Factors associated with the early detection of foot-and-mouth disease during the 2001 epidemic in the United Kingdom

Abstract

An essential objective of an effective foot-and-mouth disease (FMD) eradication campaign is to shorten the infectious period by rapidly detecting and destroying cases of disease. The purpose of our investigation was to identify factors associated with the early detection of clinical FMD during the 2001 outbreak in the United Kingdom. We performed a logistic regression analysis, using early versus late detection of disease as the outcome of interest.

During the 2001 FMD outbreak in the United Kingdom, infected premises were more likely to be detected early under the following circumstances: 1) cattle (particularly dairy) were infected rather than sheep; 2) a recently confirmed infected premises was within 3 km of the new case; and 3) the case was initially reported by the farmer, rather than a Local Disease Control Centre-initiated surveillance activity (patrol, tracing, pre-emptive cull). Our findings suggest that reporting by farmers and initiatives that increase farmer education and awareness should be encouraged.

Introduction

On February 20, 2001, foot-and-mouth disease (FMD) was confirmed in pigs in an abattoir in England. This marked the detection of a large epidemic that resulted in the identification of 2026 infected premises in Great Britain (plus 4 in Northern Ireland) and the destruction of more than 4 million animals for disease control purposes. Britain’s official status as an “FMD-free country where vaccination is not practised” was restored by the World Organisation for Animal Health (OIE) approximately 11 mo later, following a control effort that involved more than 10 000 veterinarians, soldiers, and field and support staff (1,2).

Foot-and-mouth disease is extremely contagious. It spreads between herds directly via animal contact, or indirectly through movements of people, vehicles, and other fomites contaminated with the virus. Given appropriate conditions, windborne transmission may also occur (3). Shortly after a herd has been infected, typically 3–6 d in cattle, infected animals begin to excrete sufficient virus to infect other herds (4). This viral excretion may precede the appearance of clinical signs by up to 4 d (5,6).

Historically, most countries that are normally FMD-free have applied a stamping-out policy to eradicate the disease. This entails the slaughter and disposal of all susceptible animals on premises where FMD has been diagnosed, as well as of herds that have been exposed to the infection by direct or indirect contact (7). Increased biosecurity measures, including movement restrictions, are usually also implemented. For the stamping-out policy to be successful, cases of FMD must be detected accurately and rapidly, and then destroyed. During most of the 2001 epidemic in the UK, cases were diagnosed solely on the basis of clinical signs in order to expedite the application of stamping-out measures. Laboratory testing was generally performed after the implementation of control measures.

Cases must be detected accurately, because there are costs associated with both false-positive and false-negative diagnoses (slaughter, disposal, and compensation costs; and costs associated with a larger epidemic related to onward spread, respectively). Subsequent to the 2001 United Kingdom (UK) epidemic, some cases were determined to have been false-positives (8). A previously published analysis demonstrated that premises designated as “infected” were more likely to have been diagnosed accurately (with positive laboratory results) during this epidemic under the following circumstances: 1) cattle rather than sheep were suspected of FMD infection, 2) cases were initially reported by the farmer rather than the disease control authorities, 3) lesions were less than 3 d old, 4) the report was filed outside the peak of the epidemic, or 5) the report originated from certain Local Disease Control Centres (LDCCs) (9). Speed, as well as accuracy, is important; cases must be detected and destroyed before they have the opportunity to infect animals in other herds. It has previously been reported that there was considerable variation in the time from infection to reporting of suspect cases (10).

The purpose of this investigation was to identify factors associated with the early detection of FMD-infected premises during the 2001 epidemic in the UK. Factors that we considered included the following: species infected; the timing of the case in relation to the national epidemic curve; the method of detection (found by LDCC-initiated surveillance activities or not); the proximity of the premises to other, previously confirmed, infected premises; and, finally, which LDCC was responsible for implementing the disease control procedures in that area. Logistic regression models were developed to explore the association of these variables with early detection, whilst controlling for other intervariable relationships.

Materials and methods

For each infected premises, the date of report of suspicion, the date of confirmation, the origin of report (farmer, patrol, etc.), the LDCC involved, the species infected, the age of oldest lesion, the laboratory results, and the map reference were obtained from the database compiled by the epidemiology team at the National Disease Control Centre (2,11). We accessed the database in November 2002.

All analyses were conducted using data analysis and statistical software (Stata Statistical Software: Release 8.0. 2003; StataCorp L.P. College Station, Texas, USA, unless otherwise noted). Any incompatibilities within the data (for example, if cattle were listed as the species infected, but it was not recorded that there were any cattle on the farm) were investigated, using the “Final Epidemiology Report” database (9). This 2nd database was compiled by the field epidemiology teams at each LDCC, using information from the Disease Control System (11), forms completed at the time of investigation, and local knowledge. If more detailed information could be gleaned from the Final Epidemiology Report database that clarified the inconsistency, the data were adjusted.

We performed a logistic regression analysis, using early versus late detection of disease as the outcome of interest. Only infected premises that had the presence of FMD virus confirmed by the laboratory were used in the analysis, because if FMD was not actually present, it could not be found “early” or “late.” Internationally agreed laboratory procedures, as described in the OIE Manual of Standards for Diagnostic Tests and Vaccines (2000) (12), were used during the epidemic (13). Five laboratory- positive premises were excluded from the analysis because the origin of the report was missing.

If the age of oldest lesion was less than 3 d, the premises was considered to have been detected “early,” otherwise it was classified as detected “late.” As antibodies may typically be first detected 3–5 d following the appearance of clinical signs (3), the 33 premises detected due to positive serological results were considered to have been detected “late.” To the best of our knowledge, no FMD cases were detected by virological methods in the absence of clinical signs.

The predictor variables that we considered are presented in Table 1. The method of disease detection was dichotomized into active and passive surveillance. Many cases were identified as a result of a report of suspected disease by a farmer or other member of the public. Because the disease control authorities did not initiate the visit to these premises, we referred to this method of detection as passive surveillance. Other cases of FMD came to the attention of the authorities during patrol visits, tracing visits, or pre-emptive culls, all visits that were initiated by the LDCC. These cases were considered as being detected by active surveillance. This terminology, in reference to surveillance during an epidemic, has been used before (9,14).

Table 1.

Description of variables evaluated for an association with early detection, and results of the univariable analysis. The outcome is a laboratory-confirmed infected premises being detected early (oldest lesions < 3 d) during the UK 2001 foot-and-mouth disease epidemic

Predictor variable	Classification	Premises detected early	Premises detected late	Crude odds ratio (95% CI)	P-valuea
Method of detection	Passive surveillance	674	414	1	P < 0.0001
	Active surveillance	91	142	0.39b (0.29, 0.53)
Premises within 3 km of another infected premises confirmed within previous 2 weeks	No	207	208	1	P = 0.0001
	Yes	558	348	1.61 (1.26, 2.05)
Stock infected	Dairy	313	98	5.70 (4.21, 7.72)	P < 0.0001
	Beef	304	189	2.87 (2.19, 3.76)
	Pigs	0	5	N/A
	Sheep	148	264	1
Time in the epidemic	1st month (to Mar 18)	139	116	0.79 (0.59, 1.06)	P = 0.2496
	Peak (Mar 19–29 Apr)	376	249	1
	Tail (after 30 April)	250	191	0.87 (0.67, 1.10)
Local disease control centre	17 dummy variables				P = 0.0225

95% CI — 95% confidence interval.

^aFor nondichotomous variables, refers to the significance of the predictor represented by a group of dummy variables.

^bAn OR of 0.39 means that the odds of cases being detected early by active surveillance were less than half the odds of cases detected by passive surveillance.

The type of animal infected was represented by a categorical variable (Table 2). Premises with more than 1 species or type of animal infected were classified according to the species/type in which it should be the easiest to detect disease, based on clinical signs and typical husbandry practices. It was assumed that the age of oldest lesion was the same in all types of animals on the premises. For example, if both sheep and dairy cattle were infected on a single premises, dairy cattle were recorded as the species infected, because the clinical signs of FMD are usually easier to detect in dairy cattle than in sheep.

Table 2.

Categories of variables developed to represent type of animal infected on premises during the outbreak of foot-and-mouth disease in the UK. For premises on which more than 1 type of animal was infected, the “premises” was represented by the type in which it should be the easiest to detect disease. The actual types of animals infected in each category are shown in the last column.

		Category	Composition (Number of infected premises)
Most obvious clinical signs
		Dairy	Dairy only (375)
			Dairy-beef (6)
			Dairy-sheep (29)
			Dairy-sheep-pigs (1)
		Beef	Beef only (422)
			Beef-sheep (71)
		Pigs	Pigs only (4)
			Pigs-sheep (1)
		Sheep	Sheep only (412)
Least obvious clinical signs

A variable was created that represented whether or not the infected premises was within 3 km of another premises that had been declared infected within the previous 2 wk. This distance was chosen as “Protection Zones” had been created by the disease control authorities in a 3-km radius around infected premises (15). All premises within these zones with susceptible stock were subject to enhanced movement restrictions and bio-security requirements on an ongoing basis, and they also might have received patrol visits in the days and weeks immediately following detection of the infected premises that initiated the Protection Zone. The Euclidean distance between infected premises was calculated, using the premises’ XY coordinates. These calculations and other data manipulations to create this variable were performed using Statistical Analysis System (SAS) software (SAS Version 8.2; SAS Institute 2002, Cary, North Carolina, USA).

To explore how the outcome changed over the course of the epidemic, a smoothed scatterplot of the log odds of early detection, based on the report date, was created (Figure 1). The scatterplot was lowess (locally weighted) smoothed, with a bandwidth of 0.8, using Cleveland’s tricube weighting function (16). In the univariable and multivariate analysis, calendar time was represented by a categorical variable that was created, based on changes observed in the number of infected premises that were reported each day (corresponding to the early epidemic, the peak and the tail) (Figure 1).

Figure 1.

Smoothed^a scatterplot of the log odds of early detection based on the report date, overlaid on the curve representing the number of premises reported each day. Arrows show the cutpoints used to create the variable representing calendar time

^a Lowess smoothed, bandwidth 0.8, using Cleveland’s tricube weighting function (16).

The distribution of each variable was examined. Univariable associations with the outcome were assessed, using a chi-squared test for dichotomous predictors and simple logistic regression for predictors with more than 2 categories.

Because the number of predictor variables was small, all were subject to inclusion in the multivariable model. Indicator (dummy) variables were created to represent categorical variables with more than 2 levels. Variables not statistically significant were removed from the model in a manual backwards selection method. The statistical significance (cutoff at P ≤ 0.05) of each variable was assessed by either the Wald test (dichotomous variables) or the likelihood ratio test (nondichotomous predictors).

To assess confounding, odds ratios were monitored as each variable was removed from the model. If the odds ratio of a remaining predictor variable(s) changed by more than 20%, it was considered to have a confounding relationship with the excluded variable. After the model had been determined, all possible 2-way interaction terms of the main-effects were created and their significance assessed, using the likelihood ratio test.

Calendar time did not remain in the model as either a significant main-effect or a confounder. To further explore the effect of time, 2-way interaction terms between the main effects and the variable representing time were created. Their contribution to the model was assessed, using the likelihood ratio test. To further investigate the relationship of active surveillance with the outcome over time, the proportions of the specific categories of active surveillance (patrol, pre-emptive culls, tracings) were stratified by time and tabulated against the outcome (Table 3).

Table 3.

Number of foot-and-mouth disease infected premises detected by different categories of active surveillance, stratified by time to detection (early versus late)^a and calendar time

	Infected premises detected during:
	1st month (to Mar 18)		Peak (Mar 19–29 Apr)		Tail (after 30 April)		Overall
Detected by:	#	%	#	%	#	%	#	%
Patrol	10		22		73		105
Early	5	50	12	55	33	45	50	47
Late	5	50	10	45	40	55	55	52
During pre-emptive cull	1		12		78		91
Early	0	0	6	50	19	24	25	27
Late	1	100	6	50	59	76	66	73
Tracing	29		5		3		37
Early	11	38	3	60	1	33	16	43
Late	18	62	2	40	2	67	21	57
Pearson χ2b	1.14	(P = 0.57)	0.15	(P = 0.93)	8.54	(P = 0.01)	8.64	(P = 0.01)
Total detected by active surveillance	40		39		154		233
Early	16	40	21	54	54	35	91	39
Late	24	60	18	46	100	65	142	61

^aEarly detection is defined as oldest lesions < 3 d, everything else is considered late detection

^bTesting the association between the category of surveillance that detected the case and early detection

The overall fit of the model was assessed by the Pearson chi-squared, deviance chi-squared, and Hosmer-Lemeshow goodness-of-fit tests. The impact of various covariate patterns on the model was assessed by looking for potential outliers (those with a large positive or negative standardized residual) and by calculating the delta-beta value of each covariate pattern. The delta-beta is a measure of the difference between the observed set of regression coefficients and the set that would be obtained if the observations in the covariate pattern of interest were deleted (16).

To explore why LDCC did not remain in the multivariable model as a significant predictor, 3 additional logistic regression models were constructed. In each of these models, LDCC was the sole predictor variable and the outcome variables were as follows: 1) active versus passive surveillance; 2) whether or not the premises was within 3 km of another infected premises confirmed in the previous 2 wk; and 3) whether sheep were the species infected or not. The predictive ability of these 3 models was compared, using the Akaike (AIC) and Bayesian (BIC) statistics (16).

Results

Data were used from all infected premises from which samples were submitted to the laboratory and positive results were returned, except for 5 premises that had missing data (n = 1321). Thirty-eight of these infected premises were identified by the detection of antibodies to the FMD virus in blood samples taken from sheep. The samples were taken either during a pre-emptive cull, because of routine testing prior to lifting restrictions, or as part of special test campaigns performed for surveillance purposes. The mean age of the oldest lesion on the remaining 1283 premises was 2.82 d [standard deviation (s) = 1.83], with a minimum age of 1 d and a maximum of 14 d (Figure 2). Seven hundred and sixty-five (57.9%) premises were detected early (oldest lesion < 3 d); the mean age of the oldest lesion on these premises was 1.80 d (s = 0.40). The remaining 556 (42.1%) premises were detected late (oldest lesions ≥ 3 d), with the mean age of oldest lesion of these premises equal to 4.33 d (s = 2.06).

Figure 2.

Distribution of the age of oldest lesion in premises with laboratory-confirmed infection of foot-and-mouth disease in the 2001 epidemic in the UK (n = 1283; premises found by serosurveillance are not included).

With the exception of the variables that represented time, all of the predictor variables were significantly associated with the outcome in the univariable analysis (Table 1). The 5 premises with infected pigs were excluded from the multivariable analysis, because they predicted the outcome perfectly (all were detected late), so the final logistic regression model was based on 1316 observations with 35 different covariate patterns (Table 4). The overall model likelihood ratio chi-square was highly significant (G = 178.11, 8 d.f., P < 0.0001). The Hosmer-Lemeshow goodness of fit test indicated that the model fitted the data adequately (χ² = 4.69, 6 d.f., P = 0.58). However, the Pearson chi-squared goodness of fit test approached statistical significance (χ² = 37.00, 26 d.f., P = 0.075) and the deviance chi-squared goodness of fit test was statistically significant (χ² = 40.69, 26 d.f., P = 0.03). Both of these test results indicated a possible problem with fit.

Table 4.

Results of the logistic regression analysis comparing laboratory confirmed infected premises detected early (oldest lesion < 3 d) to those detected late (oldest lesion ≥ 3 d) during the 2001 foot-and-mouth disease epidemic in the UK (n = 1316 premises^a)

Variable	Odds ratio	P-value	95% CI
Species infecteda
Dairy	4.93b	< 0.01	3.60, 6.76
Beef	2.67	< 0.01	2.01, 3.54
Sheep	1.00
Premises within 3 km of premises detected in previous 2 weeks?
Yes (Local)	1.69	< 0.01	1.30, 2.18
No (Not local)	1.00
Time in the outbreak
1st month (to Mar 18)	1.16	0.38	0.83, 1.64
Peak (Mar 19–29 Apr)	1.00
Tail (after 30 April)	1.26	0.15	0.92, 1.73
Method of detection
Active surveillance	1.08	0.82	0.54, 2.17
Passive surveillance	1.00
Interaction (time in the outbreak X method of detection)
Active surveillance X 1st month	0.66	0.41	0.23, 1.80
Active surveillance X Tail	0.35	0.01	0.15, 0.78

The overall model deviance was 805.65, and the likelihood ratio chi-square was 178.11 with 8 d.f., P < 0.0001. 95% CI — 95% confidence interval.

^aThe 5 premises with pigs infected were not included because there was no variability in the outcome (all premises with pigs infected were detected late).

^bAn OR of 4.93 means that the odds of an infected premises being detected early were almost five times greater if dairy cattle were infected than if sheep were infected.

No extreme outliers were observed. The model was rerun with observations from the 3 covariate patterns with the largest delta-beta values excluded (results from this model not shown). The goodness-of-fit tests indicated that this reduced model fitted the data (Hosmer-Lemeshow test: χ² = 5.43, 7 d.f., P = 0.61, Pearson chi-squared test: χ² = 25.71, 23 d.f., P = 0.31, deviance chi-squared test: χ² = 28.85, 23 d.f., P = 0.19). In this model, all but 1 of the coefficients were slightly closer to the null compared with the full model. The coefficient representing active surveillance moved slightly away from the null. However, the overall interpretation of the results did not change.

The Local Disease Control Centre did not remain in the final model as its coefficient became insignificant upon the addition of any other predictor to the model (except the variable representing time). There was evidence of confounding between the method of detection and the species infected. None of the 2-way interaction terms between the main-effect variables were significant; however, there was significant interaction between the time in the epidemic (not a statistically significant main-effect variable, but examined because it was of interest) and type of surveillance.

Infected premises were more likely to be detected early (within 3 d of showing clinical signs) at any time during the outbreak if there was another recently confirmed infected premises within 3 km, or cattle (dairy or beef) were infected rather than sheep. The latter association was stronger if dairy rather than beef cattle were infected. During the tail of the outbreak only, premises were more likely to be detected early if the disease was found by passive versus active surveillance (Table 5). The nature of the active surveillance activities that detected disease shifted over the course of the outbreak, from primarily tracing visits in the 1st month to more pre-emptive culls and patrol visits later in the outbreak. There was a significant association between time to detection and the specific type of active surveillance activity, with disease being most likely to be detected late during preemptive culls (Table 3).

Table 5.

Interpretation of the significant interaction between the method of detection and time in the epidemic. The outcome is early detection (infected premises with oldest lesions < 3 d) during the 2001 foot-and-mouth disease epidemic in the UK

Variable		OR	95% CI
Method of detection	1st month (to Mar 18)	0.71	0.34, 1.49
Active versus passive	Peak (Mar 19–29 Apr)	1.08	0.54, 2.17
surveillance	Tail (after 30 April)	0.38a	0.24, 0.59

95% CI — 95% confidence interval.

^aAn OR of 0.38 means that, during the tail of the epidemic, the odds of an infected premises being detected early were 0.38 times as great for premises detected by active surveillance, compared with premises detected by passive surveillance.

In the logistic regression models constructed by using LDCC as the predictor variable, there was a significant association between LDCC and (1) the type of surveillance (P < 0.0001), and (2) whether or not the premises was within 3 km of another infected premises confirmed in the previous 2 wk (P < 0.0001). The association between LDCC and whether or not sheep were the species infected was not statistically significant (P = 0.86).

The variable LDCC better predicted active versus passive surveillance (AIC = 1201.99, BIC = 1279.74) than the variable representing whether there was another recently confirmed infected premises nearby (AIC = 1549.24, BIC = 1626.99). The poorest predictor of LDCC was the variable that signified whether or not sheep were infected (AIC = 1645.63, BIC = 1728.58).

Discussion

If FMD is to be eradicated by stamping-out, it is essential that cases are detected and destroyed rapidly in order to shorten the infectious period of each herd and minimize onward spread (17). This study found that early detection of FMD during the 2001 epidemic in the UK was associated with cattle (particularly dairy) being infected rather than sheep, the presence of a recently confirmed infected premises nearby, and detection by passive surveillance. The difference between species has previously been reported; it was attributed to the disease in sheep typically being more subtle and to cattle being more frequently or intensely observed (11,18). The difference between dairy and beef cattle has not been reported previously; it is probably due to differences in husbandry practices, with dairy cattle being milked twice-daily and, thus, typically, being observed more closely than beef cattle.

The association of early detection with the presence of a recently confirmed infected premises within 3 km probably represents the effect of increased farmer vigilance as the disease entered their immediate vicinity. Such premises were under enhanced restrictions and were also more likely to be inspected by patrols. Presumably, the restrictions and visits heightened awareness about the disease and educated producers about typical clinical signs, enabling them to recognize FMD earlier.

Laboratory-positive FMD cases were more likely to be detected early by passive than active surveillance. As lesions begin to heal and/or become secondarily infected, they become less characteristic of FMD. Therefore, this finding suggests that passive surveillance detects the more obvious cases of disease, whereas active surveillance mops-up the more difficult cases. The role of active surveillance as a “mop” has been suggested before to explain the finding that, given a suspect case of FMD in the 2001 UK epidemic, the final diagnosis was more accurate if the case had been detected originally by active rather than passive surveillance (9).

Our results showed that active surveillance often detected animals with FMD lesions that were 3 or more days old, but we could not discern what proportion of these visits targeted premises likely to harbor older cases of FMD. The data available allowed us to differentiate between broad active surveillance activities, namely patrols, pre-emptive culling, and tracings (Table 3). However, each of these activities involved visits with different purposes, which could not be distinguished from the data. Some active surveillance visits were more apt to find FMD cases with older lesions, such as trace-in investigations (to discover possible sources of infection) and some patrol visits. Many other active surveillance visits aimed to detect disease early, such as those related to trace-out investigations (to detect spread from an infected premises), most patrol visits, and pre-emptive culls. However, our finding that 73% of the laboratory-positive cases detected during pre-emptive culls involved animals with older lesions suggests that many of those visits, ostensibly to detect cases early in the disease process, in fact detected older cases of FMD (Table 3).

Relatively few cases were identified by active surveillance until the tail of the outbreak, probably because resources were overwhelmed during the peak of the epidemic. Therefore, the lack of association between the type of surveillance and early detection during the 1st month and peak of the outbreak might have been due to lack of statistical power. Additionally, the lack of association might have been related to the shift in the activities of active surveillance over time, from primarily tracing visits to more pre-emptive culls and patrol visits.

Although the log odds of early detection appeared to vary over calendar time in a descriptive analysis (Figure 1), this variation was not statistically significant in either the univariable or multivariable analysis. The time to detection might have been expected to decrease over the course of the epidemic as awareness increased and the disease control procedures begin to function more smoothly. This was not observed, perhaps because the larger national epidemic actually consisted of several smaller epidemics occurring in different geographical regions of Great Britain (4). As FMD was first detected in each of these regions at different times between February and September, awareness and efficacy of operations would have increased at different times in each region over the course of the national epidemic.

Although LDCC had a significant relationship with the outcome in the univariable analysis, it did not remain in the final model because its coefficient became insignificant upon the addition of any other predictor to the model (except the variable representing calendar time). This suggested that it had an indirect relationship with early detection and, further, that each of the other predictors in the model were intervening variables in the causal pathway between LDCC and the outcome (16). In other words, the association between LDCC and early detection was mostly caused by the following: 1) the difference in species distribution in the region covered by each LDCC, 2) the different surveillance practices between LDCCs, and 3) a differing distribution of FMD cases occurring within 3 km and 2 wk of each other in different LDCCs.

Our findings suggest that the difference in time to detection between LDCCs was mostly due to differing surveillance practices. During the epidemic, each LDCC had the authority to develop unique policies and protocols regarding surveillance. The criteria that determined which premises received patrol or tracing visits, and how often, varied between LDCCs, as did the protocol followed during these visits (if animals were examined from a distance or actually handled, and how many and which animals were examined at each visit). Further research is needed to establish which surveillance protocols resulted in the earliest and most accurate detection of FMD.

The accuracy of the ages assigned to lesions is critical to the interpretation and usefulness of this analysis. During the epidemic, lesions were aged by 6 veterinarians in the National Disease Control Centre (NDCC) epidemiology team, based on the reports received from the field. On premises that had been visited by specialists from the FMD World Reference Laboratory at Pirbright, the specialists’ estimate of the age of oldest lesion was used (11). When performed by experienced clinicians, as was generally the case during this epidemic, the ageing of lesions up to 5 d old should be accurate to plus or minus 1 d. The accuracy of the estimation declines for lesions older than 5 d (19).

The accuracy of ageing disease is also dependent on the animal with the oldest lesion being examined. Whilst the veterinarians in the field typically examined several animals prior to and during culls on infected premises, it was not possible to examine every animal slaughtered (11). Animals were more likely to remain uninspected when large sheep flocks rather than cattle herds were being culled, simply because the sheep flocks consisted of several hundred animals, whereas most cattle herds were much smaller. Therefore, it is reasonable to assume that the animal(s) with the oldest lesion was more likely to have remained unexamined on premises with infected sheep than on premises with other infected species. This would lead to an underestimation of the time to detection on sheep premises and differential misclassification of the outcome in our study, in turn resulting in a conservative estimate of the odds ratio representing the relationship between “stock infected” and the outcome.

We dichotomized the outcome variable (FMD detected “early” or “late”) rather than model the age of lesion directly, such as by constructing a linear regression model. Premises were considered to have been detected “early” when the oldest lesion was less than 3 days old, because this is a valid and realistic target for FMD detection during an outbreak. Because of the inaccuracy inherent in determining the actual age of the oldest lesions, premises were more likely to be classified accurately as “early” or “late,” rather than by the actual age of lesion. Further, the decision to dichotomize the outcome allowed the inclusion of data from the premises detected by serological testing and on which clinical signs or lesions were never observed.

To minimize bias, farms that did not have infection confirmed by laboratory testing were omitted from the analysis. Some of the omitted farms did not have samples submitted to the laboratory, others tested laboratory-negative. We assumed that the results of our study should be applicable to all farms truly infected with FMD during this outbreak, regardless of whether they were laboratory tested or not. The decision as to whether or not samples were submitted was based primarily on available laboratory capacity at the time.

There was evidence of overdispersion in the final model, because the deviance chi-squared value was greater than its degrees of freedom. This might be an indication that the model had not accounted for clustering in the data, or that an important variable was not included in the model (16). Factors that might have been important but were unaccounted for because data were not available include the number of animals affected on the premises, the location of the animals on the premises (housed, or grazing nearby or remotely), and the natural variation of the clinical signs between animals of the same species.

These results should be useful for the development of recommendations and contingency plans for the management of future outbreaks. It appears that active surveillance plays a critical role in mopping-up cases missed by passive surveillance, but it should not be relied upon for early detection of disease. Rather, programs that enhance producer education and awareness should be developed and promoted. Farmers are in the best position to be the first to detect any deviations from normal, as they observe stock daily and are the most knowledgeable about the usual condition of the animals under their care. Producer training should be prioritized even when resources are scarce, although it may be targeted at beef and sheep operations where disease is more likely to be found late.