- Email: [email protected]
The 1997–1998 classical swine fever epidemic in The Netherlands — a survival analysis
Preventive Veterinary Medicine 42 (1999) 235±248
The 1997±1998 classical swine fever epidemic in The Netherlands Ð a survival analysis H.J. Benarda,*, K.D.C. StaÈrkb, R.S. Morrisb, D.U. Pfeifferb, H. Moserc a
National Centre for Disease Investigation, Ward Street, P.O. Box 40742, Upper Hutt, New Zealand b EpiCentre, Massey University, Private Bag 11222, Palmerston North, New Zealand c National Inspection Service for Livestock and Meat, Postbus 40010, 6504 AA Nijmegen, The Netherlands Accepted 21 June 1999
Abstract The aim of this analysis was to characterise the temporal pattern of infection during the 1997/98 classical swine fever (CSF) epidemic in The Netherlands and hence identify and quantify risk factors for infection in different enterprise types and areas. Survival analysis and Cox proportional hazards regression were used to describe the epidemic. Substantial differences in temporal survival patterns (herd breakdown rate) were found between areas where different control policies operated. Factors with a significant influence on the infection hazard of individual herds included: sow numbers as a percentage of total sows and fatteners (HR 3.38 for mixed herds (0.1±60% sows) vs. fattening herds (0% sows) and HR 2.74 for breeding herds (60±100% sows) vs. fattening herds), the number of `transport contacts per month' (>0.3 vs. <0.3; HR 4.11), pig density (pigs/ km2) in the area (HR1000 pigs 1.48) and herd size (HR100 pigs 1.01). Pre-emptive slaughter in an area appeared to be associated with lower subsequent disease levels. Higher frequency of transport contacts for welfare slaughter during the epidemic, however, well regulated and controlled, was associated with a substantially higher risk of becoming infected. The positive association of a higher pig density with CSF indicates the potential importance of local spread as a factor in disease transmission and emphasizes that dilution of the pig population can contribute to reduction in CSF occurrence. This analysis suggests however, that if pre-emptive slaughter can promptly be applied effectively in an area after initial diagnosis, pig density is then not a significant factor. Mixed and breeding herds had a higher probability of becoming infected than fattening herds, possibly due to different types and frequencies of inter-herd contacts. These contacts continue to some extent during the epidemic, despite the standstill of animal movements. # 1999 Elsevier Science B.V. All rights reserved. *
Corresponding author. Tel.: 64-4-526-5600; fax: 64-4-526-5601 E-mail address: [email protected] (H.J. Benard) 0167-5877/99/$ ± see front matter # 1999 Elsevier Science B.V. All rights reserved. PII: S 0 1 6 7 - 5 8 7 7 ( 9 9 ) 0 0 0 7 8 - 1
236
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
Keywords: Survival analysis; Cox proportional hazards regression; Classical swine fever; The Netherlands; Risk factors; Temporal pattern
1. Introduction On 4 February 1997, classical swine fever (CSF) was diagnosed on a farm in the village Venhorst, The Netherlands. A protection zone of 10-km radius surrounded by a surveillance zone of 2-km width was established around the index case. These measures exceeded the requirements of the European Union Council Directive 80/217/EEC and amendments, which require a protection zone of 3 km within a surveillance zone of 10 km. The additional measures were taken because of the high pig density in the area. A complete standstill of pig movements within these zones was applied, with the effect that transmission via introduced pigs and pig transport trucks was no longer occurring. These transmission mechanisms are considered important routes of infection between herds (Horst et al., 1996; Kramer et al., 1995). Early in the epidemic (20 February 1997), welfare slaughter measures were put in place to resolve overcrowding problems within the affected area. This resulted in a large number of movements over the subsequent months until mid-September 1997. In the area and period under study 18,532 such movements occurred. Pre-emptive slaughter was effectively enforced from May 1997 onwards. Due to limitations in rendering capacity, priorities were set based on working in from the edges of the control zones towards the centre of the outbreak area. This meant that around newly infected herds located in the centre of the epidemic area, pre-emptive slaughter was not used in the early stages of this control policy. The contribution to the risk of herd infection caused by stock-transport visits for welfare slaughter was unknown and has never been quantified before. Welfare-slaughter contacts are classified as a one-on-one contact (from one farm directly to the killing facility) and were carried out under supervision and with strict hygiene measures. More quantitative information with respect to the contribution to risk of infection was also needed for other risk factors, because, tracing the source of infection was not successful in many cases. The source of infection was often described as local or neighbourhood spread or unknown, as reported previously for other epidemics (Laevens et al., 1998; Saatkamp et al., 1996). The data collected during this epidemic allowed transmission risks to be explored in more depth. The aim of this study was to characterise the temporal pattern of infection during the 1997/98 CSF epidemic in The Netherlands and to identify and quantify risk factors for infection in different enterprise types and areas. 2. Material and methods The pattern of the CSF epidemic in The Netherlands was analysed for the period 4 February 1997±3 October 1997.
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
237
Table 1 Pig population and disease characteristics for the three different regions (A±C)a Variable
Region
2
Area size (km ) Total herds Infected herds Cumulative incidence Herds pre-emptively slaughtered All pigs removed (welfare slaughter) Herd density (herds/km2) Pig density (pigs/km2) Mean herd size (pigs) Onset epidemic/standstill
All
A
B
C
515 1467 266 0.18 343 514 2.8 3138 1102 4/2/97
648 1421 71 0.05 327 418 2.2 2667 1216 28/3/97
309 931 32 0.03 124 414 3.0 2624 871 12/2/97
1472 3819 369 0.10 794 1346 2.6 2822 1088 4/2/97
a
For the time period 4/2/97±3/10/97 affected by the 1997/98 CSF epidemic in The Netherlands.
The area of investigation was divided into three different regions to establish units with similar disease dynamics (onset of the epidemic, control measures applied, pig density, herd density, herd size and cumulative incidence; Table 1). It is possible that regional differences in population structure, herd density and pig density influence the spread of the disease (Saatkamp et al., 1996). Furthermore, control measures were not applied uniformly in different regions. A breeding ban (prohibition on mating of pigs) was in place from June 1997 in virtually the entire area for which the analysis was conducted. Culling of young piglets started in May 1997 and contributed to the reduction of the pig population. Each region was divided into several smaller areas, identified by block codes used in the control zones for serological screening purposes (Fig. 1). In total, 3819 herds were located in the area under examination. Of these, 9.7% became infected, 20.8% were pre-emptively slaughtered and 35.2% had all pigs removed during the period 4 February 1997±3 October 1997. Registered herds which were without pigs at the time (270 herds, 7.1%) were excluded. Herds that became censored or infected on or before day 43 of the epidemic were also excluded from this analysis, because most of these herds could have been exposed to infection prior to standstill measures being put in place. Day 43 was the date on which the last herd CSF diagnosis occurred which was traced to an infection date earlier than the diagnosis on the index farm. The definition of a case (event, non-survival) was the confirmed infection of a herd. If the event did not occur, the herd was considered to be censored. A censored herd was a herd present at the end of the interval or removed during the interval due to pre-emptive slaughter or removal of all pigs for welfare slaughter, and for which there was no evidence that it became infected. The survival time in days was calculated by subtracting the start date of the epidemic (4 February 1997) from the date the herd was lost to follow-up or the event occurred. The event date for infected herds was the date of laboratory diagnosis. The date of confirmed diagnosis was taken as the event date of interest, because under field circumstances this date was immediately available for analysis, whereas the true date of infection was not
238
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
Fig. 1. Map of the area involved in the 1997/98 CSF outbreak in The Netherlands, divided into three regions (A±C) and 16 blocks (C01-C13, N10, N12, N14). Locations of infected pig herds are indicated as dots on the map.
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
239
reliably available and estimation of the date of infection would have been speculative to a varying degree. 2.1. Putative risk factors Factors that may be associated with the risk of infection, for which data were available and evaluated in this study were: enterprise type, herd size, transport contacts for welfare slaughter and pig and herd density. Data was extracted from three different databases: the CSF outbreak database (PESTSQL), the Animal Health Service database for farm and herd information (RBD) and the database from LASER for transport contacts for welfare slaughter during the epidemic. 2.1.1. Pig and herd density The pig and herd densities were calculated by dividing the total number of pigs or herds in the different areas (block codes) by the area size (km2). These values were interpreted as a normally distributed continuous variable (pig density: mean 3089, median 2833, min. 1458, max. 4808). 2.1.2. Herd size The total number of sows, fatteners and replacement stock present at the start of the epidemic was used in this analysis to represent herd size. It was less subject to potential bias resulting from events arising from the epidemic such as increased herd size due to overcrowding of pigs, or a decrease in herd size due to lack of replacement stock, preemptive slaughter or infection of herds. It was assumed that these numbers would represent the actual places available rather than the number of pigs present at the time of recording. The total herd size was calculated as (number of fatteners) (number of replacement stock) (number of sows *6). The factor six incorporated the number of piglets and weaners on farm, calculated as five times the standing sow population. Herd size had a log-normal distribution (mean 1187, median 957, min. 1, max. 28 260). Herd structure was also evaluated by examining the number of sows and the number of fatteners within a pig herd separately. 2.1.3. Enterprise type Herds were categorised into different herd types according to the percentage of sows present, calculated as 100*sows/(number of sows number of fatteners). Because the distribution was trimodal, three groups were formed. A fattening herd was defined as one which had 0% sows (cumulative incidence (C.I.) 0.05); a breeding herd as one with 60± 100% sows (C.I. 0.13) and a mixed herd as 0.1±60% sows (C.I. 0.20). 2.1.4. Transport contacts for welfare slaughter The number of transport contacts for animal-welfare reasons was counted for each herd. On one day, several transport contacts could have taken place per herd, and each individual contact was counted. Because the time at risk was not equal for all herds (due to censoring or infection), the number of transport contacts was transformed to a transport
240
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
contact rate by dividing it by the number of months at risk. It was therefore controlled for the time of exposure. The variable was log-normally distributed (mean 0.24, median 0.19, min. 0, max. 4.5) and so data were grouped into two categories split at the third quartile (0.3 transport contacts per month). The first category had a C.I. of 0.06 and the second category 0.25. 2.2. Statistical analysis Survival analysis was performed in two steps, using univariable analysis first for screening (Kaplan-Meier survival curves) and then examining putative risk factors in a multivariable Cox proportional hazards model (Cox PH model). 2.2.1. Kaplan-Meier survival curves An estimation of the probability of herds becoming infected in different regions over different time periods was calculated. Risk factors that were potentially responsible for differences in probabilities were tested and their relative importance quantified. The survival function S(t) describes in this case the probability that a herd survived uninfected for longer than a specified time t (Kleinbaum, 1996). The probability of herds escaping infection was displayed graphically as a survival curve (SPSS V.8.0.0, SPSS, Chicago, USA). The data were stratified according to the values of the variables. The survival curves of these subgroups were compared using the log-rank test to detect the influence of the grouping variable (Kleinbaum, 1996). 2.2.2. Cox's proportional hazards regression model Continuous variables (herd size and pig density) that were categorised for the univariable analyses were entered as continuous variables in the multivariable model. For categorical variables, each was compared with a reference category. The Cox PH analysis requires the proportional hazards assumption to be met. This assumption was tested for all variables using two methods. First, it was determined if log(ÿlog) survival curves were approximately parallel for all levels of each categorical explanatory variable. Second, interaction terms between time and the variables under consideration were assessed for the statistical significance as part of the Cox model (Kleinbaum, 1996; Hess, 1995). The stratified Cox model was for variables where the proportional hazards assumption was not being met as a result of the above examination. It estimates different baseline hazard functions for each stratum; however, the coefficients are the same for each stratum. The fitted stratified Cox model will therefore yield different survival curves for each stratum (Kleinbaum, 1996). If two variables were highly collinear, the more biologically appropriate of the two was included in the model in order to obtain stable estimates of the coefficients. In the final models, biologically plausible one-way interaction terms were considered. A backward-stepwise selection in SPSS V.8.0.0 (SPSS, Chicago, USA) was used for variable selection in the Cox PH model. The criterion for removal of the variables from the model was P 0.1.
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
241
The likelihood-ratio test (ÿ2 log L) (also called ``deviance'') was used to compare the relative fit of the model with a model containing only an intercept. The hazard of the continuous variables pig density and herd size was converted into hazard ratios and their 95% confidence intervals as described by Hosmer and Lemeshow (1989). 3. Results 3.1. Survival analysis 3.1.1. Temporal pattern of the epidemic in different areas The Kaplan-Meier survival curve for the whole observation period and all regions showed an acceleration of the breakdown rate from about day 60 onwards and a marked flattening (reduction of breakdown rate) from approximately day 170 onwards (Fig. 2a). When the survival curve over the whole region was split up into three regions, a significant difference could be observed between the levels of disease in region A vs. region B (log-rank test statistic 119.78, 1 df, P < 0.00001; Fig. 2a) and also vs. region C. The pattern of disease in region A was further investigated by plotting the survival curve per block. In Fig. 3, region A was stratified by block and significant differences in breakdown probability between block C03 and the rest was observed. Between block C03 and the closest group (block C02), the log-rank test statistic was 35.97 (1 df, P < 0.00001). Herds are displayed as crosses on the survival curve at the date when they became censored. Reasons for censoring were pre-emptive or welfare slaughter of herds. Although block C03 had a high infection rate, censoring was not as extensive as in the other blocks and occurred later. The predominant reason for this was that pre-emptive slaughter did not take place to the extent it did in other blocks. 3.1.2. Putative risk factors for infection 3.1.2.1. Pig and herd density. Survival curves (Fig. 4a) for pig density showed that the probability of becoming infected increased at higher pig densities per km2. Herds in areas with pig densities over 3704 pigs per km2 had a significantly higher chance of becoming infected (log-rank test statistic 90.42, 1 df, P < 0.00001) than in areas with densities up to 3205 pigs per km2 (categories 1 and 2 collapsed). When excluding block C03 from the analysis (very high pig density (4713 pigs/km2) and cumulative incidence (0.35)), the highest pig densities (category 3) were still more at risk than the lower densities (categories 1 and 2) (log-rank test statistic 8.36, 1 df, P < 0.0038). The same trend as for pig density was visible when examining herd density (data not shown). 3.1.2.2. Herd size. Small herds (category 1, 500 pigs) had a relatively low breakdown rate (Fig. 4b). Herds with a total number of pigs of 500±1000 (category 2) had a significantly higher risk of breaking down than category 1 herds (log-rank test statistic
242
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
Fig. 2. Unadjusted (a) and adjusted (b) Kaplan-Meier survival curves for pig herds in the study area of the 1997/ 98 CSF epidemic in The Netherlands, herds stratified as follows: (a) all herds and split per region; (b) stratified per region.
18.77, 1 df, P < 0.00001), but a significantly lower risk of breaking down than herds with more than 1000 pigs (categories 3±5) (log-rank test statistic 21.70, 1 df, P < 0.00001). When the number of sows alone was evaluated for each farm, a similar trend of an increase in breakdown rate with increasing sow numbers was seen (data not shown). The effect of the number of fatteners present on survival displayed no clear influence of herd size (data not shown).
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
243
Fig. 3. Kaplan-Meier survival curves for pig herds in the study area of the 1997/98 CSF epidemic in The Netherlands, stratified per block in region A (censored herds, ).
3.1.2.3. Enterprise type. In the survival curves based on number of sows as a percentage of total sows and fatteners (Fig. 4c) a clear difference between all categories was observed. Predominantly breeding herds (sow percentage 60±100) had a significantly lower risk of breaking down with CSF than mixed herds (sow percentage 0.1±60) (logrank test statistic 8.34, 1 df, P 0.039) and had a significantly higher risk than fattening herds (sow percentage 0) (log-rank test statistic 71.55, 1 df, P 0.00001). 3.1.2.4. Transport contacts per month. The survival analysis of the risk factor `transport contacts per month' is displayed in Fig. 4d. Herds with a high number of `transport contacts per month' (>0.3) have a higher risk of infection than herds with a low number (<0.3) (log-rank test statistic 208.77, 1 df, P < 0.00001). 3.2. Cox's proportional hazards regression Although some variables had categories with only few observations, overall there were no obvious violations of the proportional hazards assumption in most respects. For the variable `region' however, the proportional hazards assumption did not hold for the entire period of observation. Therefore, different models were built and compared: Model I included the variable `region', disregarding problems with a low number of observations early in the epidemic in regions B and C. Model II was stratified on the variable `region' and no estimation of the risks associated with the regions can therefore be made. Third, an overall analysis excluding `region' from the model was run (Model III). All models included pig density, enterprise type, herd size (total pigs), `transport contacts per month' and the interaction term (herd size) * (pig density). In all three models, the one-way interaction terms dropped out after the first step. Herd size, pig density, enterprise type, `transport contacts per month' and the variable `region' were significant in all models and removing any one of them resulted in a significant increase
244 H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248 Fig. 4. Kaplan-Meier survival curves for pig herds in the study area of the 1997/98 CSF epidemic in The Netherlands, for four risk factors, stratified as follows: (a) pig density (pigs/km2): (1) 1458±2463, (2) 2700±3205, (3) 3704±4808; (b) herd size (pigs): (1) 1±500, (2) 501±1000, (3) 1001±1500, (4) 1501±2000, (5) >2000; (c) enterprise type based on sow numbers as a percentage of total sows and fatteners (%): (1) fattening herds (0), (2) mixed herds (0.1±60), (3) breeding herds (>60); (d) `transport contacts per month': (1) <0.3, (2) >0.3.
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
245
of the model deviance. The results from the stratified Cox PH model (II) are shown in Table 2. The results of Model I and III were similar (data not shown). When the area (block C03) where pre-emptive slaughter was not applied as rigorously as in the other areas was excluded from the analysis, pig density and the one-way interaction term were the only variables which were removed from the stratified Cox model (data not shown). Model III was run for each region individually. These within-region-models were each run with forced entry of the variables under study, excluding region, to be able to compare the influence of these variables within the regions. The Cox PH model for region A compared well with Model III, as demonstrated by similar size of all coefficients (data not shown). The only variables in the Cox PH model for B and C that were still significant were enterprise type and `transport contacts per month'. 3.3. Adjusted survival curves Fig. 2b shows the adjusted survival curve stratified by region for Model II. When comparing Fig. 2b with Fig. 2a, it can be seen that the breakdown rate for region A after controlling for the other variables, was reduced but was still substantially higher than in other regions. Regions B and C had very similar breakdown rates. 4. Discussion Factors with a significant influence on the infection hazard of individual herds included: sow numbers as a percentage of total sows and fatteners (HR 3.38 for mixed herds (0.1±60% sows) vs. fattening herds (0% sows) and HR 2.74 for breeding herds (60±100% sows) vs. fattening herds, the number of `transport contacts per month' (>0.3 vs. <0.3) (HR 4.11), pig density (pigs/km2) in the area (HR1000 pigs 1.48) and herd size (HR100 pigs 1.01). Pre-emptive slaughter appeared to be associated with lower subsequent disease levels. This can be inferred from comparing the occurrence of censored and breakdown herds on the survival curves of Block C03 with surrounding blocks in the univariable analysis (Fig. 3) and the differences in slope of survival curves between the regions (Fig. 2a) from day 170 onwards. Pre-emptive slaughter was rigorously enforced in May and June and this would have had its effect from 3 to 4 weeks after implementation. Regions B and C were on the outskirts of the core epidemic area and pre-emptive slaughter was enforced with higher priority in these than in the core area, which falls in region A. Other authors have suggested that pre-emptive slaughter can make a substantial difference in determining the extent of the epidemic in a high pig density area (Anonymous, 1998; Elbers et al., 1998). Saatkamp et al. (1996) concluded from simulation results that in regions with high pig densities, pre-emptive slaughter was necessary to achieve eradication of CSF. Herd size also is an influential factor, as previously shown (Terpstra, 1987; Lorenz and Boorberg, 1986; Vannier et al., 1986). It appears that the number of sows was of greater importance than the number of fatteners. This was supported by the fact that breeding and
246
Explanatory variables
Regression coefficient b
SE
Wald statistic
df
P-value
Hazard ratio (HR) exp( )
95% Confidence interval for HR
%Sows/sows and fatteners: 0.1±60% vs. 0% %Sows/sows and fatteners: 60±100% vs. 0% Transports/month: >0.3 vs. <0.3 Pig density/km2 (*1000 pigs) Herd size (*100 pigs)
1.217 1.007 1.412 0.392 0.006
0.148 0.153 0.121 0.060 0.003
67.44 43.09 136.80 42.58 3.33
1 1 1 1 1
0.00 0.00 0.00 0.00 0.07
3.38 2.74 4.11 1.48 1.01
2.53±4.52 2.03±3.70 3.24±5.20 1.32±1.66 1.00±1.01
a
n: 3372 (321 events) 90.5% censored; ÿ2 log L: 4124.29; (c2 271.08, 5 df, P < 0.00001); R2: 6.2%.
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
Table 2 Results of the stratified Cox proportional-hazard regression of the infection hazard for pig herds in the study area affected by the 1997/98 CSF epidemic in The Netherlandsa
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
247
mixed herds had a higher probability of becoming infected than fattening herds, possibly due to the different types and frequencies of inter-herd contacts associated with a different management system. Several sources mention fattening herds as the herds most at risk in other epidemics (Kramer et al., 1995; Lorenz and Boorberg, 1986), but this does not seem to have been true in this epidemic. Larger fattening herds might have multiple animal movements that could increase the risk of infection if such movements occur. However, when an area has been under movement restrictions for months (as in this epidemic), indirect person contacts may become a more-important transmission route, resulting in a higher risk for mixed and breeding herds (as found in this study). The Cox model showed that `transport contacts per month' had a significant effect on the spread of disease. Some farms with movements for welfare slaughter were later detected as infected through blood samples taken immediately prior to collection of pigs. Transport contacts for welfare slaughter during the epidemic, however well regulated and controlled, were associated with substantially higher risk of becoming infected and should therefore only be implemented if strictly necessary. In this case, there was strict supervision and rigorous hygiene measures, but additional protective steps may have to be considered. The timing for welfare slaughter to take place should be balanced between animal-welfare issues (overcrowding, overweight), rendering capacity and alternatives for market support (as discussed by Vanthemsche, 1996). From the analysis of the within-region models and the analysis excluding the core of the epidemic, the conclusion can be drawn that pig density is of more importance when pre-emptive slaughter cannot be effectively applied from the start. The EU report on the CSF epidemic concluded that it was difficult to control an outbreak of CSF in a highdensity pig area (Anonymous, 1998). Stegeman et al. (1998) found that local spread was an important factor in disease transmission, which emphasizes that dilution of the pig and herd population can contribute to the reduction in disease occurrence. The Dutch Minister of Agriculture (Van Aartsen, 1997) has proposed a reduction of the pig population by 25% (a decision that could help reduce the risk of spread of CSF). Indications from the current analysis are that `transport contacts per month' and enterprise type would remain as significant risk factors (even if pre-emptive slaughter is initiated immediately at the onset of the epidemic), but pig density would not. This could be related to breeding or mixed herds having far more contacts through persons than fattening herds. Some of these contacts will continue during the epidemic, despite the standstill of animal movements. The pseudo R-square remained relatively small in all final models. This is an indication that there are other unmeasured variables, which are important for explaining variation of CSF infection risk for individual farms. Acknowledgements The authors are grateful to the National Inspection Service for Livestock and Meat (RVV), The Netherlands for offering the opportunity for this analysis. We acknowledge the assistance of the tracing and epidemiology team operating in the Crisis Centre during this epidemic for making data available. We also thank the Ministry of Agriculture and
248
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
Forestry and AgriQuality New Zealand for making the practical work experience and data analysis possible. The EpiCentre, Massey University, New Zealand contributed through valuable discussions with Mark Bosson. We thank the referees for their constructive comments. References Anonymous, 1998. Veterinary Inspections: Final Report on a Mission Concerning the Evolution of Classical Swine Fever in The Netherlands. 23±27 March 1998, EU report on Netherlands swine fever epidemic. Elbers, A.R.W., Stegeman, J.A., Moser, H., De Jong, M.C.M., Ekker, H.M., Smak, J.A., De Leeuw, P.W., 1998. Effectiveness of preventive culling of pig herds during the Dutch CSF epidemic in 1997. In: Proceedings of the 15th IPVS Congress, Birmingham, p. 271. Hess, K.R., 1995. Graphical methods for assessing violations of the proportional hazards assumption in Cox regression. Statistics in Medicine 14, 1707±1723. Horst, H.S., Dijkhuizen, A.A., Huirne, R.B.M., Steenkamp, J-B.E.M., Thrusfield, M.V., Goodall, E.A., 1996. Eliciting expert opinions on risk factors concerning introduction of virus: application of conjoint analysis. In: Proceedings of the Society for Veterinary Epidemiology and Preventive Medicine, Glasgow, pp. 8±17. Hosmer, D.W., Lemeshow, S., 1989. Applied Logistic Regression. Wiley, New York. Kleinbaum, D.G., 1996. Survival Analysis: A Self-Learning Text. Springer, New York. Kramer, M., Ahl, R., Teuffert, J., Kroschewski, K., SchluÈter, H., Otte, J., 1995. Classical Swine Fever in Germany Ð some epidemiological aspects. In: Proceedings of the Society for Veterinary Epidemiology and Preventive Medicine, Reading, pp. 110±118. Laevens, H., Deluyker, H., Koenen, F., Van Caenegem, G., Vermeersch, J.P., De Kruif, A., 1998. An experimental infection with a Classical Swine Fever virus in weaner pigs. II. The use of serological data to estimate the day of virus introduction in natural outbreaks. Vet. Quarterly 20(2), 46±49. Lorenz, R.J., Boorberg, I., 1986. Incidence rate of hog cholera infection in various categories of pig herds. In: Proceedings of the Fourth International Symposium on Veterinary Epidemiology and Economics, Singapore, pp. 233±235. Saatkamp, H.W., Dijkhuizen, A.A., Geers, R., Huirne, R.B.M., Noordhuizen, J.P.T.M., Goedseels, V., 1996. Simulation studies on the epidemiological impact of national identification and recording systems on the control of Classical Swine Fever in Belgium. Prev. Vet. Med. 26, 119±132. Stegeman, A., Elbers, A.R.W., De Smit, H., Bouma, A., De Jong, M.C.M., 1998. Rate of transmission of Classical Swine Fever virus between herds by various routes. In: Proceedings of the 15th IPVS Congress, Birmingham, p. 269. Terpstra, C., 1987. Epizootiology of Swine Fever. Vet. Quarterly 9 (Suppl. 1), 50S±60S. Van Aartsen, J.J., 1997. Letter to the Second Chamber of Parliament from the Minister of Agriculture. Nature Management and Fisheries. 10 July 1997, The Hague, The Netherlands. Vannier, P., Colcanap, M., Carnero, R., Costes, C., Josse, J., Tillon, J.P., 1986. Study of the origin of an epizootic of Classical Swine Fever. J. Vet. Medicine B 33(4), 294±302. Vanthemsche, P., 1996. Classical Swine Fever 1993±1994 Belgium. The Pig J. 37, 43±53.
The 1997±1998 classical swine fever epidemic in The Netherlands Ð a survival analysis H.J. Benarda,*, K.D.C. StaÈrkb, R.S. Morrisb, D.U. Pfeifferb, H. Moserc a
National Centre for Disease Investigation, Ward Street, P.O. Box 40742, Upper Hutt, New Zealand b EpiCentre, Massey University, Private Bag 11222, Palmerston North, New Zealand c National Inspection Service for Livestock and Meat, Postbus 40010, 6504 AA Nijmegen, The Netherlands Accepted 21 June 1999
Abstract The aim of this analysis was to characterise the temporal pattern of infection during the 1997/98 classical swine fever (CSF) epidemic in The Netherlands and hence identify and quantify risk factors for infection in different enterprise types and areas. Survival analysis and Cox proportional hazards regression were used to describe the epidemic. Substantial differences in temporal survival patterns (herd breakdown rate) were found between areas where different control policies operated. Factors with a significant influence on the infection hazard of individual herds included: sow numbers as a percentage of total sows and fatteners (HR 3.38 for mixed herds (0.1±60% sows) vs. fattening herds (0% sows) and HR 2.74 for breeding herds (60±100% sows) vs. fattening herds), the number of `transport contacts per month' (>0.3 vs. <0.3; HR 4.11), pig density (pigs/ km2) in the area (HR1000 pigs 1.48) and herd size (HR100 pigs 1.01). Pre-emptive slaughter in an area appeared to be associated with lower subsequent disease levels. Higher frequency of transport contacts for welfare slaughter during the epidemic, however, well regulated and controlled, was associated with a substantially higher risk of becoming infected. The positive association of a higher pig density with CSF indicates the potential importance of local spread as a factor in disease transmission and emphasizes that dilution of the pig population can contribute to reduction in CSF occurrence. This analysis suggests however, that if pre-emptive slaughter can promptly be applied effectively in an area after initial diagnosis, pig density is then not a significant factor. Mixed and breeding herds had a higher probability of becoming infected than fattening herds, possibly due to different types and frequencies of inter-herd contacts. These contacts continue to some extent during the epidemic, despite the standstill of animal movements. # 1999 Elsevier Science B.V. All rights reserved. *
Corresponding author. Tel.: 64-4-526-5600; fax: 64-4-526-5601 E-mail address: [email protected] (H.J. Benard) 0167-5877/99/$ ± see front matter # 1999 Elsevier Science B.V. All rights reserved. PII: S 0 1 6 7 - 5 8 7 7 ( 9 9 ) 0 0 0 7 8 - 1
236
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
Keywords: Survival analysis; Cox proportional hazards regression; Classical swine fever; The Netherlands; Risk factors; Temporal pattern
1. Introduction On 4 February 1997, classical swine fever (CSF) was diagnosed on a farm in the village Venhorst, The Netherlands. A protection zone of 10-km radius surrounded by a surveillance zone of 2-km width was established around the index case. These measures exceeded the requirements of the European Union Council Directive 80/217/EEC and amendments, which require a protection zone of 3 km within a surveillance zone of 10 km. The additional measures were taken because of the high pig density in the area. A complete standstill of pig movements within these zones was applied, with the effect that transmission via introduced pigs and pig transport trucks was no longer occurring. These transmission mechanisms are considered important routes of infection between herds (Horst et al., 1996; Kramer et al., 1995). Early in the epidemic (20 February 1997), welfare slaughter measures were put in place to resolve overcrowding problems within the affected area. This resulted in a large number of movements over the subsequent months until mid-September 1997. In the area and period under study 18,532 such movements occurred. Pre-emptive slaughter was effectively enforced from May 1997 onwards. Due to limitations in rendering capacity, priorities were set based on working in from the edges of the control zones towards the centre of the outbreak area. This meant that around newly infected herds located in the centre of the epidemic area, pre-emptive slaughter was not used in the early stages of this control policy. The contribution to the risk of herd infection caused by stock-transport visits for welfare slaughter was unknown and has never been quantified before. Welfare-slaughter contacts are classified as a one-on-one contact (from one farm directly to the killing facility) and were carried out under supervision and with strict hygiene measures. More quantitative information with respect to the contribution to risk of infection was also needed for other risk factors, because, tracing the source of infection was not successful in many cases. The source of infection was often described as local or neighbourhood spread or unknown, as reported previously for other epidemics (Laevens et al., 1998; Saatkamp et al., 1996). The data collected during this epidemic allowed transmission risks to be explored in more depth. The aim of this study was to characterise the temporal pattern of infection during the 1997/98 CSF epidemic in The Netherlands and to identify and quantify risk factors for infection in different enterprise types and areas. 2. Material and methods The pattern of the CSF epidemic in The Netherlands was analysed for the period 4 February 1997±3 October 1997.
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
237
Table 1 Pig population and disease characteristics for the three different regions (A±C)a Variable
Region
2
Area size (km ) Total herds Infected herds Cumulative incidence Herds pre-emptively slaughtered All pigs removed (welfare slaughter) Herd density (herds/km2) Pig density (pigs/km2) Mean herd size (pigs) Onset epidemic/standstill
All
A
B
C
515 1467 266 0.18 343 514 2.8 3138 1102 4/2/97
648 1421 71 0.05 327 418 2.2 2667 1216 28/3/97
309 931 32 0.03 124 414 3.0 2624 871 12/2/97
1472 3819 369 0.10 794 1346 2.6 2822 1088 4/2/97
a
For the time period 4/2/97±3/10/97 affected by the 1997/98 CSF epidemic in The Netherlands.
The area of investigation was divided into three different regions to establish units with similar disease dynamics (onset of the epidemic, control measures applied, pig density, herd density, herd size and cumulative incidence; Table 1). It is possible that regional differences in population structure, herd density and pig density influence the spread of the disease (Saatkamp et al., 1996). Furthermore, control measures were not applied uniformly in different regions. A breeding ban (prohibition on mating of pigs) was in place from June 1997 in virtually the entire area for which the analysis was conducted. Culling of young piglets started in May 1997 and contributed to the reduction of the pig population. Each region was divided into several smaller areas, identified by block codes used in the control zones for serological screening purposes (Fig. 1). In total, 3819 herds were located in the area under examination. Of these, 9.7% became infected, 20.8% were pre-emptively slaughtered and 35.2% had all pigs removed during the period 4 February 1997±3 October 1997. Registered herds which were without pigs at the time (270 herds, 7.1%) were excluded. Herds that became censored or infected on or before day 43 of the epidemic were also excluded from this analysis, because most of these herds could have been exposed to infection prior to standstill measures being put in place. Day 43 was the date on which the last herd CSF diagnosis occurred which was traced to an infection date earlier than the diagnosis on the index farm. The definition of a case (event, non-survival) was the confirmed infection of a herd. If the event did not occur, the herd was considered to be censored. A censored herd was a herd present at the end of the interval or removed during the interval due to pre-emptive slaughter or removal of all pigs for welfare slaughter, and for which there was no evidence that it became infected. The survival time in days was calculated by subtracting the start date of the epidemic (4 February 1997) from the date the herd was lost to follow-up or the event occurred. The event date for infected herds was the date of laboratory diagnosis. The date of confirmed diagnosis was taken as the event date of interest, because under field circumstances this date was immediately available for analysis, whereas the true date of infection was not
238
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
Fig. 1. Map of the area involved in the 1997/98 CSF outbreak in The Netherlands, divided into three regions (A±C) and 16 blocks (C01-C13, N10, N12, N14). Locations of infected pig herds are indicated as dots on the map.
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
239
reliably available and estimation of the date of infection would have been speculative to a varying degree. 2.1. Putative risk factors Factors that may be associated with the risk of infection, for which data were available and evaluated in this study were: enterprise type, herd size, transport contacts for welfare slaughter and pig and herd density. Data was extracted from three different databases: the CSF outbreak database (PESTSQL), the Animal Health Service database for farm and herd information (RBD) and the database from LASER for transport contacts for welfare slaughter during the epidemic. 2.1.1. Pig and herd density The pig and herd densities were calculated by dividing the total number of pigs or herds in the different areas (block codes) by the area size (km2). These values were interpreted as a normally distributed continuous variable (pig density: mean 3089, median 2833, min. 1458, max. 4808). 2.1.2. Herd size The total number of sows, fatteners and replacement stock present at the start of the epidemic was used in this analysis to represent herd size. It was less subject to potential bias resulting from events arising from the epidemic such as increased herd size due to overcrowding of pigs, or a decrease in herd size due to lack of replacement stock, preemptive slaughter or infection of herds. It was assumed that these numbers would represent the actual places available rather than the number of pigs present at the time of recording. The total herd size was calculated as (number of fatteners) (number of replacement stock) (number of sows *6). The factor six incorporated the number of piglets and weaners on farm, calculated as five times the standing sow population. Herd size had a log-normal distribution (mean 1187, median 957, min. 1, max. 28 260). Herd structure was also evaluated by examining the number of sows and the number of fatteners within a pig herd separately. 2.1.3. Enterprise type Herds were categorised into different herd types according to the percentage of sows present, calculated as 100*sows/(number of sows number of fatteners). Because the distribution was trimodal, three groups were formed. A fattening herd was defined as one which had 0% sows (cumulative incidence (C.I.) 0.05); a breeding herd as one with 60± 100% sows (C.I. 0.13) and a mixed herd as 0.1±60% sows (C.I. 0.20). 2.1.4. Transport contacts for welfare slaughter The number of transport contacts for animal-welfare reasons was counted for each herd. On one day, several transport contacts could have taken place per herd, and each individual contact was counted. Because the time at risk was not equal for all herds (due to censoring or infection), the number of transport contacts was transformed to a transport
240
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
contact rate by dividing it by the number of months at risk. It was therefore controlled for the time of exposure. The variable was log-normally distributed (mean 0.24, median 0.19, min. 0, max. 4.5) and so data were grouped into two categories split at the third quartile (0.3 transport contacts per month). The first category had a C.I. of 0.06 and the second category 0.25. 2.2. Statistical analysis Survival analysis was performed in two steps, using univariable analysis first for screening (Kaplan-Meier survival curves) and then examining putative risk factors in a multivariable Cox proportional hazards model (Cox PH model). 2.2.1. Kaplan-Meier survival curves An estimation of the probability of herds becoming infected in different regions over different time periods was calculated. Risk factors that were potentially responsible for differences in probabilities were tested and their relative importance quantified. The survival function S(t) describes in this case the probability that a herd survived uninfected for longer than a specified time t (Kleinbaum, 1996). The probability of herds escaping infection was displayed graphically as a survival curve (SPSS V.8.0.0, SPSS, Chicago, USA). The data were stratified according to the values of the variables. The survival curves of these subgroups were compared using the log-rank test to detect the influence of the grouping variable (Kleinbaum, 1996). 2.2.2. Cox's proportional hazards regression model Continuous variables (herd size and pig density) that were categorised for the univariable analyses were entered as continuous variables in the multivariable model. For categorical variables, each was compared with a reference category. The Cox PH analysis requires the proportional hazards assumption to be met. This assumption was tested for all variables using two methods. First, it was determined if log(ÿlog) survival curves were approximately parallel for all levels of each categorical explanatory variable. Second, interaction terms between time and the variables under consideration were assessed for the statistical significance as part of the Cox model (Kleinbaum, 1996; Hess, 1995). The stratified Cox model was for variables where the proportional hazards assumption was not being met as a result of the above examination. It estimates different baseline hazard functions for each stratum; however, the coefficients are the same for each stratum. The fitted stratified Cox model will therefore yield different survival curves for each stratum (Kleinbaum, 1996). If two variables were highly collinear, the more biologically appropriate of the two was included in the model in order to obtain stable estimates of the coefficients. In the final models, biologically plausible one-way interaction terms were considered. A backward-stepwise selection in SPSS V.8.0.0 (SPSS, Chicago, USA) was used for variable selection in the Cox PH model. The criterion for removal of the variables from the model was P 0.1.
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
241
The likelihood-ratio test (ÿ2 log L) (also called ``deviance'') was used to compare the relative fit of the model with a model containing only an intercept. The hazard of the continuous variables pig density and herd size was converted into hazard ratios and their 95% confidence intervals as described by Hosmer and Lemeshow (1989). 3. Results 3.1. Survival analysis 3.1.1. Temporal pattern of the epidemic in different areas The Kaplan-Meier survival curve for the whole observation period and all regions showed an acceleration of the breakdown rate from about day 60 onwards and a marked flattening (reduction of breakdown rate) from approximately day 170 onwards (Fig. 2a). When the survival curve over the whole region was split up into three regions, a significant difference could be observed between the levels of disease in region A vs. region B (log-rank test statistic 119.78, 1 df, P < 0.00001; Fig. 2a) and also vs. region C. The pattern of disease in region A was further investigated by plotting the survival curve per block. In Fig. 3, region A was stratified by block and significant differences in breakdown probability between block C03 and the rest was observed. Between block C03 and the closest group (block C02), the log-rank test statistic was 35.97 (1 df, P < 0.00001). Herds are displayed as crosses on the survival curve at the date when they became censored. Reasons for censoring were pre-emptive or welfare slaughter of herds. Although block C03 had a high infection rate, censoring was not as extensive as in the other blocks and occurred later. The predominant reason for this was that pre-emptive slaughter did not take place to the extent it did in other blocks. 3.1.2. Putative risk factors for infection 3.1.2.1. Pig and herd density. Survival curves (Fig. 4a) for pig density showed that the probability of becoming infected increased at higher pig densities per km2. Herds in areas with pig densities over 3704 pigs per km2 had a significantly higher chance of becoming infected (log-rank test statistic 90.42, 1 df, P < 0.00001) than in areas with densities up to 3205 pigs per km2 (categories 1 and 2 collapsed). When excluding block C03 from the analysis (very high pig density (4713 pigs/km2) and cumulative incidence (0.35)), the highest pig densities (category 3) were still more at risk than the lower densities (categories 1 and 2) (log-rank test statistic 8.36, 1 df, P < 0.0038). The same trend as for pig density was visible when examining herd density (data not shown). 3.1.2.2. Herd size. Small herds (category 1, 500 pigs) had a relatively low breakdown rate (Fig. 4b). Herds with a total number of pigs of 500±1000 (category 2) had a significantly higher risk of breaking down than category 1 herds (log-rank test statistic
242
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
Fig. 2. Unadjusted (a) and adjusted (b) Kaplan-Meier survival curves for pig herds in the study area of the 1997/ 98 CSF epidemic in The Netherlands, herds stratified as follows: (a) all herds and split per region; (b) stratified per region.
18.77, 1 df, P < 0.00001), but a significantly lower risk of breaking down than herds with more than 1000 pigs (categories 3±5) (log-rank test statistic 21.70, 1 df, P < 0.00001). When the number of sows alone was evaluated for each farm, a similar trend of an increase in breakdown rate with increasing sow numbers was seen (data not shown). The effect of the number of fatteners present on survival displayed no clear influence of herd size (data not shown).
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
243
Fig. 3. Kaplan-Meier survival curves for pig herds in the study area of the 1997/98 CSF epidemic in The Netherlands, stratified per block in region A (censored herds, ).
3.1.2.3. Enterprise type. In the survival curves based on number of sows as a percentage of total sows and fatteners (Fig. 4c) a clear difference between all categories was observed. Predominantly breeding herds (sow percentage 60±100) had a significantly lower risk of breaking down with CSF than mixed herds (sow percentage 0.1±60) (logrank test statistic 8.34, 1 df, P 0.039) and had a significantly higher risk than fattening herds (sow percentage 0) (log-rank test statistic 71.55, 1 df, P 0.00001). 3.1.2.4. Transport contacts per month. The survival analysis of the risk factor `transport contacts per month' is displayed in Fig. 4d. Herds with a high number of `transport contacts per month' (>0.3) have a higher risk of infection than herds with a low number (<0.3) (log-rank test statistic 208.77, 1 df, P < 0.00001). 3.2. Cox's proportional hazards regression Although some variables had categories with only few observations, overall there were no obvious violations of the proportional hazards assumption in most respects. For the variable `region' however, the proportional hazards assumption did not hold for the entire period of observation. Therefore, different models were built and compared: Model I included the variable `region', disregarding problems with a low number of observations early in the epidemic in regions B and C. Model II was stratified on the variable `region' and no estimation of the risks associated with the regions can therefore be made. Third, an overall analysis excluding `region' from the model was run (Model III). All models included pig density, enterprise type, herd size (total pigs), `transport contacts per month' and the interaction term (herd size) * (pig density). In all three models, the one-way interaction terms dropped out after the first step. Herd size, pig density, enterprise type, `transport contacts per month' and the variable `region' were significant in all models and removing any one of them resulted in a significant increase
244 H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248 Fig. 4. Kaplan-Meier survival curves for pig herds in the study area of the 1997/98 CSF epidemic in The Netherlands, for four risk factors, stratified as follows: (a) pig density (pigs/km2): (1) 1458±2463, (2) 2700±3205, (3) 3704±4808; (b) herd size (pigs): (1) 1±500, (2) 501±1000, (3) 1001±1500, (4) 1501±2000, (5) >2000; (c) enterprise type based on sow numbers as a percentage of total sows and fatteners (%): (1) fattening herds (0), (2) mixed herds (0.1±60), (3) breeding herds (>60); (d) `transport contacts per month': (1) <0.3, (2) >0.3.
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
245
of the model deviance. The results from the stratified Cox PH model (II) are shown in Table 2. The results of Model I and III were similar (data not shown). When the area (block C03) where pre-emptive slaughter was not applied as rigorously as in the other areas was excluded from the analysis, pig density and the one-way interaction term were the only variables which were removed from the stratified Cox model (data not shown). Model III was run for each region individually. These within-region-models were each run with forced entry of the variables under study, excluding region, to be able to compare the influence of these variables within the regions. The Cox PH model for region A compared well with Model III, as demonstrated by similar size of all coefficients (data not shown). The only variables in the Cox PH model for B and C that were still significant were enterprise type and `transport contacts per month'. 3.3. Adjusted survival curves Fig. 2b shows the adjusted survival curve stratified by region for Model II. When comparing Fig. 2b with Fig. 2a, it can be seen that the breakdown rate for region A after controlling for the other variables, was reduced but was still substantially higher than in other regions. Regions B and C had very similar breakdown rates. 4. Discussion Factors with a significant influence on the infection hazard of individual herds included: sow numbers as a percentage of total sows and fatteners (HR 3.38 for mixed herds (0.1±60% sows) vs. fattening herds (0% sows) and HR 2.74 for breeding herds (60±100% sows) vs. fattening herds, the number of `transport contacts per month' (>0.3 vs. <0.3) (HR 4.11), pig density (pigs/km2) in the area (HR1000 pigs 1.48) and herd size (HR100 pigs 1.01). Pre-emptive slaughter appeared to be associated with lower subsequent disease levels. This can be inferred from comparing the occurrence of censored and breakdown herds on the survival curves of Block C03 with surrounding blocks in the univariable analysis (Fig. 3) and the differences in slope of survival curves between the regions (Fig. 2a) from day 170 onwards. Pre-emptive slaughter was rigorously enforced in May and June and this would have had its effect from 3 to 4 weeks after implementation. Regions B and C were on the outskirts of the core epidemic area and pre-emptive slaughter was enforced with higher priority in these than in the core area, which falls in region A. Other authors have suggested that pre-emptive slaughter can make a substantial difference in determining the extent of the epidemic in a high pig density area (Anonymous, 1998; Elbers et al., 1998). Saatkamp et al. (1996) concluded from simulation results that in regions with high pig densities, pre-emptive slaughter was necessary to achieve eradication of CSF. Herd size also is an influential factor, as previously shown (Terpstra, 1987; Lorenz and Boorberg, 1986; Vannier et al., 1986). It appears that the number of sows was of greater importance than the number of fatteners. This was supported by the fact that breeding and
246
Explanatory variables
Regression coefficient b
SE
Wald statistic
df
P-value
Hazard ratio (HR) exp( )
95% Confidence interval for HR
%Sows/sows and fatteners: 0.1±60% vs. 0% %Sows/sows and fatteners: 60±100% vs. 0% Transports/month: >0.3 vs. <0.3 Pig density/km2 (*1000 pigs) Herd size (*100 pigs)
1.217 1.007 1.412 0.392 0.006
0.148 0.153 0.121 0.060 0.003
67.44 43.09 136.80 42.58 3.33
1 1 1 1 1
0.00 0.00 0.00 0.00 0.07
3.38 2.74 4.11 1.48 1.01
2.53±4.52 2.03±3.70 3.24±5.20 1.32±1.66 1.00±1.01
a
n: 3372 (321 events) 90.5% censored; ÿ2 log L: 4124.29; (c2 271.08, 5 df, P < 0.00001); R2: 6.2%.
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
Table 2 Results of the stratified Cox proportional-hazard regression of the infection hazard for pig herds in the study area affected by the 1997/98 CSF epidemic in The Netherlandsa
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
247
mixed herds had a higher probability of becoming infected than fattening herds, possibly due to the different types and frequencies of inter-herd contacts associated with a different management system. Several sources mention fattening herds as the herds most at risk in other epidemics (Kramer et al., 1995; Lorenz and Boorberg, 1986), but this does not seem to have been true in this epidemic. Larger fattening herds might have multiple animal movements that could increase the risk of infection if such movements occur. However, when an area has been under movement restrictions for months (as in this epidemic), indirect person contacts may become a more-important transmission route, resulting in a higher risk for mixed and breeding herds (as found in this study). The Cox model showed that `transport contacts per month' had a significant effect on the spread of disease. Some farms with movements for welfare slaughter were later detected as infected through blood samples taken immediately prior to collection of pigs. Transport contacts for welfare slaughter during the epidemic, however well regulated and controlled, were associated with substantially higher risk of becoming infected and should therefore only be implemented if strictly necessary. In this case, there was strict supervision and rigorous hygiene measures, but additional protective steps may have to be considered. The timing for welfare slaughter to take place should be balanced between animal-welfare issues (overcrowding, overweight), rendering capacity and alternatives for market support (as discussed by Vanthemsche, 1996). From the analysis of the within-region models and the analysis excluding the core of the epidemic, the conclusion can be drawn that pig density is of more importance when pre-emptive slaughter cannot be effectively applied from the start. The EU report on the CSF epidemic concluded that it was difficult to control an outbreak of CSF in a highdensity pig area (Anonymous, 1998). Stegeman et al. (1998) found that local spread was an important factor in disease transmission, which emphasizes that dilution of the pig and herd population can contribute to the reduction in disease occurrence. The Dutch Minister of Agriculture (Van Aartsen, 1997) has proposed a reduction of the pig population by 25% (a decision that could help reduce the risk of spread of CSF). Indications from the current analysis are that `transport contacts per month' and enterprise type would remain as significant risk factors (even if pre-emptive slaughter is initiated immediately at the onset of the epidemic), but pig density would not. This could be related to breeding or mixed herds having far more contacts through persons than fattening herds. Some of these contacts will continue during the epidemic, despite the standstill of animal movements. The pseudo R-square remained relatively small in all final models. This is an indication that there are other unmeasured variables, which are important for explaining variation of CSF infection risk for individual farms. Acknowledgements The authors are grateful to the National Inspection Service for Livestock and Meat (RVV), The Netherlands for offering the opportunity for this analysis. We acknowledge the assistance of the tracing and epidemiology team operating in the Crisis Centre during this epidemic for making data available. We also thank the Ministry of Agriculture and
248
H.J. Benard et al. / Preventive Veterinary Medicine 42 (1999) 235±248
Forestry and AgriQuality New Zealand for making the practical work experience and data analysis possible. The EpiCentre, Massey University, New Zealand contributed through valuable discussions with Mark Bosson. We thank the referees for their constructive comments. References Anonymous, 1998. Veterinary Inspections: Final Report on a Mission Concerning the Evolution of Classical Swine Fever in The Netherlands. 23±27 March 1998, EU report on Netherlands swine fever epidemic. Elbers, A.R.W., Stegeman, J.A., Moser, H., De Jong, M.C.M., Ekker, H.M., Smak, J.A., De Leeuw, P.W., 1998. Effectiveness of preventive culling of pig herds during the Dutch CSF epidemic in 1997. In: Proceedings of the 15th IPVS Congress, Birmingham, p. 271. Hess, K.R., 1995. Graphical methods for assessing violations of the proportional hazards assumption in Cox regression. Statistics in Medicine 14, 1707±1723. Horst, H.S., Dijkhuizen, A.A., Huirne, R.B.M., Steenkamp, J-B.E.M., Thrusfield, M.V., Goodall, E.A., 1996. Eliciting expert opinions on risk factors concerning introduction of virus: application of conjoint analysis. In: Proceedings of the Society for Veterinary Epidemiology and Preventive Medicine, Glasgow, pp. 8±17. Hosmer, D.W., Lemeshow, S., 1989. Applied Logistic Regression. Wiley, New York. Kleinbaum, D.G., 1996. Survival Analysis: A Self-Learning Text. Springer, New York. Kramer, M., Ahl, R., Teuffert, J., Kroschewski, K., SchluÈter, H., Otte, J., 1995. Classical Swine Fever in Germany Ð some epidemiological aspects. In: Proceedings of the Society for Veterinary Epidemiology and Preventive Medicine, Reading, pp. 110±118. Laevens, H., Deluyker, H., Koenen, F., Van Caenegem, G., Vermeersch, J.P., De Kruif, A., 1998. An experimental infection with a Classical Swine Fever virus in weaner pigs. II. The use of serological data to estimate the day of virus introduction in natural outbreaks. Vet. Quarterly 20(2), 46±49. Lorenz, R.J., Boorberg, I., 1986. Incidence rate of hog cholera infection in various categories of pig herds. In: Proceedings of the Fourth International Symposium on Veterinary Epidemiology and Economics, Singapore, pp. 233±235. Saatkamp, H.W., Dijkhuizen, A.A., Geers, R., Huirne, R.B.M., Noordhuizen, J.P.T.M., Goedseels, V., 1996. Simulation studies on the epidemiological impact of national identification and recording systems on the control of Classical Swine Fever in Belgium. Prev. Vet. Med. 26, 119±132. Stegeman, A., Elbers, A.R.W., De Smit, H., Bouma, A., De Jong, M.C.M., 1998. Rate of transmission of Classical Swine Fever virus between herds by various routes. In: Proceedings of the 15th IPVS Congress, Birmingham, p. 269. Terpstra, C., 1987. Epizootiology of Swine Fever. Vet. Quarterly 9 (Suppl. 1), 50S±60S. Van Aartsen, J.J., 1997. Letter to the Second Chamber of Parliament from the Minister of Agriculture. Nature Management and Fisheries. 10 July 1997, The Hague, The Netherlands. Vannier, P., Colcanap, M., Carnero, R., Costes, C., Josse, J., Tillon, J.P., 1986. Study of the origin of an epizootic of Classical Swine Fever. J. Vet. Medicine B 33(4), 294±302. Vanthemsche, P., 1996. Classical Swine Fever 1993±1994 Belgium. The Pig J. 37, 43±53.