Community ecology
rsbl.royalsocietypublishing.org
Environmental risk factors associated with bovine tuberculosis among cattle in highrisk areas
Research
B. Winkler and F. Mathews
Cite this article: Winkler B, Mathews F. 2015 Environmental risk factors associated with bovine tuberculosis among cattle in high-risk areas. Biol. Lett. 11: 20150536. http://dx.doi.org/10.1098/rsbl.2015.0536
Received: 23 June 2015 Accepted: 23 September 2015
Subject Areas: ecology, health and disease and epidemiology Keywords: habitat, badgers, cattle, ecology, epidemiology, landscape-scale
Author for correspondence: F. Mathews e-mail: [email protected]
Hatherly Laboratories, Biosciences, College of Life and Environmental Sciences, University of Exeter, Prince of Wales Road, Exeter EX4 4PS, UK Our research shows that environmental features are important predictors of bovine tuberculosis (bTB) in British cattle herds in high-prevalence regions. Data from 503 case and 808 control farms included in the randomized badger culling trial (RBCT) were analysed. bTB risk increased in larger herds and on farms with greater areas of maize, deciduous woodland and marsh, whereas a higher percentage of boundaries composed of hedgerows decreased the risk. The model was tested on another case – control study outside RBCT areas, and here it had a much smaller predictive power. This suggests that different infection dynamics operate outside high-risk areas, although it is possible that unknown confounding factors may also have played a role.
1. Introduction Bovine tuberculosis (bTB) is a significant economic burden to agriculture, particularly in the UK where the number of new breakdowns remains high. Within high-risk areas, the spatial heterogeneity in the risk of both new and recurrent breakdowns remains largely unexplained [1]. The movement of infected cattle plays an important role in the range expansion of the disease [2]. However, recent work modelling transmission pathways suggests that the environment plays an important role in the within farm maintenance and short-distance spread of the disease [2]. The European badger (Meles meles) is an important wildlife reservoir of bTB in the UK [1], and the farm environment can become contaminated owing to the presence of infected badgers [3] and/or cattle [4]. It has been suggested that the importance of environmental factors to bTB epidemiology has increased since the foot and mouth outbreak, possibly owing to greater contamination of badgers by infected cattle [5]. Reducing exposure to environmental contamination could therefore play a fundamental role in managing bTB. This may extend beyond simply excluding badgers from cattle feeding areas, to wider landscape management which influences habitat use by both badgers and cattle. For example, increased density of hedges and the presence of buffer strips on field margins have been linked with reduced risk of bTB in cattle herds [6]. The aim of our study is to identify environmental variables that influence the risk of cattle acquiring bTB, in order to explore the potential for landscape management to contribute to bTB control.
2. Material and methods Electronic supplementary material is available at http://dx.doi.org/10.1098/rsbl.2015.0536 or via http://rsbl.royalsocietypublishing.org.
We analysed data collected between 1998 and 2004 as part of the TB99 case– control study associated with the randomized badger culling trial (RBCT). Within the 10 trial areas of the RBCT, all breakdowns (whether confirmed or not) triggered a survey of potential farm-level risk factors [7]. In addition, for each breakdown, the
& 2015 The Author(s) Published by the Royal Society. All rights reserved.
AIC
DAIC
Akaike weight
deciduous wood (ha), marsh (ha), rough pasture (ha), maize (ha), internal boundary hedges (%), grazing silage/
1600.66
0.00
0.56
deciduous wood (ha), marsh (ha), maize (ha), internal boundary hedges (%), grazing silage/hay aftermath (y/n), feeding silage (y/n), herd size category, enterprise type, N. cattle moving on, incident number
1603.40
2.38
0.17
deciduous wood (ha), marsh (ha), rough pasture (ha), internal boundary hedges (%), grazing silage/hay aftermath (y/n), feeding silage (y/n), herd size category, enterprise type, N. cattle moving on, incident number
1603.50
2.40
0.17
deciduous wood (ha), marsh (ha), rough pasture (ha), maize (ha), internal boundary hedges (%), grazing silage/
1604.10
3.45
0.10
hay aftermath (y/n), feeding silage (y/n), herd size category, enterprise type, N. cattle moving on, incident numbera
hay aftermath (y/n), feeding silage (y/n), herd size category, enterprise type, N. cattle moving on, incident number, cull areas a
Incident number—case control design variable.
same survey was conducted at one to three control herds within the same trial area (including, where possible, one contiguous herd). Control herds had no bTB test reactors in the previous 12 months, and were selected to represent the range of herd sizes within the trial area. In total, we analysed data from 503 case and 806 control farms. The ability of habitat and herd management data to predict bTB breakdown status was analysed using generalized linear modelling with a binomial error structure in R v. 3.1.0 [8]. All models included the case– control design variable as a fixed factor. In addition, they accounted for the RBCT treatment ( proactive and reactive badger culling or control), because breakdown risk among farms recruited some years after the onset of the study could have varied according the treatment regime. We used an information-theoretic approach to model selection, as this is designed to capture real-world complexity while minimizing the risk of making spurious associations [9]. We screened all environmental variables and a subset of herd management predictors. These were selected on the basis of results obtained previously with similar datasets [7,10]. Univariate logistic regression with a relaxed inclusion criterion ( p , 0.10) was used. See the electronic supplementary material for complete list and descriptive statistics. We repeated the analysis including only control herds that did not have a previous breakdown to account for any possible residual effect of a breakdown before the 12 month selection period. The results were not different from those obtained using the full dataset (see the electronic supplementary material). The relative predictive ability of the models was compared using Akaike’s information criterion (with delta AIC 4) [9] (R v. 3.1.0, MuMIn package). Inferences were made based on model-averaged predictions and were computed as a weighted mean for the set of best models. We then tested the consistency of the variables in predicting a bTB outbreak on a separate case– control dataset, the CCS05. This study was conducted in 2005 – 2006 and focused on four areas where the number of bTB breakdowns in cattle herds ranged from medium to high (Carlisle, Carmarthen, Stafford and Taunton). It included 400 case farms that were randomly selected from farms that suffered bTB outbreaks (confirmed or not). Two control farms were randomly selected in the same region for each case farm, one matching the case farm in herd size and type. The same criteria as in the TB99 study were used to define control herds.
3. Results The risk of bTB breakdown increased on farms with greater areas of deciduous woodland, maize, marsh and rough pasture, and in herds that were larger, fed silage and were dairy units. The risk decreased on farms that had a greater percentage of hedges in boundaries, that grazed cattle on fields that had been cut for silage or hay and had greater numbers of cattle moving off the holding. The models explaining the risk of bTB breakdown in the TB99 dataset are presented in table 1 and the predictor weights, model-averaged odds ratio and confidence interval (CI) for variables in the top models are shown in table 2. No difference to the results was observed according to whether or not RBCT treatment was included in the model. The pseudo-R 2, that indicates the goodness of fit of the top TB99 model, was 0.21 and the AUC 0.71 (a measure of the predictive ability of the model) [11]. When testing the same variables in a second dataset, the CCS05, many of the same predictors appeared in the top-ranking models and had similar weightings (table 3). However, seasonally wet soils (approximately corresponding to ‘marsh’ in TB99) and percentage of hedgerows appeared in less than half the top models. Also the direction of the association with deciduous woodland was reversed. Full outcomes for the CCS05 dataset and differences between the two datasets are shown in the electronic supplementary material. The positive predicted value of the top model when applied to the new dataset was 61.5%, and the negative predicted value was 31.0%, indicating that almost two-thirds of the case herds and one-third of control herds were correctly classified (AUC 0.63), suggesting poorer predictive ability compared with the use of the same model in the TB99 areas.
4. Discussion Our research shows that environmental features—farmland habitat and herd management—are important predictors of
Biol. Lett. 11: 20150536
model
2
rsbl.royalsocietypublishing.org
Table 1. Akaike information statistic ranking logistic regression models of the odds of bovine tuberculosis on cattle farms. (AIC, Akaike information criterion; DAIC, amount of support for the model relative to the top-ranking model; Akaike weight, probability of the candidate model being the ‘best’ out of all those considered.)
Table 2. Predictor weights and odds ratios of variables appearing in the top models from logistic regression of bovine tuberculosis breakdown risk.
univariate odds ratio
odds ratio from multivariate model
95% CI for multivariate odds ratio
deciduous wood (10 ha)
4
1.00
1.40
1.32
1.08– 1.62
marsh (10 ha)
4
1.00
1.79
1.70
1.11– 2.60
rough pasture (10 ha) internal boundary hedge (%)
3 4
0.83 1.00
1.10 0.66
1.07 0.63
1.00– 1.15 0.47– 0.83
maize (10 ha) grazing silage hay aftermath
3 4
0.83 1.00
1.40 0.71
1.20 0.56
1.01– 1.44 0.43– 0.73
4 4
1.00 1.00
2.98
2.20
1.45– 3.32
0.33 0.73
0.50 0.88
0.33– 0.75 0.66– 1.19
—
—
—
—
—
1.83 0.88
1.24 1.01
0.91– 1.69 0.62– 1.64
0.48 0.96
0.46 0.95
0.30– 0.71 0.93– 0.97
control reactive
— 0.96
— 0.95
— 0.90– 1.20
proactive
1.07
1.07
0.70– 1.43
(yes/no) feeding silage (yes/no) herd size category: small (,50 cattle) medium (50– 150 cattle) large (.150 cattle)a cattle enterprise type: beefa
4
1.00
dairy sheep other cattle moving on (10 cattle)
4
1.00
cull areas:
1
0.10
a
a
Levels with no odds ratio were used as the reference level.
Table 3. Predictor weights of the variables in the logistic regression models of the TB99 and CCS05 datasets.
variable
no. models in which variable appears in the TB99 dataset (out of 6)
predictor weight TB99 dataset
no. models in which variable appears in the CCS05 dataset (out of 35)
predictor weight CCS05 dataset
deciduous wood
6
1.00
32
0.96
marsh (TB99)—seasonally wet soil (CCS05)
6
1.00
17
0.50
rough pasture
4
0.81
27
0.88
internal boundary hedge (%) maize (y/n)a
6 4
1.00 0.80
8 27
0.20 0.88
feeding silage (y/n) herd size category
6 6
1.00 1.00
18 35
0.53 1.00
cattle enterprise type
6
1.00
35
1.00
cattle moving on
6
1.00
19
0.61
a
Maize was included as a binomial variable (grown/not grown) when both models were compared, but remained a numeric variable (ha) in the main TB99 analysis. bTB in high-prevalence areas. Broadly, characteristics of higher intensity production, such as larger herd size, maize production, use of silage and reduced hedgerow abundance were linked with elevated infection risk (though note that
area of rough pasture was also linked to a small increase in risk). In areas of mixed infection risk outside bTB ‘hotspots’, many of the same predictors (including herd size, enterprise type, maize production, deciduous woodland area and the
Biol. Lett. 11: 20150536
predictor weight
rsbl.royalsocietypublishing.org
variable
no. models in which variable appears (out of 4)
3
Ethics. This was a secondary analysis of data collected for other projects. Full permission for use of these data was provided by the UK’s Animal and Plant Health Agency. None of the data analysed in our work involved the use of animals. Data accessibility. The complete TB99 and CCS05 datasets on which this project was based are freely available from the UK’s Animal and Plant Health Agency. The subset of data used on our analyses, together with the R script, is available from the Open Repository Exeter (http://as.exeter.ac.uk/library/resources/openaccess/). Authors’ contributions. B.W. analysed and interpreted the data, drafted the article and approved the final version to be published. F.M. designed the project, interpreted the data, revised the article and approved the final version to be published.
Competing interests. We have no competing interests. Funding. B.W. held a Daphne Jackson Fellowship and was supported by the BBSRC and the University of Exeter.
Acknowledgements. We thank Andrew Mitchell for helpful comments on the manuscript, and the Animal and Plant Health Laboratory for providing access to the datasets.
References 1.
2.
3.
Bessell PR, Orton R, White PCL, Hutchings MR, Kao RR. 2012 Risk factors for bovine tuberculosis at the national level in Great Britain. BMC Vet. Res. 8, 51. (doi:10.1186/1746-6148-8-51) Brooks-Pollock E, Roberts GO, Keeling MJ. 2014 A dynamic model of bovine tuberculosis spread and control in Great Britain. Nature 511, 228– 231. (doi:10.1038/nature13529) Courtenay O et al. 2006 Is Mycobacterium bovis in the environment important for the persistence of bovine tuberculosis? Biol. Lett. 2, 460 –462. (doi:10. 1098/rsbl.2006.0468)
4.
5.
6.
Neill S, Hanna J, O’Brien J, McCracken R. 1988 Excretion of Mycobacterium bovis by experimentally infected cattle. Vet. Rec. 123, 340– 343. (doi:10. 1136/vr.123.13.340) Vial F, Miguel E, T Johnston W, Mitchell A, Donnelly CA. 2013 Bovine tuberculosis risk factors for British herds before and after the 2001 foot-and-mouth epidemic: what have we learned from the TB99 and CCS2005 studies? Transbound. Emerg. Dis. 62, 505 – 515. (doi:10.1111/tbed.12184) Mathews F, Lovett L, Rushton S, Macdonald DW. 2006 Bovine tuberculosis in cattle: reduced risk on
7.
8.
9.
wildlife-friendly farms. Biol. Lett. 2, 271 –274. (doi:10.1098/rsbl.2006.0461) Johnston WT et al. 2005 Herd-level risk factors associated with tuberculosis breakdowns among cattle herds in England before the 2001 foot-andmouth disease epidemic. Biol. Lett. 1, 53 –56. (doi:10.1098/rsbl.2004.0249) R Core Team. 2014 R: a language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing. Burnham KP, Anderson DR, Huyvaert KP. 2010 AIC model selection and multimodel inference in
4
Biol. Lett. 11: 20150536
industry is currently undergoing particularly marked alterations owing to market and regulatory changes. Average dairy herd sizes rose by 36% from 1990 to 2003 in England, with the most marked changes being in the south. In the same period, the area planted with maize in South West England increased fourfold [16]. Badgers favour maize as a food source: in the South West of England 72% of land owners report badger damage to cereal crops (oats, maize, barley and wheat) [17]. Contamination of maize by badger faeces and urine may therefore present a possible route of infection. Maize may also play a role by altering badger population sizes and their nutritional status. It is vital for food security that holistic approaches to bTB control are implemented. These need to consider landscape composition, herd management and the use of the environment by badgers. They must also be tailored to the local situation. For example, the 70% increase in risk of breakdown observed for every 10 ha of marsh area in the TB99 study may be linked with exposure to liver fluke (Fasciola hepatica) which can affect the sensitivity of bTB tests. The parasite is transmitted by an amphibious snail, and focal screening in wet areas and/or exclusion of cattle from wet land may therefore be warranted to address the dual problems of bTB and liver fluke [18]. Further studies should try to pinpoint disease hotspots within farms, synthesizing data on cattle grazing management, habitat and distribution of badger setts and pathways.
rsbl.royalsocietypublishing.org
abundance of hedgerows) appeared in the top models. However, the predictive value of some of the habitat features, and of feeding silage, was reduced. This change in relative importance may reflect differences in disease dynamics, with the movement of infected cattle being the main risk factor [1]. While the TB99 dataset—on which we based our models— is derived from farms in the South West of England the comparator CCS05 dataset is more geographically dispersed. Whereas all the TB99 farms fall within land classes 1 and 4 [12], only one region in CCS05 (Taunton) is in this group. Some of the habitat variables included in the top models for the South West may therefore be of less relevance elsewhere: for example, hedgerows are rarely used as field boundaries in some geographical areas. It is also possible that the relative importance of badger –cattle and cattle – badger transmission (and the interactive effects of land management which could modify this transmission risk) differs in land classes where badger density is lower [13]. The use of the landscape by both badgers and cattle affects the likelihood of successful bTB transmission between the two. The distribution of badger setts is influenced by habitat, with higher sett densities occurring in areas with greater areas of broadleaved woodland, improved grassland and high hedgerow density, and lower densities being found in heather moorland [14]. Therefore, we expected a reduced risk of breakdown to be associated with areas of rough and moorland grazing. By contrast, the risk increased in both the TB99 and CCS05 datasets. This may be because the category of ‘rough grassland’ included a wide range of habitat types some of which may be favourable to badgers. Like our earlier research [6], we found reduced risks of bTB on farms with greater hedgerow abundance. Badger latrines and urination sites are often associated with woodland edges, hedgerows and walls [15], and it is therefore likely that the availability of these resources limits the contamination of pasture. How closely cattle graze to the boundary features will depend on management practices and grazing pressure, and further work is warranted to assess whether reduced cattle access to these features is effective in reducing bTB risk. Much of the variability in landscape management is tightly tied with herd size and enterprise type (for example, large herd sizes are associated with large fields and lower hedgerow densities). Our study demonstrates that the trend in livestock farming towards larger herd sizes, and the use of silage and field maize for the maintenance of high-productivity animals, is likely to have consequences for bTB control. The dairy
tuberculosis to cattle. J. Zool. 272, 311– 320. (doi:10.1111/j.1469-7998.2006.00271.x) 16. Clothier L. 2009 Analysis of recent data on dairy cows in England and implications for the environment. Defra Agricultural Change and Environment Observatory Research Report No. 14. 17. Moore N, Whiterow A, Kelly P, Garthwaite D, Bishop J, Langton S, Cheesman C. 1999 Survey of badger Meles meles damage to agriculture in England and Wales. J. Appl. Ecol. 36, 974 –988. (doi:10.1046/j. 1365-2664.1999.00454.x) 18. Claridge J et al. 2012 Fasciola hepatica is associated with the failure to detect bovine tuberculosis in dairy cattle. Nat. Commun. 3, 853. (doi:10.1038/ ncomms1840)
5
Biol. Lett. 11: 20150536
13. Judge J, Wilson GJ, Macarthur R, Delahay RJ, McDonald RA. 2014 Density and abundance of badger social groups in England and Wales in 2011 –2013. Sci. Rep. 4, 3809. (doi:10.1038/ srep03809) 14. Newton-Cross G, White PCL, Harris S. 2007 Modelling the distribution of badgers Meles meles: comparing predictions from field-based and remotely derived habitat data. Mamm. Rev. 37, 54 –70. (doi:10.1111/j.1365-2907.2007. 00103.x) 15. Delahay RJ, Ward AI, Walker N, Long B, Cheeseman CL. 2007 Distribution of badger latrines in a highdensity population: habitat selection and implications for the transmission of bovine
rsbl.royalsocietypublishing.org
behavioral ecology: some background, observations, and comparisons. Behav. Ecol. Sociobiol. 65, 23– 35. (doi:10.1007/s00265-010-1029-6) 10. Johnston WT et al. 2011 Herd-level risk factors of bovine tuberculosis in England and Wales after the 2001 foot-and-mouth disease epidemic. Int. J. Infect. Dis. 15, e833 –e840. (doi:10.1016/j. ijid.2011.08.004) 11. Hosmer DW, Lemeshow S, Sturdivant RX. 2013 Applied logistic regression, 3rd edn. Hoboken, NJ: John Wiley & Sons. 12. Bunce RGH, Barr CJ, Whittaker HA. 1981 Land classes in Great Britain: preliminary descriptions for use of the Merlewood method of land classification. Grange-overSands, UK: Institute of Terrestrial Ecology.
rsbl.royalsocietypublishing.org
Environmental risk factors associated with bovine tuberculosis among cattle in highrisk areas
Research
B. Winkler and F. Mathews
Cite this article: Winkler B, Mathews F. 2015 Environmental risk factors associated with bovine tuberculosis among cattle in high-risk areas. Biol. Lett. 11: 20150536. http://dx.doi.org/10.1098/rsbl.2015.0536
Received: 23 June 2015 Accepted: 23 September 2015
Subject Areas: ecology, health and disease and epidemiology Keywords: habitat, badgers, cattle, ecology, epidemiology, landscape-scale
Author for correspondence: F. Mathews e-mail: [email protected]
Hatherly Laboratories, Biosciences, College of Life and Environmental Sciences, University of Exeter, Prince of Wales Road, Exeter EX4 4PS, UK Our research shows that environmental features are important predictors of bovine tuberculosis (bTB) in British cattle herds in high-prevalence regions. Data from 503 case and 808 control farms included in the randomized badger culling trial (RBCT) were analysed. bTB risk increased in larger herds and on farms with greater areas of maize, deciduous woodland and marsh, whereas a higher percentage of boundaries composed of hedgerows decreased the risk. The model was tested on another case – control study outside RBCT areas, and here it had a much smaller predictive power. This suggests that different infection dynamics operate outside high-risk areas, although it is possible that unknown confounding factors may also have played a role.
1. Introduction Bovine tuberculosis (bTB) is a significant economic burden to agriculture, particularly in the UK where the number of new breakdowns remains high. Within high-risk areas, the spatial heterogeneity in the risk of both new and recurrent breakdowns remains largely unexplained [1]. The movement of infected cattle plays an important role in the range expansion of the disease [2]. However, recent work modelling transmission pathways suggests that the environment plays an important role in the within farm maintenance and short-distance spread of the disease [2]. The European badger (Meles meles) is an important wildlife reservoir of bTB in the UK [1], and the farm environment can become contaminated owing to the presence of infected badgers [3] and/or cattle [4]. It has been suggested that the importance of environmental factors to bTB epidemiology has increased since the foot and mouth outbreak, possibly owing to greater contamination of badgers by infected cattle [5]. Reducing exposure to environmental contamination could therefore play a fundamental role in managing bTB. This may extend beyond simply excluding badgers from cattle feeding areas, to wider landscape management which influences habitat use by both badgers and cattle. For example, increased density of hedges and the presence of buffer strips on field margins have been linked with reduced risk of bTB in cattle herds [6]. The aim of our study is to identify environmental variables that influence the risk of cattle acquiring bTB, in order to explore the potential for landscape management to contribute to bTB control.
2. Material and methods Electronic supplementary material is available at http://dx.doi.org/10.1098/rsbl.2015.0536 or via http://rsbl.royalsocietypublishing.org.
We analysed data collected between 1998 and 2004 as part of the TB99 case– control study associated with the randomized badger culling trial (RBCT). Within the 10 trial areas of the RBCT, all breakdowns (whether confirmed or not) triggered a survey of potential farm-level risk factors [7]. In addition, for each breakdown, the
& 2015 The Author(s) Published by the Royal Society. All rights reserved.
AIC
DAIC
Akaike weight
deciduous wood (ha), marsh (ha), rough pasture (ha), maize (ha), internal boundary hedges (%), grazing silage/
1600.66
0.00
0.56
deciduous wood (ha), marsh (ha), maize (ha), internal boundary hedges (%), grazing silage/hay aftermath (y/n), feeding silage (y/n), herd size category, enterprise type, N. cattle moving on, incident number
1603.40
2.38
0.17
deciduous wood (ha), marsh (ha), rough pasture (ha), internal boundary hedges (%), grazing silage/hay aftermath (y/n), feeding silage (y/n), herd size category, enterprise type, N. cattle moving on, incident number
1603.50
2.40
0.17
deciduous wood (ha), marsh (ha), rough pasture (ha), maize (ha), internal boundary hedges (%), grazing silage/
1604.10
3.45
0.10
hay aftermath (y/n), feeding silage (y/n), herd size category, enterprise type, N. cattle moving on, incident numbera
hay aftermath (y/n), feeding silage (y/n), herd size category, enterprise type, N. cattle moving on, incident number, cull areas a
Incident number—case control design variable.
same survey was conducted at one to three control herds within the same trial area (including, where possible, one contiguous herd). Control herds had no bTB test reactors in the previous 12 months, and were selected to represent the range of herd sizes within the trial area. In total, we analysed data from 503 case and 806 control farms. The ability of habitat and herd management data to predict bTB breakdown status was analysed using generalized linear modelling with a binomial error structure in R v. 3.1.0 [8]. All models included the case– control design variable as a fixed factor. In addition, they accounted for the RBCT treatment ( proactive and reactive badger culling or control), because breakdown risk among farms recruited some years after the onset of the study could have varied according the treatment regime. We used an information-theoretic approach to model selection, as this is designed to capture real-world complexity while minimizing the risk of making spurious associations [9]. We screened all environmental variables and a subset of herd management predictors. These were selected on the basis of results obtained previously with similar datasets [7,10]. Univariate logistic regression with a relaxed inclusion criterion ( p , 0.10) was used. See the electronic supplementary material for complete list and descriptive statistics. We repeated the analysis including only control herds that did not have a previous breakdown to account for any possible residual effect of a breakdown before the 12 month selection period. The results were not different from those obtained using the full dataset (see the electronic supplementary material). The relative predictive ability of the models was compared using Akaike’s information criterion (with delta AIC 4) [9] (R v. 3.1.0, MuMIn package). Inferences were made based on model-averaged predictions and were computed as a weighted mean for the set of best models. We then tested the consistency of the variables in predicting a bTB outbreak on a separate case– control dataset, the CCS05. This study was conducted in 2005 – 2006 and focused on four areas where the number of bTB breakdowns in cattle herds ranged from medium to high (Carlisle, Carmarthen, Stafford and Taunton). It included 400 case farms that were randomly selected from farms that suffered bTB outbreaks (confirmed or not). Two control farms were randomly selected in the same region for each case farm, one matching the case farm in herd size and type. The same criteria as in the TB99 study were used to define control herds.
3. Results The risk of bTB breakdown increased on farms with greater areas of deciduous woodland, maize, marsh and rough pasture, and in herds that were larger, fed silage and were dairy units. The risk decreased on farms that had a greater percentage of hedges in boundaries, that grazed cattle on fields that had been cut for silage or hay and had greater numbers of cattle moving off the holding. The models explaining the risk of bTB breakdown in the TB99 dataset are presented in table 1 and the predictor weights, model-averaged odds ratio and confidence interval (CI) for variables in the top models are shown in table 2. No difference to the results was observed according to whether or not RBCT treatment was included in the model. The pseudo-R 2, that indicates the goodness of fit of the top TB99 model, was 0.21 and the AUC 0.71 (a measure of the predictive ability of the model) [11]. When testing the same variables in a second dataset, the CCS05, many of the same predictors appeared in the top-ranking models and had similar weightings (table 3). However, seasonally wet soils (approximately corresponding to ‘marsh’ in TB99) and percentage of hedgerows appeared in less than half the top models. Also the direction of the association with deciduous woodland was reversed. Full outcomes for the CCS05 dataset and differences between the two datasets are shown in the electronic supplementary material. The positive predicted value of the top model when applied to the new dataset was 61.5%, and the negative predicted value was 31.0%, indicating that almost two-thirds of the case herds and one-third of control herds were correctly classified (AUC 0.63), suggesting poorer predictive ability compared with the use of the same model in the TB99 areas.
4. Discussion Our research shows that environmental features—farmland habitat and herd management—are important predictors of
Biol. Lett. 11: 20150536
model
2
rsbl.royalsocietypublishing.org
Table 1. Akaike information statistic ranking logistic regression models of the odds of bovine tuberculosis on cattle farms. (AIC, Akaike information criterion; DAIC, amount of support for the model relative to the top-ranking model; Akaike weight, probability of the candidate model being the ‘best’ out of all those considered.)
Table 2. Predictor weights and odds ratios of variables appearing in the top models from logistic regression of bovine tuberculosis breakdown risk.
univariate odds ratio
odds ratio from multivariate model
95% CI for multivariate odds ratio
deciduous wood (10 ha)
4
1.00
1.40
1.32
1.08– 1.62
marsh (10 ha)
4
1.00
1.79
1.70
1.11– 2.60
rough pasture (10 ha) internal boundary hedge (%)
3 4
0.83 1.00
1.10 0.66
1.07 0.63
1.00– 1.15 0.47– 0.83
maize (10 ha) grazing silage hay aftermath
3 4
0.83 1.00
1.40 0.71
1.20 0.56
1.01– 1.44 0.43– 0.73
4 4
1.00 1.00
2.98
2.20
1.45– 3.32
0.33 0.73
0.50 0.88
0.33– 0.75 0.66– 1.19
—
—
—
—
—
1.83 0.88
1.24 1.01
0.91– 1.69 0.62– 1.64
0.48 0.96
0.46 0.95
0.30– 0.71 0.93– 0.97
control reactive
— 0.96
— 0.95
— 0.90– 1.20
proactive
1.07
1.07
0.70– 1.43
(yes/no) feeding silage (yes/no) herd size category: small (,50 cattle) medium (50– 150 cattle) large (.150 cattle)a cattle enterprise type: beefa
4
1.00
dairy sheep other cattle moving on (10 cattle)
4
1.00
cull areas:
1
0.10
a
a
Levels with no odds ratio were used as the reference level.
Table 3. Predictor weights of the variables in the logistic regression models of the TB99 and CCS05 datasets.
variable
no. models in which variable appears in the TB99 dataset (out of 6)
predictor weight TB99 dataset
no. models in which variable appears in the CCS05 dataset (out of 35)
predictor weight CCS05 dataset
deciduous wood
6
1.00
32
0.96
marsh (TB99)—seasonally wet soil (CCS05)
6
1.00
17
0.50
rough pasture
4
0.81
27
0.88
internal boundary hedge (%) maize (y/n)a
6 4
1.00 0.80
8 27
0.20 0.88
feeding silage (y/n) herd size category
6 6
1.00 1.00
18 35
0.53 1.00
cattle enterprise type
6
1.00
35
1.00
cattle moving on
6
1.00
19
0.61
a
Maize was included as a binomial variable (grown/not grown) when both models were compared, but remained a numeric variable (ha) in the main TB99 analysis. bTB in high-prevalence areas. Broadly, characteristics of higher intensity production, such as larger herd size, maize production, use of silage and reduced hedgerow abundance were linked with elevated infection risk (though note that
area of rough pasture was also linked to a small increase in risk). In areas of mixed infection risk outside bTB ‘hotspots’, many of the same predictors (including herd size, enterprise type, maize production, deciduous woodland area and the
Biol. Lett. 11: 20150536
predictor weight
rsbl.royalsocietypublishing.org
variable
no. models in which variable appears (out of 4)
3
Ethics. This was a secondary analysis of data collected for other projects. Full permission for use of these data was provided by the UK’s Animal and Plant Health Agency. None of the data analysed in our work involved the use of animals. Data accessibility. The complete TB99 and CCS05 datasets on which this project was based are freely available from the UK’s Animal and Plant Health Agency. The subset of data used on our analyses, together with the R script, is available from the Open Repository Exeter (http://as.exeter.ac.uk/library/resources/openaccess/). Authors’ contributions. B.W. analysed and interpreted the data, drafted the article and approved the final version to be published. F.M. designed the project, interpreted the data, revised the article and approved the final version to be published.
Competing interests. We have no competing interests. Funding. B.W. held a Daphne Jackson Fellowship and was supported by the BBSRC and the University of Exeter.
Acknowledgements. We thank Andrew Mitchell for helpful comments on the manuscript, and the Animal and Plant Health Laboratory for providing access to the datasets.
References 1.
2.
3.
Bessell PR, Orton R, White PCL, Hutchings MR, Kao RR. 2012 Risk factors for bovine tuberculosis at the national level in Great Britain. BMC Vet. Res. 8, 51. (doi:10.1186/1746-6148-8-51) Brooks-Pollock E, Roberts GO, Keeling MJ. 2014 A dynamic model of bovine tuberculosis spread and control in Great Britain. Nature 511, 228– 231. (doi:10.1038/nature13529) Courtenay O et al. 2006 Is Mycobacterium bovis in the environment important for the persistence of bovine tuberculosis? Biol. Lett. 2, 460 –462. (doi:10. 1098/rsbl.2006.0468)
4.
5.
6.
Neill S, Hanna J, O’Brien J, McCracken R. 1988 Excretion of Mycobacterium bovis by experimentally infected cattle. Vet. Rec. 123, 340– 343. (doi:10. 1136/vr.123.13.340) Vial F, Miguel E, T Johnston W, Mitchell A, Donnelly CA. 2013 Bovine tuberculosis risk factors for British herds before and after the 2001 foot-and-mouth epidemic: what have we learned from the TB99 and CCS2005 studies? Transbound. Emerg. Dis. 62, 505 – 515. (doi:10.1111/tbed.12184) Mathews F, Lovett L, Rushton S, Macdonald DW. 2006 Bovine tuberculosis in cattle: reduced risk on
7.
8.
9.
wildlife-friendly farms. Biol. Lett. 2, 271 –274. (doi:10.1098/rsbl.2006.0461) Johnston WT et al. 2005 Herd-level risk factors associated with tuberculosis breakdowns among cattle herds in England before the 2001 foot-andmouth disease epidemic. Biol. Lett. 1, 53 –56. (doi:10.1098/rsbl.2004.0249) R Core Team. 2014 R: a language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing. Burnham KP, Anderson DR, Huyvaert KP. 2010 AIC model selection and multimodel inference in
4
Biol. Lett. 11: 20150536
industry is currently undergoing particularly marked alterations owing to market and regulatory changes. Average dairy herd sizes rose by 36% from 1990 to 2003 in England, with the most marked changes being in the south. In the same period, the area planted with maize in South West England increased fourfold [16]. Badgers favour maize as a food source: in the South West of England 72% of land owners report badger damage to cereal crops (oats, maize, barley and wheat) [17]. Contamination of maize by badger faeces and urine may therefore present a possible route of infection. Maize may also play a role by altering badger population sizes and their nutritional status. It is vital for food security that holistic approaches to bTB control are implemented. These need to consider landscape composition, herd management and the use of the environment by badgers. They must also be tailored to the local situation. For example, the 70% increase in risk of breakdown observed for every 10 ha of marsh area in the TB99 study may be linked with exposure to liver fluke (Fasciola hepatica) which can affect the sensitivity of bTB tests. The parasite is transmitted by an amphibious snail, and focal screening in wet areas and/or exclusion of cattle from wet land may therefore be warranted to address the dual problems of bTB and liver fluke [18]. Further studies should try to pinpoint disease hotspots within farms, synthesizing data on cattle grazing management, habitat and distribution of badger setts and pathways.
rsbl.royalsocietypublishing.org
abundance of hedgerows) appeared in the top models. However, the predictive value of some of the habitat features, and of feeding silage, was reduced. This change in relative importance may reflect differences in disease dynamics, with the movement of infected cattle being the main risk factor [1]. While the TB99 dataset—on which we based our models— is derived from farms in the South West of England the comparator CCS05 dataset is more geographically dispersed. Whereas all the TB99 farms fall within land classes 1 and 4 [12], only one region in CCS05 (Taunton) is in this group. Some of the habitat variables included in the top models for the South West may therefore be of less relevance elsewhere: for example, hedgerows are rarely used as field boundaries in some geographical areas. It is also possible that the relative importance of badger –cattle and cattle – badger transmission (and the interactive effects of land management which could modify this transmission risk) differs in land classes where badger density is lower [13]. The use of the landscape by both badgers and cattle affects the likelihood of successful bTB transmission between the two. The distribution of badger setts is influenced by habitat, with higher sett densities occurring in areas with greater areas of broadleaved woodland, improved grassland and high hedgerow density, and lower densities being found in heather moorland [14]. Therefore, we expected a reduced risk of breakdown to be associated with areas of rough and moorland grazing. By contrast, the risk increased in both the TB99 and CCS05 datasets. This may be because the category of ‘rough grassland’ included a wide range of habitat types some of which may be favourable to badgers. Like our earlier research [6], we found reduced risks of bTB on farms with greater hedgerow abundance. Badger latrines and urination sites are often associated with woodland edges, hedgerows and walls [15], and it is therefore likely that the availability of these resources limits the contamination of pasture. How closely cattle graze to the boundary features will depend on management practices and grazing pressure, and further work is warranted to assess whether reduced cattle access to these features is effective in reducing bTB risk. Much of the variability in landscape management is tightly tied with herd size and enterprise type (for example, large herd sizes are associated with large fields and lower hedgerow densities). Our study demonstrates that the trend in livestock farming towards larger herd sizes, and the use of silage and field maize for the maintenance of high-productivity animals, is likely to have consequences for bTB control. The dairy
tuberculosis to cattle. J. Zool. 272, 311– 320. (doi:10.1111/j.1469-7998.2006.00271.x) 16. Clothier L. 2009 Analysis of recent data on dairy cows in England and implications for the environment. Defra Agricultural Change and Environment Observatory Research Report No. 14. 17. Moore N, Whiterow A, Kelly P, Garthwaite D, Bishop J, Langton S, Cheesman C. 1999 Survey of badger Meles meles damage to agriculture in England and Wales. J. Appl. Ecol. 36, 974 –988. (doi:10.1046/j. 1365-2664.1999.00454.x) 18. Claridge J et al. 2012 Fasciola hepatica is associated with the failure to detect bovine tuberculosis in dairy cattle. Nat. Commun. 3, 853. (doi:10.1038/ ncomms1840)
5
Biol. Lett. 11: 20150536
13. Judge J, Wilson GJ, Macarthur R, Delahay RJ, McDonald RA. 2014 Density and abundance of badger social groups in England and Wales in 2011 –2013. Sci. Rep. 4, 3809. (doi:10.1038/ srep03809) 14. Newton-Cross G, White PCL, Harris S. 2007 Modelling the distribution of badgers Meles meles: comparing predictions from field-based and remotely derived habitat data. Mamm. Rev. 37, 54 –70. (doi:10.1111/j.1365-2907.2007. 00103.x) 15. Delahay RJ, Ward AI, Walker N, Long B, Cheeseman CL. 2007 Distribution of badger latrines in a highdensity population: habitat selection and implications for the transmission of bovine
rsbl.royalsocietypublishing.org
behavioral ecology: some background, observations, and comparisons. Behav. Ecol. Sociobiol. 65, 23– 35. (doi:10.1007/s00265-010-1029-6) 10. Johnston WT et al. 2011 Herd-level risk factors of bovine tuberculosis in England and Wales after the 2001 foot-and-mouth disease epidemic. Int. J. Infect. Dis. 15, e833 –e840. (doi:10.1016/j. ijid.2011.08.004) 11. Hosmer DW, Lemeshow S, Sturdivant RX. 2013 Applied logistic regression, 3rd edn. Hoboken, NJ: John Wiley & Sons. 12. Bunce RGH, Barr CJ, Whittaker HA. 1981 Land classes in Great Britain: preliminary descriptions for use of the Merlewood method of land classification. Grange-overSands, UK: Institute of Terrestrial Ecology.
