Study design

I used multivariate regression analysis to study factors associated with trends and spatial variations in UVC within and among Swedish administrative units (Fig. 1). Trends in UVC were studied relative to changes in traffic and ungulate harvest at (i) national level between 1970 and 1999, (ii) within counties between 1985 and 1999, and (iii) at the district level between 1986 and 1997. Spatial patterns in UVC, harvest statistics, landscape parameters, road density and traffic volume were studied among (i) 22 counties over a period of 15 years (1985–1999), (ii) 95 districts (from seven counties) covering 12 consecutive years (1986–1997), and (iii) 311 parishes (in 36 districts in three counties) over a period of five years (1993–1997).

Figure 1.

Spatially nested levels of data: Trends in ungulate-vehicle collisions were studied at national level during 1970–1999, at county level during 1985–1999 and at district level during 1986–1997. Spatial patterns were studied among 22 Swedish counties, 95 moose hunting districts within seven counties, and among 311 parishes within three counties during 1993–1997.

These spatially nested levels were due to differences in the extent and the resolution of available data, but allowed for comparisons of large-scale data with long temporal continuity (i.e. national and county level) with small-scaled data covering a greater spatial variation (i.e. district and parish level). The county level is the only reliable spatial resolution for roe deer hunting statistics, whereas moose harvest data are also available at the district and parish levels. Moose populations are typically managed at the district level, but individual hunting teams operate within smaller areas that encompass one or a few individual moose home ranges. The average size of a moose home range (∼10 km²/moose; Cederlund & Okarma 1988) was set at the minimum size for the selection of parishes. Mitigation measures such as wildlife fences or underpasses are best evaluated at the parish level due to the limited extend of available data on their occurrence. At the national and county levels, most data on road traffic, ungulate harvest and collisions were readily available from statistical yearbooks and official databases. At the district and parish levels, all parameters were quantified from official databases and statistics that I linked to digital maps on land cover, administrative borders and roads.

Ungulate-vehicle collisions

Swedish car drivers are obliged to report any vehicle accident involving ungulates to the police, irrespective of where or when it occurred. Since 1970, the SNRA have kept UVC records in a national Road Accident Database. Prior to 1985, only summary statistics were available, but for the period 1985–1999, the database provides information on the exact location of each accident. After 1999, the database was limited and now only records accidents involving human injury or death, which is < 5% of all UVC. The Road Accident Database does not include information on the age or sex of the animals involved. If not stated otherwise, I refer to UVC-densities as the annual frequency of accidents with moose and roe deer per hundred kilometres of public road.

Hunting statistics

Data on the annual harvest of moose and roe deer were used to index trends and spatial variations in the abundance of these species. Although game bag statistics tend to less accurately reflect trends in ungulate populations than hunter observations or pellet counts (Ericsson & Wallin 1999, Solberg & Sæther 1999), only the annual game bag data covered a sufficiently large area or time period required for the analysis. Hunting statistics for moose and roe deer at the national and county levels were provided by the Swedish Association for Hunting and Wildlife Management. At the district and parish levels, moose harvest data were compiled from County Administrative Board records. No roe deer hunting statistics exist at the district or parish level. To obtain a comparable measure of moose harvest across administrative units, I calculated the annual density of moose shot per 1,0 ha of land, quantified from digital topographic maps.

Table 1.

Variables used in the stepwise forward regression analyses at the three spatial levels county (I), district (II) and parish (III); see also Fig. 1.

Road and traffic data

The length of public roads and averaged traffic volumes at the national and county levels were combined from data in Statistical Yearbooks of Sweden (Statistics Sweden 1970–2000), from Edwards, Nilsson, Thulin & Vorwerk (1999), and from digital road maps provided by the SNRA. For each area, I calculated the overall traffic volume (‘TRAFFIC’) as the summed kilometres driven by all vehicles combined during an average day or year. The density (km/km²) of public, i.e. state administered roads, was separately calculated for roads with different speed limits (‘ROAD<70, ROAD70, ROAD90, ROAD110’; Table 1). These speed categories largely coincide with road standards such as local roads, county roads, national roads and international roads, but do not necessarily relate to traffic volume.

Exact information on the location of wildlife fences was available for only three counties and was therefore considered at the parish level only. I calculated the proportion of unfenced public roads (‘UNFENCED’), and the density of road underpasses and bridges (‘PASSAGE’) per hundred kilometres of road. The passages were constructed for private or local roads crossing larger highways or motorways, but may occasionally be used by wildlife as well (A. Seiler, unpubl. data).

Land cover

Data on the composition of land cover at the county level were obtained from Swedish Statistical Yearbooks (Statistics Sweden 2000). Within districts and parishes, land cover composition was determined from digital topographic maps provided by the Swedish National Land Survey. Major land cover types were distinguished as forest (including deciduous and coniferous forest), wetland (peat bogs, mires), open land (mostly agricultural land), urban areas and mountains (see Table 1).

Statistics

Land cover proportions were transformed using the arcsine transformation to compensate for skewed distributions (Zar 1999). I used univariate and stepwise multivariate regression analyses, as well as analyses of covariance to evaluate the influence of the selected independent environmental variables on the number and density of UVC. To reduce intercorrelation between the predictor variables, I omitted those variables from further use that were highly (R > 0.75) correlated with others (Zar 1999). As the proportion of open land inversely relates to the proportion of forest, I excluded this category from further regression analysis. If required by the logic of the relationship, for example when testing the correlation between road length and collision numbers, intercepts were set at zero. In stepwise multiple regression analyses, the required minimum F-value to enter a new independent variable was set at F = 4. In all analyses, the significance level was set at P < 0.05. In repeated statistical tests on the same dataset, I used the sequential Bonferroni correction (Rice 1989) to adjust the significance level and reduce the risk of Type I errors. Calculations were performed using the statistical software package STATISTIC A (StatSoft 1999).

Table 2.

Multiple regression results for trends in vehicle collisions with moose and roe deer in relation to traffic intensity (billion kilometres driven by all motor vehicles) and the nationwide annual harvest in both species between 1970 and 1999 in Sweden. For moose: overall model P < 0.0001, R² = 0.85, F(2,27) = 77.92; for roe deer: overall model P < 0.0001, R² = 0.94, F(2,27) = 218.84.

Trends in moose and roe deer-vehicle collisions

Between 1970 and 1999, UVC in Sweden showed significant variation (Fig. 2). The steadily increasing traffic (about 1.5% increase per year) and changes in the annual harvest in moose and roe deer as an index of ungulate abundance explained 85% and 94% of the trends in MVC and RVC, respectively (Table 2). In both species, harvest statistics appeared as the primary correlate with the number of collisions over time, while traffic explained most of the residual variation (moose: univariate R² = 0.72, N = 30, P < 0.0001; roe deer: univariate R² = 0.15, N = 30, P = 0.036).

Practically all variation in annual MVC between 1970 and 1980, when MVC increased approximately 6.6 fold from 902 incidents to 5,951, can be explained by the concurrent increase in moose harvest and traffic intensity (multivariate adjusted R² = 0.99, N = 11, P < 0.0001). After the peak in 1980, however, traffic continued to increase, whereas moose harvest and MVC decreased slightly but remained correlated (univariate R² = 0.22, N = 19, P = 0.026). However, during both the increasing and decreasing phases, changes in MVC seemed to precede changes in moose harvest by two years (for the entire period: R² = 0.81, N = 28, P < 0.0001; for the increase phase 1970–1980: R² = 0.98, N = 17, P< 0.0001; for the decrease phase 1981–1999: R² = 0.69, N = 19, P < 0.0001). In roe deer, no time lag was observed in the correlation between collision numbers and the annual harvest. The annual harvest of roe deer was significantly correlated with the extent of collisions between 1970 and 1993, when RVC numbers rose 30.5 times; from 925 to 24,797 (R² = 0.93, N = 24, P < 0.0001). After 1993, harvest numbers dropped significantly, but RVC remained at a high level, probably due to the continued increase in traffic (univariate R² = 0.79, N = 6, P < 0.0001). However, the ratio between collision numbers and the annual harvest in roe deer increased with a significantly higher rate per year (6.8 ± 13.0% SD) than expected from the increase in traffic (2.3 ± 3.0% SD; t = 29.71, df = 56, P < 0.0001). For moose, there was no difference between the increase rates (mean increase in MVC ratio: 2.2 ± 12.5% SD per year; Fig. 3).

Figure 2.

Trends in ungulate harvests correlate significantly with the number of police reported vehicle collisions with moose (A; R² = 0.77, N = 30, P < 0.0001) and roe deer (B; R² = 0.81, N = 30. P < 0.0001) during 1970–1999 in Sweden.

Figure 3.

Development in the ratio of vehicle collision numbers to the annual harvest in roe deer and moose compared with the annual change in traffic volume in Sweden. The mean annual increase rate in the collision/harvest ratio in roe deer (6.8%) was significantly higher than the change in traffic (2.3%; t = 26.37, df = 56, P < 0.0001), whereas there was no difference in increase rate between traffic (2.3%) and the moose ratio (2.2%).

Table 3.

Results of stepwise forward regression models on the density of vehicle collisions with moose and roe deer at the three levels of spatial resolution county, district and parish. For all models, F-to-enter was set at 4. Coefficients and Revalues refer to full multivariate models. For description of the included variables, see Table 1.

The strong relationships between trends in UVC, ungulate harvest and traffic observed at the national level were not observed at finer scales. Only in three of 22 counties (county LM, H and X), and in none of the 90 hunting districts, were univariate regressions between MVC densities and moose harvest significant after a sequential Bonferroni correction. In roe deer, only four of 22 counties revealed a significant relationship between harvest and accident numbers over time (county LM, N, P and Y). There was no effect of traffic on the trends in UVC within counties or districts.

Spatial variations

County level

Geographic differences in the abundance of moose and roe deer, indexed on the basis of game bags, appeared as the primary correlates with UVC at the county level, explaining 57% of the variation in collision densities among counties (Table 3). This pattern has been consistent over time; in any year between 1985 and 1999 counties with a larger harvest in roe deer or moose also suffered a higher frequency of RVC or MVC than counties with smaller game bags (Fig. 4). Linear regression of annual county harvests and UVC densities yielded P-values of < 0.007 in moose, and < 0.010 in roe deer (all of which were significant in a sequential Bonferroni test).

Maximum average density of MVC (8.6/100 km) occurred in county S, which also showed the highest average game bag (5.5 moose/1,000 ha land; Table 4). A minimum of 1.7 and 2.4 MVC/100 km were reported in the very southern and northern counties (LM and BD; see Fig. 1), where average moose harvests were below 1.1 moose/1,000 ha. Similarly, RVC densities were smallest in the northernmost counties (BD and AC; < 0.5 RVC/100 km) where roe deer harvests were marginal, but peaked with 50 RVC/100 km in the southeastern counties K and H with harvests >25 deer/1,000 ha. Differences in the density of public roads explained a significant part of the variation in the ratio of collision numbers to harvest (moose: R² = 0.95, df = 21, P< 0.0001; roe deer: R² = 0.51,df=21,P = 0.0001; intercepts set at zero). In moose, the ratio varied from 1.5% to 19.3% in individual years and counties, averaging 5.8% at national level (see Table 4). In roe deer, the ratio ranged within 2–92.6% (see Table 4).

Figure 4.

Relationships between the average number of moose-vehicle collisions/100 km public road and the average annual moose harvest/1000 ha at the three spatial scales county (A; N = 22), district (B; N = 95) and parish (C; N = 311) during 1985–1999, 1986–1977 and 1993–1997, respectively. The correlation is significant at county and district levels (counties: R² = 0.60, df = 21, P < 0.0001; districts: R² = 0.23, df = 94, P < 0.0001), but not at parish level.

Table 4.

Descriptive statistics for all variables included in multiple regression analyses. For explanation of variables, see Table 1.

County land cover correlated with ungulate harvests, but was subordinate in its effect on UVC and did not enter the multivariate models. Moose harvests were largest in counties dominated by forests (R² = 0.44, df = 21, P = 0.0008), whereas roe deer harvests increased with the proportion of urban habitats (R² = 0.45, df = 21, P = 0.0007) and decreased with an increasing proportion of wetlands (R² = 0.64, df = 21, P > 0.0001).

Trends

The significant change in ungulate-vehicle collisions that occurred in Sweden during the past three decades can mainly be attributed to the increase in ungulate density and vehicle traffic. Game bag statistics, as an index of ungulate abundance, were the primary correlate with collision numbers in this analysis, whereas traffic explained a significant part of the residual variation and retained collision numbers at high levels while population sizes declined. Trends in moose and roe deer populations have been strongly biphasic due to changes in food availability, predation and harvest regimes. Large-scale changes in forestry created favourable habitat conditions for moose, which, coupled with restrictive and selective hunting quotas and a virtual lack of large predators, accelerated the growth of the population during the 1970s (Lavsund & Sandegren 1989). By the winter of 1981/82, the Swedish moose population was estimated at approximately 315,000 animals, which was probably one of the highest national moose densities recorded worldwide (Lavsund & Sandegren 1989). The considerable damage to forestry and agriculture due to high moose densities and the drastic increase in MVC called for a substantial reduction in the Swedish moose population (Cederlund & Markgren 1987). Consequently, hunting quotas peaked with more than 175,0 moose shot in autumn 1982. Quotas were reduced to approximately 100,000 animals in the following years (Swedish Association for Hunting and Wildlife Management, database). Moose-vehicle collisions followed a similar pattern, although changes in MVC preceded changes in the harvest by two years. Also the Swedish roe deer population went through a significant increase and expansion during the 1980s and early 1990s (Swedish Association for Hunting and Wildlife Management, database). The increase was likely a response to a series of mild winters, coupled with the outbreak of Sarcoptic mange that significantly reduced the population of the red fox Vulpes vulpes (Lindström, Andrén, Angelstam, Cederlund, Hörnfeldt, Jäderberg, Lemmell, Martinsson, Sköld & Swenson 1994, Cederlund & Liberg 1995) which is the main predator on roe deer fawns. In 1993, roe deer harvest peaked at 382,000 animals. Since then, roe deer densities and game bags dropped substantially, due in part to the recovering populations of red fox and Eurasian lynx Lynx lynx (Andrén, Liberg & Sand 1999). The number of RVC remained consistent over the early years, but stabilised after 1993 at > 20,000 per year.

As with moose, the deviation between trends in RVC and roe deer harvest is probably a result of the steadily increasing traffic. In contrast to moose, however, the ratio of collisions to harvest in roe deer has increased three times faster than traffic. This suggests a non-linear relationship between traffic volume, harvest and accident numbers or the influence of a fourth, maybe behavioural, factor that makes it difficult to foresee future trends in RVC. Also in The Netherlands, the increase in RVC during 1985–1995 was five times as high as the increase in roe deer densities or traffic volume (Groot-Bruinderink & Hazebroek 1996). A similar pattern was described by Oosenbrug, Mercer & Ferguson (1991), who observed that a 46% increase in traffic and a 23% increase in moose harvest produced an 89% increase in MVC during 1983–1989 in Newfoundland. One complication in projecting moose collision numbers from harvest and traffic statistics is the time lag of 1–2 years between actual population densities and harvest statistics. The time lag occurs because the numbers of hunting licenses in any given year are based on previous years' game bags and population surveys (e.g. Cederlund & Markgren 1987). My results suggest a time lag of two years between MVC and moose harvest, but no time lag in roe deer. Roe deer hunting in Sweden is not regulated by licenses as is moose hunting, so that game bags may directly relate to changes in roe deer densities. Rather than predicting ungulate collision numbers from current hunting statistics, it may thus be more appropriate to use accident statistics in the planning of future hunting quotas. Using road kill statistics (corrected for traffic volumes) as a large-scale index of wildlife populations has been suggested previously (Jahn 1959, McCaffery 1973, Hicks 1993).

Spatial patterns

Strong relationships between trends and spatial patterns in ungulate densities and UVC have been found in many studies (e.g. Beilis & Graves 1971, Désiré & Recorbet 1987, Oosenbrug et al. 1991, Jaren, Andersen, Ulleberg, Pedersen & Wiseth 1991), but not in all (e.g. Del Frate & Spraker 1991, Groot-Bruinderink & Hazebroek 1996). McCaffery (1973), for example, observed significant correlations between antlered buck harvest and road kills in white-tailed deer Odocoileus virginianus in 28 of 29 management areas in Wisconsin. Puglisi, Lindzey & Beilis (1974) reported positive relationships between county deer population estimates, deer harvest and collision numbers among 15 counties in Pennsylvania. My findings suggest that county UVC and district MVC were significantly related to harvest statistics.

Adaptive management of ungulate populations may appear as an implicative measure against UVC, albeit it is hardly considered as a possible option (e.g. Putman 1997, Keller, Alvarez, Bekker, Cuperus, Folkeson, Rosell Pagés & Trocmé 2002). Large-scale reductions in ungulate densities may likely have a counteractive effect on UVC, but it will be difficult to find political or public support for such action. Local reductions, on the other hand, can more easily be achieved, but may not have the desired effect. In their survey of mitigation measures applied across the United States, Romin & Bissonette (1996) reported that only one of two states that tried to reduce UVC and local deer populations through intensified hunting indicated success. Joyce & Mahoney (2001) documented that the risk of MVC along the Trans-Canada Highway in Newfoundland was elevated in areas with both scarce and dense moose populations, but reduced in intermediate areas. The authors expressed doubt as to whether a simple management for reduced moose densities would decrease MVC. Also Almkvist et al. (1980) questioned the effect of local density control on traffic safety, as they did not find a significant relationship between local moose harvests, i.e. density, and the local frequency of MVC in Sweden during the 1970s. Similarly, I observed that there was no relation between moose harvest and the density of collisions at parish level, although it was significant at the district and county level. Despite the strong relationship between trends in harvest, traffic and UVC at the national level, there was no consistent relationship within counties and, especially, nor within moose hunting districts.

The more complex patterns at higher spatial resolution are likely due to the emerging influence of local factors such as the animal's habitat use and behaviour, proximity to and juxtaposition of preferred habitat near roads, road and roadside specifications, vehicle speed, or the presence of mitigation measures (e.g. Bashore et al. 1985, Hartwig 1993, Finder et al. 1999, Hubbard, Danielson & Schmitz 2000, Madsen, Strandgaard & Prang 2002). UVC are usually spatially and temporally aggregated, which implies that patterns observed at broad scale may not necessarily apply at a finer scale. Also, local density measures may produce a misleading picture of local collision risks as density necessarily refers to comparatively large areas encompassing many individuals' home ranges. Harvest data may provide a reasonable index of moose abundance at broad scales, encompassing several management areas or populations, but will fail to reflect local variations in abundance at finer scales, if the spatial extent covers but a few individual home ranges and hunting licences are decided upon other criteria than local animal abundance.

At the parish level, UVC densities were significantly related to preferred habitat, road density and occurrence of underpasses. For moose, clearcuts and young forests provide important staple forage. Wetlands, marsh-lands and moist coniferous forests are preferred habitats especially during snow free periods (e.g. Markgren 1974, Bergström & Hjeljord 1987, Cederlund & Okarma 1988) as they provide sources of protein and sodium (e.g. Jordan 1987). Conversely, roe deer in Sweden are more often found in rural habitats, agricultural land and clearcuts, while feeding on grasses and herbs primarily during the summer months (Cederlund et al. 1980). This may explain why the RVC density was increased with the proportion of urban areas at the parish and district levels.

In addition, collision densities were positively correlated with the density of roads with speed limits of 90 km/hour (i.e. major national roads), but inversely related to minor roads (70 km/hour) and motorways (110 km/hour). At lower speeds, drivers presumably have more reaction time to prevent such collisions. Roads with high speed limits and traffic volumes are more often protected by fences or extensive roadsides than minor roads. Also, heavy traffic may deter animals from entering the roadway and thereby prevent UVC from occurring (e.g. Skölving 1987, Berthoud 1987).

Mitigation measures

I found a significant inverse correlation between the density of road passages and the density of both MVC and RVC among parishes. This suggests that even conventional road bridges and tunnels may reduce animal-vehicle collisions. The use of such passages by animals has been documented elsewhere (e.g. Rodriguez, Crema & Delibes 1996, Yanes, Velasco & Suarez 1995, Clevenger & Waltho 2000). Conversely, Hubbard et al. (2000) observed that the density of bridges increased the likelihood of UVC in Iowa, probably because bridges were associated with travel corridors that funnel deer across the roads (e.g. Finder et al. 1999).

In contrast to my expectations, there was no effect of fencing on the density of UVC at parish level. Earlier studies conducted by the SNRA indicated that fencing could reduce the risk of MVC by more than 80% and RVC by up to 55% (Skölving 1987, Niklasson & Johansson 1987). However, because only 4.5% of all public roads within the parishes I studied were fenced, the expected reduction in the number of collisions would be 3.6% in MVC and 2.5% in RVC. These values are well within the confidence limits of the variation in UVC densities among parishes. In addition, where fences terminate or get interrupted by local roads or water courses, accident rates tend to be increased (Ward 1982, Niklasson & Johansson 1987, Foster & Humphrey 1995). It is not uncommon that moose break through or jump over fences if they do not find a safer passage across the road (e.g. Nilsson 1987, Seiler, Cederlund, Jemelid, Grängstedt & Ringaby 2003). Thus the overall effect of fencing on UVC numbers may be smaller than previously suggested. Also at the national level, the effect of fencing is likely only marginal. In the past 25 years, approximately 5,000 km (34%) of motorways and major national roads have been fenced against wildlife (A. Sjölund, SNRA, pers. comm.). About half of all UVC recorded during this period, occurred on such roads. If all fences were equally efficient, we may thus expect a 13.6% reduction in the risk of moose-vehicle collisions, and a 9.35% reduction in the risk of roe deer collisions nationwide. Also these reductions are smaller than the variation in UVC from year to year. On the other hand, the proportion of UVC involving human injury or death in Sweden decreased from 5% in 1983 to 2.2% in 1999 (SNRA, database). This decrease contrasts the growing rate of human injuries in other, non-wildlife related road accidents (+20%) and may approve for some beneficial effect of fencing, particularly at high-speed roads. Nevertheless, fencing is probably still the most efficient and practical way to keep large ungulates off the road.