Poultry farmer response to disease outbreaks in smallholder farming systems

Avian influenza outbreaks have been occurring on smallholder poultry farms in Asia for two decades. Farmer responses to these outbreaks can slow down or accelerate virus transmission. We used a longitudinal survey of 53 small-scale chicken farms in southern Vietnam to investigate the impact of outbreaks with disease-induced mortality on harvest rate, vaccination, and disinfection behaviors. We found that in small broiler flocks (≤16 birds/flock) the estimated probability of harvest was 56% higher when an outbreak occurred, and 214% higher if an outbreak with sudden deaths occurred in the same month. Vaccination and disinfection were strongly positively correlated with flock size and farm size, respectively. Small-scale farmers – the overwhelming majority of poultry producers in low-income countries – tend to rely on rapid sale of birds to mitigate losses from diseases. As depopulated birds are sent to markets or trading networks, this reactive behavior has the potential to enhance onward transmission. One sentence summary A cohort study of fifty three small-scale poultry farms in southern Vietnam reveals that when outbreaks occur with symptoms similar to highly pathogenic avian influenza, farmers respond by sending their chickens to market early, potentially exacerbating the effects of the outbreak.


Introduction
Livestock production systems have been a major driver of novel pathogen emergence events over the past two decades (1-3). The conditions enabling the emergence and spread of a new disease in the human population partly depend on human behavioral changes, like hygiene improvements or social distancing, in the face of epidemiological risks (4). The same observation applies to disease emergence and spread in livestock populations as farmers adapt their farm management to maximize animal production and welfare while limiting cost in a constantly changing ecological and economic environment (5).
Poultry farming generates substantial risk for disease emergence. It is now the most important source of animal protein for the human population and the industry is changing rapidly (6). The link between poultry sector expansion and pathogen emergence is exemplified by the worldwide spread of the highly pathogenic form of avian influenza (AI) due to the H5N1 subtype of influenza A, after its initial emergence in China in 1996 (2,7). Highly Pathogenic Avian Influenza (HPAI) causes severe symptoms in the most vulnerable bird species (including chicken, turkey, and quail), with mortality rates as high as 100% reported in broiler flocks (8). HPAI causes outbreaks in humans, and the risk that the pathogen makes the leap to a human pandemic is a persistent threat to public health (9). While HPAI does not persist in poultry populations in most affected countries, it has become endemic in parts of Asia and Africa and is periodically reintroduced into other areas like Europe and North America (10,11). In affected countries, major factors influencing HPAI epidemiology appear to be farm disinfection, poultry vaccination, and marketing of potentially infected birds through trade networks, all of which depend on farmers' management decisions (12)(13)(14)(15)(16).
It is still unclear how and to what extent changes in outbreak risk or mortality risk affect the behavior of poultry farmers. An anthropological study in Cambodia showed that high levels of farmer risk awareness associated with HPAI did not translate into major changes in their farming practices (17). Qualitative investigations conducted in Vietnam, Bangladesh, China, and Indonesia reported that farmers sometimes urgently sell or cull diseased poultry flocks as a way to mitigate economic losses, but evidence of this behavior's onward epidemiological impact was not available (12,(18)(19)(20)(21). Additionally, it is unknown whether poultry farmers increase application of disinfection practices or vaccination rates against avian influenza in response to disease outbreaks occurring in their flocks. Changes in farm management caused by variations in epidemiological risk have not been quantified for any livestock system that we are aware of, primarily because of the lack of combined epidemiological and behavioural data in longitudinal studies of livestock disease (22). Ifft  modelled the effect of cattle mortality and production performance on the frequency of sales and culling in New Zealand dairy farms (24). One limitation of these two studies is that the dynamics were observed over long time steps (1 year), which does not allow for a precise estimation of the timing of farmer response after the occurrence of disease outbreaks and the potential feedback effect of this response onto the resulting outbreaks or epidemics.
We present a cohort study of small-scale poultry farms where we aimed to characterize the effect of disease outbreaks on livestock harvest rate and on two prevention practices, vaccination and farm disinfection. This longitudinal farm survey was conducted on small-scale poultry farms in the Mekong river delta region of southern Vietnam (25). HPAI has been endemic in this region since its initial emergence in 2003-2004 (26). Small-scale poultry farming is practiced by more than seven million Vietnamese households, mostly on a scale of fewer than 100 birds per farm (27). Farm production in this sector is severely impacted by infectious poultry diseases. Aside from HPAI, Newcastle disease, fowl cholera, and Gumboro are all endemic despite the availability of vaccines for their control (28).

Data collection
An observational cohort study of small-scale poultry farms was conducted in Ca Mau province in southern Vietnam (25,29). Fifty poultry farms from two rural communes were initially enrolled and three additional farms were subsequently added to the sample in order to replace three farmers who stopped their poultry farming activity. The two communes were chosen based on their high poultry density and their history of past HPAI outbreaks (29). Study duration was 20 months, from June 2015 to January 2017. Monthly Vietnamese-language questionnaires were used to collect information on (1) number of birds of each species and production type, (2) expected age of removal from the farm, (3) number of birds introduced, removed, and deceased in the last month, (4) clinical symptoms associated with death, (5) vaccines administered, (6) type of poultry housing used, and (7) disinfection activity. Each farm's poultry were classified into "flocks", defined as groups of birds of the same age, species, and production type (25). The main poultry species raised in these farms were chickens, ducks and Muscovy ducks.

Selection of data and covariates
We fit models with three different dependent variables: a "harvest model" of the probability of harvesting (i.e. selling or slaughtering) flocks at a particular production stage (data points are flock-months), an "AI vaccination model" of the probability of performing AI vaccination on flocks which had never received AI vaccination (data points are flock-months), and a "disinfection model" of the probability of disinfecting farm facilities (data points are farm-months). For the two first models, we focused our analysis on broiler chicken flocks. Chicken was the predominant species in the study population, the overwhelming majority of chicken flocks were broilers, and their age-specific harvest was easier to predict than the harvest of layer-breeder hens. Additionally, only six layer-breeder chicken flocks were vaccinated against AI during the study period. More details on data selection are provided in the supplementary materials and methods 1.
Outbreak categorical variables were included in each model, corresponding to an outbreak occurring in the corresponding farm in the same month, one month prior, and two months prior.
An outbreak was defined as the death of at least two birds of the same specieson the same farm, in the same month, with similar clinical symptomsas this may indicate the presence of an infectious pathogen on the farm. For the harvest model, only outbreaks in chickens were considered. For the AI vaccination model, outbreaks in chickens and outbreaks in any other species were included as two separate covariates. For the disinfection model, outbreaks in any of the species present in the farm were considered. In chickens, outbreaks with "sudden deaths" (i.e. the death of chickens in less than 24h after the onset of clinical symptoms) are considered as suspicions of highly pathogenic avian influenza (30). Therefore, we created two sub-categorical variables for outbreaks in chickens, with sudden deaths (OS, "outbreaks sudden") and with no sudden deaths (ONS, "outbreaks not sudden"). For both the harvest and AI vaccination models, we assumed the effect of outbreaks on the dependent variable may be affected by the size of the considered flock (n). Consequently, we included this interaction term in the analysis.
The three dependent variables are likely affected by several farm-, flock-, and time-related factors, justifying the inclusion of several control covariates in the multivariate models. For the harvest model, we used the observed age of the chicken flocks (t) and the reported "intended age at harvest" (t*) and used δt = tt* as the key independent variable indicating whether a flock was observed before or after its intended harvest time. For the AI vaccination model, the age of the chicken flocks t was considered as influencing the likelihood of vaccination. Other control covariates included flock size, age, calendar time (T), housing, vaccination status, introduction of other flocks onto the farm, and farm size (i.e. a farm's poultry population size). The farm poultry population was broken down by species (chicken, duck, and Muscovy duck) and type of production (broiler and layer-breeder). Summary statistics of variables are displayed in the Table   1 and all control covariates are listed in the supplementary materials and methods 2.

Multivariate modelling
We assumed that the events of interest, namely harvest, AI vaccination, and disinfection were drawn from a binomial distribution and used a logistic function to link their probability to a function of the independent covariates. Some of the included effects are non-linear in nature, and we needed to account for the intra-farm autocorrelation of the dependent variables. We therefore used a mixed-effects general additive model (MGAM) implemented in R with the "mgcv" package (31). This enabled us to model the combined effect of δt, t*, and flock size (n) on harvest time; the effect of t and n on AI vaccination; and the effect of calendar time (T) on all the dependent variables, as penalized thin plate regression splines (32). We specifically chose these variables because they are presumably the most important factors influencing the dependent variables and their effect could possibly be highly non-linear. All other covariates were included as parametric regression terms. We also modelled the individual effects of farms on the dependent variables as random effects.
The complete models linking the logit Yij of probability of realization of an event and the set of explanatory variables, for a flock-month i (harvest, vaccination for AI) or a farm-month i (disinfection) in a farm j, are described by the following set of equations: Harvest model (flock-month level): The model parameters are α the model intercept; β the parametric coefficients; f a thin-plate spline function; X k the general notation for variables with linear effects; X O-m , X OS-m , X ONS-m and X OD-m , categorical variables denoting presence or absence of an outbreak in the same farm m months prior in any species (O), in chickens with sudden deaths (OS), in chickens with no sudden deaths (ONS), and in different species (OD) respectively; n the flock size; t the current age of the flock; t* the age at maturity of the flock anticipated by the farmer; δt the difference between current age and age at maturity; T the calendar time; φ the farm random effect; ε the residual error term. Interaction terms between outbreak categorical variables and flock size log(nij) were added in the Harvest and AI vaccination models.
Some variables with a highly skewed distribution were transformed (either log-or square-roottransformed). Details are provided in the supplementary materials and methods 3. Excessive multi-collinearity between covariates was assessed by estimating their variance inflated factor using the "usdm" R package (33). We fitted the complete models using the whole set of covariates using restricted maximum likelihood estimation. We then used a backward-forward stepwise selection, based on Akaike Information Criterion (AIC) comparison, to eliminate the variables with non-significant effects (34). Arguably, one farmer is likely to maintain the same farm management from one month to the next despite changes in influential covariates. Therefore, for each model, we tested the presence of farm-level time autocorrelation in the model deviance residuals. If there was sufficient evidence for the presence of autocorrelation, we implemented the same model fitting protocol with an additional intra-farm AR-1 time autocorrelation term on the dependent variable (32). More details on testing and accounting for time autocorrelation are provided in the supplementary materials and methods 4.
All analyses and graphical representations were performed with R version 3.6.1 (35).

Ethical statement
The collaboration between the investigators (authors) and the Ca Mau sub-department of livestock production and animal health (CM-LPAH) was approved by The Hospital for Tropical Diseases in Ho Chi Minh City, Vietnam. The CM-LPAH, which the equivalent of an ethics committee for studies dealing with livestock farming, specifically approved this study; at the province-level.

Results
Fifty three farms were monitored from June 2015 to January 2017.  A total of 1656 broiler chicken flock-months were available for analysis. They belonged to 391 chicken flocks present on 48 farms. Occurrences of outbreaks in monitored farms in the same month, 1 month prior, and 2 months prior are summarized in Table 1. Additional descriptive statistics on control covariates are also displayed in Table 1 Table 2. Fitted spline functions cannot be elegantly summarized by their coefficients and are most adequately represented with graphs (in Figures 2 and 3).  16 20 Proportion harvested (%)  The harvest model explained 34.2% of the observed deviance. Probability of harvest of chicken broiler flocks of the same farm was not shown to be auto-correlated in time As the interaction term between flock size (n) and outbreak occurrence is significant (p < 0.01) but difficult to interpret (displayed in Table S1) weeks away from being designated as mature. The probability of harvest increased steeply from δt = -10 to δt = 0. For δt > 0 (flocks past their age at maturity), the probability of harvest was consistently high but lower than 100% and did not depend on age. Larger flocks had a steeper increase in harvest probability as a function of δt; once past the mature age (δt > 0), the estimated probability of harvest for large flocks was higher (41% -55%) than for small flocks (25% -41%) ( Figure 2).  However, the small sample sizeonly seven observations of outbreaks with sudden deaths in large flocksmeans that we do not have sufficient statistical power to support this association. In the last six months of data collection, farmers were asked to indicate the destination of harvested birds.
Based on these partial observations, flocks harvested during or one month after outbreaks in chickens (OS or ONS) were more likely to be sold to traders and less likely to be slaughtered at home ( Table 3). The likelihood of harvest was also positively correlated with the number of other broiler chickens present on the farm (Table S1, Table 2). Blue histograms correspond to the number of observed flockmonths in the different classes of δt (scaled to their maximum, 139 in the top graphs and 157 in the bottom graphs).  Slaughter at home 36 20 11 Gift 5 8 11 Feed farmed pythons 5 1 22 Other 7 3 0 In the AI vaccination model, 71.9% of the observations' deviance was explained. The probability for broiler chicken flocks on the same farm to be vaccinated against AI was not found to be auto-correlated in time. The likelihood of broiler chicken vaccination against AI strongly increased with flock size; probability of vaccination was almost zero for flocks of 16 birds or fewer and nearly 100% for flocks of more than 200 birds (Figure 3.A). Vaccination was preferentially performed at 4.3 weeks of age (Figure 3 Table 2). Farm disinfection appeared to have a seasonal component. It was least likely in October-November and most likely in the January-April period ( Figure 3D). It was not found to be affected by the occurrence of outbreaks on farms (no decrease in AIC when including outbreak occurrence).

Discussion
Regions like the Mekong river delta combine high human population density, wildlife biodiversity, and agricultural development. As such, they are considered hotspots for the emergence and spread of novel pathogens (36). The high density of livestock farmed in semi-commercial operations with limited disease prevention practices further increases the risk of spread of emerging pathogens in livestock and their transmission to humans (16). In-depth studies of poultry farmers' behavioral responses to disease occurrence in animals are needed to understand how emerging pathogensespecially avian influenza virusesmay spread and establish in livestock populations and how optimal management policies should be designed. To the best of our knowledge, this study is the first to provide a detailed and quantified account of the dynamics of livestock management in small-scale farms and its evolution in response to changing epidemiological risks shortly after disease outbreaks occur. While our analysis was performed on a geographically restricted area, the decision-making context of the studied sample of farmers applies to a wide range of poultry producers in low-and middle-income countries. Small-scale poultry farming, combining low investments in infrastructure, no vertical integration, and subject to limited state control on poultry production and trade, is common in most regions affected by avian influenza, in Southeast Asia, Egypt, and West Africa (37)(38)(39)(40).
In our cohort study, owners of small chicken broiler flocks resorted to early harvesting of poultry, also referred to as depopulation, as a way to mitigate losses from infectious diseases. The revenue earned from the depopulation of flocks might be low, either because birds are still immature or because traders use disease symptoms as an argument to decrease the sale price.
Nevertheless, depopulation allows the farmer to avoid a large revenue loss resulting from diseaseinduced mortality or the costs of management of sick or dead birds. More importantly, farmers avoid the cost of feeding chickens at high risk of dying and prevent the potential infection of subsequently introduced birds. Our results also suggest that the depopulation period, which lasts approximately 2 months, is followed by a "repopulation" period during which farmers lower their harvest rate, possibly to increase their pool of breeding animals in order to repopulate their farm.
The epidemiological effect of chicken depopulation is likely twofold: on the one hand it may slow the transmission of the disease on the farm, since the number of susceptible and infected animals is temporarily decreased; on the other hand, since most poultry harvested during or just after outbreaks were sold to itinerant traders or in markets, depopulation increases the risk of dissemination of the pathogens through trade circuits (41). There is epidemiological evidence that poultry farms can be contaminated with HPAI through contact with traders who purchase infectious birds and that infectious birds can contaminate other birds at traders' storage places and in live bird markets (12,14,42). and who may offer a lower price per bird. When farm production increases, farmers tend to rely on pre-established agreements with traders, middlemen, or hatcheries on the sale dates in order to reduce these transaction costs, giving them little possibility to harvest birds at an earlier time (44).
While government-supported vaccination programs have been proposed as a suitable tool to control AI in small scale farms with little infrastructure (45), in this cohort AI vaccination was almost exclusively performed in large flocks kept indoors or in an enclosure. Vaccination against AI is believed to be inexpensive for farmers as vaccines are supplied for free by the sub-department of animal health of Ca Mau province and performed by local animal health workers. However, vaccination may still involve some fixed transaction cost as farmers have to declare their flocks to the governmental veterinary services beforehand. Also it is possible that small flocks, being less likely to be sold to distant larger cities (46), are less likely to have their vaccination status controlled, making their vaccination less worthwhile from the farmers' perspective. Crucially, it is these smaller flocks that are more likely to be sold into trading network during outbreaks.
Finally, farmers' willingness to expand their production, invest in farm infrastructure, and implement AI prevention are likely correlated. Farms with a large breeding-laying activity tend to invest more in preventive actions (disinfection and vaccination) compared to farms specialized in broiler production. This may reflect a higher individual market value of layer-breeder hens compared to broiler chicks, making their protection more beneficial.
While vaccination against AI and disinfection appear to depend on individual farmer attitude, as shown by the significance of the farm random effects, they still vary over time when viewed across all farms (Figure 1). Contrary to harvesting behavior, these preventive actions have a seasonal component (Figure 3.C and 3.D) indicating a willingness to maximize the number of vaccinated broiler chickens and the protection against other diseases during the January-March period. This would be consistent with epidemiological observations as AI transmission increases in the January-February period (26,47). Farm disinfection has a significant temporal autocorrelation component and is unaffected by disease outbreaks, indicating that farmers are slower at adapting this practice to changing conditions.
The last 23 years of emerging pathogen outbreaks and zoonotic transmissions failed to prepare us for the epidemiological catastrophe that we are witnessing in 2020. Multiple subtypes of avian influenza viruses have crossed over into human populations since 1997 (3,11), all resulting from poultry farming activities. Small-scale poultry farming is likely to be maintained in low-and middle-income countries as it provides low-cost protein, supplemental income to rural households, and is supported by consumer preference of local indigenous breeds of poultry (37,48,49). If we ignore the active role that poultry farmers play in the control and dissemination of