation of viruses at markets may take place, posing a threat to the poultry production sector and to public health.


Algorithm for sample size calculation
Using the following algorithm, the minimum number of traders to be interviewed per LBM was defined in order to identify with a probability of 95% at least 90% of the "sites" visited by traders to purchase or sell poultry, assuming that each trader visited 5 sites, and each site was visited on average by 5% of traders operating in that LBM.
A LBM with N traders was simulated. Each trader visited nL sites, and each site was visited by a proportion P of the traders operating in that LBM. A site could be either another LBM or the poultry farms in a district (second administrative division).
First, the total number of sites L visited by the population of traders was calculated as: The formulation of the denominator ensures that it lied between 1/N (each site was visited by one trader only) and 1 (each site was visited by all traders). L was rounded to the nearest integer.
For each trader, nL different sites were randomly drawn. Although the L sites were visited by an average of NP traders, the random allocation of these sites meant the actual number of traders visiting each site varied.
s traders were then randomly sampled. The proportion p of sites visited by traders operating in this LBM that were identified through this sample was computed.
This algorithm was repeated 10,000 times, and the proportion of simulations for which p was equal or higher than a defined threshold T was assessed.
Here, nL=5, P=0.05 and T=0.9. Sample sizes for which the probability to detect at least a proportion T of all sites visited by traders was higher than 0.95 are reported in Supplementary   Table S1.

Basic reproduction number calculation
The dominant eigenvalue of the next generation matrix was the estimate of the basic reproduction number 0, m k R in LBM k 1 . Each element rz,x of the matrix was the expected number of chickens becoming pre-infectious after having spent x hours in the LBM due to direct or indirect contacts with a primary case that became pre-infectious after having spent z hours in the LBM. It was given by:  (Supplementary Fig. S1). bird was initially set as infected, and the number of susceptible birds entering into markets every day was chosen to be high (10,000) in order to avoid extinction once the disease reached its endemic equilibrium. Therefore, in this specific setting, a viral introduction resulting in a major outbreak in the host population would also lead to the disease becoming endemicdisease invasion was here similar to disease endemicity. The probability of disease endemicity (or invasion), for a given demographic profile, was the proportion of simulations for which at least one bird was infected or virus was present in the environment 100 days after viral introduction.

Description of the multivariate analysis results
The PCA was performed for LBMs based on variables related to (i) the number and sources of chickens sold, and (ii) the egocentric network characteristics. For both analyses, the two first components were selected. For the former analysis, the two first components accounted for 40% and 20% of the data variability, and for the latter analysis they accounted for 46% and 23%. LBMs with a high score for each of these components are described below. The converse is true for LBMs with a low score.

Practices of interviewed traders
The practices of all interviewed traders are summarised in Supplementary Table S2.

Partitions of LBMs under different assumptions
LBMs were partitioned according to the number and sources of chickens sold under different assumptions: crude (Supplementary Tables S3-S4) or simulated trader populations   (Supplementary Tables S5-S6 Table S8).