3.3.1 Study population
The study was conducted within the Manawatu Health District of New Zealand’s North Island. The population consists of 155,072 people who live in a mixture of rural and urban dwellings including both costal and inland areas and is centered around the city of Palmerston North, which has a population of 75,000 people.
3.3.2 Case definition
A case was defined as a laboratory-confirmed sporadicC. jejuni case in the Manawatu Health District between March 1st 2005 and February 29th 2008. Outbreak-related cases, except the index case, were excluded. Data from the national public health surveillance system were obtained for cross-validation with notified case numbers in the study region.
3.3.3 Case ascertainment
Cases were prospectively recruited via the regional diagnostic laboratory, which tests all human pathological samples from the region. Samples were tested with a commer- cial enzyme immunoassay (ProSpecT, Remel, USA), and positive fecal swabs (Amies Carchoal transport swabs, Copan, Italy), were sent to the Hopkirk Research Institute, for isolation of presumptiveCampylobacter. In brief, swabs were cultured on modified Cefoperazone Charcoal Desoxycholate agar plates (Fort Richard, Auckland, NZ) and in Bolton Broth (Lab M, Bury, England) and incubated in 42◦C in a microaerophilic atmosphere for two days. A single colony resembling Campylobacter species was sub- cultured to Columbia Horse Blood Agar (Fort Richard, Auckland, NZ) and incubated microaerophilically at 42◦C for two days before DNA preparations were made. Isolates were confirmed as C. jejuni by polymerase chain reaction. If multiple samples were collected from a single patient, only one isolate was selected.
3.3.4 Genotyping
After speciation, MLST of C. jejuni isolates was performed to assign isolates to a sequence type (ST) using seven house-keeping genes: aspA (aspartase A), glnA (glu- tamine synthase),gltA(citrate synthase),glyA(serine hydroxymethyltransferase),pgm
(phosphoglucomutase), tkt (transketolase) and uncA (ATP synthase alpha subunit) based on the method outlined by Dingle et al., 2001 (Dingle, Colles et al. 2001).
3.3.5 Epidemiological surveillance
Three years of human data were acquired systematically - initially collected using rou- tine Public Health Service methods, superseded by a sentinel surveillance approach with data collection targets. Case information between February 2005 and June 2006 was acquired using both mailed questionnaires and telephone interviews. To enhance the quality of the surveillance data gathered, from July 1st 2006 contact of notified cases of campylobacteriosis was via telephone with a target of 95 % of cases interviewed and 95 % of data sets completed. The interviews were conducted using the standard structured EpiSurv case report form used for enteric diseases in New Zealand (Lake, Whyte et al. 2005). The questionnaire gathered demographic variables, risk factor information for
variables within the incubation period, as well as case characteristics, such as whether the case was hospitalised. Exposure history covered aspects such as occupation, recent travel, and indicators of person-to-person transmission. Public health officers were trained to ensure consistency and standardization of interviews.
3.3.6 Data handling and statistical analysis
Epidemiological data collected was merged with the genotyping information in an Ac- cess database, which was linked into a Bionumerics database for analysis. Data was quality controlled and validated by techniques such as consistency check, range check and batch totals. R, version 2.7.0, was used for statistical analysis. Distributions were compared using two-sidedχ2 tests and Fisher’s exact test for count data.
All STs isolated from human cases were assigned a probability that they were ac- quired from individual animal reservoirs using the asymmetric island model (Wilson, Gabriel et al. 2008), which is in detail discussed in Chapter 7. Data for this assign- ment was collected from parallel structured studies at the sentinel surveillance site over the same time-period (French and Molecular Epidemiology and Veterinary Public Health Group Hopkirk Institute; Mullner, Jones et al. 2009). STs with a probability of at least 70 % to have originated from a single source were assigned to that source, thereby creating four different groups of human cases: Poultry-associated ST - 474, other poultry-associated STs, ruminant associated STs and STs not associated with a particular source, of which the latter two were pooled in the analysis due to sample size considerations. The spatial location of each notified human case was given at the meshblock level: the smallest regions defined for the New Zealand census, each contain- ing between zero and about 200 individuals. The spatial resolution is therefore much higher in urban areas. The relative risk of being a case of campylobacteriosis in each meshblock was described using a Bayesian hierarchical model. Relative risk surfaces were prepared for human cases attributable to the common poultry-associated STs, and for non-poultry associated STs.
Let Yi,t represent the number of notified cases of campylobacteriosis in meshblock iand week t. We assume that Yi,t ∼P oisson(niλi,t), where ni is the usually resident
population of meshblocki(obtained from the most recent census) and λi, t represents the expected risk at this point in time and space. Next, we separate the risk into its spatial and temporal components through the relationlog(λi,t) =Rt+Ui, whereRtand Ui are the purely temporal and purely spatial components of the risk respectively. For
the spatial component a Gaussian Markov Random Field prior is assumed (also called a Gaussian intrinsic autoregression) in which the risk in each meshblock is assumed to be similar to the mean risk of the neighbouring meshblocks. More formally, we assume the following full conditional distributions
Ui ∼N( X j∈n(i) Uj |n(i)|, 1 κu |n(i)| ), (3.1)
where n(i) is the set of meshblocks that are neighbours to meshblock i. For the temporal component we assume a Gaussian second order random walk prior: that the change in risk from week t to weekt+ 1 will be similar to the change in risk from week
t−1 to week t, ie. givenR1, . . . ,Rt,
Rt+1−Rt∼(Rt−Rt−1, 1
κR
). (3.2)
We assume flat priors for R1 and R2 so that the temporal component can ab- sorb the baseline level of risk. For the hyperparameters κR and κU we assume the
mildly informative and conjugate gamma-distributed priors: κRGamma(5,1/100) and
κU Gamma(1,1/2). These priors were chosen to aid convergence of the Markov chain,
but to be still sufficiently flexible to allow the model to arrive at the appropriate level of variation for the data.
Multiple covariates were analysed using the defined source associations of strains and odds ratios (OR) were calculated. To reduce the number of covariates and the collinearity among different variable for farm animal exposure items measured on the questionnaire were summarized on biological grounds. The variable ‘farm ani- mal contact’ therefore comprised ‘occupational exposure to farm animals’ and ‘farm animal contact’. Age was modeled in two groups: five years and under and over five years of age. This was because differences in risk for pre-school children (children first attend school on their fifth birthday in New Zealand), compared to other age- groups, had been observed in other studies and was evident from surveillance data (http://www.nzpho.org.nz/). We constructed multivariable logistic regression models to investigate the relationship between infection with poultry associated strains and a set of exposure variables. Variables for inclusion in the final models were chosen by exploratory univariable analysis.