GUI
We created a novel graphical user interface (GUI) to dynamically illustrate the spread of an epidemic disease of cattle and used the GUI to investigate farmer behaviour during a simulated disease epidemic. The GUI showed a series of maps, with infected herds plotted, week by week. The outbreak started in southern-central France (epidemic stage 1), was introduced to the UK (epidemic stage 2), then spread throughout the UK (epidemic stages 3–8). The GUI presented a common outbreak experience at each stage of the epidemic, in terms of distance to the nearest infected farm, for each participating farmer regardless of the location of their farm. Table 1 describes the scenarios presented to the farmers for each week of the scenario. An example screenshot from the GUI can be seen in Supplementary Fig. 1 and the GUI can be accessed online (https://feed.warwick.ac.uk/map.html).
Elicitation exercise (farmer interviews)
We conducted interactive online interviews with GB cattle farmers, which consisted of three parts: (i) demographic questions, (ii) hypothetical disease scenario using the GUI and (iii) online questionnaire of validated psychosocial and behaviour change measures. The interview script (i and ii) and the online survey (iii) are in Supplementary Text 2 and Supplementary Text 3, respectively. For the disease scenario, we gave farmers a detailed description of the hypothetical disease and a vaccine that was available to them. The vaccine cost £50 per animal, had to be given to all cattle on the same day and was 100% protective after five days.
The farmers then proceeded with the scenario using the GUI and were asked at weekly intervals whether they would vaccinate or do any other preventative measures. Farmers answered the questions in the online questionnaire directly after the disease scenario, which used validated measures to investigate trust in others, psychological proximity to others and the COM-B behaviour change framework. Trust in others was measured using Likert-scale statements24,58, psychological proximity to others was measured using the Inclusion of Other in the Self (IOS) scale28,59 and the COM-B factors were investigated using Likert-scale statements based on the framework guide60. The survey web page generated an individual code for each respondent, which was used to anonymously link the responses with the interview responses once all interviews were completed.
We pilot tested the GUI, interview and online questionnaire on three dairy farmers from three counties and two countries in GB. For the research interviews, we then recruited GB cattle farmers from two sources: (i) farmers who took part in previous research24 and confirmed interest in participating in further research, and (ii) via advertisement through multiple cattle interest organisations. The interviews lasted up to one hour, were online (Microsoft Teams) and farmers received a £40 voucher for their time.
Analysis of elicitation data
We analysed the interview results using R statistical software v4.2.161. To investigate associations between a farmer being an Early (epidemic stages 1–2), Mid (epidemic stages 3–5) or Late (epidemic stages 6–8 or never) vaccinator (Table 1) and their psychosocial and COM-B factor scores, we used a multinomial logistic regression model. We fit the regression model to 500 bootstrapped datasets in the following way. Starting at the null model, each covariate was tested in the model and the covariate with the lowest p-value retained. At each iteration, any covariates in the model with a p-value greater than 0.05 were removed before all remaining covariates were retested in the model and the process was stopped when no remaining covariate had a p-value less than 0.05 when added to the model. We calculated the stability of each covariate (the proportion of multinomial logistic regression models fit to the 500 bootstrapped datasets that selected the covariate) to mitigate overfitting in this wide dataset52. We calculated stability thresholds for different probabilities of a variable being a true positive using the methods by Green et al.62. Selected covariates had a stability higher than a threshold where covariates had a 15% probability of not being associated with when farmers vaccinated. We calculated odds ratios of the stable covariates by refitting the model to the bootstrapped datasets with the candidate covariates limited to those above the stability threshold and taking the mean of the coefficients. Finally, model fit was assessed by fitting the multinomial logistic regression model to the full dataset and comparing the observed and expected data using a Hosmer-Lemeshow goodness-of-fit test and decile plots63, and by comparing predicted vaccination class from the full and 10 x 10-fold cross validated models.
Using k-means clustering40, we grouped farmers based on their scores for the covariates selected by stability thresholds for a 10% and 15% probability of not being associated with when farmers vaccinated respectively. We used visual inspection of a plot of total within-cluster sum of squares against number of clusters to determine the number of clusters that gave the best fit (Supplementary Fig. 2). These psychosocial groupings were incorporated into the mathematical livestock disease transmission model.
Mathematical transmission model of infectious livestock disease
We simulated an epidemic process in Great Britain amongst holdings with cattle to ascertain the impact of differing population compositions with regards to behavioural stances on intervention usage on an emergent outbreak of a fast-spreading pathogen. Within this subsection we overview: (i) the data sources used to inform cattle demography, (ii) the epidemiological model framework that was conceptually based on a swift, locally spreading pathogen with no long-range movement of animals, (iii) our implementation of vaccination, (iv) expanded details on the eight behavioural configurations and (v) the simulation protocol used to assess the scenarios of interest.
Livestock data description
We used the Cattle Tracing System database to procure average holding cattle herd sizes throughout 2020. The Cattle Tracing System contains virtually complete records of the births, deaths, and movements of individual cattle in Great Britain since 200164.
These data contained 59,774 holdings. Cattle herd sizes ranged from 1 to 7634, with a median of 58, interquartile range of 16–155, and 97.5th percentile of 622. There was regional variation in the number of holdings and cattle. Most populous regions included the south-west of England (particularly Devon), south-west Wales, north-west England (most notably Cumbria) and Dumfries and Galloway in south-west Scotland (Fig. 4).
Epidemiological model
We formulated the infectious disease dynamics as a stochastic, spatially explicit, holding level Susceptible-Exposed-Infectious-Removed (SEIR) model. With the epidemiological unit being a holding (rather than at the individual animal level), we inherently assumed that once infection had entered the cattle herd on the holding the within-premises disease spread occurred rapidly, leaving the whole cattle herd infected.
Time was discretised into daily time steps. The daily probability of a susceptible holding j becoming infected by an infectious holding i obeyed:
$${p}_{ij}= 1 - exp\left({\lambda }_{ij}\right)$$
where,\({\lambda }_{ij}={\xi N}_{i}^{\psi }{N}_{j}^{\varphi }K\left({d}_{ij}\right)\)
In detail, the force of infection between two herds, \({\lambda }_{ij}\), was a nonlinear function of the transmissibility of cattle \(\left(\xi \right)\), the number of cattle on the infectious holding (term \({N}_{i}^{\psi }\)) and the number of cattle on the susceptible holding (term \({N}_{j}^{\varphi }\)). For the herd size exponents, \(\psi\) and \(\varphi\), we used cattle epidemiological parameter estimates inferred from the 2001 UK FMD epidemic for Cumbria65.
The fourth contributor to the force of infection was the transmission kernel K, which was a function of the Euclidean distance between holdings i and j (\({d}_{ij}\), measured in metres). We applied a power-law transmission kernel, with a maximum range of 50km:
\(K\left({d}_{ij}\right)=\frac{{k}_{1}}{1 + (\frac{d}{{k}_{2}}{)}^{{k}_{3}}},\) for \({0 \le d}_{ij}\le 50000;\)Otherwise, \(K\left({d}_{ij}\right)=\) 0.
We recognise that our parameterisation of the force of infection is an amalgamation of values that were inferred for an FMD-like pathogen in different contexts. Yet, for demonstrating the utility of the model framework, in particular how one may use behavioural-associated data gathered from an elicitation study, our chosen parameterisation suited the investigative purposes of our study. Notably, in the absence of additional controls beyond infected holdings, the simulated outbreaks were extensive irrespective of the seed infection location.
Upon a holding becoming infected, we assumed a latent period of five days (based on epidemiological and veterinary records from the 2001 UK FMD outbreak66). Thereafter, the entire livestock population at that holding was considered infectious for a period of eight days (days 6–13 after infection). We assumed all infected holdings provided notification of infection nine days after the initial infection event, meaning there was no under-reporting of infection, but there was a four-day delay between the holding becoming infectious and subsequent notification of infection. At the end of the infectious period (13 days after infection) the cattle herd and holding were considered removed from the population.
See Table 3 for an overview of the epidemiological model values.
Table 3
Summary of the livestock disease model epidemiological and intervention parameter notation, descriptions and values.
Notation | Description | Value |
---|
Epidemiological parameters |
---|
\({\lambda }_{ij}\) | Infectious pressure on susceptible holding j from infectious holding i. | Variable |
\({N}_{i}\) | Number of cattle on premises i | Variable |
\(\xi\) | Transmissibility of cattle. | 10^6 |
\(\psi\) | Exponent on the cattle population on an infectious holding, for calculating the infectious pressure. | 0.42 |
\(\varphi\) | Exponent on the cattle population on a susceptible holding, for calculating the infectious pressure. | 0.41 |
\({k}_{1}\) | Transmission kernel normalisation constant (to 2 s.f.) | 1.2x10^-8 |
\({k}_{2}\) | Transmission kernel distance length scaling | 2000 |
\({k}_{3}\) | Transmission kernel exponential parameter on the distance component | 2 |
tincub | Time elapsed until end of incubation period (relative to the day of infection) | 5 days |
tnotif | Time elapsed until notification (relative to the day of infection) | 9 days |
tremoval | Time elapsed until removal (relative to the day of infection) | 13 days |
Intervention parameters |
veff | Vaccine efficacy | 100% |
vdelay | Delay in vaccine effectiveness | 5 days |
Implementation of vaccination
To correspond with the intervention descriptions in the interview study, we modelled the administration of vaccine to livestock (relevant parameter values are contained in Supplementary Table 3).
We present the idealised situation of having a vaccine available with 100% effectiveness in blocking infection. As per the description of the vaccine product in the interview script, we included a lag for the vaccine inducing an immune response (fixed at 5 days, based on measures of FMD virus titres in milk from inoculated cows in the days post-inoculation67). As the vaccine could be administered to an infected population during its latent phase (thus prior to onset of symptoms and subsequent notification of infection), in these circumstances it was feasible for a cattle herd at a holding to be vaccinated but to still become infected.
Intervention behaviour configurations
We tested eight different behavioural group population compositions, referred to as configurations, each governing the proportion of the population that would implement interventions at a given time (with respect to the outbreak situation). Note that in all configurations controls are applied at holdings with confirmed infection. As the basis for naming the different groupings, we used the conceptual framework to categorise cooperation suggested by Bshary and Bergmüller18.
Uncooperative
Controls only applied at holdings with confirmed infection (cattle removed). No holdings applied vaccination, irrespective of the epidemiological situation.
Strong parasitism
All holdings administered vaccines in their herds upon infection being confirmed within 50km (approx. 30 miles) of their holding. This configuration resembles a situation where all individuals wait to see what is happening, how the infection spreads and as such they are exploiting others or benefit from what happens to others.
Weak parasitism
All holdings administered vaccines in their herd upon infection being confirmed within 320km (approx. 200 miles) of their holding. Similar to the strong parasitism scenario, where all individuals observe the epidemiological situation and as such they are exploiting others or benefit off what happens to others, although in this instance all individuals are more precautionary.
Mutual cooperation
All holdings vaccinated their herds prior to pathogen emergence (no outbreak occurs). Represents a scenario where all individuals cooperate to produce the maximum epidemiological benefit to all.
Cooperation-Parasitism-Free riders (Coop-Parasitism-FR)
Uniform partitioning of the population across four intervention stance groups. A quarter of holdings never vaccinated, irrespective of the epidemiological situation (free-riders); a quarter of holdings vaccinated their herd upon infection being confirmed within 50km of their holding (strong parasitism); a quarter of holdings vaccinated their herd upon infection being confirmed within 320km of their holding (weak parasitism); a quarter of holdings vaccinated their herds prior to pathogen emergence (cooperators).
Cooperation-Parasitism (Coop-Parasitism)
Uniform partitioning of the population across the three intervention timing groups. A third of holdings vaccinated their herd upon infection being confirmed within 50km (strong parasitism) of their holding; a third of holdings vaccinated their herd upon infection being confirmed within 320km (weak parasitism) of their holding; a third of holdings vaccinated their herds prior to pathogen emergence (cooperators).
Trust-Expectancy
Partitioning of holdings into four behavioural groups, using the empirical estimates for psychosocial profile clusters from the model comprising the two most stable variables (Fig. 1). The four groups covered combinations of two trust groups (high, low) and two “expectancy” groups, their ability to physically intervene (high, low).
Herd size dependent
Partitioning of holdings into three behavioural groups, using the empirical estimates for psychosocial profile clusters from the model comprising the five most stable variables (Fig. 2). Specific to the Herd size dependent configuration, herd size determined the probability of the holding being assigned to each of the three behavioural groups (Fig. 5).
For the Trust-Expectancy and Herd size dependent configurations, we used the interview results to parameterise: (i) the split of holdings between behavioural groups; (ii) in each behavioural group the partitioning of the holdings between the different intervention timings (see Figs. 1–2 and Supplementary Tables 1–2).
Simulation outline
We considered the fast-spreading pathogen first emerging in a spatially localised area of Great Britain from a low case level. Therefore, in each simulation replicate we seeded infection in a randomly selected cluster of three premises (we selected one premises at random and found the two premises that were closest in terms of Euclidean distance). A replicate terminated upon there being no premises in an infected state.
For each behaviour configuration, to explore the sensitivity of epidemiological and economic outcomes to the geographical location of initial infected premises we assessed 89 different seed region scenarios. We ran 500 replicates for each scenario, comprising a behaviour configuration and seed infection region.
To assess the implications of differing psychosocial and geographical attributes on epidemiological outcomes, we tracked the percentage of holdings infected and outbreak duration. To evaluate the economic implications of behavioural attributes on intervention usage, relative to the baseline control strategy, we computed threshold intervention unit costs (the maximum cost per intervention unit where the total intervention cost equalled the costs saved from averted infections).
Given our use of a large-scale spatially explicit model, for our simulation procedure we employed an optimised gridding approach (the conditional subsample algorithm) as described in Sellman et al.68. We performed all model simulations and produced plots in Julia v1.8. The code repository for the study is available at https://github.com/EdMHill/FEED_farmer_disease_management_heterogeneity.