Temporal features
The variation in numbers of case farms across the different years included in the present study (Figure 1), as well as the high incidence of AEP in Norway in 2012, suggests that the factor(s) associated with the disease do not appear continuously with the same intensity. The clear seasonal pattern found suggests that horses are more likely to be exposed to the factor, or combination of factors, associated with AEP during a part of the year when they are not on pasture, assuming that the incubation period between exposure and development of clinical signs is not several months. An association between the number of horses of a particular breed and incidence of AEP is unlikely, which together with absence of breed predilection in other studies [3],[4] again indicates that the causes of AEP are to be found in the horses’ environment rather than in a genetic predisposition. Moreover, case farms often had more than one affected horse, which, unless affected horses were related, is in favour of environmental exposure.
Spatial features
One very interesting result was the large excess of case farms and cases in Norway compared with Sweden, despite the larger estimated Swedish horse population (Sweden, n = 362,700 horses in 2010; Norway, n = 125,000 horses in 2012) [22],[23]. This geographical pattern of uneven case load was already noted in a previous study reporting AEP data from 2007-2009 [4], and remained in this study with all known cases of AEP from 1995 to 2012 included. Interestingly, all cases reported in 2010-2012 were from Norway. Management variables, meteorological variables or other factors with local or regional differences between horses, which could explain outbreaks of AEP remain to be hypothesised and studied, as further discussed below.
Further, the areas where many cases of AEP aggregated (Figure 2) correspond to some, but not all, areas where the human population is dense. This is not unexpected as three-quarters of the total horse population and two-thirds of all establishments with horses in Sweden are located either in city areas or in areas adjoining urban areas [23]. The three counties with largest subpopulations of horses in 2010 (approx. 50,000 horses each) are Stockholm, Västra Götaland and Skåne [23] (Figure 2) and 42% of all Swedish horses thus live there. However, the outbreaks of AEP in this study were not concentrated to these three regions, or even to the five most horse-dense counties (Stockholm, Skåne, Blekinge, Gotland, Halland, Figure 2), as might have been expected if risk factors had been evenly distributed across the whole horse population or if a dense horse population in itself would constitute a risk factor. Risk factors for AEP at the individual horse level have been discussed by Gröndahl et al. [4]. A farm with many horses could, obviously, have more cases than a farm with only a few horses. There were no data suggesting that farms with larger horse populations were located in certain areas or that such farms were affected during certain time periods, which, had this been the case, could possibly have explained the higher number of cases for some areas and during some periods.
Spatio-temporal features
The distribution of case farms in time and space led us to form a hypothesis of space-time interaction clustering of AEP case farms (and cases). Space-time clustering was supported by the results from the spatial scan statistic at case level and at case farm level by the space-time K-function, although some clusters detected by the scan statistic included a larger area and longer time period than suggested by the distances with increased risk found with the K-function. Also, the proportion of case farms that were included in a space-time cluster was low (eleven out of 118). In this scan statistic, all case farms for which information on number of affected horses was missing were excluded, which is a limitation of the study. The second scan statistic, at case farm level, included these 18 farms, 17 of which were in Norway and affected in 2012; nine of these 17 farms were located in Rogaland County (Figure 2). However, no clusters of high rates of case farms were identified.
The first scan statistic evaluated if affected horses were clustered or not. The second scan statistic evaluated if affected farms were clustered or not adjusting for any within-farm clustering of case horses, i.e. that one horse would be more likely to be affected if there already was one affected horse on the farm. Although the number of affected horses per farm often was >1, the median was two and therefore the overall within-farm clustering could be regarded as low compared to many infectious diseases, Nevertheless, two of the clusters identified with the first scan statistic included only one farm each with many case horses. It would have been interesting to study exposure to potential risk factors in the farms identified by the first scan statistic to learn more about a potential aetiology of AEP and to compare to results from other studies of risk factors for AEP [4]. Further, an investigation of the case farms in the spatio-temporal cluster with fewer cases than expected, determining what possible exposures that were absent during this time period, would have been interesting.
Because of the lack of data on the background population, methods for only case-data had to be used to evaluate space-time interaction clustering. By using case-only methods, the interpretation of the results cannot be done in relation to the density of the horse population where clusters are detected, and the incidence will not be possible to estimate. Even though cases have been reported in less horse-dense areas, and despite that the true incidence is not known, we suggest that AEP should be considered a relatively rare disease. However, the long-term loss of performance and high mortality described [4] still merits the disease to be classified as a serious equine health problem in the areas affected and research for further knowledge on aetiology, prophylactic and therapeutic management is warranted.
The global tests of the overall spatio-temporal features of the case data produced different results. The positive space-time K-function uses the Euclidian distances between case locations, i.e. two farms are “close” if they are in the same geographical area. The negative kNN on the other hand, does not consider the distances between case locations, i.e. the neighbour case farm can be geographically located on the other side of the road or far away. This difference in what type of closeness the two methods evaluate could be one reason for the different results, and suggests that the exposure is geographically limited.
The local test for detection of space-time interaction clustering, i.e. the scan statistic, produced some clusters of a size that was, roughly, in agreement with the temporal distances of two months as well as the spatial distances of approximately 20 km seen in the test with the K-function (Figure 3), however three clusters had a considerable larger radius in kilometres (Table 1). Both the larger and smaller clusters detected with the scan statistic could possibly be attributed to factors such as natural geographical structures or barriers, shared network of feed suppliers, experienced similar climatic conditions, etcetera during the temporal length of the cluster. Using case-only methods also means the interpretation of the results cannot be done in relation to the density of the horse population where clusters are detected, and the incidence will not be not possible to estimate, although AEP should be considered a rare disease also when cases have been reported in less horse-dense areas. Nevertheless, AEP is relevant to study because of the high mortality and lack of knowledge of the aetiology.
Outbreaks of infectious diseases typically show space-time clustering when disease is spread from one herd or animal to its neighbours. But non-infectious disease may also appear in outbreaks if the aetiology is related to transient and local causes, e.g. environmental exposure. The findings suggests either a variable incubation time between exposure and development of clinical signs, or a multifactorial background, or else that the same exposure phenomenon may appear at various times of the year but with a strong predilection for certain periods of the year. Previous studies have discussed a forage-related aetiology for AEP [3],[4], specifically use of wrapped forage; however, this hypothesis remains to be proved. The hypothesis is based on the fact that affected cases usually had been fed wrapped forage (haylage or silage), as reported in more than 95% of the case farms in the present study (data not shown). A shift from hay to wrapped forage occurred in many horse farms in Norway and Sweden in the 1990s, concurrent with the emergence of AEP in these countries. In the present study, there was no space-time interaction during the period 1995-2000, examined using the space-time K-function. The first case horse level scan statistic, on the other hand, did detect two clusters during this time period. However, the second scan statistic, at case farm level, did not. Many of the first affected case farms included in this study were early adaptors to the use of wrapped forage, and on some farms only horses fed a certain batch of wrapped forage become affected with AEP [4]. Today, wrapped forage is used partly or exclusively as roughage fed to horses during the winter by over 50% of horse keepers in Norway and Sweden [22],[24], and in up to 90% of the bigger establishments in Sweden [24], but our data do not suggest a corresponding increase in the incidence of AEP, so factors other than simple exposure to wrapped forage have to be part of a “forage hypothesis”.
One environmental exposure possibly shared by case farms, and which may be coupled to forage quality, is the meteorological conditions during grass growth, at harvest, during storage time or during feed-out. Local meteorological conditions affecting pasture grass and thus triggering disease in susceptible animals have been discussed in relation to spatio-temporal clustering of cases of equine grass sickness [25]. Weather (including temperature, rainfall, air humidity) and sun irradiation at harvest largely influences the wilting rate of the cut crop and a slow wilting process may result in increased microbial growth or toxin production in the crop already before baling and wrapping. In one study, forage microbial load in wrapped forages on commercial horse farms in autumn and in the following spring did not show increased mould growth after winter storage [26]. Preliminary results from a study of forages in 13 farms affected with polyneuropathy suggest that weather conditions during forage harvest were unfavourable (wet and damp), and the forages often showed presence of soil contamination, grass roots and decaying plant material (personal communication, Gröndahl, G). Both findings are considered risk factors for a good hygienic quality of the end product. It would have been interesting to study risk factors in climatic conditions around the harvest date at the harvest location for the forage of each case farm in the present study, and also at the farm location at the time of the outbreak, but such background data were unavailable.
The three methods used to test for space-time clustering assume that the density of the background population is stable, or is changing at a rate that is consistent through space [13]. If this is not the case, this may cause bias where non-existent clusters of cases are detected. There was no information suggesting that the horse population in Norway or Sweden should have changed very differently in different regions during the study period. The population growth was assumed to be homogeneous in space and through time. The Swedish horse population was estimated to have increased by 10-20% from 2004 to 2010 in a study by the Swedish Board of Agriculture [23]. The spatial resolution was number of horses per county in Sweden. In Norway, the only survey conducted did not include the geographical distribution of the horse population [22].
To reduce bias caused by any heterogeneous changes in the background population, i.e. detection of non-existent clustering [27],[28], the tests with positive results (space-time K-function and scan statistic at case horse level) were repeated with subsets of the data covering 6-year periods, or geographically including only southern Norway (K-function). By using three different methods, and in addition repeating some analyses for subsets of the data, the risk of type I error, i.e. incorrectly discarding H0 of no clustering, increased. However, the results that were positive were significant at a rather strict α-level. Further, the Jacquez kNN test included a correction for multiple testing.
A limitation of the study was the small sample size in relation to the size of the study area and length of the study period. This may have caused lack of power and increased risk of type II-error, and could explain why the scan statistic at case farm level and Jacquez kNN test result were negative. The statistical power of techniques to detect space-time clustering has been described as low to moderate [29], which in the present study could have led to false negative results.
With the passively reported case data, one possible reason for the higher caseload in Norway could be larger under-reporting in Sweden. During the last 6 years of the study period a research project, of which the current study is a part, has included a number of articles in magazines for veterinary practitioners and horse owners in both Norway and Sweden. Because of the typical clinical manifestation of AEP any veterinarian working in horse care should be able to diagnose the disease independently or at least after consulting a referral clinic. Nevertheless, there were likely cases of AEP that were never diagnosed, reported and included in this study and possibly this under-reporting was larger in some areas for unknown reasons. It has been shown that the space-time K-function was not invalidated by under-reporting if the under-reporting in time, e.g. during certain calendar months, was independent of the under-reporting in space, e.g. by a particular veterinary practice [25]. We reason it unlikely that the under-reporting in the present study should be linked to time or space, or that under-reporting is larger in Sweden than in Norway.
If geographic proximity to data collection centres in Oslo and Uppsala increased the inclination to consult on or report AEP cases, this might be a possible confounder for geographical clustering. Nevertheless, we think that larger outbreaks were recognised in the study regardless of location, because of the public awareness of the disease, the high degree of social networking in the equine community in Scandinavia, and the commotion in the community generally noted at outbreaks of this severe disease.