Poultry trading behaviours in Vietnamese live bird markets as risk factors for avian influenza infection in chickens

Abstract Vietnamese poultry are host to co‐circulating subtypes of avian influenza viruses, including H5N1 and H9N2, which pose a great risk to poultry productivity and to human health. AIVs circulate throughout the poultry trade network in Vietnam, with live bird markets being an integral component to this network. Traders at LBMs exhibit a variety of trading practices, which may influence the transmission of AIVs. We identified trading practices that impacted on AIV prevalence in chickens marketed in northern Vietnamese LBMs. We generated sequencing data for 31 H9N2 and two H5N6 viruses. Viruses isolated in the same LBM or from chickens sourced from the same province were genetically closer than viruses isolated in different LBMs or from chickens sourced in different provinces. The position of a vendor in the trading network impacted on their odds of having AIV‐infected chickens. Being a retailer and purchasing chickens from middlemen was associated with increased odds of infection, whereas odds decreased if vendors purchased chickens directly from large farms. Odds of infection were also higher for vendors having a greater volume of ducks unsold per day. These results indicate how the spread of AIVs is influenced by the structure of the live poultry trading network.

systems, often without robust biosecurity (Fournié et al., 2016;Webster, 2004). Live bird markets (LBMs) are a traditional aspect of these systems that facilitate the storage and sale of live poultry including chickens, ducks, quail and pigeons. As a consequence, LBMs play a significant role in the maintenance and spread of AIVs and thus pose a zoonotic risk to poultry workers and consumers, and to temporary workers enlisted during stamping out programmes (Bridges et al., 2002;Mounts et al., 1999). LBMs have been a primary target for AIV control strategies; during a zoonotic outbreak of H7N9 in China in 2013, closure of LBMs was shown to be remarkably effective in reducing the risk of human infection by up to 99% (Yu et al., 2014). Control strategies in LBMs have also been shown to significantly reduce AIV detection in chickens: the most effective strategies include monthly rest days that involve routine market closure followed by slaughter of unsold poultry, a ban to the sale of live quail and a ban to overnight storage of live poultry (Kung et al., 2003;Lau et al., 2007;Leung et al., 2012). However, although rest days are effective at breaking the viral amplification cycle in LBMs, they do not prevent reintroduction of virus. Indeed, rest days/nights are an important component of long-term AIV control but are not sufficient alone to eliminate infection (Kung et al., 2003). Furthermore, risk factor studies in LBMs have shown that having a greater variety of poultry species, including ducks being sold alongside other species, having poor sanitary conditions, storing poultry in floor pens instead of cages and having ≥1 wholesaler trading in LBMs, all increase the odds of having AIV-infected poultry and/or having AIV-contaminated environments (Kim et al., 2018;Kirunda et al., 2015;Santhia et al., 2009;Sayeed et al., 2017;Wang et al., 2018).
Vietnam has enzootic H5N1 and H9N2 and is at risk of incursion by H7N9 due to a shared border with China (Thuy et al., 2016).

Poultry traders are an integral component of poultry production in
Vietnam. They transport poultry from farms to LBMs, shaping a live poultry trading network through which AIVs may spread. Traders' practices may thus impact on the likelihood of introducing AIVs in LBMs and also facilitate the amplification of AIV circulation within marketed chicken populations (Fournié et al., 2012). However, a quantitative assessment of the association between poultry management practices and AIV prevalence in marketed chickens is lacking.

To address this gap, AIV infection status of chickens in Vietnamese
LBMs was assessed and the practices of traders offering them for sale characterized. This allowed us to assess the extent to which those practices may impact on the risk of viral circulation in LBMs.

| Sample collection
Eight live bird markets (six retail and two wholesale) in four provinces of northern Vietnam were included in our study, which was conducted between 2 October 2017 and 3 December 2017. Markets were selected if they had previously been confirmed positive for AIV in chickens within the past 12 months according to FAO-supported surveillance conducted by the National Centre for Veterinary Diagnostics (NCVD) and the Department of Animal Health (DAH, Hanoi). Markets were also selected if they were open seven days per week, facilitated the trade of live chickens, ducks and pigeons, and had more than 10 poultry traders operating in them. There were no quail (live or dead) at any of the LBMs, although this was not by design. Each LBM was sampled daily for seven consecutive days. On each day the LBM was sampled, the first 10 traders to arrive who contained at least five chickens in their flocks were recruited for the study and oropharyngeal swabs were collected from 5 chickens in each of their respective flocks, which were then pooled together.
Selected traders were then asked about their recent trading practices in a closed-ended questionnaire (Supplementary information Data S1). A total of 493 pools were collected from 2,465 chickens, and swabs were pooled in 2 ml virus transport medium (VTM) (Eagle's minimum essential medium supplemented with gentamicin, penicillin, streptomycin, bovine serum albumin, fungizol and HEPES solution) per trader and linked to traders and their questionnaire responses (on some days, less than 10 traders were sampled). Of the 493 sampled poultry traders, seven were removed during univariable and multivariable analysis due to incomplete feedback.
Environmental swabs were taken from three discrete areas of markets to determine the level of influenza virus A contamination of LBM environments. These discrete areas represented different poultry-related work activities which had previously been recommended to be included in routine monitoring and surveillance programmes for avian influenza viruses in LBMs (Indriani et al., 2010): (a) slaughter area including equipment used for slaughtering birds, (b) waste area including bins and containers used for disposing of bird waste such as feathers and (c) poultry stall including cages and the vicinity where birds were stored during LBM trading hours. Three swab samples were taken from each area and pooled each day (three separate pools representing three sampled areas generated per day).
All pooled swabs were maintained in cold chain for transportation to NCVD, Hanoi, where they were stored at −70°C until further processing.

| Sample screening and virus isolation
Virus RNA was extracted from pooled swab VTM using the QIAamp Thermal cycling conditions were as follows: 50°C for 15 min, 95°C for 2 min, then 40 cycles of 95°C for 10 s and 60°C for 30 s.
To ensure we could obtain sequencing data from our samples, we employed next-generation sequencing (NGS) on PCR products generated from viral RNA taken directly from pooled swab samples (Passage 0) and from infected allantoic fluid (Passage 1). Embryonated hens' eggs were inoculated with VTM from pooled swabs which had a Ct < 27 for H9 or any Ct for H5 and "unknown" subtype-positive samples. Allantoic fluid was harvested after 48 hr of incubation and confirmed for influenza virus A by haemagglutination (HA) assay. Viral RNA was extracted from allantoic fluid as above. In total, 50 samples meeting the above criteria were passaged in eggs, and this yielded 34 samples positive for HA activity. Passage 0 and Passage 1 samples were both subjected to NGS, and where possible, sequencing data for Passage 0 were used in phylogenetic analysis. Sequencing data were generated for a total of 33 viruses (31 H9N2 and 2 H5N6).

| Next-generation sequencing
Multisegment RT-PCR was conducted on viral RNA yielded directly from the VTM of pooled swabs and from inoculated allantoic fluid.
Briefly, this involved multigene amplification using the SuperScript TM III One-Step RT-PCR kit (Life Technologies) and the MBTUni12/13 universal primer set with specificity towards the conserved untranslated regions (UTRs) of each influenza virus gene (Zhou et al., 2009). These PCR products were used to generate DNA libraries

| Phylogenetic analysis
Alignment and analysis of nucleotide and deduced amino acid sequences were conducted using MEGA7 (Kumar, Stecher, Stecher, & Tamura, 2016). Neighbour-joining trees with 1,000 bootstrap replicates were also generated using MEGA7, and reference sequences for use in analysis alongside sequencing generated in this study were downloaded from the NCBI and GISAID databases.

| Statistical analysis
All statistical analyses were conducted using RStudio 2016. Data from hard copy questionnaires were entered into a Microsoft Access database. Logistic regressions were used to conduct univariable analysis of explanatory variables where Influenza virus A infection status of each pool of 5 chicken swabs was used as the response variable.
Explanatory variables with p < .05 were explored for collinearity by computing VIF values with the vif() function in the "car" package.
All variables with p < .05 from univariable analysis had VIF < 5 so were kept for subsequent stepwise variable selection. A final model of explanatory variables with LBMs as random effects was used in multivariable analysis. Final selection of explanatory variables was conducted by backward stepwise variable selection in R. Mantel tests were conducted in the R package, "ecodist", where virus isolates with whole-genome sequencing data were included (Mantel, 1967). For each of the virus isolates, the ORF of each gene segment was concatenated and a dissimilarity matrix, or genetic distance matrix, was constructed from the pairwise nucleotide differences in MEGA (Kumar et al., 2016). Additional dissimilarity matrices were also constructed from the explanatory variables of same dimension as the genetic distance matrix, and related to the characteristics of the poultry from which the viruses were isolated. We refer to them as sample characteristic matrices M. For any of those matrices, an element m ij = 1 if strains i and j are from samples with the same characteristic (e.g. poultry sold in the same market, poultry originating from the same type of premise, farm or market, and poultry originating from the same province), if not, mij = 0 (e.g. poultry sold in different markets, poultry originating from different types of premises and from different provinces).
There were no samples with detectable H7 influenza virus. Influenza virus prevalence varied greatly between LBMs with the two wholesale LBMs having the least amount of detectable influenza virus (Table 1). Of the 154 pooled environmental swabs, 70 (45%) were confirmed positive for influenza virus A. The proportion of positive pooled samples was similar across the different market areas that were sampled: poultry stall area (38%, n = 27); waste area (34%, n = 24); and slaughter area (27%, n = 19) (birds were not slaughtered in one LBM, for which it was not therefore possible to collect swabs samples for slaughter or waste sites).

| Phylogenetic analysis
Next-generation sequencing (NGS) of M gene-positive samples yielded whole-genome sequence data for 12 H9N2 viruses and partial genomes for 19 H9N2 and 2 H5N6 viruses. H9N2 viruses sequenced in this study were most closely related to previously sequenced H9N2 viruses from Vietnam (Thuy et al., 2016) (Figure 1 and Figure   S1). For example, BLASTn analysis of the PB1 gene of A/chicken/ Vietnam/1DO10/2017 from this study was most closely related to A/ chicken/Vietnam/H7F-BG4-383 with nucleotide homology of 98%.
These viruses retained the G57-like genotype, a prevalent genotype of H9N2 viruses in China known to be donors of all six internal genes to zoonotic H7N9 and H10N8 viruses (Pu et al., 2015).
We assessed whether the genetic distance between viral isolates was associated with their sampling location and the origin of chickens (i.e. the LBM that chickens were sampled in, the LBM/ farm type that poultry originated from or the province that poultry originated from). To do this, we utilized the 12 Vietnamese H9N2 viruses which we had full-genome sequence data for and concatenated their open reading frames. The genetic distance between any two of the 12 fully sequenced H9N2 isolates decreased if these two isolates originated from the same LBM (Mantel test, r = −.41, p = .004), or sampled chickens were sourced in the same province (r = −.37, p = .031) (see Table S1 for genotype distribution between sampled LBM and province source). In the light of this, we were able to classify viruses into seven different sub-genotypes using a > 98% nucleotide difference cut-off for each gene segment (for viruses where full-genome sequencing data were available) ( Figure 2). From this, we could see that several strains which originated from the same LBM were also grouped into the same genotype; genotype VN4 contained three viruses from LBM Pho Hien, and genotype VN5 contained three viruses from LBM Do. HA and NP genes had the greatest maximum nucleotide pairwise distance with 6.7% and 6.9%, respectively, followed by NS with 5.7%, NA with 5%, PB1 with 4%, PB2 with 3.9%, M with 1.6% and PA with 1.5%.

| Molecular characteristics of virus isolates
All H9N2 viruses were low pathogenic avian influenza (LPAI) viruses due to the presence of a dibasic cleavage motif RSSR/G in the haemagglutinin (HA) glycoprotein. However, the partial sequencing data for the HA genes of the H5N6 viruses contained the polybasic cleavage motif RRKR/G, classifying them as highly pathogenic avian influenza (HPAI) viruses (Chen et al., 1998)

| Risk factors associated with influenza virus A infection in chickens
Univariable analysis was used to identify potential risk factors related to the origin and management of poultry by traders, and subsequently included in multivariable analysis. Thirteen of 19 explanatory variables were identified as having a significant association with influenza virus A infection in chickens (Table 2) In the final multivariable model, the sampled LBM was used as a random effect because poultry traders were naturally grouped into the eight selected LBMs. Three risk factors and one protective factor were identified. The risk factors included sourcing poultry from middlemen, selling poultry to consumers and having a greater number of ducks unsold per day (Table 3). The protective factor was selling more chickens per day.

| Summary of poultry vendor practices
To put the identified risk factors into a broader context, we  A previous study by Fournié et al. (2012) have shown that it is possible to identify specific and distinct trader profiles of LBM sellers in Vietnam. As such, traders are classified as retailers or wholesalers based on who they primarily sell poultry to; retailers primarily sell directly to consumers, whereas wholesalers primarily sell to other poultry vendors within the trading network. In our study, we show retailers experienced higher odds of infection due to their trading practices. The retailers in our study were those who sourced their birds from middlemen, sold a relatively small volume of chickens and primarily sold directly to consumers. The risk factors associated with influenza virus A infection in chickens, selling only to consumers and buying from middlemen, can therefore be linked to retailers, which highlights their potential role in disseminating virus through the poultry trade network. In contrast, the wholesalers in our study were those who sourced from large commercial farms, sold a large volume of chickens and primarily sold to other vendors.

| D ISCUSS I ON
The protective risk factor of selling more chickens is associated with the practices of wholesalers and identifies this group of poultry traders as relatively low risk.
When considering the potential impact on AIV dissemination that these traders can have, it is important to take note of the position that vendors have in the poultry trade network. Vendors who have strong connections to a network of contacts operating in and around LBMs would be expected to have a more pronounced role in disseminating AIVs, whereas vendors holding a loose link to a network of contacts may have a reduced impact on AIV dissemination (Fournié et al., 2016). Thus, middlemen are mobile, highly connected poultry traders that travel between farms and LBMs to purchase and sell birds, mixing poultry from many different sources. As a consequence, they facilitate a network of LBMs that are tractable to the circulation of influenza viruses (Fournié et al., 2013(Fournié et al., , 2012. The identification of middlemen supplying poultry to traders as a risk factor for influenza virus infection could be explained by their mobility and propensity to mix poultry, and their high connectivity to the poultry trade network. Likewise, retailers could be associated with higher odds of infection because they may purchase birds that have "changed hands" multiple times, promoting the amount of time spent by birds within the trade network and facilitating the mixing of birds from different sources. All the LBMs included in this study were open seven days a week, which would allow for greater connectivity between traders as they have more opportunity to interact at LBMs, potentially increasing the risk posed by retailers in particular.
Although we did not explicitly capture the structure of the trade network in our study, the trading practices that we assessed can be used as indicators for the position of traders within the trade network.
In Vietnam, outbreaks of highly pathogenic avian influenza in spatially dispersed communes were shown to be closely linked to practices in agri-livestock farming systems, which involve communities producing rice, and domestic aquatic birds and chickens (Pfeiffer, Minh, Minh, Martin, Epprecht, & Otte, 2007).  (Indriani et al., 2010). Avian influenza viruses are frequently detected in shared poultry water , wooden tabletops, cages, bins and floors (Indriani et al., 2010). In our study,  (Nguyen et al., 2014).
Both large commercial farms and backyard flocks are included in emergency response H5 vaccination programmes (Domenech et al., 2009), and discretionary use of routine anti-H5 vaccines is practiced within some commercial farms in provinces believed to be high risk.
Going forward, vaccination in farms in highly connected trade networks where high-risk traders operate, as identified in this study, may be beneficial in mitigating AIV dissemination.
The primary limitation to our study was that poultry and poul- In conclusion, we have identified poultry trade practices that impact the risk of influenza virus A infection in chickens, and we have been able to attribute these practices to certain types of poultry trader. Being able to identify a specific type of poultry trader responsible for impacting AIV dissemination due to their poultry trading practices is novel and could be useful in future surveillance and control programmes. H9N2 viruses continue to cause significant poultry outbreaks and expand their global distribution within poultry producing countries. It is therefore increasingly important to monitor trends in H9N2 epidemiology, by using both active and passive surveillance systems that are already in place for H5 pandemic pre-

paredness. Surveillance of AIVs is particularly important in countries
where there is co-circulation of multiple subtypes. Prevention and control of zoonotic risks associated with endemic AIVs require continued surveillance efforts, and cost-effective targeted approaches to identify and protect high-risk poultry traders in highly connected trade networks.

ACK N OWLED G EM ENTS
We would like to thank the veterinary and laboratory staff at NCVD UK, for the use of the NGS platform. We would also like to thank Ms Nghiem Nguyen Minh Trang at the Oxford University Clinical Research Unit in Hanoi, Vietnam, for supporting the study.

E TH I C S S TATEM ENT
The approval for the study was obtained from the Department of Animal Health (DAH), Hanoi, and from each sub-department (SDAH) with jurisdiction over the provinces where selected LBMs were situated. FAO-supported surveillance of avian influenza viruses in LBMs is routinely conducted in northern Vietnam by NCVD and DAH, Hanoi, who enabled the implementation of this study.

CO N FLI C T O F I NTE R E S T S
The authors declare no competing interests.