Full habitat models for relatively common species

We use a set of statistical models to relate the relative abundance of each species measured at ABMI sites to three sets of variables: vegetation types, human footprint and geographic location. The analyses are done separately for two overlapping regions: North or Forested Region (including Boreal Forest, Foothills, Canadian Shield and Parkland Natural Region), and South or Prairie Region (includes Grassland and Parkland Natural Regions, and the Dry Mixedwood Natural Subregion in the Boreal Natural Region).

• Relative abundance for vascular plants, bryophytes, lichens, and mites is the probability that the species will occur in a 50 m x 50 m ABMI quadrat, or a soil sample in the case of mites. For birds, relative abundance is the number of birds detected at a point count, adjusted for the species’ detection distance.
• Vegetation types in the boreal and foothills regions are based on several province-wide GIS layers of vegetation variables, and include main forest stand types by broad age classes and several categories of open vegetation (grass, shrub, open wetland, open water, barren). For the vegetation figures, the broad age classes are interpolated to 20-yr age classes. In the south, ‘vegetation types’ are actually mapped soil types, such as loamy, sandy, etc.
• Several categories of human footprint are distinguished in the models, including ‘successional’ (forestry, temporary linear features), ‘alienating’ (agriculture, industrial, urban area, roads), as well as linear versus non-linear features. A flexible curve is used for the footprint relationships so that the models can show different responses when there is partial footprint at a site compared to 100% footprint. This allows the models to represent edge effects. The scale of the footprint information used in the intactness models is 64 ha for birds and 1 ha for other taxa.
• Flexible relationships with latitude and longitude are also included in the models to represent the geographic distribution of the species.

The analysis is conducted in a ‘model selection framework’, which means that the models for each species are only as complex as the data for that species support. For example, the models for a rarer species with few records may not be able to separate different age classes or even broad stand types. All age classes or several stand types will therefore all have the same average value in the vegetation figures. This does not mean that there are truly no differences between these types for that species. It just means that we do not yet have enough data to estimate abundances in all those individual vegetation types.

We cannot distinguish different open vegetation types in the boreal and foothills—grass, shrubs and open wetland—because these are rare types in our terrestrial sampling. We also have few samples in very young or very old forest, because these are also rare. Our results might miss species that have a strong association with one of these rare vegetation types.

This modeling requires moderate sample sizes, so we can only do the analyses for species that have at least 20 records in a natural region. ABMI has some information on many other rarer species. We can provide simpler summaries of our records for these species.

Figures show mean relative abundances, with 90% confidence intervals.

The bird modeling differed from other taxa from some aspects because the data set is more extensive (see Collaborations). The combined data set needed to be standardized due to the differences in survey protocol among the data sets. Algorithms to deal with different protocols in bird surveys and methods for estimating predictive uncertainty were developed in collaboration with the Boreal Avian Modelling Project (BAM) team (Sólymos et al. 2013). BAM also provided data management help for organizing bird point count data. The minimum number of detections required for model building for birds was 25.

The bioclimatic variables used in the modeling were calculated at a 4-km resolution using monthly climate normals of temperature and precipitation averaged over 1961-1990. The monthly climate normals are based on instrument-measured climate data that were interpolated by PRISM (Daly et al., 2002) and WorldClim (Hijmans et al., 2005). The western North American portion of these data are described by Wang et al. (2011).

​Full details of habitat modeling can be found in the ABMI Species Website Manual.

There are many limitations and caveats about the use of these habitat models. Click on each for further details:

• For vascular plants, bryophytes, lichens and mites, we are only measuring the occurrence of the species at 0-4 of the 4 quadrats or soil samples per site. The number of individuals, percent cover, total volume or biomass, reproductive status, etc. of the species is not part of this measure -- a single sapling of a tree species is an occurrence as much as complete cover of large trees. There is also no adjustment for the probability that the species was present but was missed by the field technicians. That probability is higher in quadrats with complex vegetation, species rich habitats and for groups of species that are hard to distinguish in the field, all of which could affect the habitat models.

  For birds, the measure is based on counts of birds heard in the field by trained observers in some studies, or heard on recordings made using various technologies in other studies (including ABMI). These are mainly, but not exclusively, singing male birds. Adjustments are made for different detection distances in the different studies, habitats and for different species, but these adjustments all have statistical uncertainty. The number of vocalizing birds may also not directly reflect the total number of birds or reproductive success in different habitat types.  There are also several groups of birds, such as waterfowl or raptors, where point counts are not a good survey technique. We provide only simple summary information for these species, not full habitat models.

  For mammals, ABMI is switching from snow transects to camera-based indices, but results from cameras are not available yet. The current models based on snow-track transects reflect how far the species travels in different habitat or human footprint types during winter. This can be very different from the habitats that they use for foraging, or during summer. Additionally, ABMI transects were along linear features that could be travelled by snowmobile, so the results are confounded by how those features affect winter travel by the species.

• The confidence intervals on individual habitat values for a species are often wide, and should be integral to any use of the habitat models. There are two main reasons for wide uncertainty in our results: 1) We estimate values for many habitat and human footprint types, particularly in the North, with many combinations of broad stand types and age classes. Ideally we would have hundreds of observations of a species to estimate so many values. Instead, we use a fairly complicated modeling procedure to extract reasonable values for many species with lower sample sizes by simplifying the complexity of habitat types. For all but the most common species, the resulting habitat models are simplified (some habitat types are assumed to be the same) and estimated with large error. 2) ABMI sites are systematically located, and the Alberta landbase often contains a fine-scale mix of several habitat types and human footprint. Individual sites, even 1-ha areas, therefore usually fall in more than one habitat and/or human footprint type. Our models are designed to separate out the effects of each individual habitat or human footprint type as best as possible, but the analysis is necessarily imperfect with limited sample sizes. A highly specialised species, for example, will almost inevitably show some predicted abundance in habitat types that it does not use, when these co-occur with preferred habitat types.

• We have limited ability to separate effects of each vegetation and human footprint type, and spatial and climate variables. In addition to the challenges generating uncertainty in the individual values, ABMI is not an experimental study. We do not design and manipulate habitat and human footprint types, and simply use what is sampled in the province. As a result, habitat and human footprint types are partially confounded with geographic location and climate variables. Our modeling separate out these effects as best as possible, but without impossible province-wide experiments, we can never be sure how well we have estimated the separate habitat, human footprint and geographic/climate effects.

• Our model selection approach weights different models in a model set based on the support provided by each species’ data. However, we use the same set of models for all species. We currently do not have the resources to develop separate model sets for individual species to benefit from the differences in natural history.

• All taxa have potential for errors in species identification, including the mammal track surveys and the bird recording interpretation. Taxonomic issues are common with the plants, bryophytes, and lichens, where many species groups are difficult to distinguish and taxonomy changes frequently. This adds to uncertainty in models for some species.

• ABMI’s vegetation map is compiled from various sources, not all of which cover the entire province. Some layers are less accurate than others. Also, even if individual polygons are accurate, their boundaries may not be precise, and many ABMI sites fall on the edges of vegetation polygons. Resulting error in the vegetation types at sites increases error in the habitat models. Some habitat types may have additional challenges. Grass and shrub sites in the North (forested region), for example, may be true permanent open habitats, but the same designation is often given to early seral stages of forested sites after fire. Grass and shrub sites in the prairies are also poorly mapped. Mixedwood stands are also poorly mapped, because they represent a gradient between deciduous and upland conifer stands, which can often change as a stand ages.

• Few species show strong responses to soil type in our models, suggesting that these groups may not be well-defined or relevant to many species. We have too few sites to treat lowland sites in the South (prairie region) separately.

• Human footprint is mapped more precisely than vegetation, but we use only broad groups of footprint types in the modeling. Industrial sites and urban or rural residential areas, for example, are not reliably distinguished. Values are particularly difficult to estimate reliably for linear features, because these usually occupy only a small portion of the area of a site. We also have few sites with older forestry cutblocks. Values for those habitat types combine our rough estimates with an assumed rate of convergence of aging forestry and natural stands.

• Many sites in the South (prairie region) are grazed to varying degrees but this information is not available. Thus grazing effects cannot be included in our models, and if grazing differs among native habitat types the effects confound those difference. Any wide-scale human effects, like pollution away from visible human footprint or recent climate change, are also not included in our modeling. Additionally, we only include effects of footprint at the local scale—the 1-ha scale we sample at, or larger sampling areas for birds and mammals. Due to sample size constraints and confounding among variables, we do not include any additional effects that might be caused by human footprint in the larger surrounding area.