Prevalence of and risk factors for influenza in southern Ontario swine herds in 2001 and 2003

Abstract

This research included 2 prevalence studies and a risk-factor investigation conducted in 2001 at 93 sites with sows only, finishers only, or both. In 2001, 1300 serum samples from sows in 65 herds and 720 serum samples from finisher pigs in 72 herds were tested for antibodies to swine influenzavirus (SIV) of H1N1 subtype with an enzyme-linked immunosorbent assay (ELISA). In 2003, 1140 serum samples from sows in 76 herds were tested for antibodies to SIV of H3N2 subtype with a hemagglutination-inhibition assay based on A/Swine/Colorado/1/77 and A/Swine/Texas/4199-2/98 isolates. The apparent pig-level H1N1 seroprevalence in 2001 was 61.1% and 24.3% in sows and finishers, respectively. The apparent pig-level seroprevalence in 2003 for H3N2 A/Sw/CO/1/77 and A/Sw/TX/4199-2/98 in sows was 0.6% and 0.7%, respectively. The factors associated with sow-herd H1N1 positivity included pig or farm density at different geographic levels, an external source of breeding pigs, number of animals on site, and decreasing proximity to other barns. Higher-parity sows had higher odds of seropositivity, but there was significant random variability in this association among herds. The odds of finisher-herd SIV positivity were higher with large herd size, high pig farm density, and farrow-to-finish type of farm. Finisher herds were SIV-positive only if source sow herds were positive. Simultaneously, 45% of finisher herds were SIV-negative although sow source herds were positive.

Introduction

The influenzavirus is a single-stranded negative-sense RNA virus classified in the family Orthomyxoviridae. Type A influenzavirus is further characterized into subtypes according to the combination of its surface hemagglutinin (HA) and neuraminidase (NA) antigens. Many of these viruses can infect pigs under experimental (1) and field (2,3) conditions; however, only type A subtypes H1N1, H3N2, and H1N2 have been isolated from pigs on a regular basis. Viruses belonging to the same subtype may show antigenic and genetic differences owing to various mechanisms of emergence.

In a swine herd, influenzavirus infection in its epidemic form is recognized clinically as an acute febrile respiratory tract infection characterized by high morbidity and low case-fatality rates (4). In contrast, endemic influenzavirus infection may spread more slowly (5,6) and without clinically apparent signs (7). Swine influenzavirus (SIV) has also been reported as a pathogen contributing to the porcine respiratory disease complex in finisher pigs (8). There are case reports of clinical disease in people caused by SIV after exposure to pigs (4) and elevated titers of antibody to SIV due to occupational exposure (9); therefore, SIV may be of public health concern.

In North America, SIV has been endemic, and until the late 1990s most SIV infections in swine were caused by the H1N1 subtype. However, with the emergence of the triple-reassortant H3N2 subtype in the United States in 1998, and its subsequent reassortments with the classic H1N1 subtype to produce H1N2, significant losses in US swine herds have been reported (10). Passive monitoring in Ontario enabled the identification of a recent introduction into the swine population of human and avian influenzavirus strains H4N6, H3N2, H3N3, and H1N2 (3,11–13). Such introduction may be of public health concern because it could result in new influenzavirus strains that might be capable of spreading from human to human and lead to a pandemic in humans (14).

Sow and finisher herds differ in terms of population dynamics (number and types of additions and withdrawals), number of influenza-susceptible animals at any point in time, and several management procedures unique to the age group. Consequently, it was presumed that influenza prevalence and risk factors are different for sow and finisher herds. At the individual animal level, recent findings indicate that sows could have a long-lasting serologic response after exposure to SIV (15). Thus, parity number may be a cumulative measure of historical exposure to SIV. Furthermore, parity is easily obtainable from most herds, generally reflects age, and as a sow-level variable may be used to explore characteristics of random variability among herds.

There had been no recent estimates of the prevalence of influenza in swine in Ontario; therefore, the primary objectives of our study were to estimate the apparent and true prevalence at the herd and pig levels and to apply cluster analysis to describe within-herd prevalence for different strains of SIV. The secondary objective was to estimate the risk factors for SIV positivity in sows and finishers at the herd level. The final objective was to explore the nature of the association between parity and SIV seropositivity in the sow study population.

Materials and methods

Sampling frame

Our study was a part of a larger active monitoring system established as a 4-y project called the Ontario Swine Sentinel Project. This project began in 2001 as a network of 112 sentinel swine operations, with the primary goal of monitoring infections and practices of public health significance in pigs. The study population at the swine-operation level was a combination of operations selected randomly (in a previous study; Anne Deckert, Laboratory for Foodborne Zoonoses, Public Health Agency of Canada, Guelph, Ontario: personal communication, 2005), conveniently (ease of access), and purposively (areas of low pig density or membership in a multisite system). Farms that temporarily or permanently dropped out of the study in a year were replaced by conveniently selected farms.

For the purpose of this study, the following definitions were applied. A swine operation was defined as a swine business in 1 or more geographically separate locations that included 1 or more defined age groups and was operated under single ownership. A swine farm (or premise) was defined as a contiguous land location, including all structures housing pigs. A swine herd was defined as a group of pigs of a defined age category (sows or finishing pigs) housed in all structures on a sampling site. A sampling site was defined as a swine farm visited during the cross-sectional study at which pigs of a defined age category were sampled. Ontario was considered an area.

Swine operations included in this monitoring system were almost exclusively sampled during summer months. On each farm, 30 breeding-age animals (sows and gilts) were selected for blood collection by systematic random sampling in 2001 and by stratified systematic random sampling for 5 gilts and 25 sows of greater parity from the gestation barn in 2003. In addition, when collecting blood from the breeding-age pigs, we collected blood from 30 pigs in the cohort of finishers closest to market weight (approximately 80 to 120 kg of live bodyweight) housed in the finisher farm under the same ownership, as this was convenient. All samples were taken from the orbital venous sinus. Blood was centrifuged and serum separated and stored in a bank at −20°C until tested.

The protocol had been reviewed and accepted by the Animal Care Committee, University of Guelph, Guelph, Ontario.

Data collection

At each visit a questionnaire from the Ontario Swine Sentinel Project was administered to the owner or manager of the facility to assess the current management and production practices, biosecurity measures, demographic characteristics, medication protocols, and parity distribution of the sows. The questionnaire covered an array of practices and information that was believed to be potentially associated with the presence of agents causing diseases of public health interest. During the visit, the location of the farms was recorded with the use of a hand-held global positioning system receiver (eTrex Legend; Garmin, Olathe, Kansas, USA). Maps of Ontario for 2001 at the levels of census division (CD) and census consolidated subdivision (CCS) were downloaded from the Statistics Canada Web site (www.statcan.ca). A CD represents a geographic unit intermediate between a municipality and a province and consists of several CCSs; a CCS is usually formed as a group of municipalities. The total area in square kilometers was calculated for each polygon, and the polygons were then dissolved to the CD and CCS levels. The numbers of pig farms and pigs at the CD and CCS levels were obtained from the 2001 Census of Agriculture (16). Spatial and attribute data were obtained through the University of Guelph Library Web site (www.lib.uoguelph.ca) under the Data Liberation Initiative Licence Agreement. The density of pig farms and pigs at the CD and CCS levels were then calculated from the total area and the numbers of pig farms or pigs at the 2 levels. Subsequently, the location of each farm was merged with the density attributes at both levels. Data management was done in ArcGIS 8.2 (Environmental Systems Research Institute, Redlands, California, USA).

Inclusion criteria

In 2001, 65 sow and 72 finisher herds were selected and sampled for antibodies to H1N1 SIV between May 7 and August 20. The herds were sampled from 93 farms nested within 76 swine operations. In 2003, 76 sow herds were selected for testing for the presence of antibodies to 2 strains of H3N2 SIV between April 30 and September 8. From the bank of serum samples, we randomly selected 20 sow and 10 finisher samples per herd in 2001 and 15 sow samples per herd in 2003 and submitted them to the Animal Health Laboratory (AHL), University of Guelph, for testing. The inclusion criterion for the prevalence studies was availability of a sample in the serum bank at the time of submission to the laboratory. Herd visits were scheduled according to 2 sets of criteria: (a) the accessibility of finisher pigs of the appropriate age on a farm and the availability of a producer for an interview, both of which were likely not associated with SIV prevalence; and (b) biosecurity requirements, including the absence of contact with pigs by the researchers for a specified period. Hence, if a systematic difference existed between herds included in the larger study in a corresponding year and the population considered for the prevalence study, it was likely negligible. In total, 1300 sow and 720 finisher samples were tested for antibodies to H1N1 SIV, and 1140 sow samples were tested for antibodies to 2 strains of H3N2 SIV.

Diagnostic tests

We used a commercial indirect enzyme-linked immunosorbent assay (ELISA) for H1N1-specific antibodies (HerdChek Swine Influenza Antibody Test Kit; IDEXX Laboratories, Westbrook, Maine, USA) to test the available sow and finisher pig serum samples in 2001 for H1N1 SIV and hemagglutination-inhibition (HI) assays for H3N2-specific antibodies to test the available sow serum samples in 2003 for the 2 H3N2 SIV strains. The ELISA was performed according to the manufacturer’s instructions, and a resulting sample-to-positive (S/P) ratio ≥ 0.4 was considered a positive result.

The viruses used in the HI assays were A/Swine/Colorado/1/77 and A/Swine/Texas/4199-2/98. These isolates have routinely been used in testing for antibodies to swine H3N2 influenzavirus in the AHL and represent either the genotypes occasionally isolated in Ontario or the index isolate of triple-reassortant H3N2 (HA, NA, and PB1 of human influenzavirus origin; M, NP, and NS of classic SIV origin; and PA and PB2 of avian influenzavirus origin) that caused an epidemic in US swine herds in the late 1990s. The serum samples were tested in serial 2-fold dilutions starting at 1:8 and 1:20 for the Colorado and Texas strains, respectively; reactions at dilutions ≥ 1:32 and ≥ 1:40, respectively, were considered to be positive results.

Sensitivity and specificity were assumed to be fixed at 98.8% and 91.6%, respectively, for the H1N1 ELISA. These estimates, based on a challenge study in finisher pigs in which the gold-standard status was known vaccination against H1N1 SIV (17), were lower than the manufacturer’s claim but in line with the expectation that the specificity of the HI test would be higher than the specificity of the ELISA at the expense of sensitivity (18). Sensitivity and specificity were also assumed to be equally valid for the Colorado and Texas variants and fixed at 89.9% and 88.9%, respectively, for the HI test. These estimates were based on Bayesian analysis of 376 pig serum samples tested simultaneously with the H3N2 ELISA and the HI test for the Colorado, Texas, or both H3N2 variants (19).

Prevalence estimation

Apparent and true prevalence was estimated at the pig level. Apparent prevalence and 95% confidence intervals (CIs) were calculated by an exact method based on binomial distribution (20) in Stata 8 (StataCorp, College Station, Texas, USA). The true prevalence at the pig level was calculated with the use of the Rogan–Gladen estimator (21); calculation of the variance of this estimator accounted for the sample size that generated estimates of the sensitivity and specificity. The 95% CI was calculated with use of the normal approximation. Herd-level and, in the case of H3N2, area-level sensitivity and specificity were calculated in Freecalc 2 (22) assuming sampling from a hypergeometric distribution. For the calculation of herd-level sensitivity and specificity, we assumed that the ELISA and HI assays had fixed sensitivity and specificity. We further assumed that the estimates of the HI assay characteristics reported in the validation study were valid estimates of test accuracy for both strains of H3N2. In addition, we explored a scenario in which H3N2 HI specificity was assumed to be 99% and all other parameters were kept equal. The assumed minimum expected within-herd prevalence was 30% for all age categories. Sample size was fixed by the study design, and the assumed population size was 100 animals. An optimal cut-off number of positive animals that would classify the herd as positive was determined by comparing the estimates of sensitivity and specificity and selecting a number that optimized both. Herds having more than the cut-off number of animals testing positive for each combination of strain and age category were considered to be “true-positive”. A herd was declared “apparently positive” if any animal was test-positive for a particular strain. For the H3N2 strains, the calculated estimates of sensitivity and specificity were used as the inputs to calculate the area-level sensitivity (22) at a between-herd prevalence of 5%.

Evaluation of within-herd prevalence

Within-herd prevalence in the sow and finisher herds was examined descriptively. We did agglomeration (23)-based hierarchical cluster analysis on within-herd prevalence, calculating the distance between clusters as the squared Euclidian distance between centroids as well as by other methods available in SAS (SAS Institute, Cary, North Carolina, USA) and comparing the results. The optimal number of clusters was decided on biologic grounds (published evidence of diseased finisher herds with both a fast SIV spread during an outbreak and a slower SIV spread, and apparently disease-free herds), with statistical considerations (24).

Assessment of risk factors

Binary logistic regression was used to analyze univariable and multivariable models for H1N1 influenzavirus antibody positivity at the herd level. The outcome variable was the “true” status. Only covariates that reflected management practices related to an external source of animals, proximity to other barns, within-barn mixing patterns, level of biosecurity, and demographic information were used as potential covariates in a series of univariable models. Covariates significant at P < 0.10 on the basis of the likelihood ratio χ² in the univariable analyses were considered as potential candidates for the multivariable model, which was built manually. Quadratic effects of quantitative variables and biologically plausible interactions were evaluated. The choice of the best model was based on the number of observations and on the Akaike Information Criterion (AIC). The multivariable model was fitted for finisher-herd status only, since we had a low number of antibody-negative sow herds. The fit of the final model for finisher-herd positivity was evaluated by the Hosmer–Lemeshow test. Outliers, leverage, and potentially influential points on the basis of 1 per observation were examined by Pearson residuals and dfbeta statistics. Models were refitted after deletion of influential observations, and parameters were compared. In addition, spatial clustering of Pearson residuals was examined descriptively and by the spatial scan statistic. Models were built in SAS 8.2.

We used the same model-fitting strategy to analyze the sow-level data and considered the same herd-, CD-, and CCS-level covariates as for the sow-herd-level logistic regression. In addition, parity was used as the only animal-level factor. Univariable models for the sow H1N1 SIV status were fitted as logistic regression models within the generalized linear mixed-effects model framework (25) with the use of farm as a random effect and predictive quasi-likelihood with a 1st-order approximation (PQL1) in SAS 8.2. Candidate multivariable models were fitted within the generalized linear mixed-effects model framework with farm and parity as random effects. Parity was centered at the median. Quadratic effects of quantitative variables and biologically plausible interactions were evaluated. The final random coefficient logistic regression model was fitted by means of PQL1 in SAS 8.2 and MLwiN 2.0 (Institute of Education, London, England), predictive quasi-likelihood with a 2nd-order approximation (PQL2) in MLwiN, maximum likelihood (ML) in Stata, and the Bayesian method with Markov-chain Monte Carlo estimation (MCMC) in MLwiN. The PQL estimations were based on reiterative generalized least squares, the ML estimate was based on the adaptive quadrature, and the MCMC estimation was based on a chain length of 2 million iterations with 500 000 burn (number of initial iterations that were discarded) and default parameters for prior distribution and sampling type. Coefficients from the final model based on different estimation procedures were compared. Statistical significance of the random coefficient for parity was assessed by a 2-sided Wald test except for the ML method, where the likelihood ratio test was used. The choice of the final model was based on AIC criteria and evaluated by checking the distributional assumptions of level 1 (sow) and level 2 (herd) residuals and influential statistics in MLwiN. Farm-specific estimates of logit for gilts (parity = 0) and slopes from a conditional model (SAS 8.2) were produced and summary statistics calculated according to the 3-group membership based on cluster analysis.

Results

Descriptive characteristics of the study population

Of the 99 swine operations included in the study, 53 were visited in both years, 23 only in 2001, and 23 only in 2003. This dynamic study population was due to (a) submission of samples from a serum bank established before the study period for the Ontario Swine Sentinel Project was completed for both years and (b) dropout and inclusion of new farms. The sows and finisher pigs at the 76 operations (defined by ownership) included in 2001 were sampled on 93 farms (defined by location). Table I shows the number of tested animals, herds, and farms in 2001 and 2003. In 2001, both age groups (sows and finisher pigs) were sampled at 61 swine operations and 44 swine farms.

Table I.

Pig- and herd-level apparent and true seroprevalence of selected strains of swine influenzavirus (SIV) in Ontario, 2001–2003

	Herds				Herds per farm, a 2001
	2001		2003		Sows only H1N1	Sows and finishers H1N1	Finishers only H1N1
Variable	Sows H1N1	Finishers H1N1	Sows H3N2, CO	Sows H3N2, TX
Pig-level prevalence
No. of samples tested	1300	720	1140	1140	420	1320	280
No. of herds/farms tested	65	72	76	76	21	44	28
No. of positive samples	794	175	7	8	250	676	43
Prevalence, % (and 95% CI)
Apparentb	61.1 (58.4–63.7)	24.3 (21.2–27.6)	0.6 (0.2–1.2)	0.7 (0.3–1.4)	NC	NC	NC
Truec	58.3 (55.4–61.2)	17.6 (15.0–20.2)	0.0 (0.0–3.0)	0.0 (0.0–3.1)	NC	NC	NC
Herd-level prevalence
No. of samples tested per herd/farm	20	10	15	15	20	30	10
No. of positive herds/farms
Apparent	57	34	7d	6d	19	38	9
Truee	54	29	0	0	18	36	7
Prevalence, % (and 95% CI)
Apparent	87.7 (77.2–94.5)	47.2 (35.3–59.3)	9.2 (3.8–18.1)	7.9 (3.0–16.4)	NC	NC	NC
Truee	83.1 (71.7–91.2)	40.3 (28.9–52.5)	NA	NA	NC	NC	NC
Herd-level test assumptions
Expected pig prevalence (%)	30	30	30	30	NA	NA	NA
Cut-off for herd true positivity	> 3 animals	>2 animals	> 2 animals	> 2 animals	NA	NA	NA
Herd sensitivity (%)	97.1	75.9	94.4	94.4	NA	NA	NA
Herd specificity (%)	76.6	79.2	49.2	49.2	NA	NA	NA
Expected herd prevalence (%)	NA	NA	5	5	NA	NA	NA
Cut-off for area true positivity	NA	NA	> 0 herds	> 0 herds	NA	NA	NA
Areaf sensitivity (%)	NA	NA	100 (98.3g)	100 (98.3g)	NA	NA	NA
Areaf specificity (%)	NA	NA	0 (NA)	0 (NA)	NA	NA	NA

CO — Colorado (A/Swine/Colorado/1/77); TX — Texas (A/Swine/Texas/4199-2/98); CI — confidence interval; NC — not calculated (since prevalence would also reflect a ratio between sows and finisher pigs sampled on site); NA — not applicable.

^aReflects the type of samples collected on a particular farm (location) rather than farm type. Type of samples on multisite operations was influenced by farm type and by the presence of finisher pigs of the appropriate age on a premise on the sampling occasion. The description of the study population does not reflect this description. For example, only sows could have been sampled on the farrow-to-finish site if finisher pigs of the appropriate age were not present on the sampling occasion.

^bBased on assumptions of binomial distribution.

^cBased on simple proportion estimates.

^dEleven farms.

^eBased on a herd test.

^fOntario.

^gAssuming that the animal-level specificity was 99% (conservative scenario) and other parameters were held constant. With animal-level specificity of 88.9%, the herd specificity was low, partly influencing high area sensitivity (100%). With 99% test specificity, herd specificity increased, which decreased area sensitivity to 98.3%. Consequently, even with an unrealistically high specificity, there was a 98% probability that the Ontario study sows, and the population they represented, were free from exposure to the 2 test H3N2 strains at a between-herd prevalence of 5% and a within-herd prevalence of 30% in 2003. Geographic representativeness and an outbreak of the reassortant H3N2 in 2005 increased our confidence in the absence of the Colorado and Texas strains in the 2003 target population.

In the 2001 study, 61% of the finisher herds were at farrow-to-finish farms, 35% at finishing-only farms, and 4% at nursery and finishing farms. The mean and median numbers of finisher pigs were 1182 and 900, respectively, the minimum number was 30, and the maximum number was 6000. Of the sow herds, 86% were at farrow-to-finish farms, 8% at farrowing-only farms, and 6% at farrowing and nursery farms. The mean and median numbers of sows were 516 and 300, respectively, the minimum number was 25, and the maximum number was 3000.

In the 2003 study, 84% of the sow herds were at farrow-to-finish farms, 9% at farrowing-only farms, and 7% at farrowing and nursery farms. The mean and median numbers of sows were 478 and 267, respectively, the minimum number was 25, and the maximum number was 2700.

Estimates of prevalence

As Table I shows, H1N1 SIV was endemic in the study herds in 2001, with true herd prevalence of 83.1% and 40.3% for sows and finishers, respectively. In 2003, none of the herds were true-positive for the Colorado and Texas strains of H3N2. Table I also contains the characteristics of the herd-level test for sow and finisher herds in 2001 and for sow herds in 2003, as well as estimates of area-level freedom from infection with the 2 H3N2 strains, apparent herd prevalence, and true herd prevalence in both years.

When the true H1N1 status of the sow and finisher herds was cross-classified, several important findings were apparent (Table II). First, a finisher herd was positive only if the source sow herd was positive. Second, 45% of the finisher herds were H1N1-negative although the source sow herds were H1N1-positive. Third, for the off-site negative finisher herds, 90% of the source sow herds were positive, whereas for the on-site negative finisher herds, only 61% of the sow herds were positive (P > 0.10). Vaccination for SIV was not reported for any herd.

Table II.

Cross-classification of herd true H1N1 SIV status in swine operations at which sows and finisher pigs were tested in 2001

	Finisher herds
Sow herds	Positive	Negative	Total
Positive	28	23	51
Negative	0	10	10
Total	28	33a	61

^aOf these herds, 10 were located separately from the source sow herds (off-site); 27% of the 33 (n = 9) were on-site with negative sow herds, 42% (n = 14) were on-site with positive sow herds, 3% (n = 1) were off-site with negative sow herds, and 27% (n = 9) were off-site with positive sow herds. For the off-site negative finisher herds, 90% of the source sow herds were positive, whereas for the on-site negative finisher herds, only 61% of the sow herds were positive (P > 0.10).

The median and interquartile range for within-herd prevalence were 70% and 55%, respectively, in the sow herds and 0% and 50%, respectively, in the finishing herds. All 7 farms apparently positive for the Colorado strain had only 1 positive animal each. For the Texas strain, 2 farms had 2 positive animals, and 4 others had 1 positive animal each.

The within-herd prevalence of H1N1 for sows and finisher pigs had a 3-modal distribution, with modes at 0%, 50%, and 100% at both ages (Figure 1). In the cluster analysis, most of the distance calculation methods classified a prevalence of < 30% into cluster I, which, for our purposes, may be considered “disease free”. Therefore, a minimum prevalence of 30% should be used to calculate the sample size needed to declare freedom from influenza in swine herds.

Figure 1.

Frequency distribution of rounded (to the nearest 10) within-herd apparent prevalence of infection with swine influenzavirus (SIV) in Ontario sows (upper left) and finisher pigs (lower left) tested in 2001 for antibodies to an H1N1 strain by enzyme-linked immunosorbent assay and in sows tested in 2003 for antibodies to the H3N2 Colorado strain (upper right) and the H3N2 Texas strain (lower right) by hemagglutination-inhibition assay.

On the basis of biologic plausibility, as confirmed by an examination of statistical criteria, the sow and finisher herds were classified into 3 distinct clusters according to their within-herd apparent prevalence for H1N1 SIV: for the sow herds, 0% to 35% (cluster I, “disease free”), 40% to 75% (cluster II, medium prevalence), and 80% to 100% (cluster III, high prevalence) (Figure 2); and for the finisher herds, 0% to 30%, 40% to 70%, and 80% to 100%. Using solely statistical criteria, it might, however, be preferable to divide these low-prevalence herds into 2 subclusters, for a total of 4 clusters for sow and finisher herds: the examined statistics in part showed a higher degree of levelling-off at 4 clusters than at 3 clusters when plotted against the potential number of clusters, a criterion recommended by some authors (25).

Figure 2.

Clustering of the within-herd apparent prevalence for the H1N1 SIV subtype in the sow herds. The prevalence increases from the upper left toward the lower left.

Statistical models

Table III reports the results of the univariable models for H1N1 SIV positivity in the sow and finisher herds. The odds of sow-herd positivity were higher for herds with an external source of gilts and with other farms or pig farms nearby. There was a curvilinear association of herd positivity with the number of sows on the farm. The expected probability of herd positivity was lower for herds with minimum and maximum numbers of animals; the highest expected probability was at approximately 1400 sows. The odds of herd positivity increased as the density of pigs and pig farms increased at the CD and CCS levels for sow herds but only at the CCS level for finisher herds. In addition, the odds of finisher-herd positivity were higher if the source sow herds were positive and if they were not a part of an integrated company, whereas the odds were lower if there were only finisher herds on the sampling sites and if the herds were managed all-in/all-out (AIAO) by site.

Table III.

Factors associated with H1N1 SIV positivity of the sow and finisher herds according to univariable estimates of odds ratio (OR)

Risk factora	OR	95% CIb	Pb	n
For sow herds
External source of gilts (70%)	4.62	(1.14–20.91)	0.03	59
Distance to nearest pig barn (km)			0.01	60
< 1 (28%)	7.32	(1.49–55.15)
1–3 (42%)	10.18	(1.51–204.38)
> 3 (referent) (30%)	—	—
Number of sowsc			0.05	65
Pig density (100 pigs/km2)
CCS (2.2, 382.2)	1.02	(1.01–1.05)	0.02	62
CD (2.0, 259.3)	1.05	(1.02–1.09)	< 0.01	62
Pig farm density (0.1 farms/km2)
CCS (0.01, 0.81)	32.46	(3.60–1422.26)	< 0.01	65
CD (0.007, 0.29)	317.35	(13.87–> 9999)	< 0.01	65
For finisher herds
Positive source sow herd (83%)	15.96	(2.36–+∝)d	< 0.01	61
AIAO pig flow in finisher barn by site (12%)	0.15	(0.01–0.92)e	0.04	59
Level of integration			0.03	72
Independent (68%)	9.68	(1.68–184.93)
Family loops (15%)	9.21	(1.14–198.34)
Company (referent) (17%)	—	—
Increase in number of finisher pigs by 1000	1.62	(1.03–2.75)	0.04	72
Finishers only (35%)	0.33	(0.93–0.10)	0.04	72
Pig density (100 pigs/km2)
CCS (2.2, 382.2)	1.48	(0.97–2.29)	0.07	69
CD (2.0, 259.3)	1.46	(0.84–2.61)	0.18f	69
Pig farm density (0.1 farms/km2)
CCS (0.01, 0.81)	1.38	(1.01–1.98)	0.04	72
CD (0.007, 0.29)	1.53	(0.89–2.69)	0.13f	72

CCS — Statistics Canada census consolidated subdivision; CD — Statistics Canada census division; AIAO — all-in/all-out.

^aValues in parenthesis and italics are the percentage of the study population in this category or the minimum and maximum values for continuous variables.

^bBased on likelihood ratios unless specified otherwise. For nominal variables with > 2 levels, the P value refers to the overall test of significance.

^cCurvilinearly associated with influenza status, increasing until reaching a maximum of 98% probability at approximately 1400 sows and then decreasing to the starting point of 63% probability at approximately 2800 sows.

^dExact P-value.

^eParameter obtained after the model was fitted with 1 “fudge”-positive AIAO farm added to the data to enable convergence, since no AIAO farms tested positive, and, consequently, likely an underestimate of the protective association for AIAO.

^fReported for comparison purposes, although the P-values were high.

Table IV reports the results of the multivariable model for finisher-herd positivity. The odds of positivity were lower if only finisher herds were on the sampling sites but increased as the number of finisher pigs on a site increased and as the density of pig farms increased in the CCS.

Table IV.

Factors associated with H1N1 SIV positivity of the finisher herds according to OR estimates from the final multivariable model^a

Risk factor	OR	95% CIb	Pb
Finishers only	0.11	(0.01–0.62)	< 0.01
Increase in number of finisher pigs by 1000	4.44	(1.90–13.07)	< 0.01
Pig farm CCS density (0.1 farms/km2)	1.41	(1.00–2.04)	0.02

^aLikelihood ratio χ² = 19.2, 3 degrees of freedom (df), P = 0.0002. Hosmer–Lemeshow goodness of fit: χ² = 5.42, 8 df, P = 0.71 (n = 72).

^bBased on likelihood ratios.

For an average sow herd from the target population, the logit of sow SIV seropositivity increased linearly with parity number (Table V). However, there was a significant variation in the intercepts (representing the parity-3 sows) among herds and also among the farm-specific slopes (representing farm-specific associations). Figure 3 depicts the herd-specific associations between the log odds of positivity and parity, based on the conditional model. Farm-specific estimates are colored by herd membership as determined by the cluster analysis.

Table V.

Factors associated with sow seropositivity to H1N1 SIV according to sow-level subject-specific model point estimates of log OR^a

	Log OR point estimate (and standard error) [with OR (and 95% CI)]b
		MLwiN			Stata (glamm) adaptive quadrature
Risk factor	SAS (glimmix) PQL1	PQL1	PQL2	MCMC
Fixed effects
Intercept	0.15 (0.28)	0.15 (0.28)	0.24 (0.33)	0.13 (0.37)	0.14 (0.34)
Parityc	0.17 (0.06) [1.19 (1.05–1.33)]	0.17 (0.06) [1.19 (1.05–1.33)]	0.20 (0.07) [1.22 (1.06–1.40)]	0.21 (0.08) [1.23 (1.05–1.44)]	0.20 (0.07) [1.22 (1.06–1.40)]
Pig farm CD density (0.1 farms/km2)d	1.49 (0.32)	1.49 (0.32)	1.66 (0.39)	1.89 (0.45)	1.85 (0.42)
Random effects
Intercept	2.97 (0.83)	2.97 (0.75)	3.79 (1.02)	5.33 (1.78)	4.41 (1.40)
Paritye	0.07 (0.04)	0.07 (0.04)	0.09 (0.04)	0.12 (0.06)	0.10 (0.06)
Covariance	0.01 (0.13)	0.01 (0.12)	−0.03 (0.15)	0.05 (0.24)	0.04 (0.21)

PQL1 — predictive quasi-likelihood with a 1st-order approximation; PQL2 — predictive quasi-likelihood with a 2nd-order approximation; MCMC — Markov-chain Monte Carlo estimation.

^aHerd-level predictors examined for inclusion in the model, together with descriptive information, are in Table III.

^bFor 862 animals and 50 herds; all observations with parity number available were included in the model.

^cCentered to parity 3. Mean parity = 3.3, median = 3.0, maximum = 11, standard deviation = 2.54. Mean farm parity ranged from 0 (if only information about gilts was known) to 5.8.

^dCentered to mean density.

^eSignificant at P ≤ 0.05 when calculated on the basis of estimates and standard errors or on the basis of the likelihood ratio test for model estimates by adaptive quadrature (2-sided P-values).

Figure 3.

Subject-specific (conditional) estimates of the association between sow parity and log odds of being apparently positive for classic H1N1 SIV. Conditional mean denotes the estimated association on an average farm and illustrates the fixed part of the model. Random slopes denote the conditional (farm-specific) estimates of the same association. The estimates are based on the random components of the same subject-specific model. In cluster I (n = 14), II (n = 16), and III (n = 20) herds, respectively, the mean expected probability of being an SIV-positive gilt (parity −3 sows) was 19%, 38%, and 81%, the proportion of herds with positive slopes was 50%, 62%, and 30%, and the mean parity was 3.0, 2.9, and 2.8.

Discussion

The apparent seroprevalence of H1N1 SIV was 61.1% and 24.3% in sows and finisher pigs, respectively, comparable to the rates in other pig-producing regions of the world (26–28), although the rates for finisher pigs have varied considerably among studies (29). The current study was conducted during the summer, which precluded the observation of possible seasonal trends (4). The distribution of SIV subtypes agreed with the historical evidence of the predominant circulation of the classic swine H1N1 virus from 1930 until the late 1990s (10,26,27,30) and serologic or virologic evidence of only occasional (26,28) to low (27,31) exposure to H3N2. In our study, herds were reclassified from apparent to “true” herd status after herd-level sensitivity and specificity were considered, and the rate of misclassification was minimized. This resulted in a small reduction in the estimates of herd prevalence.

Using characteristics of the herd test to calculate area (Ontario)-level sensitivity resulted in high confidence in the relative absence of the Colorado and Texas H3N2 strains from Ontario sow herds in 2003: the expected herd prevalence was 5% and the expected within-herd pig prevalence 30%. These findings were based on 1 or 2 positive results for sows per herd that were likely false-positive and due to imperfect test specificity. For the Ontario-level sensitivity, 2 important assumptions were violated. First, animal-level sensitivity and specificity were likely not fixed and known exactly, but when we explored the effect of specificity (by increasing it from 88.9% to 99%), the conclusion did not change. Second, the study population was not a true random sample of the target population. Consequently, our results can only strictly be applied to the study population and the population it represented. However, herd demographics and the geographic distribution of the study population, along with the fact that a Texas-like H3N2 strain was detected in at least 25 Ontario swine herds between April and September 2005 (32), support our evidence that Ontario sow herds were free from the Texas H3N2 strain in 2003 at the prespecified levels.

The primary objective of the cluster analysis was to determine the minimum expected prevalence in a “disease-free” herd since the minimum expected within-herd prevalence is commonly used to calculate the sample size needed to determine freedom from disease at the herd level (22). There was disagreement among statistical decision criteria on the best number of groups (3 or 4) within sow herds. The minimum expected prevalence of 30% was an overestimation if 1 of the subclusters within a cluster of “disease-free” herds indicated something other than false-positive results. The lower of the 2 prevalence estimates was ≤ 5%. As an example of such a scenario, influenzaviruses of wholly human and avian lineages may not transmit efficiently between pigs, and this may result in a lower within-herd prevalence. Alternatively, SIV exposure in gilt source herds may also have influenced these results. With the data at hand it was not possible to explore this question; future studies could investigate disease dynamics in herds naturally exposed to influenzaviruses of nonswine lineages and record information about SIV exposure in gilt source herds. Studies might benefit from more reliable estimates of test sensitivity and specificity, possibly by methods without a gold standard.

The secondary objective of the cluster analysis was to describe the within-herd distribution of apparent prevalence to gain knowledge about transmission of the pathogen (7). The difference between the cluster II and III herds may be explained by the time since the herd was last exposed to influenzaviruses, the rate of sow replacement, or both. The within-herd prevalence of influenza in sow herds is not well described except in an epidemic situation in which the clinical signs mimic those of finisher pigs. However, infected sows are serologically positive for up to 24 mo (15). Although cluster analysis classified the sow herds into groups that made biologic sense, linking the results of the cluster analysis with the results of the random coefficient model suggested a complex transmission of SIV in sow herds. In finisher herds, spread of SIV can be quite slow (5,7), which might explain the cluster of finisher herds in our study with a seroprevalence around 50% and may be of interest for future longitudinal observational studies and theoretical models.

Risk factor analysis was done at the herd and sow level. Owing to the low number of negative sow herds, we did not fit multivariable models for positivity at the sow-herd level. The epidemiologic aspects of SIV infection in the sow and finisher herds were likely different and possibly driven by the number of susceptible animals. Finisher herds were positive only if the source sow herd was positive, and H1N1 positivity in sow herds was a significant risk factor for finisher herd positivity in the univariable model. Furthermore, almost half (45%) of the finisher herds were H1N1-negative although the source sow herds were H1N1-positive. In addition, for the off-site negative finisher herds, 90% of the source sow herds were positive, whereas for the on-site negative finisher herds, only 61% of the sow herds were positive. Hence, it seems that transmission of infection from positive sow herds could be largely avoided by locating the herds on different premises (farms); for example, by moving weaned pigs. Alternatively, sow herds may be exposed to SIV from nursery pigs. This would be consistent with the longevity of sows in a herd and long-lasting immunity to SIV. De Jong et al (33) suggested that the supply herds (nursery pigs) or finisher herds are the most likely sources of SIV infection in the finisher herds.

The most important factor for incident cases of influenza at the farm level is the introduction of animals carrying the infection into a naïve herd (4). More pigs are raised in North America in multisite systems than in traditional single-site farrow-to-finish operations. Multisite loop systems have a directed and planned flow of animals, usually AIAO by barn. In this study, the external source of finisher pigs was defined as pigs coming from more than 1 nursery. The external source of nursery pigs was defined as pigs coming from different sow barn locations within the same system. This is perhaps why, other than sampling variation, we were unable to find any significant association between multiple sources of finishers or nursery pigs and finisher farm positivity, an association found in other studies (34).

Increasing the number of finishers by 1000 increased by 4.4 the adjusted odds of a finisher herd being positive. The direction of the association agreed with the observation of Gardner et al (35) that herd size is positively associated with an increased probability of respiratory disease. Those authors discussed biologic reasons for the effect. The most likely explanation is an increased risk of influenzavirus introduction and transmission within the herd. In the simplest scenario, the number of animals in the finisher barn is likely positively associated with the number of newly introduced animals that, if susceptible, could maintain persistent infection within the herd.

Pig farm and pig density were statistically associated with positive SIV status for finisher herds only at the CCS level. Several studies have identified density or number of pigs in the area as a significant risk factor (29,36), but little is known about the mechanism of influenza spread between farms, other than introduction of infected animals. Wind (4,15) and possibly the spread of pig slurry (15) may both play a role. The association that we found between farm or pig density and SIV status of farms agrees with the association found in other studies (36) and general knowledge about swine influenza (4). Other factors indicative of increased farm proximity were associated with a positive sow herd status. Finisher-only facilities were 0.11 times as likely to be SIV positive as farms with multiple age groups. Finisher-only sites are more likely to be completely or partially depopulated and, if they have good biosecurity practices, should be more likely to avoid the introduction of influenza infection either from another age group within the site or from some form of indirect contact. Management AIAO by site was another protective factor for influenza finisher positivity, decreasing the odds to 15% of those for finisher facilities operating on a continuous-flow or AIAO-by-room basis. Finishers only and AIAO by site were confounded when evaluated in the same model (not shown); however, the agreement was not perfect between these 2 management practices since some farms housed only finisher pigs on 1 site and ran the site by continuous flow. Therefore, flow was an important factor for maintaining an SIV-negative status in finisher barns. Alternatively, it is possible that SIV shows different epidemiologic features in finisher barns of different management styles during the fall and early winter. Since our sampling was predominantly in the summer, we cannot exclude a seasonal effect as partly responsible for the reported associations.

The conditional model was fitted for the positivity at the sow level. There was a linear association between parity and logit of apparent sow-level H1N1 seropositivity. This model was fitted to examine not only the estimates of association on a “typical” farm (represented by the fixed effect) but also the farm-specific estimates in the study population. In a typical sow herd, the odds of being SIV-positive increased by 19% as the parity increased by 1 unit. There was significant variation in the linear association between parity and SIV status seen as the variance of the slope. These findings were consistent when data were analyzed by different estimation methods.

Several groups of herds were observed on the basis of the magnitude of intercepts (logit and probability of SIV positivity in gilts) and the direction of slopes (change in logit of SIV positivity with increase in parity). The 1st group had low to medium logit for intercepts and horizontal or negative slopes. These could be herds in which exposure to SIV did not occur and the SIV-test-positive results were false-positive. Alternatively, these could be herds in which exposure to SIV occurred but sows were younger (of lower parity) at the time of exposure. The 2nd group also had low to medium logit for intercepts and positive slopes. These could be herds in which exposure to SIV occurred as a point outbreak, after which the virus stopped circulating. Consequently, only higher-parity sows would have evidence of exposure to SIV, resulting in positive slopes. The drop in the proportion of SIV-seropositive animals within herds has been associated more with herd dynamics (18) than with the disappearance of antibodies. Desrosiers et al (15) recently showed that exposed sows could maintain high titers for extended periods, which supports our results. Again, depending on the magnitude of intercepts but also the magnitude of these positive slopes, these herds would be classified as cluster I or II herds. Cluster I herds with a low intercept and positive slopes could alternatively be a consequence of higher cross-reactivity of serum from higher-parity sows. Finally, the 3rd group of herds had high intercepts and different directions of slopes. In these herds, the infection may be relatively recent and all animals equally exposed, or else the infection was occurring continuously. They generally corresponded well to cluster III herds.

Lack of perfect agreement between groupings based on cluster analysis and based on estimates from the random coefficient model was to be expected since the latter analysis accounts for the additional component of within-herd dynamics, which, at least in part, may be a consequence of historical exposures to influenzavirus in a herd. Introduction of gilts from breeding herds with varying exposure to influenzavirus may also have been a factor. Because of the study design, there is a possibility that our parity effect was confounded by either period or cohort effect; only by conducting a cohort study would this concept be accurately addressed.

A limitation of this analysis was a lack of parity information for a considerable portion of the population. Parity is generally easily determined in sow herds, and collection of this information in cross-sectional studies could enable more powerful data analysis and guide further investigations with little additional cost. In addition, other approaches to group prevalence data, including model-based clustering, should be explored. Factors identified in this study as significant are based on prevalence data and on endemic subtype. The epidemiologic aspects of epidemic versus endemic SIV subtypes need to be considered in future investigations. Another set of limitations is a consequence of assuming fixed sensitivity and specificity: this may have produced overly optimistic estimates of confidence in the absence of H3N2 strains in 2003.

We conclude that H1N1 SIV was endemic in study herds in 2001, whereas no herds were true-positive for the Colorado and Texas H3N2 strains in 2003. We recommend a minimum expected within-herd prevalence of 30% to calculate the sample size needed to determine freedom from disease at the herd level. Exposure of animals to SIV in commercial sow herds is likely more complex than population average models and estimates from fixed parameters suggest. An infected source sow herd increased the risk of positivity of a finisher herd, but management may reduce the risk of virus transmission from a sow herd to the finisher herd.