Ethics statement

Ethical approval for the study was obtained from the University of Bristol, Faculty of Health Sciences Research Ethics Committee, approval number 17048 on 16 January 2024.

Informed written consent (using e-consent hosted on REDCap) for the use of data collected via the questionnaire was obtained from respondents. A copy of the consent form and participant information sheet is given in the accompanying underlying data.

Data description

We used data from the Avian Contact Study—a questionnaire open to individuals aged 18 or over in the UK who have contact with domestic and/or wild birds to investigate the potential transmission dynamics of HPAI and inform public health policies [22]. Although the questionnaire was open to people aged 18 and above, no individuals aged18 responded, so our dataset includes only those aged 19 and over. Most respondents were poultry farmers, reporting contact with poultry, mainly chickens (82%). Respondents generally had high levels of bird exposure: 20% reported flock sizes over 10000 birds, and more than 60% reported daily contact with birds [22]. The questionnaire was launched in-person at the British Pig and Poultry Fair (15–16 May 2024) and subsequently circulated online via various poultry forums and networks by word-of-mouth (15 May to 31 October 2024). The questionnaire was divided into four sections:

1. Section A: “About you”
2. Section B: “Your contact with any type of domestic or wildlife birds”
3. Section C: “Your contact with other people”
4. Section D: “Your awareness of avian influenza (bird flu)”

Details of the eligibility criteria, questionnaire format, patient and public involvement, and key results are described in Thomas AC, Gokool S, Clapp G et al. (2024) [22].

We extracted self-reported contact data with other people to inform a branching process model. Direct contact was defined as either physical contact (such as a handshake, hug, kiss, or contact sport), conversational contact (exchanging at least a few words), or contact by distance (being within 2 metres of another person). Respondents were asked how many direct contacts they had in the last 24 hours, with the response options “1-5”, “6-10”, “11-15”, “16-20”, and “More than 20”. Respondents then provided further details about each contact, such as relationship to the contact, age group of the contact, location, duration, and type of contact (physical, conversational, or by distance). Although the categorical response included an option of “More than 20 contacts”, respondents could only provide further information for up to twenty contacts in the following questions.

Additionally, we used contact data from CoMix, a large-scale contact survey started during the COVID-19 pandemic and conducted in several European countries, including the UK, as a comparator dataset from the general population [23]. We took data relating to the final survey round collected post-pandemic in November 2022. The data processing for CoMix is described in S1 File.

Data analysis

Since the number of reported contacts was a categorical variable, we counted the number of contacts listed by each participant to obtain an integer variable. This allowed calculation of summary statistics, degree distributions, and informed the branching process model. We used the ‘Relationship to the contact’ column to count the listed contacts, due to its lower proportion of missing data. We then compared the number of listed contacts to the reported contact category. No exclusions were made if the number of listed contacts did not align with the reported contact category. The number of listed contacts was used in our model.

We calculated the average number of contacts per person per day overall and stratified by age group and compared these values to CoMix [23]. Distributions of contacts were calculated for three age groups (19–44, 45–59, 60+) representing young adults, adults, and older adults, as well as for all age groups combined. We fitted negative binomial and Poisson distributions to the data using the Maximum Likelihood Estimate method from the ‘fitdistrplus’ package version 1.2-1 in R and compared both fits using the likelihood ratio test.

Social networks commonly exhibit structural heterogeneity, and contact matrices are a widely used method for representing contact heterogeneities across age groups [24,25]. We constructed age-stratified contact matrices to illustrate the average number of contacts made between different age groups, both overall and stratified for family contacts. We grouped ages into 10-year intervals to balance the level of detail without compromising sample sizes in each age group.

Model framework

We used a branching process model to simulate early-stage outbreaks, as it can capture stochastic extinction and outbreak variability for infections with limited human-to-human transmission. We used the model to estimate the potential outbreak size and effectiveness of contact tracing and self-isolation in preventing the spread of avian influenza (Fig 1) [26–28].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Contact Tracing Schematic.

Example of a simulated transmission chain starting with person A, the index case, who was infected through contact with an infected bird. The number of secondary infections generated by each person is drawn from a negative binomial distribution. The time of infection for new cases is drawn from a gamma distribution, relative to the symptom onset of the infector. In this example, Person A infects 2 individuals, B and C, before isolating. If a contact is traced, they isolate on the day they develop symptoms and do not cause further infections. They may still be able to infect other people before showing symptoms. If a traced contact does not develop symptoms, or a contact is not traced, they would not isolate.

https://doi.org/10.1371/journal.pgph.0005463.g001

We defined outbreak size as the number of individuals in a cluster including the index case and analysed how it changed with the basic reproduction number, R₀, the average number of secondary infections generated by an infected individual in a wholly susceptible population [29]. We varied R₀ from 0 to 1.1 in increments of 0.1, consistent with estimates from previous studies on avian and pandemic influenza [13,30]. To capture contact heterogeneities with age, we divided the population into four age groups (0–18, 19–44, 45–59, 60+). Since our dataset did not include individuals under 19 years old, we did not explore scenarios where index cases were under 19, but they are included in transmission chains using contact data from CoMix.

Number of secondary cases

The number of contacts made by an infected individual was drawn from an age-specific negative binomial distribution. The number of secondary infections (R) generated by that case was calculated as the product of the number of contacts (n), a fixed infectious period (d), and the probability of transmission (, where R₀ is the basic reproduction number, is the dominant eigenvalue of the contact matrix). Mathematically, the number of secondary infections generated by an individual in age group (for age groups 19–44 years, 45–59 years and 60 + years) is given by:

, where {19–44, 45–59, 60+}

The age group of each secondary case was determined by sampling from a multinomial distribution, using probabilities from the corresponding row of the contact matrix, normalised to one. Time of infection for each secondary case relative to the symptom onset of the primary case was drawn from a shifted Gamma distribution taken to reproduce estimates from the literature (shape parameter 13, rate parameter 2, resulting in a mean of 3.5 days (SD: 1.80 days) between the symptom onset and infection) [8,13,29–32]. We adjusted the distribution to allow infections to happen up to one day prior to the onset of symptoms.

Asymptomatic infections

The probability of being asymptomatic may depend on factors such as age and the strain of the virus. However, due to limited data on human cases of H5N1 avian influenza, the proportion of asymptomatic cases and the impact of age on this probability are not well understood. For the baseline scenario, we assumed 40% of individuals developed asymptomatic infections [8,9]. However, as the proportion of asymptomatic infections has a major impact on the effectiveness of contact tracing and self-isolation, we explored a full range of scenarios from 0% of infections being asymptomatic to 100%

Contact tracing and self-isolation

We investigated the impact of contact tracing and self-isolation on the outbreak size. We assumed that 40% of contacts were traced in the baseline scenario, based on an estimate from reported figures during the COVID-19 pandemic [33]. We determined whether new cases would be traced based on their symptom status. By default, the index case and the contacts of asymptomatic cases were not traced. Contacts of symptomatic cases could be traced with probabilities varying from 0 (no contact traced) to 1 (all contacts traced). For the purposes of this model, we defined self-isolation as the separation from others of symptomatic individuals who had been traced by public health teams, beginning on the day of symptom onset, in accordance with current guidelines. Traced contacts who do not develop symptoms, and untraced contacts, would not self-isolate.

We ran the model 9000 times per parameter set with the maximum time for each simulation of 100 days.

We used R version 4.4.3 in our analyses.

Contact patterns

The distribution of numbers of contacts is shown in Fig 2A. The average number of contacts was 4.64 (standard deviation (SD): 3.40). A Negative Binomial (NB) model provided a better fit than a Poisson model based on Akaike Information Criterion (NB: 968, Poisson: 1052), Bayesian Information Criterion (NB: 974, Poisson: 1056), and likelihood ratio test (p-value < 0.001). The final NB model had a mean of 4.64 and a dispersion parameter of 4.02. Additionally, we fitted NB distributions to the number of contacts stratified by age group to inform the branching process model, shown in S2 Fig.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Contact patterns.

Contact patterns reported in the Avian Contact Study. (A) the distribution of the number of contacts per person. The red line indicates the fitted negative binomial distribution, and the pink region represents the 95% confidence interval. The fitted negative binomial distribution had a mean of 4.64 and a dispersion parameter of 4.02. (B) Average number of contacts between age groups, while (C) Average number of contacts between family members only. In both (B) and (C), darker colours indicate higher average contacts.

https://doi.org/10.1371/journal.pgph.0005463.g002

Contact matrices suggest intergenerational living in agricultural workers

We constructed age-structured contact matrices to show the average number of contacts made between different age groups, both for all contacts and family members (Fig 2B, C). We observe broadly similar patterns to previous contact surveys, with respondents interacting more frequently with individuals of a similar age. Respondents aged 19–29 reported an average of 6.8 contacts per day, with nearly 90% of their contacts occurring with individuals aged 20–59. Respondents aged 30 and above had an average of 4.4 contacts per day.

Fig 2C illustrates contact patterns among family members. Respondents reported an average of 1.8 family contacts per day, with the 19–29 age group having the lowest (mean: 1.35 contacts, SD: 0.18 contacts) and the 40–49 age group the highest (mean: 1.97 contacts, SD: 0.14 contacts). We observed three distinct diagonal patterns: younger individuals frequently interacting with their parents, older individuals with their children and grandchildren, and individuals interacting with those of a similar age, reflecting intergenerational living arrangements.

A comparison of the Avian Contact Study and CoMix is shown in Fig 3. Respondents in the Avian Contact Study had higher average contact levels than respondents of CoMix across all age groups. The smaller sample size in the Avian Contact Study compared to CoMix may explain the wider confidence intervals around the mean. Statistical comparisons of the contact patterns are provided in S2 File.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Average number of contacts from the Avian Contact Study (AvCos) and CoMix for all age groups.

Black lines indicate the 95% confidence interval, and grey error bars indicate the interquartile range (IQR).

https://doi.org/10.1371/journal.pgph.0005463.g003

Influence of age on outbreak size

As expected, outbreak size was highly dependent on the reproduction number. Infections with reproduction numbers less than 0.2 did not lead to onward transmission. Over 75% of simulations resulted in small-scale outbreaks (0–10 cases). The average outbreak size increased sharply from 4 cases at R₀ = 1–41 cases at R₀ = 1.1 (S3 Fig).

We also examined how the age group of the index case influenced outbreak dynamics. For R₀ between 0.7 and 1, half of simulations with index cases aged 60 + resulted in no secondary cases, compared to 40% of simulations for those aged 19–59. When R₀ = 1.1, 9% of simulations with index cases in the 19–59 age group led to outbreaks larger than 50 cases, compared to 6% for those with index cases aged 60 + .

For index cases in the 19–44 and 45–59 age groups, R₀ did not significantly affect the age distribution of the cases. In contrast, for index cases aged 60 + , we observed an increase in the percentage of cases in the 0–18 age group from 0 to 13% as R₀ increased. This is due to limited direct contact between the 60+ and 0–18 age groups, meaning individuals aged 0–18 were infected further down the transmission chain through interactions with other age groups (S3 and S4 Figs).

Reduced effectiveness of contact tracing and self-isolation with higher proportion of asymptomatic infections

We compared how outbreak sizes changed when 40% of infections were asymptomatic, under both no control and control (contact tracing and self-isolation) scenarios, depending on R₀ (Fig 4A). Under no control, no secondary cases were observed when R₀ was below 0.2 - the only case was the index case. Over 75% of simulations resulted in small-scale outbreaks (less than 10 cases). Contact tracing and self-isolation reduced the maximum outbreak size from 5705 cases to below 1000, and the average outbreak size from 41 cases to below 10, when R₀ = 1.1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. The impact of contact tracing and self-isolation on final outbreak size.

(A) The distribution of outbreak sizes for varying levels of R₀ under no control and control (contact tracing and self-isolation). The orange line indicates the average outbreak size, and the shaded region represents the 95% prediction interval. We assumed that 40% of cases were asymptomatic and that 40% of contacts were traced. Among those traced, all symptomatic cases were assumed to self-isolate. (B) The average outbreak size across 9000 simulations for varying levels of contact tracing and proportions of asymptomatic infections when R₀ = 1.1. Shaded regions represent the 95% confidence interval.

https://doi.org/10.1371/journal.pgph.0005463.g004

We also evaluated how the average outbreak size changed under different levels of control measures and varying proportions of asymptomatic infections (Fig 4B). When the proportion of asymptomatic infections was low, even low to moderate levels of contact tracing and self-isolation (0–40%) were sufficient to reduce the average outbreak size from 41 cases to below 10. However, the reduction in the average outbreak size became less pronounced as the proportion of asymptomatic infections increased.

Principal findings

Our model provides insights into the potential outbreak size of avian influenza among humans under different levels of transmission and a full range of proportions of asymptomatic infections to account for uncertainties around these values. Most simulations resulted in small-scale outbreaks, with outbreaks exceeding 100 cases when the basic reproduction number was above 0.8, consistent with previous estimates [34]. Contact tracing and self-isolation reduced outbreak size, but their potential effectiveness in preventing cases decreased as the proportion of asymptomatic infections increased, since they rely on symptomatic cases. Additionally, larger values of R₀ would increase outbreak size and the number of contacts, which may slow down contact tracing efforts and reduce the effectiveness of current interventions. Previous modelling studies similarly concluded that contact tracing and self-isolation alone were highly unlikely to prevent large outbreaks, particularly when asymptomatic infections were present [28,35]. Our findings support and extend these conclusions to a different pathogen.

Translating findings into action

Limited understanding of the proportion of asymptomatic H5N1 infections and the factors influencing symptomatology, combined with their potential to reduce the effectiveness of contact tracing and self-isolation, highlights that these measures, while valuable, are vulnerable. Embedding them within a layered response strategy that includes asymptomatic testing programmes, which have previously shown to be successful at detecting cases, could enhance current interventions [36].

Moreover, we observed that the index cases aged 19–59 play an important role in transmission dynamics, leading to larger outbreaks based solely on contact data. Additionally, analysis of previous H5N1 infections suggests that the infection is more prevalent among younger adults, and older age may not be a risk factor for severe outcomes [37]. Taken together, these findings suggest that younger adults may benefit more from targeted interventions.

Strengths and limitations

We used a novel dataset from the Avian Contact Study, which focused on individuals at higher risk of avian influenza and was developed with the help of public contributors to ensure relevance and accessibility [22]. To our knowledge, it is the only contact survey conducted in the UK that has focused on farming communities since the COVID-19 pandemic. This allowed us to examine contact patterns in rural, high-risk groups, revealing strong intergenerational household mixing that reflects the multigenerational nature of farming enterprises, a pattern often not observed in the general population [38]. We used age-stratified mixing patterns in our model to estimate outbreak sizes, extending previous models that did not incorporate age structure. This allowed us to capture variation in outbreak sizes across age groups and identify which groups were more likely to drive larger outbreaks.

In addition to common limitations of other contact surveys [21,39], our data did not include individuals under the age of 19, so we could not compare their contact patterns to those of the general population. Although we included 0–18 age group from CoMix in the transmission chain, this may not fully reflect the contact patterns due to differences in the cohorts and study periods.

Additionally, our analysis was based on data from individuals living in the UK, most of whom were poultry farmers. Contact patterns may differ in low- and middle-income countries, where intergenerational living and higher contact rates among older adults are more common [40]. Furthermore, a study by Eze et al. comparing epidemic preparedness and response capacity across 186 countries identified countries with lower response capacity, most of which are low- and middle-income countries [41]. Hence, our findings may not be generalised to other countries with different contact structures or limited diagnostic and contact tracing resources.

Although we included age in our model, we could not explore its relationship with all the parameters. For instance, the influence of age on symptomatology is not well understood. As a result, we kept the proportion of asymptomatic infections constant across all age groups. Our model assumed that people with symptoms would isolate on the same day, although this may not be the case in practice with people adopting a ‘wait and see’ approach [42].

Areas for future work

Our sample was relatively homogeneous in the occupational group, with most respondents being poultry farmers. In addition to farmers, who make up 60% of the agricultural workforce in the UK, 12% are casual workers involved in seasonal work and often exhibit different living arrangements [43]. Despite their high exposure risk and potential influence on outbreak dynamics, it can be challenging to engage with this group due to language and trust barriers. Hence, future research is required to further characterise contact patterns within agricultural demographics and develop effective strategies to engage with groups that are underrepresented in current surveys.