Determining Population Change Relative to 1970

One of the main goals of The State of Canada's Birds is to provide a species-level assessment of how bird populations in Canada have changed over the long term, with a measure of how confident we are in that assessment based on the reliability of the information available to us. The methods for determining population change and survey reliability are adapted from Blancher et al. (2009), which assessed the population status of forest birds in Ontario.

We selected the year 1970 (or as close as data allow) as a baseline because it is the earliest year for which sufficient systematic and large-scale monitoring data are available to perform reliable trend analysis for most species. However, 1970 may not be an appropriate reference point for all species. For example, some raptors and waterbirds experienced dramatic declines in the 1950s and 1960s due to the pesticide DDT, and were at very low levels in 1970. Conversely, there are other species for which systematic monitoring did not begin until 1980 or later We took these factors into consideration when doing our assessments and establishing population goals. For simplicity, reference to 1970 in the methods is meant to refer to the start of survey coverage, though it may be later for some species.

Data Sources

Survey data are needed to inform the population change categories used in our assessments. NatureCounts hosts over 450,000 regional and national population trend estimates from 24 monitoring programs.

The State of Canada's Birds primarily relies on trends derived from surveys that provide annual indices over a long-term period (see Table 1), as these allow for the most quantitative interpretation of change relative to 1970. However, other sources can provide additional insight, for example provincial breeding bird atlases or data from the Canadian Migration Monitoring Network. We also use species-specific surveys or sources from the published literature where appropriate and available.

Assessing Population Trends

To determine long-term population change for each species (from 1970 to the most recent year available), we examine all available monitoring program results and select the most reliable data source for that species (see Survey Reliability below). We designate this source as the Main Survey and use it to categorize the change into one of five categories ranging from Large Decrease to Large Increase (see Population Change Categories below). If additional surveys contribute context and information, they are designated as Supporting Surveys and used as a basis for confirming or adjusting our confidence in the assessment based on the Main Survey (see Confidence in Population Change Categories below). If no surveys are available, we designate the species as Data Deficient (see Population Change Categories for details).

Survey Reliability

We assess the reliability of each survey to determine whether it is an appropriate source of data for a given species, and which survey provides the best basis for evaluating population change relative to 1970.

We determine the reliability of each survey by considering:

• the precision of the trend (see Survey Precision below)
• the survey's coverage (see Survey Coverage below)
• the survey's suitability for the species (see Survey Suitability below)

We score each of these components on a scale of Low, Medium, or High. By default, reliability is equal to the lowest score among the three components. For example, if precision scores High, coverage scores Medium, and suitability scores High, the reliability score will be Medium. There are two uncommon exceptions:

• We score reliability one level below the lowest component score if the trend is strongly influenced by exceptional year(s) in the indices, or if the trend spans fewer than 35 years, which reduces certainty in estimates of long-term change.
• We ignore the precision score if the trend shows an increase of more than 400% or a decrease of more than 80%. These values correspond to a five-fold change in the population, which is a very large, clear change, rendering precision essentially irrelevant. In this case, the reliability score is equal to the lower of the scores for coverage and suitability.

Survey Reliability: Survey Precision

We score precision based on the degree to which the available data can reliably detect population trends. For surveys analyzed using a hierarchical Bayesian model (e.g., Breeding Bird Survey, Christmas Bird Count, migration monitoring by the Program for Regional and International Shorebird Monitoring [PRISM], Nocturnal Owl Survey, and Seabird Colony Monitoring Program) precision reflects the width of the 95% credible interval around the trend. Similarly, for surveys analyzed using frequentist models (e.g., Great Lakes Marsh Monitoring Program), precision is based on the width of the 95% confidence interval. A 95% credible or confidence interval is the range of values that have a 95% chance of including the true population trend. The width of a credible/confidence interval is the difference between the upper and lower limits.

We score precision as follows:

• High: width of credible/confidence interval around the trend is less than 3.5% per year, corresponding to potential for detecting a decline of 30% over a period of 20 years
• Medium: width of credible/confidence interval around the trend is between 3.5% per year and 6.7% per year, corresponding to potential for detecting a decline of 50% over a period of 20 years
• Low: width of credible/confidence interval around the trend is greater than 6.7% per year, which is insufficient to detect a decline over a period of 20 years

Precision estimates are not available for some surveys, such as those that do not have annual coverage. This includes the Great Lakes Decadal Colonial Waterbird Census, some species-specific surveys, and sometimes information from published literature. We cannot graph these surveys or score their reliability, but we describe insights from them in the text.

Survey Reliability: Survey Coverage

We score coverage based on the degree to which a given survey samples the population (or range) of a species. For example, coverage for the Breeding Bird Survey (BBS) is calculated using a grid of cells, each one measuring 1-degree latitude by 1-degree longitude. Coverage represents the proportion of a given species' breeding population (BirdLife International 2016, Will et al. 2018) that occurs within the grid cells that contain BBS routes included in the trend analysis for that species (Smith et al. 2023). Where coverage estimates are unavailable, we estimate coverage based on a visual assessment or expert knowledge of the proportion of the survey area relative to the species' range, either in the breeding season (e.g., Eastern Waterfowl Survey) or in the non-breeding season (e.g., Christmas Bird Count). We also estimate the proportion of the population passing through the survey area on migration, where relevant (e.g., Migration PRISM).

We score coverage as follows, based on Blancher et al. (2009):

• High: greater than 50% of a species' population or range covered
• Medium: between 25 and 50% of a species' population or range covered
• Low: Less than 25% of a species' population or range covered

For most species, we use surveys based in Canada. However, we use continental-scale analyses in cases where they best reflect the Canadian population (e.g., Christmas Bird Count data for some species that breed primarily in Canada but winter mostly in the United States, like American Tree Sparrow).

Survey Reliability: Survey Suitability

The suitability score considers the survey's strengths, weaknesses, and potential biases in estimating trends for a given species over time. Survey standardization and design are particularly important. Suitability can only be high if the timing, conditions, and extent of effort follow defined protocols and are comparable across years.

The suitability score can be reduced if there is risk of bias. This refers to the potential for trend estimates that are not a function of actual changes in abundance. For example, in many areas, there has been increasing effort to document winter roosts of owls and crows. This has resulted in higher numbers of these species on Christmas Bird Counts [CBC] over time, but it is difficult to know whether the positive trends are related to changing abundance or greater effort. The CBC can also be vulnerable to bias for species that are wintering farther north than they did previously, and now have a greater proportion of their population within the area covered by the survey (Soykan et al. 2016).

Note that some survey limitations, like limited geographic sampling of a population or a mismatch in seasonal or daily timing with a species' peak in activity, do not influence suitability as they are addressed through coverage or precision scores. Although such factors can greatly reduce the detectability of a species, they are typically consistent across years, and therefore unlikely to bias the trend. For example, the BBS is a roadside survey and therefore samples a smaller proportion of the population of species that favour remote landscapes. However, roadside avoidance (or attraction) should not bias BBS trend estimates unless this behaviour has changed over time.

The one exception is situations where a species tends to be detected in such extremely low numbers that it is prudent to consider the survey less suitable, regardless of precision or coverage. For example, Eastern Whip-poor-will is a nocturnal species that is only rarely detected among the first few stops of a BBS route. Although the BBS covers a moderate portion of Eastern Whip-poor-will's breeding range and the trend estimate has high precision, the detectability of the species on the BBS is so poor that suitability is considered low, as precision and coverage scores do not adequately reflect the survey's limitations.

Taking these factors together, we score suitability as follows, adapted from Blancher et al. (2009):

• High: standardized survey with species-appropriate timing and little risk of bias, likely to result in accurate estimates of the direction and magnitude of population change for this species.
• Medium: trend estimate for a species is likely accurate with respect to direction, but magnitude may be poorly estimated, given limitations among one or more of survey standardization, detectability, and risk of bias.
• Low: estimates of both magnitude and direction of trend for a species may be inaccurate, based on multiple concerns among survey standardization, detectability, and risk of bias, or a particularly serious limitation in one of the three factors.

In situations where two or more surveys tie for having the greatest reliability, we prioritize the one with higher suitability, with coverage as a second tie-breaker if needed.

Population Change Categories

We assign long-term population change since 1970 to one of six categories as follows, based on Blancher et al. (2009):

• Large Decrease: decrease of 50% or greater
• Moderate Decrease: between a 25% and 49.9% decrease
• Little Change: between a 24.9% decrease and 32.9% increase
• Moderate Increase: between a 33% and 99.9% increase
• Large Increase: increase of 100% or greater
• Data Deficient: insufficient data to determine population trend

The values in the increase and decrease categories are purposely asymmetrical. Values of population change on a percent-scale are not symmetrical around zero because a population can increase far beyond 100%, but can never decrease past 100% because that represents extinction. The values in the Large and Moderate categories act opposite to each other. For example, a Large Increase is required to recover from a Large Decrease, or stated another way, a population that has declined by 50% must increase by 100% to return to its original level.

The Data Deficient category is used sparingly in cases where no trend data is available, or if all available data sources are of low reliability and indicate opposite trajectories (e.g., Moderate Decrease vs Moderate Increase). If directionality is the same or neutral (e.g., Little Change), we identify the survey with greater suitability (and coverage, if necessary as a tie-breaker) as the Main Survey and its trend is used to assign the corresponding population change category.

While the current focus of The State of Canada's Birds is long-term population change at the national scale, we present both long- and short-term trends where available. Short-term changes in population indicate whether the recent rate of change is increasing, decreasing, or remaining consistent relative to the long-term trend. This context can be useful for conservation planning or prioritization purposes, and may warn of the potential for longer-term changes.

Confidence in Population Change Categories

Whereas each survey is assigned a reliability score reflecting the degree to which its results can be trusted, the species' change category is assigned a confidence score representing the overall degree of trust in the assigned population change category.

We score confidence as follows, adapted from Blancher et al. (2009):

• Very High: population change category almost certainly correct, based on corroboration of two or more sources, or extremely large changes in the population
• High: population change category likely to be correct
• Medium: population change category somewhat uncertain but likely within one category of population change
• Low: population change category uncertain, but likely within two categories of population change
• Very Low: population change category highly uncertain, possibly off by as many as three categories of population change

For species with only a Main Survey identified, the default confidence level is equal to the reliability of that survey, except in the case of large magnitudes of population change. If the magnitude of the change shows an increase of more than 400% or a decrease of more than 80%, we increase the confidence score by one level (i.e., Low to Medium, Medium to High, High to Very High).

For species with multiple surveys identified, we use the matrix below (Table 2) to assign the confidence score based on the degree of agreement or disagreement between the Main Survey and the most reliable Supporting Survey. We first use the column on the left (Step 1 in Table 2 below; Reliability of Main Survey) to find the correct cell based on the Main Survey, then adjust accordingly based on the metrics from the Supporting Survey (Step 2 in Table 2 below). We have highlighted an example in the table below, which applies to Downy Woodpecker. Note that confidence associated with Little Change is treated differently than the other categories because of the lack of directionality in the category.

Canadian Responsibility

We determine the Canadian responsibility for each species by calculating the proportion of the global population that occurs in Canada annually (i.e. the total population in Canada; see Population Estimates section above).

We score Canadian responsibility as follows:

• Very Low: less than 1% of the global population
• Low: between 1 and 19.9% of the global population
• Medium: between 20 and 49.9% of the global population
• High: between 50 and 74.9% of the global population
• Very High: between 75 and 100% of the global population

Population Status in Relation to Goal

We assess the current population status of each species in relation to its goal (see Population Goals above). Current, in this case, means the most recent annual estimate available, unless there is strong evidence that the latest annual index may not be representative of the current status (e.g., reduction in survey coverage that year due to wildfires). If this is the case, we define current as the average of the most recent 5 years.

Categories for population status in relation to goal are:

• Above Goal Range: population is above its defined upper limit, which is when ecological or human conflict is likely to occur, indicating the need for management
• Within Goal Range: population is above the defined lower limit and below the upper limit (if applicable)
• Below Goal Range: population is below its defined lower limit, indicating the need for conservation efforts to increase abundance
• Data Deficient: there is insufficient data to establish a goal or determine population status in relation to goal. In some cases, SARA goals have been set despite not having sufficient monitoring data to track the status of the population. We consider the status of these species, relative to their goals, to be Data Deficient.
• Not Applicable: a goal was not determined for non-native species

Seasonality in Canada

We provide an overview of how reports of each species vary throughout the year based on checklists from datasets hosted by NatureCounts. A barchart displays the weekly proportion of checklists on which the species was reported. All bars are scaled relative to the maximum for the species, so they cannot be used to compare the abundance of one species to another. It is important to note that these plots may not truly represent abundance for species that live in remote places or are otherwise difficult to detect at certain times of year.

We also provide statements that summarize the probable nesting dates for each species at a national level, following Rousseu and Drolet (2015). For more information, see the Nesting Calendar Query Tool.

First Stage: Smoothing the Species' Population Trajectories

We first smooth each species population trajectory so that our analyses: 1) retain the most important medium- and long-term patterns of change in the species' populations, 2) account for uncertainty in each species' data, 3) interpolate estimates for years with missing data for a given species (i.e., fill-in missing years), and 4) standardize the level of annual fluctuations among species. We smooth each species' population trajectory using a generalized additive model (GAM), so that our analyses are focused on the long- and medium-term patterns in population change (i.e., our primary interests are not related to annual fluctuations). We use a hierarchical Bayesian GAM that accounts for the uncertainty of each annual estimate of relative abundance and generates both a smoothed non-linear population trajectory (e.g., red and blue lines in Figure 2) and a series of annual estimates of change since the previous year (i.e., annual population change rates, or the slopes of the red and blue lines between sequential years in Figure 2). Smoothing the trajectories also allows us to fill in missing years for species that may not have data in all years (e.g., some isolated seabird colonies cannot be monitored every year). Finally, this smoothing reduces some of the variation among surveys, caused by differences in the underlying statistical models used to estimate trends and trajectories from different programs. There are important differences among the underlying statistical models used in many of the regularly published estimates of species-level monitoring data. For example, the Christmas Bird Count (CBC) uses models that incorporate a log-linear regression approach to estimating the time-series of annual estimates. As a result, the CBC model is particularly well designed to estimate long-term trends (i.e., rates of change) but not as sensitive to changes in trends (e.g., bends in the population trajectory). Similarly, estimates from the annual waterfowl breeding-ground surveys use models that do not smooth the time-series at all. Instead, they estimate each year's estimate independently of all other years. This structure is particularly well suited for modeling the annual fluctuations in waterfowl populations, but allows estimates of annual abundance to vary much more than many other models. Smoothing estimates from these two programs makes their information more comparable.

In more detail, the input data for the GAM smoothing model are the estimated population trajectories from each species' main survey. These population trajectories are the collection of annual indices of abundance (e.g., an average expected count during a Breeding Bird Survey in each year). We first log-transform these annual indices of abundance \(\left(\hat{i}_{s,t}\right)\) and an estimate of their standard errors for a given species-\(s\) and year-\(t\) \(\left(\hat{\sigma}_{s,t}\right)\). We estimate log-scale standard errors by assuming that the available uncertainty intervals were approximately log-normally distributed. Specifically, we log-transform the upper and lower 95% credible or confidence limits of the annual indices, estimate the difference between them and divide by 3.9. \[\hat{i}_{s,t} = \log\left(n_{s,t}\right)\] \[\hat{\sigma}_{s,t} = \frac{\log\left(n^{uci}_{s,t}\right) - \log\left(n^{lci}_{s,t}\right)}{3.9}\]

To account for the uncertainty in each species estimated log-scale annual indices, the model treats the estimated log-scale annual indices as realizations of a normal distribution, centered on the true population status in that year \(\left(i_{s,t}\right)\). \[\hat{i}_{s,t} \sim \text{Normal}\left(i_{s,t}, \hat{\sigma}_{s,t}\right)\]

We then model the true population status in each year as a function of an intercept \(\left(\alpha_s\right)\), a non-linear smooth component \(\left(\delta_{s,t}\right)\), and some random error \(\left(\varepsilon_{s,t}\right)\) that represents the annual fluctuations around the smooth component. \[i_{s,t} = \alpha_s + \delta_{s,t} + \varepsilon_{s,t}\]

The random error has a mean of 0 and is normally distributed with an estimated standard deviation \(\left(\varepsilon_{s,t} \sim \text{Normal}\left(0, \sigma_{\varepsilon_s}\right)\right)\). The non-linear smooth component of the time-series \(\left(\delta_{s,t}\right)\) is estimated using a low-rank, thin-plate regression spline (Wood 2017) on year. We scale the spline basis to include sum-to-zero identifiability constraints to improve the model estimation. We limit the number of knots to 1 knot for approximately every 3 years of data in the dataset. This number of knots controls the upper limit on the complexity of the smooth and ensures that the smooth will track the long- and medium-term population fluctuations. This Bayesian approach to estimating the smooth population trajectory penalizes the complexity (i.e., the wiggliness of the line), so that complex non-linear patterns are only apparent if they are well supported by the data.

From this model and the estimated smooth, we generate an estimate of the species annual rates of population change \(\left(\rho_{s,t}\right)\) by differencing subsequent annual estimates of the smooth component \(\left(\delta_{s,t}\right)\). In any year after the first year of the time series, for example 1990, the estimated annual rate of population change \(\rho_{s,1990}\) is the difference between the smoothed trajectory in that year (1990) and the smooth trajectory in previous year (1989): \[\rho_{s, 1990} = \delta_{s,1990} - \delta_{s,1989}\] The estimated annual rate of change is not defined for the first year of a species' time series, because we have no data from the previous year. The posterior mean estimates \(\left(\hat{\rho}_{s,t}\right)\) and standard errors \(\left(\sigma_{\rho_{s,t}}\right)\) of these annual rates of change for each species and year provide the input data for the composite indicator models in the second stage of the modeling.

Second Stage: Group Indicator Model

The composite indicator model estimates the average annual rate of change across the group of species for each year and then accumulates those average annual changes through time to estimate the composite indicator. This model also uses a measurement error structure to propagate the uncertainty of each species estimate from the first stage of the modeling into the second stage.

This second stage model uses the estimated annual rates of change for each species and year \(\left(\hat{\rho}_{s,t}\right)\) and the standard errors of these estimates \(\left(\sigma_{\rho_{s,t}}\right)\) from the first stage model as data. It accounts for the uncertainty in the estimates from the first stage by modeling the estimated rates of change \(\left(\hat{\rho}_{s,t}\right)\), random draws from a normal distribution, centered on the true rate of change \(\left(\rho_{s,t}\right)\) and with a standard deviation based on the estimate from the first stage model \(\left(\sigma_{\rho_{s,t}}\right)\). \[\hat{\rho}_{s,t} \sim \text{Normal}\left(\rho_{s,t}, \sigma_{\rho_{s,t}}\right)\]

This approach propagates the uncertainty from the first stage model into the second stage model and ensures that species with more uncertainty in their estimated annual change (i.e., higher values of \(\sigma_{\rho_{s,t}}\)), will have a less precise influence on the mean rate of change across species.

These true rates of change are then modeled as a function of an intercept term that represents the mean rate of change for each year across all species with data \(\left(\mu_t\right)\) and an error term that represents the variation among species in each year \(\left(\omega_{s,t} \sim \text{Normal}\left(0, \sigma_{\omega_t}\right)\right)\). \[\rho_{s,t} = \mu_t + \omega_{s,t}\]

The final composite indicator values \(\left(I_t\right)\) are then compiled by setting \(I_1 = 0\) in the first year and letting \(I_t= I_{t-1} + \mu_t + X_t\) where \(X_t \sim \text{Normal}(0,\sigma_{\omega_t})\), so that the annual estimates of mean rates of change accumulate through time into the full time series of indicator values and the inclusion of random noise based on the estimated among species variation in a given year \(\left(\text{Normal}\left(0, \sigma_{\omega_t}\right)\right)\) ensures that the full uncertainty is propagated through time.

Finally, we convert these log-scale indicator values into estimates of the average percent change since the base year as a direct exponential retransformation. \[\text{Percent Change since 1970} = 100\left(e^{I_t} - 1\right)\]

Aerial Insectivore Indicator Example

The indicator for the sub-group of aerial insectivores that includes the swallows, swifts, and nightjars is a compilation of the annual estimates for 13 bird species (Figure 2). The indicator line (black line with gray 95% credible interval) can be roughly interpreted as the average of the 13 species-specific lines (red lines for the 11 species that have decreased overall, and blue lines for the two species that have increased). Each of the species' lines influences the indicator to a degree dependent on its precision (the transparency of the species' lines in Figure 2 reflect the uncertainty of the species' data, more transparent lines for species with higher uncertainty). So this indicator runs approximately through the centre of the species lines, but is not simply the mean of all species lines. For example, the White-throated Swift and Bank Swallow lines have less influence compared to a simple average because their annual estimates are less precisely estimated than those for most other species in this group.

Variation Among Species

The indicators represent the average change in population across all of the species in the group. For many of the groups here, these averages are informative but may also contain a great deal of variation among species. To show some of the variation among individual species' trends within each indicator line, we present info-graphics composed of groupings of coloured bird icons showing the number of species in each of three categories of long term (since approximately 1970) changes in population (Table 3). The categories are asymmetrical to respect the same asymmetry as the scaling of the vertical axis (i.e., a 33% increase is required to recover from a 25% decrease). The variation shown in these info-graphics highlights the appropriate interpretation of these indicators-as indicators of the average or overall status of the group, not as indicators of the status of each species within the group. Our indicators give the best overall estimate of the group's status, but do not reflect the trends for all species in a group equally well; a stable indicator may reflect a group in which most or all species have stable trends, or it may reflect a group with an equal number of species with large increases and large decreases. Almost all of the indicators in the report, regardless of their overall pattern of change, include both species that are increasing and species that are decreasing. For example, populations of aerial insectivores as a group have decreased, but not all aerial insectivore populations have decreased (Figure 2).