Methods
Part One of the study re-analysed the dataset of 10,000 simulated outbreaks of FMD in New Zealand generated during the study by Garner and colleagues (2016), to calculate the third quartile values for the numbers of IPs and the observed EDR values at days 14, 21, 28 and 35 of the response (post first-detection). The purpose was to use these values to define a time-varying series of triggers that operated within specified time periods within the ISP modelling platform. These time periods were defined as response days (post first detection) 11-14, 15-21, 22-28 and 29-35 inclusive.
In Part Two of the study, the threshold values for the numbers of IPs and EDR for each time period were specified as a complex EDI trigger within the ISP platform, and the New Zealand Standard Model of FMD was initialized to simulate a further set of FMD incursions into New Zealand. The underlying farm denominator dataset was based on a September 2015 extract of AgriBase (Sanson 2005), a national farms database in New Zealand, owned and operated by AsureQuality Limited, a state-owned enterprise. The model was set up to randomly introduce FMD into farms in the upper North Island, within an area termed the “Auckland Mega-region”, which was created by combining the Northland, Auckland, Waikato and Bay of Plenty regions.
Before each introduction, several other variables were randomly varied (see Table 1). These included whether the FMD virus was of a type that could be transmitted by the wind, the number of personnel of various roles that were available for response duties, the number of direct and/or indirect contacts that a tracer could process per shift, and the number of farms that a surveillance veterinarian could visit per day. Once detected by passive surveillance, each outbreak was controlled by standard stamping-out (SO) measures, including tracing of movements, quarantine and depopulation of IPs, movement controls and active surveillance by patrol veterinarians. Each simulated outbreak continued until eradication or to a maximum of 365 days if not eradicated.
Data generated during each simulated outbreak was stored in a Sqlite3 database. The main outputs of the model for each iteration were whether the EDI trigger fired and if so when, the number of farms infected each day, the number of IPs detected each day, the number of farms depopulated per day, and the number of personnel used in response duties per day by role type. From these, further outputs were derived, including the farm type of the primary case, the day of first detection, the total number of IPs detected, and the duration of each outbreak (day of last detection – day of first detection + 1). In addition, there were some variables that were able to be measured such as the farm and livestock densities around the primary and index cases (see Table 2).
For the purposes of evaluating the performance of the EDI trigger prospectively, ‘large’ outbreaks were defined as the final number of IPs being in the upper quartile (i.e. > 75thpercentile) of all outbreaks in the Part Two simulations, and ‘long’ outbreaks were classified as having duration within the upper quartile (> 75th percentile) of epidemic lengths for the Part Two simulations. Performance was evaluated by calculating the sensitivity (the proportion of large / long outbreaks during which the trigger fired [Se]), specificity (the proportion of small / short outbreaks during which the trigger did not fire [Sp]), positive predictive value (the proportion of trigger firings which resulted in large / long outbreaks [PPV]) and negative predictive value (the proportion of outbreaks for which the EDI trigger did not fire which ended up as small / short outbreaks [NPV]) against both IPs and duration using 2x2 contingency tables. Sensitivity analysis of these performance measures was conducted by re-classifying the outbreaks into large or long using the 70th and 80th percentiles.
Statistical analysis included logistic regression of the factors that were associated with the trigger firing, with the independent variables being cattle, sheep, pig and farm densities within a 5x5 km square centred on the primary case, whether airborne spread could occur or not, the numbers of personnel available by role and the time of first detection. Fitting the model was by backwards, stepwise elimination of non-significant variables (p > 0.05) based on the Wald test. Logistic regression modelling was conducted on the largeand long variables to see if the trigger firing was associated with large or long outbreaks. All analyses were conducted using R v3.5.3.