Analysis of Meat Juice ELISA Results and Questionnaire Data to Investigate Farm-Level Risk Factors for Salmonella Infection in UK Pigs

Authors


Richard Piers Smith. Centre of Epidemiology and Risk Analysis, VLA Weybridge, Addlestone, UK. Tel.: +44 1932 359465;
Fax: +44 1932 359429;
E-mail: r.smith@vla.defra.gsi.gov.uk

Summary

The study set out to explore risk factors for Salmonella infection in pigs, based on seroprevalence amongst slaughtered pigs, using a large study population of holdings and a comprehensive list of farm characteristics. Farm data were collected from pig quality assurance schemes and supplemented by a postal questionnaire. These data were used with meat juice serology results from ongoing abattoir Salmonella surveillance, for a multivariable risk factor analysis, modelling the ELISA sample to positive ratio directly (ELISA ratio). The study population contained 566 farms, covering a geographically representative spread of farms within the United Kingdom, with a mean average of 224 sample results per holding over a 4-year period. The model highlighted that temporal factors (quarterly and yearly cycles) and monthly meteorological summaries for rainfall, sunshine and temperature were associated with Salmonella presence (< 0.01). The ELISA ratio was found to be highest in autumn and lowest in spring and summer, whereas yearly averages showed a greater degree of variation than seasonal. Two feed variables (homemix and barley) were found to be protective factors, as was a conventional, rather than organic or freedom foods, farm enterprise type. The number of annual pig deliveries and dead stock collections, and the main cause of pig mortality on the farm were found to be associated with Salmonella infection. Scottish farms had a lower ELISA ratio than other regions, and an increased number of pig farms within a 10-km radius was associated with a higher ELISA ratio. The study demonstrated that the analysis of routinely collected data from surveillance and quality assurance schemes was cost-effective, with sufficient power to detect modest associations between Salmonella and exposure variables. The model results can be used to inform on-farm Salmonella control policies and could target-specific geographical regions and seasons to assist the efficiency of surveillance.

Impacts

  • • This manuscript explains how routinely collected data from abattoir surveillance, quality assurance schemes, and from a postal questionnaire, could be used for a cost-effective study to detect risk factors for Salmonella infection in pigs.
  • • The study included a large number of serum samples from pigs located on 566 farms, which provided sufficient power to detect even weak associations.
  • • A number of farm characteristics and management practices, including seasonal cycles; feed types used; frequency of pig deliveries and the density of pig farms within a 10-km radius, were associated (< 0.01) with Salmonella ELISA sample to positive ratio.

Introduction

Salmonella enterica is a zoonosis and different serovars can be carried by livestock raised for food production. Human salmonellosis is characterized by diarrhoea and can be transmitted through foodborne routes (O’Brien, 2005). The importance of pigs as vectors of Salmonella has been shown by a large abattoir study where the prevalence of Salmonella in pig caecal samples, collected in Great Britain, was high (23.4%), when compared with both cattle and sheep (1.4% and 1.1% respectively) (Milnes et al., 2007). In a European Union baseline survey in 2007, a similar level (21.2%) of Salmonella was isolated from mediastinal lymph node samples from United Kingdom (UK) pigs at slaughter (Anon, 2008). Infection in pigs can cause a range of clinical signs, from scouring to fever and death, but is often sub-clinical and so, is difficult for farmers to monitor and detect. Although it is unknown how many cases of human salmonellosis are attributed to eating pig products, of the 13 213 laboratory confirmed cases in the UK identified in 2007, 13% were related to the serovar S. Typhimurium, which is the predominant type detected in samples from UK pigs (Defra, 2007; VLA, 2007).

Many studies have tried to ascertain the factors that influence Salmonella prevalence, and identify on-farm control measures to reduce the Salmonella burden in pigs. Recent studies in the UK have highlighted associations with factors such as herd size; outdoor rearing of pigs; flooring type; and farm location (VLA, 2004; Pritchard et al., 2005; Smith et al., in press). These findings have been supported by European, Canadian and American studies (Funk et al., 2001; Nollet et al., 2004; Farzan et al., 2006). Seasonal peaks and troughs of Salmonella prevalence have been identified by studies, with a two peaked annual cycle apparent, which may be related to meteorological conditions such as environmental temperature (Funk et al., 2001; Hald and Andersen, 2001). However, a number of the studies above were limited to a small and potentially unrepresentative subset of the pig farm population, which may not have had sufficient statistical power to detect modest associations between Salmonella infection and putative risk factors. Other studies analysed only a small number of variables and so may have missed more important risk factors or not estimated the true affect of a variable by accounting for the affect of other variables.

Schemes are present in the UK that routinely collect data on Salmonella in pigs and farm management characteristics, from a large number of farms. In June 2002, the UK Zoonoses Action Plan (ZAP, now called the Zoonoses National Control Plan) monitoring programme was designed to run in conjunction with Quality Assurance schemes (QAS), to estimate the burden of Salmonella from a sample of slaughtered pigs (Armstrong, 2003). The scheme was based on a design by the Danish pig industry that had contributed to a reduced Salmonella prevalence (Nielsen et al., 2001). The QAS routinely collect details on the structure and management of pig farms to ensure a level of health and welfare standards are met.

This study reports how data from these sources were used, along with a postal questionnaire, to implement a cross-sectional study to analyse the effect of a large number of explanatory factors (biosecurity, farm demographics, meteorology) on Salmonella seroprevalence for QAS-registered holdings, in Great Britain and Northern Ireland, that submitted finisher pigs to the ZAP scheme. By using a comprehensive list of variables, from a large population of pig holdings, a more detailed picture of the risk factors for Salmonella seroprevalence results would be generated, that would detect factors with even modest (variables with model coefficients close to zero) associations.

Materials and Methods

Data on explanatory factors were collected from a number of sources and combined into a single dataset, for analysis in the model. Datasets were collected from three QA schemes [Approved British Pigs (ABP); Genesis Quality Assured (GQA); and Quality Meat Scotland (QMS)] and from the ZAP scheme. The data were coded and linked to map reference co-ordinates according to the previous method (Smith et al., in press). These coordinates were also used to identify the Nomenclature of Units for Territorial Statistics (NUTS) geographical region for that holding. NUTS have four subdivisions and NUTS 1, equivalent to government office regions, were used rather than other sources of clustering, such as county, as they are more stable over time and less subject to boundary changes. It was also believed that the categories represent more biologically sensible categories in terms of the country’s animal species population. Meteorological data of monthly regional summaries, including actual and anomaly (difference from long-term averages) records, were gathered (http://www.metoffice.gov.ukclimateukindex.html) and linked to the dataset by the region of farm and the month of sample collection. A supplementary questionnaire was designed to collect information on a number of covariates previously identified as significantly associated with Salmonella presence and rated as key to Salmonella presence in the UK by a number of experts. These included pig stocking levels (Farzan et al., 2006); feeding practices (Lo Fo Wong et al., 2004); housing systems (Nollet et al., 2004); biosecurity (Beloeil et al., 2004) and geographical location (Benschop et al., 2008), to supplement those routinely collected by the QAS (Table 1). Full details are available on request. The questionnaire was posted, along with a covering letter, to all 2064 farms listed under the three QAS, asking for the farmer’s voluntarily completion of the questionnaire, which was to be returned in a supplied envelope.

Table 1.   Variables generated from data collected by quality assurance schemes and a postal questionnaire
Variable category
NUTS 1 region
Coordinates (X, Y)
Pig farm density at 3 and 10 km radii
Season of sampling
Quality assurance scheme
Enterprise type
Reared on contract
Production system (batch/continuous)
Any pig production outdoor
Flooring
No. each pig type
Other farm animal species present
Mixing of pigs
Isolation of sick pigs (freq, where)
Types of feed fed to weaners, growers, finishers and sows
Drinking system and water source
Cleaning and disinfection of pig houses and drinking system
No. pig deliveries/collections
No. and type of other farm visitors
Delivery procedures
Boot dip usage
Health conditions present
Top three causes of pig mortality
Top three causes of pig treatment
Regional summaries of meteorological factors
Temporal cycles

The ZAP data were limited to results collected up to 4 years prior to the completion date of the postal questionnaire, to allow a comparison of temporal trends over a number of years. Variables for temporal trends and seasonal effects were designed by adding sinusoidal components (sine and cosine terms) for 3, 6 and 12-month periods to create quarterly, half-yearly and yearly cycles (Chatfield, 2003). These cycles may account for seasonal trends or any reduction of ELISA ratio through the years of the study population caused by the control of Salmonella through the ZAP scheme.

For the ZAP scheme, small pieces of skeletal muscle (from diaphragm/neck) were removed from pigs at the abattoir and placed in meat juice (MJ) tubes which were frozen and then thawed to collect the fluid (Nielsen et al., 1998; Armstrong, 2003). The MJ sample was tested at a single UK laboratory by a mix-ELISA serological test (Guildhay VETSIGN™Kit, Guildford, UK) for a ‘host’ response of antibodies to Group B and C1Salmonella (Nielsen Nielsen et al., 1998), 1998). Salmonella infection in pigs produces an immune response, which includes the production of antibodies. These are detected by the ELISA from which a sample to positive ratio (ELISA ratio) is calculated, which is related to the titre of circulating antibodies (Sorensen et al., 2004; Hill et al., 2008). Three samples were randomly collected from every batch of pigs sent to slaughter on any particular date in accordance with the sampling regime agreed on May 2003 (BPEx, personal communication). For routine surveillance, a cut-off point is applied to the ELISA ratio to provide a binary outcome but for this study the ELISA ratio was used directly to allow an analysis of a linear relationship.

Data analysis

A Boxcox plot was used to verify whether the ELISA ratio results required transformation and what type of transformation was necessary to approximate normality. All negative and zero ELISA ratios were coded to 0.005, which was half of the lowest recorded result, prior to transformation. Relationships between this transformed outcome and explanatory factors were analysed by univariable mixed linear regression (stata 10; Stata corp. LP, College Station, TX, USA), with the farm holding identifier selected as a random effect to allow for dependence between observations within the same premises. All continuous variables were plotted on a graph and assessed for normality and whether transformation was necessary. Explanatory factors with more than two levels were also tested to see whether they should be split into multiple dichotomous variables. For example, a variable with levels for each NUTS region was tested at the univariable level, as well as binary variables for each individual region, to see which factor was more significant/fitted the model better. Explanatory variables for which the association on univariable analysis yielded a P-value of 0.25 or more were omitted from the multivariable model.

Because of the large number of factors under examination, variables were entered into the models manually using a forward stepwise method. The variable with the lowest P-value was entered first into the model, and each subsequent variable was then independently introduced into this model before selecting the next variable with the lowest P-value and repeating the process. Becauseof the large dataset size, a P-value of 0.01 was set as the significance threshold and this stepwise method continued until no further variables could be identified whose addition generated a P-value of less than 0.01. Records with missing data for the selected variables were dropped from the model. All rejected variables were added separately into the final model to ensure no significant variables had been omitted.

Likelihood ratio tests were used to compare models of the same population size to determine whether the included variable significantly improved the model. The Wald’s chi-squared test and Akaike Information Criterion were also examined to ensure model fit. The standardized residuals were plotted against the fitted values to examine signs of heteroscedasticity and a histogram of model residuals was plotted to evaluate normality, to ensure the standard model assumptions were met. Each variable that entered the final model was compared against the model residuals and a Bartlett test completed to assess homogeneity of variances (R version 2.7.1; R Development Core Team, Vienna, Austria). Explanatory variables, that were perfectly collinear with variables already included in the model, were dropped automatically by the stata package. All variables in the final models were assessed for biologically plausible interactions; however, because of a small number of positive samples in some strata, this was not possible in all cases.

The farm holding records with map references were plotted as points onto a map of the UK using ArcGIS 9.1 (ESRI, Redlands, CA, USA).

Results

Between 6 June 2007 and 30 October 2008, a total of 566 questionnaires were returned and successfully linked to the ZAP database. These questionnaires consisted of 305 ABP, 171 GQA and 90 QMS registered holdings. The 554 holdings that provided the necessary information to generate map coordinates were presented on a map (Fig. 1). This shows the distribution of participating farms around Great Britain and Northern Ireland, and the large difference in farm density between regions such as Yorkshire and the Humber and North West England (Smith et al., in press). A chi-square comparison between the holdings present in the study population and the total QAS population indicated fewer farms in East England and more in Scotland and the South West (< 0.05). The holdings linked to a total of 119 906 ZAP samples, with a mean average of 224 samples per holding (range 1–1671). Plots of the ELISA ratio results (Figs 2 and 3) indicate that a seasonal average ranged from 0.25 (autumn) to 0.22 (spring and summer) and a comparison of mean values showed that this was significant (F = 29.09, < 0.001), and also the mean ELISA ratio differed greatly (F = 12.75, < 0.001) between each year of sampling. The majority of ELISA ratio results were close to zero (60.9% were below 0.10) and a Boxcox plot verified that a logarithmic transformation was required to approximate normality.

Figure 1.

 Distribution of participating pig holding locations by quality assurance scheme (n = 554).

Figure 2.

 Mean meat juice ELISA ratio results, with 95% confidence intervals, by season of sampling, for 566 pig holdings. Dotted line indicates mean ELISA ratio.

Figure 3.

 Mean meat juice ELISA ratio results, with 95% confidence intervals, by year of sampling, for 566 pig holdings. Dotted line indicates mean ELISA ratio.

The results of the linear regression model are presented in Tables 2 and 3, with Table 2 presenting the strongly significant variables detected from the univariable screening of the variables and Table 3 presenting the final variables that entered the multivariable model. Thirteen variables entered the final model and the model population was reduced to 474 holdings, as a result of missing data. The model had a significant Wald’s chi-squared result (< 0.001) and a likelihood ratio test for the inclusion of the random effect was also significant (< 0.001). The ‘season’ variable was dropped from the model as it was collinear with the temporal cycles, and ‘scheme’ was dropped as it was perfectly collinear with region.

Table 2.   Variables strongly associated (< 0.05) with Salmonella from univariable mixed linear regression of logged meat juice ELISA ratio results collected from slaughtered pigs
VariableLevelCoefficientP-valueNo. farms
  1. *’Anomaly’ is the difference from long-term averages.

QA schemeABPBaseline 305
GQA0.458<0.001171
QMS−0.658<0.00190
NUTS RegionOtherBaseline 469
Scotland−0.824<0.00193
Pig farm density within 3 km radiusContinuous0.085<0.001554
Pig farm density within 10 km radiusContinuous0.021<0.001554
Season that sample was collected fromSpringBaseline n/a
Summer−0.169<0.001n/a
Autumn−0.133<0.001n/a
Winter−0.099<0.001n/a
Farm enterprise – conventionalNoBaseline 51
Yes−0.741<0.001515
Farm enterprise – freedom foodsNoBaseline 480
Yes0.583<0.00186
Pigs reared on contract at farmNoBaseline 292
Yes0.385<0.001254
Cattle present on farmNoBaseline 373
Yes−0.2820.001193
Number of cattle currently presentContinuous−0.0010.019537
Sheep present on farmNoBaseline 419
Yes−0.2410.011147
Cats present on farmNoBaseline 554
Yes0.6360.02512
Pigs mixed at weaner groupNoBaseline 142
Yes−0.368<0.001390
Pigs mixed at other timeNoBaseline 376
Yes0.2090.046114
Pigs never mixedNoBaseline 494
Yes0.477<0.00172
Weaners fed fermented feedNoBaseline 366
Yes−0.6920.0448
Weaners fed homemixNoBaseline 285
Yes−0.623<0.00189
Weaners fed concentratesNoBaseline 128
Yes−0.1820.031246
Weaners fed barleyNoBaseline 190
Yes−0.2720.002184
Percentage of barley in weaner feedPercentage−0.014<0.001533
Growers fed homemixNoBaseline 280
Yes−0.572<0.001126
Growers fed wheatNoBaseline 180
Yes−0.302<0.001226
Percentage of wheat in grower feedPercentage−0.0050.002540
Growers fed barleyNoBaseline 185
Yes−0.404<0.001221
Percentage of barley in grower feedPercentage−0.019<0.001545
Finishers fed fermented feedNoBaseline 501
Yes−0.5910.00325
Finishers fed homemixNoBaseline 385
Yes−0.539<0.001141
Finishers fed barleyNoBaseline 269
Yes−0.2840.001257
Percentage of barley in finisher feedPercentage−0.014<0.001532
Sows fed fermented feedNoBaseline 282
Yes−0.8990.0098
Sows fed homemixNoBaseline 187
Yes−0.660<0.001103
Percentage of wheat in sow feedPercentage−0.0060.012542
Sows fed barleyNoBaseline 134
Yes−0.440<0.001156
Percentage of barley in sow feedPercentage−0.016<0.001541
Pig water source: MainsNoBaseline 188
Yes0.2180.014366
Pig water source: BoreholeNoBaseline 376
Yes−0.2230.013178
Any nipple drinkers usedNoBaseline 170
Yes0.2490.006377
No. pig deliveries0–5Baseline 343
6–110.702<0.00183
>110.483<0.001132
No. live pig collections0Baseline 21
1–50.8140.00121
6–110.6260.00974
>110.5700.007441
No. dead stock collections0–5Baseline 114
>60.380<0.001442
No. vermin controller visits0Baseline 237
>00.1930.026295
No. any other deliveries0–11Baseline 564
>11−1.5120.0262
Enzoonotic pneumonia status (last 12 months)NegativeBaseline 302
Positive0.1870.026264
PMWS status (last 12 months)NegativeBaseline 250
Positive0.365<0.001316
PRRS status (last 12 months)NegativeBaseline 436
Positive0.435<0.001130
Glassers status (last 12 months)NegativeBaseline 473
Positive0.3160.00593
Swine dysentery status (last 12 months)NegativeBaseline 535
Positive−0.3640.04931
Clinical salmonellosis status (last 12 months)NegativeBaseline 528
Positive0.5620.00138
No health conditions present (last 12 months)NoBaseline 445
Yes−0.3180.002121
Primary cause of pig mortality in the last 12 monthsOtherBaseline 278
Respiratory or wasting0.510<0.001266
No. sows (log. converted)Continuous0.0360.030438
Any homemix fedNoBaseline 334
Yes−0.548<0.001144
Any wet feedingNoBaseline 448
Yes−0.434<0.00165
Any compound feedingNoBaseline 67
Yes0.500<0.001446
Any solid flooring in finisher housesNoBaseline 275
Yes0.462<0.001249
Monthly maximum temperature anomaly for farm’s region (oC)*Continuous0.023<0.001505
Monthly minimum temperature actual for farm’s region (oC)Continuous0.0030.013505
Monthly minimum temperature anomaly for farm’s region (oC)*Continuous0.032<0.001505
Monthly mean temperature anomaly for farm’s region (oC)*Continuous0.031<0.001505
Monthly rainfall actual for farm’s region (mm)Continuous0.001<0.001505
Monthly rainfall anomaly for farm’s region (mm)*Continuous<0.0010.004505
Monthly sunshine actual for farm’s region (h)Continuous<0.001<0.001505
Monthly sunshine anomaly for farm’s region (h)*Continuous0.001<0.001505
Quarterly cycleCos−0.051<0.001566
Sin−0.038<0.001566
Yearly cycleCos−0.070<0.001566
Sin0.060<0.001566
Table 3.   Multivariable mixed linear regression of logged meat juice ELISA ratio results collected from slaughtered pigs [n = 109,912 samples (474 holdings)]. The standard deviation of the random effect was 0.74 [0.69–0.80 (95% confidence intervals)]
VariableLevelCoefficientP-value
  1. *’anomaly’ is the difference from long-term averages.

NUTS regionScotland−0.747<0.001
OtherBaseline 
Pig farm density within 10 km radius 0.017<0.001
Farm enterpriseConventional−0.518<0.001
Non-conventionalBaseline 
Primary cause of pig mortality in the last 12 months Respiratory or wasting0.290<0.001
OtherBaseline 
Monthly mean temperature anomaly for farm’s region (oC)* 0.024<0.001
Monthly rainfall actual for farm’s region (mm) 0.001<0.001
Monthly sunshine actual for farm’s region (h) 0.0010.001
Finishers fed homemixYes−0.377<0.001
NoBaseline 
Percentage of barley in grower feed −0.0070.003
No. pig deliveries>11/year0.2890.001
6–11/year0.439<0.001
0–5/yearBaseline 
No. dead stock collections>6/year0.2450.007
0–6/yearBaseline 
Yearly cycleCos−0.100<0.001
Sin0.042<0.001
Quarterly cycleCos−0.046<0.001
Sin−0.041<0.001
Constant −2.866<0.001

Discussion

In total, over a quarter (27%) of the QAS population participated in the study and on average each holding was linked to over 200 ZAP samples, providing a large dataset for analysis. The geographical spread of the study holdings indicated that the population was generally representational of the quality assured pig farms in the UK, with similar high-density clusters in Eastern England (mean average of 28 farms within 10 km), Yorkshire and the Humber (21 farms) and in the North East of Scotland (11 farms) (Smith et al., in press).

In the final model, both yearly and quarterly cycles were found to be significant and improved the final model, with the highest mean ELISA ratio in autumn and the lowest in spring. Large differences to long-term averages in the mean temperature, and high actual rainfall and hours of sunshine were identified as risk factors. These results agree with a previous study which presented increased temperature variability as associated with Salmonella prevalence (Funk et al., 2001). Air temperature has been linked to pig stress, which in turn can increase the shedding of Salmonella and can lower immunity (Hald and Andersen, 2001). The meteorological results came from monthly averages from weather stations within each of the regions, whereas the temporal cycles may represent the influence of specific local or daily weather conditions.

The selected spatial factors showed that pigs in Scotland have a lower logarithmic ELISA ratio and thus farms in Scotland have a lower seroprevalence of Salmonella. This may be because the farms in Scotland are more likely to use certain management procedures (e.g. all indoor production; home-mixing) and, because of their geographical isolation, are more likely to purchase animals from similarly low seroprevalence Scottish farms. The range of neighbouring pig farms within 10 km varied greatly (from 0 to 73) and farms with a higher Salmonella prevalence have been shown to be more clustered in space than low prevalence farms by other studies in the UK and Denmark (Benschop et al., 2008; Clough et al., 2009). In these studies, positive farms were more congregated in space than would be expected, possibly caused by local spread and transmission of disease. Location and farm density were identified by a review of UK pig Salmonella, which noted that the ‘type, number and density of pig holdings in a 2-km radius is crucial’ (Pritchard et al., 2005).

It has been described in other studies that health conditions, especially respiratory and wasting diseases such as porcine reproductive and respiratory syndrome, and post-weaning multisystemic wasting syndrome, may have interacted with Salmonella, possibly by lowering the immune system or increasing transmission by sneezing or shedding Salmonella in larger numbers and for a longer period of time, and this relationship was also identified in the model (Schwartz, 1999; Wills et al., 2000; Beloeil et al., 2004, 2007).

A larger number of pig deliveries were also shown to be a risk factor, and the introduction of pigs onto a farm was agreed to be the most likely cause of pig infection by an international expert workshop (Stark et al., 2002). A larger number of pig deliveries may indicate a larger number of suppliers, which has been shown to be a risk factor when farms recruit pigs from more than three herds in comparison to herds that breed their own replacements or recruit from a maximum of three herds (Lo Fo Wong et al., 2004). A higher number of dead-stock collections might indicate that the farms have a greater amount of health condition problems, possibly caused by Salmonella or from health conditions associated with Salmonella infection. These factors may also be a risk simply because the increased number of vehicles entering the farm can facilitate the spread of Salmonella. To decrease the risk from deliveries and visitors, biosecurity measures such as wearing farm-specific clothing and footwear; the routine use of bootdips; ensuring deliveries are only made at the farm perimeter, and closing the farm to all but essential external vehicles should be utilized (Pritchard et al., 2005; Beloeil et al., 2007).

Managing a farm as a conventional pig enterprise was found to be protective, and this may be because the other types of enterprise (organic, freedom foods) utilize a higher degree of outdoor production (only 5% of the conventional farms had any outdoor production in comparison with 33%), and these enterprises have been shown to have significantly higher Salmonella seroprevalence in pigs (Gebreyes et al., 2008). Procedures to control Salmonella transmission which are used in indoor production are harder to implement outdoor and so the pigs may be at an increased risk of infection from wildlife and the environment (Jensen et al., 2006).

Feed has been identified in numerous studies as a factor that influences Salmonella infection. Specific feed types can disrupt the microbial ecosystem in the gut, especially feed with a high level of acid, which can inhibit Salmonella and encourages Gram-positive bacteria which favour acidic environments and can out-compete Salmonella (Lo Fo Wong et al., 2004; Papenbrock et al., 2005). The use of home mix feed was found to be protective, which has been indicated in an earlier British pig study (VLA, 2004) and the use of purchased feed, rather than that mixed on farms, was a significant risk factor for Salmonella in other studies at the multivariable (Benschop et al., 2008) and univariable (Rajic et al., 2007b) levels. A reason for this could be that home mixed feed is usually coarser than purchased feed, with a larger particle size, and these factors influence the growth of competitive gut flora by affecting the acid and starch content in the gut. Purchased feed is also likely to have been pelleted, which has also been indicated as associated with a higher Salmonella prevalence (Leontides et al., 2003; Lo Fo Wong et al., 2004). However, in a longitudinal study of the use of fermented feed, no significant effect was shown, indicating that Salmonella may be able to bypass the stomach environment via the tonsils (Van Winsen et al., 2001, 2002). The use of other feed types, such as a higher percentage of barley in the diet fed to growers, was found to be protective, which concurs with the findings from other studies (Kelliher, 2002; Jorgensen, 2003).

Collecting information from only one time-point for each holding may have introduced error into the analysis as the management of the farm may have changed in the 4-year period, from which samples were collected. The 4-year period was decided upon to provide a suitable dataset to analyse the temporal variation in the data, but an improvement to this study design would be to collect data on any changes to the farm over the period. The cross-sectional study design also meant that we were unable to distinguish between risk factors associated with the infection or persistence of Salmonella. The analysis may also have identified risk factors through reverse-causation, with explanatory factors associated with Salmonella which have been instigated as a response to Salmonella presence, rather than contributing towards Salmonella presence. The large sample size and large number of explanatory variables may also have identified factors associated with Salmonella by chance, because of the large amount of statistical power, although the significance level was lowered to account for this.

Utilizing a study population drawn from the QAS may have provided selection bias to the results, as although the QAS are believed to contain around 50% of all the pig holdings and 90% of the pigs in the UK, it is unknown whether the farms are representational of the remaining farms. Anecdotal evidence suggests that non-assured farms are more likely to be smaller, non-conventional holdings. The non-assured population may utilize different pig management to the QAS holdings and so they may have a different set of factors that are associated with Salmonella. The use of a postal questionnaire may also have provided selection bias, as holdings that responded may be more aware of Salmonella, and Salmonella control, and thus more eager to assist our research. The use of a questionnaire that included questions relating to time periods may also have introduced some recall bias, and it could be theorized that a well-managed and organized farm would have been more likely to be able to use recorded information to answer the questions, whereas a disorganized farm would have been less likely to recall instances over the time period.

The use of serological samples from the ZAP study was a key component of this study, as they provided outcome data from a large number of pig farms, with a large number of samples collected from 4 years. The MJ ELISA is a useful screening tool for surveillance as the test is cost-effective, quick and does not require more specialized microbiological skills (Bohaychuk et al., 2005). However, the use of serological samples and modelling the ELISA ratio directly, without conversion to a binary positive/negative outcome, may limit the interpretation of the findings when considering the infection status of pigs, as the results represent previous exposure, rather than current infection. As the ELISA ratio is an indicator of previous infection, this may have caused information bias in the temporal results. The ELISA ratio is influenced by the strength of the Salmonella challenge and the time since infection, but immune reactions vary by individuals and are affected by many other factors, such as stress. A high ELISA ratio does not necessarily coincide with a more recent infection and a high mean average ELISA ratio in the autumn does not indicate that pigs were infected in the autumn (Tizard, 2004). The ELISA test benefits from detecting life time infection, even if it is subclinical, but only detects a number of known serovars with potentially differential abilities to detect infection by different serovars (Funk et al., 2005). However, studies have shown a significant correlation between serology results and caecal prevalence, and although farm results can fluctuate between visits and sampling occasions, the test has been shown to be useful in identifying farms with a Salmonella problem (Sorensen et al., 2004; Rajic et al., 2007a).

The study provided a comprehensive risk factor analysis and examination of the spatial and temporal trends of Salmonella seroprevalence, with a study population large enough to detect even factors with modest associations to the ELISA ratio. Large sample sizes can provide greater statistical power and provide narrower confidence intervals for estimated associations and so are more likely to detect if a significant difference is present in the data. Even though association may be weak, it may still have a significant impact on Salmonella presence in the study population if it is present in a large proportion of the population. Specifically, the model results suggest that measures are needed to control Salmonella infection on farms utilizing outdoor production and to protect pigs from the effects of large variations in weather conditions and an intervention study would be required to test this finding. The model also highlighted a region of the UK that may require more intensive surveillance and control to limit the transmission of Salmonella. The utilisation of data collected routinely via the QAS and ZAP schemes, as well as a one-off postal questionnaire provided a cost-effective means to design and analyse a large risk factor study.

Acknowledgements

The authors would like to thank the participating farmers and schemes, and BPEx for their support. Colleagues in CERA are thanked for their help in data entry and handling. Defra are also thanked for funding the study under projects OD0215/ OZ0323.

Ancillary