Abstract
The H3N2 swine influenza virus appeared for the 1st time in the province of Manitoba during the fall of 2004. The purpose of this study was to characterize how swine influenza moved through the province in time and space, and to determine if there are any significant patterns associated with this movement. Herds with outbreaks of H3N2 swine influenza were located by using a Geographic Information System and analyzed by using spatial analysis software. Descriptive and spatial statistics, including the Nearest Neighbor Index, Cuzick and Edwards’ test, Spatial Scan statistic, k Nearest Neighbor test, Knox’s test, and Mantel’s test, were used to analyze the outbreak. There was clustering of the outbreak in the region surrounding Steinbach, Manitoba, an area that is densely populated with swine. It is hypothesized that density of swine farms was a factor in the clustering and movement of this swine influenza outbreak.
Résumé
Épidémiologie spatiale d’un flambée d’influenza porcin H3N2. Le virus de l’influenza porcin H3N2 est apparu pour la 1ère fois dans la province du Manitoba au cours de l’automne 2004. Le but de cette étude était de caractériser la façon dont l’influenza porcin s’est déplacé dans le temps et l’espace à travers la province et de déterminer s’il pouvait y avoir des modèles significatifs associés à ce déplacement. Les troupeaux présentant des flambées d’influenza porcin H3N2 ont été localisés en utilisant un système d’information géographique et analysés à l’aide d’un logiciel d’analyse spatiale. Des statistiques descriptives et spatiales dont l’index du voisin le plus près, le test d’agrégation globale par proches voisins, la méthode de Scan pour la localisation des clusters spatio-temporels, le test du voisin le plus près, le test de Knox et le test de Nantel ont été utilisés dans l’analyse de cette flambée. Il y avait un groupement de l’éclosion dans le voisinage de Steinbach au Manitoba, une région à dense population porcine. Nous avons émis l’hypothèse que la densité des porcheries constituait un des facteurs associés au groupement et à la dissémination de cette éclosion d’influenza.
(Traduit par Docteur André Blouin)
Introduction
The H3N2 swine influenza virus is a subtype of influenza A virus that has recently emerged in North America. Prior to 1998, the H3N2 subtype had a low prevalence in swine in the USA (1) but, since then, the H3N2 virus has swept across the USA (2) and, in 2005, it became established throughout Canada (3). A swine-like H3N2 influenza virus has also been isolated from turkeys (4). Pigs have sialic acid receptors that are capable of binding both human and avian influenza viruses and, it is theorized, are involved in the development of new strains of influenza virus (5). The activity of swine and all influenza viruses is of concern because of the viruses’ zoonotic capacity and the possibility that a new strain of influenza virus could cause a human influenza pandemic.
One of the methods for monitoring the activity of disease is with Geographical Information Systems (GIS) and software that can analyze disease and its relationships in space by using spatial epidemiology. Spatial epidemiology or spatial data analysis is a process with 3 steps: data visualization, data description and exploration, and data modeling (6). Data visualization involves placing data on a map and subjectively looking for meaningful patterns, as John Snow did in producing the famous Broad Street spot map in the cholera outbreak in London, England, in 1849 (7). This map showed a significant clustering of cholera cases in the St. James Parish and, particularly, in the area surrounding the Broad Street pump. However, the weakness of data visualization is that it cannot indicate whether an aggregation of cases is due to a large number of cases or a large number of individuals in an area.
Data description and exploration overcome the weakness of data visualization through the use of spatial analysis statistics that look for spatial structure in the data and determine whether a spatial structure is real, significant, and clustered, rather than random and nonsignificant. A cluster of disease is defined as closely grouped cases of disease with a well-defined distribution pattern in relation to time, place, or both (8). This clustering or grouping occurs in an amount that is greater than what could be expected by chance alone.
Spatial analysis statistics (Figure 1) can be divided into focused tests and general tests, based on the criterion of whether the location of clustering of a disease is known or unknown (9). Focused clustering tests are used when the location of a cluster is known or suspected. These statistics test for an elevated risk of disease around the source; they can also be used to develop models about the progress of a cluster or for estimating the risk of a disease at different distances around the cluster (10). These tests are not commonly used in veterinary epidemiology. General clustering tests are used when the location of a cluster is not known; they are further subdivided into global and localized tests (11). Global clustering tests evaluate the relationship between cases and controls; clustering is present if cases are closer to other cases than they are to controls. Alternatively, clustering can be determined by distance between cases (12). Localized clustering tests divide the study area into subsets, called quadrats; they indicate significantly elevated rates of disease in these quadrats and identify these areas as a cluster.
Figure 1.
Overview of spatial analysis statistics.
Other clustering tests include temporal statistics, which detect clusters of disease in time. Space-time interaction tests detect clustering in both space and time by looking for nearby cases that occur at about the same time.
The final step in spatial epidemiology is data modeling. Data modeling is used in testing hypotheses and in health utilization planning; it determines if there is a correlation between variables of interest and variables on a map. Cook and Pocock (13) modeled heart disease in Great Britain on explanatory variables that included environmental and socioeconomic factors. Disease incidence in a region was found to be spatially correlated to that in neighboring regions, and this was accounted for in the model. In this project, data modeling was not performed.
The purpose of this paper is to report on how an outbreak of H3N2 swine influenza in the province of Manitoba, which started during the fall of 2004, moved through time and space.
Materials and methods
The Manitoba Agriculture, Food and Rural Initiatives’ Veterinary Services Laboratory diagnosed H3N2 swine influenza during the fall of 2004. No previous diagnoses of H3N2 swine influenza had been made previously in Manitoba. Veterinarians from 4 organizations collaborated by sharing data on the locations and characteristics of the farms diagnosed as having H3N2 swine influenza. These organizations were Elite Swine Incorporated, Hytek Limited, The Puratone Corporation, and Sheridan, Heuser, Provis Swine Health Services.
A case was defined as “the outbreak of H3N2 swine influenza in a swine herd diagnosed by either real time polymerase chain reaction (RT-PCR), serologic testing, clinical signs, or a combination of these 3 methods.” An RT-PCR for swine influenza virus was performed on nasal swabs, samples of lung tissue, or both, and analyzed by using gel electrophoresis and photography. Pairs of primers with oligonucleotide sequences specific for the H3 hemagglutinin gene and the N2 neuraminidase gene of the H3N2 swine influenza virus were used (A. Hamel, personal communication). Serologic testing was used in cases where the H3N2 swine influenza virus was not diagnosed by RT-PCR and as an adjunct to other serological tests for other respiratory pathogens. An ELISA (HerdChek Swine Influenza H3N2 Antibody Test Kit; IDEXX Laboratories, Westbrook, Maine, USA) with a specificity for antibodies to H3N2 swine influenza virus reported to be between 98.9% and 99.2% was used. Acute and convalescent blood samples were drawn 3 to 4 wk apart and an increase in antibody titer to H3N2 swine influenza virus was used to declare a herd positive. Clinical signs used were those of classical swine influenza, which included a rapid onset of and recovery from the outbreak, a dramatic “barking” cough, high fever, serous nasal discharge, lethargy, and anorexia. Clinical signs were used in cases where the H3N2 swine influenza virus could not be isolated by RT-PCR, and where the veterinarian was confident, based on testing for other respiratory pathogens (including H1N1 swine influenza virus), that the outbreak was due to H3N2 swine influenza virus.
The longitude and latitude of farms with H3N2 swine influenza were plotted from the end of the driveway of the farm. Either a Global Positioning System (GPS) handheld unit and software (Emap and MapSource; Garmin Limited, Olathe, Kansas, USA) or a Geographic Information System (GIS) mapping software (ArcMap 9.1; ESRI, Redlands, California, USA) was used to locate the farm on an aerial map, based on the Dominion Land Survey Land Description (DLS) coordinate, and to visually plot the end of the farm driveway from the aerial map.
A case file was compiled; farm names were removed and replaced with a code to ensure confidentiality. Data collected included site type (sow or nursery/finisher), number of barns per site, herd size (number of sows or number of pigs per farm), DLS coordinate, latitude, longitude, and date of onset of clinical signs. The farm outbreaks were also classified as animal spread or area spread, based on whether the outbreak was thought by the veterinarian to have been caused by incoming animals from an infected herd or by the swine influenza virus moving through the area.
A composite file of all (n = 1506) swine farms located in Manitoba was obtained from Manitoba Pork Council, with farm names removed from this list and replaced with a code to ensure confidentiality. The location of farms in this file was indicated by the Rural Municipality (RM) and the Land Description; from the Land Description, a coordinate was geocoded that would be accurate within 0.4 km. This composite file served 2 purposes: First, it provided a denominator to calculate the incidence of disease by farm in each RM; this information was stored in an incidence file, which was used in the calculation of the Spatial Scan statistic. Second, it enabled a control file to be established from the composite file by identifying, based on Land Description, the farms with an outbreak of influenza, and removing them. This control file was used for the calculation of the Cuzick and Edwards’ statistic.
The case files, incidence files, and control files were compiled on a spreadsheet (Excel; Microsoft Corporation, Redmond, Washington, USA) and then imported into spatial analysis software (ClusterSeer 2; Biomedware, Crystal Lake, Illinois, USA) for spatial analysis, and into SaTScan (14) for analysis with the Bernoulli model Spatial Scan statistic.
The epidemic was visualized in time, by constructing an epidemic curve, and in space, by constructing a map sequence. The epidemic curve of the number of cases versus time was used to describe the temporal aspects of the epidemic and to help in investigating the mode of transmission. A time interval of 5 d was selected for the epidemic curve. Finally, a map sequence, or series of spot maps, was constructed to visualize the epidemic moving spatially over time, and a time interval of 2 wk was selected for the map sequence.
The epidemic data were described and explored through descriptive spatial statistics and through several spatial analysis tests. Descriptive spatial statistics, as described elsewhere (15), are used to describe the area of an outbreak and can be used to compare an outbreak with other outbreaks. The arithmetic mean center of the outbreak was calculated by taking the means of the latitudes and the means of the longitudes and plotting the point of intersection of these 2 means. The standard distance deviation is used to define the degree of the dispersion of the points in the epidemic; it is calculated by taking the square root of the summation of the variances of the latitudes and longitudes from the arithmetic mean. The Nearest Neighbor Index was calculated to determine the distribution of the points in space. The Nearest Neighbor Index is represented by the variable R, and is calculated by the ratio of the mean Euclidean distance observed between the nearest neighbor points in an area to the mean distance expected if these points were randomly distributed (16). The values for R range from 0 to 2.15, where 0 indicates clustering, 1 indicates a random distribution, and 2.15 indicates a uniform distribution with equal distance from all neighbors and maximal dispersion throughout the study area. The Z Statistic is the test of statistical significance of the Nearest Neighbor Index. These were calculated by placing all the coordinates in a spreadsheet (Excel; Microsoft Corporation) on 2 axes and calculating the Euclidean distance between each point by using the Pythagorean theorem. The minimum value function was used to indicate the nearest neighbor for each point, and the Nearest Neighbor Index and Z Score were calculated according to formulas described elsewhere (16).
Since the area of clustering was not known in this outbreak, focused clustering tests were not used. The global clustering test used was the Cuzick and Edwards’ test, which analyzes how cases and controls are distributed relative to each other; its calculations have been described elsewhere (17). The statistic was run to the 5th nearest neighbor, a Monte Carlo simulation with 999 random outcomes was generated and the Statistical Distance Test statistic was used to determine the significance. A large test statistic suggests clustering and cases will tend to have cases as neighbors; if the test statistic is small, it suggests nonsignificance and cases will tend to have controls as neighbors.
The localized clustering test used was the Spatial Scan statistic using a Poisson model, as described elsewhere (12). Each quadrat in the study area has a centroid, the defined center of the quadrat. A circular window is generated by the software and progressively scanned across the study area. If a centroid is within the window, its corresponding quadrat is included in the calculation. Once the circular window has passed across the entire study area, the radius of the circular window is enlarged and the entire process is repeated. An upper limit, usually 50% of the population, is set to prevent the entire study area from being incorporated into the circle, which would make a cluster meaningless. The Spatial Scan statistic compares the incidence of disease inside the circular window with the incidence that would be expected outside the circular window, if the disease was distributed randomly throughout space. A Monte Carlo simulation, where the cases are randomly distributed for 999 permutations, is used to help in determining the significance of the clusters. The quadrat used in this study was the RM, and the centroid used was the calculated geographic center of each RM. Municipalities that contained 10 or fewer farms were aggregated with adjacent RMs with the fewest number of farms until a minimum of 10 farms was reached. This was done in order to prevent skewing of the incidence rate. A Spatial Scan statistic using a Bernoulli model was also applied to the data. The Bernoulli model uses case and control point data, instead of rates of disease, and applies the same scanning technique to identify farms where a disease is clustering.
For clusters detected by the Spatial Scan technique, the characteristics of herds inside the cluster were compared with those of herds outside of the cluster to determine if population, as measured by herd size or number of barns, was a factor in the clustering of H3N2 swine influenza cases. This was done by comparing the average number of barns per site and the average number of animals per barn inside the cluster versus outside the cluster. Since the number of barns per site was not normally distributed, it was tested by using the Mann-Whitney test, while the significance for the number of animals per barn was tested by using Student’s t-test. Type of site (sow or nursery/finisher) and type of spread (animal or area) were compared between barns inside and outside the cluster by calculating the odds ratios for sow versus nursery/finisher site type and for animal spread versus area spread.
Space-time interaction tests performed included Mantel’s test, the Knox test, and the k Nearest Neighbor test. Mantel’s test multiplies the time distances by the spatial distances and sums these products across all case pairs (18). In a contagious disease scenario, there would be a correlation between small distances in space and time, but not between large distances. The Knox test detects space-time interactions by categorizing points into near and far in time versus near and far in space, based on critical space and time distances set by the user (19). For this study, a space cut off of 3.2 km, based on the theory that Mycoplasma pneumonia can travel this distance, and a time cut off of 4 d, equivalent to the latency period for swine influenza (1,20), were selected. The k Nearest Neighbor Test uses the relationship of the nearest neighbor rather than geographic distance, which may be affected by nonuniform population densities (18). There was an insufficient number of cases to perform any temporal tests.
Results
A total of 115 herds were diagnosed as having H3N2 swine influenza; however, 21 herds were missing time information and were not used in some analyses. Complete space and time information was available for 94 herds, and the characteristics of these 94 case herds are summarized in Table 1. The number of herds served by each veterinary organization is listed in Table 2.
Table 1.
Summary of characteristics of case farms
| n | s | |
|---|---|---|
| Total herds | 94 | — |
| Site type | ||
| Sow | 47 | — |
| Finisher | 47 | — |
| Number of barns per site (average) | 1.7 | 1.52 |
| Number of animals per barn (average) | 2755.6 | 1855.55 |
| Spread type | ||
| Animal | 31 | — |
| Area | 59 | — |
| Unknown | 4 | — |
n — number; s — standard deviation
Table 2.
Summary of cases of H3N2 by organization
| Organization | Number of cases |
|---|---|
| Elite Swine Incorporated | 27 |
| Hytek Limited | 18 |
| The Puratone Corporation | 22 |
| Sheridan, Heuser, Provis Swine Health Services | 27 |
An epidemic curve of the number of cases versus time was plotted (Figure 2). The disease first appeared in mid-September 2004 and several cases occurred during the fall of that year, but the outbreak coincided with the colder months of the year, as is typical of influenza. The epidemic curve shows 2 peaks in the epidemic, 1 at the end of January 2005 and 1 at the end of April 2005. The epidemic subsided at the end of April, but the virus continued to circulate, as several cases occurred during the spring and summer months of 2005.
Figure 2.
Epidemic curve, H3N2 swine influenza in Manitoba, 2004–2005.
The points representing infected farms on a map sequence showed the distribution of the disease over time. The 1st case of H3N2 swine influenza was in a herd southeast of Steinbach, on September 17, 2004. By the end of December 2004, 5 herds were positive in an area south and southeast of Steinbach. By mid-January 2005, there were outbreaks in herds north of Steinbach and in a herd north of Winnipeg. By the end of January 2005, there was a peak in the epidemic, with outbreaks in herds encompassing the Steinbach area and in 1 herd south of Winkler, adjacent to the US border. By mid-February 2005, there were outbreaks in western Manitoba, north of Brandon. By the end of April, there was another peak in the epidemic with a column of outbreaks extending from the Interlake region to the United States’ border. After April 2005, various outbreaks occurred north and south of Winnipeg. Figure 3 shows the affected RMs.
Figure 3.
Affected rural municipalities, H3N2 swine influenza in Manitoba, 2004–2005.
Descriptive spatial statistics were used to describe the area of the outbreak. The arithmetic mean center of the outbreak was at latitude 49.61734° north and longitude -97.31929° west, just northeast of the village of Osborne, about 30 km due south of Winnipeg (Figure 4). The standard distance deviation was equal to 1.13932 degrees, and included an area from the mean north into the south Interlake region, east into Steinbach, and south to the US border (Figure 4). The standard deviation has an oval shaped formation, because at this point of latitude, distance in latitude is longer than distance in longitude, so there is a greater distance in the north-south plane on maps using conventional projection. The Nearest Neighbor Index had a value of 0.41, which indicates a clustered distribution, and a Z Statistic of 11.13, which indicates that this clustered distribution is significant (P > 0.0005). The Nearest Neighbor Index was also calculated for the 21 herds that were missing time information; its value of 0.28 and Z Statistic of 6.1915 also indicate a clustered distribution (P > 0.0005).
Figure 4.
Mean and standard deviation, H3N2 swine influenza in Manitoba, 2004–2005.
The results of the spatial analyses are listed (Table 3). There was clustering according to the global clustering test of the Cuzick and Edwards’ statistic. The output of the spatial analysis software shows the relationship between cases and controls (Figure 5).
Table 3.
Summary of spatial data analyses
| Test | Statistic | P-value | Clusteringa |
|---|---|---|---|
| Nearest Neighbor Index | |||
| 94 herds with time data | 0.41 | > 0.0005 | S |
| 21 herds no time data | 0.28 | > 0.0005 | S |
| Cuzick and Edwards | 6.42 | 0.001 | S |
| Spatial scan | |||
| 1st: Stuartburn, La Broquerie | 0.003 | S | |
| 2nd: Cameron/Glenwood | 0.996 | NS | |
| 3rd: Woodlands, Portage La Prairie, | 1.000 | NS | |
| Armstrong | |||
| Mantel Test | 0.03 | 0.291 | NS |
| Knox Test | 5.00 | 0.004 | S |
| k Nearest Neighbor | 19.87 | 0.001 | S |
S — statistically significant; NS — non-significant
Figure 5.
Cuzick and Edwards’ map plot, H3N2 swine influenza in Manitoba, 2004–2005.
A total of 29 RMs were affected in this swine influenza outbreak. Eight of these RMs had 10 or fewer farms, according to the composite file from Manitoba Pork Council; therefore, these RMs were aggregated with a neighboring RM with the fewest farms. The Spatial Scan statistic using a Poisson model detected the 3 most likely clusters of swine influenza (Figure 6); the 1st cluster in the RMs of La Broquerie and Stuartburn was significant. The 2nd most likely cluster in the aggregated RM of Cameron and Glenwood was insignificant, and the 3rd most likely clustering in the RMs of Woodlands, Portage la Prairie, and Armstrong was also insignificant.
Figure 6.
Spatial scan (Poisson), H3N2 swine influenza in Manitoba, 2004–2005.
When the Bernoulli model was used in the Spatial Scan statistic, a significant cluster was located in an area extending from southeast to north of Winnipeg.
For the space-time interaction tests, the Mantel test was insignificant. The Mantel test statistic would be affected by the nonlinearity of the outbreak in time and space, by the skewing due to large distances involved in the outbreak, and by the nonuniform population density of farms, as there is a higher density of farms in southeastern Manitoba compared with western Manitoba. The Knox test and k Nearest Neighbor test are not affected by nonuniform population densities, and both these statistics were significant.
Characteristics of farms inside the clusters identified as significant by the Spatial Scan Poisson model were compared with those of farms outside the cluster. A total of 21 farms were located within the cluster of the RMs of Stuartburn and La Broquerie, while 73 farms were located outside the cluster. Farms inside the cluster had on average a larger number of barns per site and a larger herd size than farms outside the cluster, although both these factors were nonsignificant (Table 4). Farms inside the cluster were 1.4 times more likely to be sow barns than nursery/finisher sites (Table 5). The role of area spread in the Poisson model was no different within the cluster than outside the cluster, as farms within the significant cluster of Stuartburn and La Broquerie were 0.97 times more likely to be classified as area spread than farms outside the cluster (Table 5).
Table 4.
Comparison of characteristics of farms inside versus outside the cluster
| Inside cluster | Outside cluster | Significance | |
|---|---|---|---|
| Poisson model | |||
| Average number of barns | 2.38 | 1.45 | U = 669 (P = 0.216) |
| Average herd size | 3071.4 | 2664.7 | t = 0.88 (P = 0.379) |
| Bernoulli model | |||
| Average number of barns | 1.8 | 1.3 | U = 683.5 (P = 0.102) |
| Average herd size | 2877.1 | 2380.4 | t = 1.12 (P = 0.267) |
Table 5.
A 2 by 2 table of spread type and site type for farms inside and outside the cluster (Poisson model)
| Inside cluster | Outside cluster | |
|---|---|---|
| Spread type | ||
| Area | 13 | 46 |
| Animal | 7 | 24 |
| Site type | ||
| Sow | 12 | 35 |
| Nursery/finisher | 9 | 38 |
Characteristics of farms inside the Bernoulli model Spatial Scan cluster were compared with those of farms outside the cluster. Farms inside the cluster had a larger number of barns per site and a larger herd size compared with farms outside the cluster; however, these factors were nonsignificant (Table 4). Farms inside the cluster were 1.8 times more likely to be nursery/ finisher sites than sow barns (Table 6). Farms within the cluster were 1.7 times more likely to be classified as animal spread than farms outside the cluster (Table 6).
Table 6.
A 2 by 2 table of spread type and site type for farms inside and outside the cluster (Bernoulli model)
| Inside cluster | Outside cluster | |
|---|---|---|
| Spread type | ||
| Area | 42 | 17 |
| Animal | 25 | 6 |
| Site type | ||
| Sow | 33 | 14 |
| Nursery/finisher | 38 | 9 |
Discussion
Geographical Information Systems and spatial data analysis are powerful tools for analyzing disease outbreaks. Spatial analysis has been used to concentrate diagnostic and preventive resources in areas of disease clustering (21), to identify the presence of epidemics through surveillance (22), to generate hypotheses regarding the causes and transmission of epidemics (23), and to help in the management of foreign animal disease outbreaks (24).
An epidemic curve of the number of cases versus time was plotted to help in describing the temporal aspects of the epidemic and to help in investigating the mode of transmission. Previous studies (25,26) on an outbreak of anthrax in Australia constructed an epidemic curve based on the index cases of anthrax on farms. Based on the pattern of the outbreak, a periodicity corresponding to the incubation period of anthrax was noted, and a multiple source outbreak with the possibility of propagative spread was postulated. Selection of the time interval displayed in an epidemic curve is critical, as time intervals that are too short overemphasize randomness in the pattern, while time intervals that are too long aggregate too many groups together and hide the true pattern of the disease. Fontaine and Goodman (27) suggested a time interval for an epidemic curve of between one-fourth and one-half of the incubation/latency period, and the incubation period for swine influenza is 1 to 3 d (1). Cliff et al (20) suggested a chain length in human influenza of approximately 4 to 5 d, based on the midpoint of the average latency period plus the midpoint of the average infectious period.
The epidemic curve of this outbreak has a propagated pattern that is associated with diseases that transmit directly between individuals or through an intermediate method. There are 4 main characteristics for the propagated pattern: 1) it includes several generation periods for the agent, 2) it begins with a small number of cases that gradually increases over time, 3) it has a periodicity that coincides with the generation period for the agent, and 4) it has a rapid downslope following the outbreak due to the lack of susceptible hosts (27). The epidemic curve did not demonstrate any periodicity and this is typical of influenza, probably due to the disease’s shorter and more variable serial interval (28), which is the period of time between the analogous phases in successive cases of an infectious disease (8). However, the epidemic curve did show the other 3 characteristics comparable with a propagated pattern of disease. The propagated pattern of disease transmission indicated by this epidemic curve suggests that transmission between individuals (animal spread) and through an intermediate method (area spread) are both important in the spread of H3N2 swine influenza. Bridges et al (23) summarized 2 human influenza transmission studies that investigated outbreaks, one a hospital ward and the other a commercial aircraft. The hospital ward outbreak spread through direct contact and droplet spread, and had an epidemic curve that suggested a point source outbreak with person to person spread. The aircraft outbreak suggested airborne transmission was more likely involved, as the epidemic curve suggested a point source outbreak, although droplet transmission could not be ruled out.
There is a possibility of selection bias in this study, as the veterinarians in the participating organizations are principally located in southeastern Manitoba and may be more likely to work-up cases in this region of the province than cases that are more remotely located. Producers in distant areas of the province may be less likely to summon a veterinarian or utilize the veterinary diagnostic laboratory to investigate a disease outbreak. This selection bias could result in H3N2 swine influenza appearing to be more prevalent in the southeastern region of the province.
There are advantages and limitations in the use of every spatial statistic. The advantages of the Cuzick and Edwards’ test are its ability to change the nearest neighbor order and its high relevance to veterinary medicine because of the nonuniform distribution of animal populations, as information on population numbers and disease rates is not always readily available (15). Since the Spatial Scan statistic is able to show the location of clusters, as well as test for significance, and is able to correct for multiple testing, it can detect clusters without the bias of first determining size and location of clusters (12). There are several disadvantages to the Mantel’s test: it assumes that time and space in a disease outbreak are linear, it is biased by large distances, and it cannot adjust for variable population densities (18). The Knox test is advantageous because no controls are needed to calculate the statistic, and no knowledge of population distribution is needed. However, the disadvantage of the Knox test is the subjectivity in setting the critical space and time distance, which should be done before calculating the statistic, as was done in this study (19). The k Nearest Neighbor Test assumes that the population does not change in time, although it still has good statistical power with nonuniform population densities. This statistic can be helpful in infectious disease situations because it can help to model a chain of infection through a population (18).
In this study, there were 8 RMs that contained 10 or fewer farms. Quadrats that have a small population have more variability in their incidence or prevalence rates compared with areas with a large population. This is referred to as a small numbers problem (29), as a difference of even 1 case can make a large difference in the calculated rates. The small numbers problem can be compensated for by probability mapping or by empirical Bayes smoothing (29). Probability mapping uses the statistical significance of rates rather than the rates themselves in an area, while empirical Bayes smoothing adjusts the rates in an area based on the size of the population in that area. Both of these techniques require a national or regional rate of disease, which was not available; therefore, these techniques could not be used in this study. Instead, the technique of areal aggregation was used to incorporate neighboring RMs of small population size together, in order to stabilize incidence rates.
Results from several of the clustering tests used in this study suggest that the recent H3N2 swine influenza outbreak in Manitoba was clustered in space and space-time. The incidence of swine influenza by farm in the RMs of Stuartburn and La Broquerie (0.14 farms/day), as shown by the Poisson model of the Spatial Scan statistic, is greater than the incidence that could be expected by chance. These municipalities are in an area of the province that is densely populated with swine and swine farms. The census division containing La Broquerie has 219.8 pigs per square km compared with the census division containing Brandon, in the western part of the province, which contains 15.6 pigs per square km, according to the 2001 census of Statistics Canada (30). These findings agree with those of Vandeputte et al (31) and Gourreau et al (32), who studied outbreaks of H1N1 swine influenza that occurred in densely populated areas of swine barns over large geographical areas in Belgium and France. This was attributed to the role of airborne spread in the transmission of the virus. However, in the Poisson Spatial Scan there was no difference between a farm outbreak classified as animal spread versus area spread for farms located inside the cluster. On the other hand, farms outside the Bernoulli Spatial Scan cluster were more likely to be classified as area spread. Both these scans had a higher average number of barns and a higher average swine population per farm inside the swine influenza clusters, but both these factors were insignificant for both scan types.
The spatial aspects of H3N2 swine influenza in this outbreak appear to model the spatial aspects of influenza in humans. Work by Cliff et al (28) with various influenza outbreaks in many countries found 2 principle characteristics of influenza movement. First, influenza epidemics always establish in densely populated areas, and second, influenza moves from densely populated areas to less densely populated areas. For example, influenza outbreaks in Iceland follow a distinctive pattern of movement from the capital region of Reykjavik, to the regional centers of Isofjordur, Akureyrar, and Seydisfjardar, and then out to the rural hinterland. This type of disease movement from high to low population density is termed hierarchical diffusion and reflects mobility and transportation access (33). In this H3N2 swine influenza outbreak, the disease was first detected in an area densely populated with swine near Steinbach; it spread around the Steinbach region, then to the regions of Brandon, the Interlake, and Morden/Winkler, and finally out into areas less densely populated with swine. However, more epidemics of swine influenza would need to be spatially analyzed to prove the hypothesis that population density is a factor in the development and movement of swine influenza. Based on this hypothesis, effective surveillance strategies for monitoring the development and movement of swine influenza should have a strong focus in densely populated areas.
Acknowledgments
The author thanks Drs. Tim Snider (Elite Swine Incorporated), Brad Lage (Hytek Limited), and Tony Nikkel (The Puratone Corporation) for their collaboration in this study. April Keedian of GeoMap Manitoba provided GIS and mapping services, while Jeff Clark of Manitoba Pork Council provided the composite file of swine herds in Manitoba. The author also thanks Dr. Lisa Lix, Biostatistical Consulting Unit, Department of Community Health Sciences, Faculty of Medicine, University of Manitoba, for guidance in this project.
References
- 1.Olsen CW, Brown IH, Easterday BC, Van Reeth K. Swine influenza. In: Straw BE, Zimmerman JJ, D’Ailaire S, Taylor DJ, editors. Diseases of Swine. 9. Ames: Blackwell Publ; 2006. pp. 469–482. [Google Scholar]
- 2.Webby RJ, Swenson SL, Krauss SL, Gerrish PJ, Goyal SM, Webster RG. Evolution of Swine H3N2 Influenza viruses in the United States. J Virol. 2000;74:8243–8251. doi: 10.1128/jvi.74.18.8243-8251.2000. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Olsen CW, Karasin AI, Carman S, et al. Triple reassortant H3N2 Influenza A viruses, Canada, 2005. Emerg Inf Dis. 2006;12:1132–1135. doi: 10.3201/eid1207.060268. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Choi YK, Lee JH, Erickson G, et al. H3N2 Influenza virus transmission from swine to turkeys, United States. Emerg Inf Dis. 2004;10:2156–2160. doi: 10.3201/eid1012.040581. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Webster RG, Sharp GB, Class EC. Interspecies transmission of influenza viruses. Am J Respir Crit Care Med. 1995;152:S25–30. doi: 10.1164/ajrccm/152.4_Pt_2.S25. [DOI] [PubMed] [Google Scholar]
- 6.Gatrell AC, Bailey TC. Interactive spatial data analysis in medical geography. Soc Sci Med. 1996;42:843–855. doi: 10.1016/0277-9536(95)00183-2. [DOI] [PubMed] [Google Scholar]
- 7.Young TK. Population Health: Concepts and Methods. 2. New York: Oxford Univ Pr; 2005. pp. 15–16. [Google Scholar]
- 8.Last JM, editor. Dictionary of Epidemiology. 4. New York: Oxford Univ Pr; 2001. pp. 31–32. [Google Scholar]
- 9.Besag J, Newell J. The detection of clusters in rare diseases. J R Statist Soc A. 1991;154:143–155. [Google Scholar]
- 10.Gangnon RE, Clayton MK. A weighted average likelihood ratio test for spatial clustering of disease. Stat Med. 2001;20:2977–2987. doi: 10.1002/sim.917. [DOI] [PubMed] [Google Scholar]
- 11.Kulldorff M. Statistical methods for spatial epidemiology: tests for randomness. In: Gatrell A, Loytonen M, editors. GIS and Health for Europe. London: Taylor and Francis; 1998. pp. 49–62. [Google Scholar]
- 12.Kulldorff M, Nagarwalla N. Spatial disease clusters: detection and inference. Stat Med. 1995;14:799–810. doi: 10.1002/sim.4780140809. [DOI] [PubMed] [Google Scholar]
- 13.Cook DG, Pocock SJ. Multiple regression in geographical mortality studies, with allowance for spatially correlated errors. Biometrics. 1983;39:361–371. [PubMed] [Google Scholar]
- 14.Kulldorff M. A spatial scan statistic. Communications in Statistics: Theory and Methods. 1997;26:1481–1496. [Google Scholar]
- 15.Ward MP, Carpenter TE. Techniques for analysis of disease clustering in space and in time in veterinary epidemiology. Prev Vet Med. 2000;45:257–284. doi: 10.1016/s0167-5877(00)00133-1. [DOI] [PubMed] [Google Scholar]
- 16.Clark PJ, Evans FC. Distance to nearest neighbor as a measure of spatial relationships in populations. Ecology. 1954;35:445–453. [Google Scholar]
- 17.Cuzick J, Edwards R. Spatial clustering for inhomogeneous populations. J R Statist Soc B. 1990;52:73–104. [Google Scholar]
- 18.Jacquez GM. A k nearest neighbour test for space-time interaction. Stat Med. 1996;15:1935–1949. doi: 10.1002/(sici)1097-0258(19960930)15:18<1935::aid-sim406>3.0.co;2-i. [DOI] [PubMed] [Google Scholar]
- 19.Knox EG. The detection of space-time interactions. J Appl Stat. 1964;13:24–30. [Google Scholar]
- 20.Cliff AD, Haggett P, Ord JK. Spatial Aspects of Influenza Epidemics. London: Pion Limited; 1986. p. 17. [Google Scholar]
- 21.Mostashari F, Kulldorff M, Hartman JJ, Miller JR, Kulaskena V. Dead bird clusters as an early warning system for West Nile Virus activity. Emerg Inf Dis. 2003;9:641–646. doi: 10.3201/eid0906.020794. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Ward MP. Clustering of report cases of leptospirosis among dogs in the United States and Canada. Prev Vet Med. 2002;56:215–226. doi: 10.1016/s0167-5877(02)00160-5. [DOI] [PubMed] [Google Scholar]
- 23.Bridges CB, Kuehnert MJ, Hall CB. Transmission of influenza: implications for control in health care settings. Clin Infect Dis. 2003;367:1094–1101. doi: 10.1086/378292. [DOI] [PubMed] [Google Scholar]
- 24.Ehlers M, Moller M, Marangon S, Ferre N. The use of Geographic Information System (GIS) in the frame of the contingency plan implemented during the 1999–2001 Avian Influenza (AI) epidemic in Italy. Avian Dis. 2003;47:1010–1014. doi: 10.1637/0005-2086-47.s3.1010. [DOI] [PubMed] [Google Scholar]
- 25.Turner AJ, Galvin JW, Rubira RJ, Miller GT. Anthrax explodes in an Australian summer. J Appl Microbiol. 1999;87:196–199. doi: 10.1046/j.1365-2672.1999.00869.x. [DOI] [PubMed] [Google Scholar]
- 26.Turner AJ, Galvin JW, Rubira RJ, Condron RJ, Bradley T. Experiences with vaccination and epidemiological investigations on an anthrax outbreak in Australia in 1997. J Appl Microbiol. 1999;87:294–297. doi: 10.1046/j.1365-2672.1999.00894.x. [DOI] [PubMed] [Google Scholar]
- 27.Fontaine RE, Goodman RA. Describing the findings. In: Gregg MB, editor. Field Epidemiology. New York: Oxford Univ Pr; 2002. pp. 88–97. [Google Scholar]
- 28.Cliff AD, Haggett P, Ord JK. Spatial Aspects of Influenza Epidemics. London: Pion Limited; 1986. p. 261. [Google Scholar]
- 29.Cromley EK, McLafferty SL. Analyzing the risk and spread of infectious diseases. In: GIS and Public Health. New York: Guilford Pr; 2002. p. 133. [Google Scholar]
- 30.2001 Agriculture Community Profiles [database on the Internet] Ottawa: Statistics Canada c2007; [Last accessed June 8, 2007]. Available from http://www25.statcan.ca:8081/AgrProfile/acphome.jsp. [Google Scholar]
- 31.Vandeputte J, Pensaert M, Castryck F. Serologische diagnose en onderzoek naar verspreiding van het varkensinfluenzavirus in Belgie. Vlaams Diergeneeskd Tijdschr. 1980;49:1–7. [Google Scholar]
- 32.Gourreau JM, Kaiser C, Hannoun C, Vaissaire J, Gayot G. Premier isolement en France duvirus de l’influenza du porc (Hsw1N1) dans un environnement pathologique plurimicrobien. Bull Acad Vet. 1980;53:181–188. [Google Scholar]
- 33.Cromley EK, McLafferty SL. GIS and Public Health. New York: Guilford Pr; 2002. Analyzing the risk and spread of infectious diseases; pp. 190–193. [Google Scholar]