The BCN Survey, which began in 1997, is a breeding season monitoring program implemented in the Chicago Wilderness region by the BCN and Audubon Chicago Region. Volunteers are recruited and trained in conjunction with land management agencies throughout the region. Approximately 180 experienced volunteer monitors now collect data through methods developed in 1997 by a team of CW land managers, birders and ornithologists. The methods are point count, transect and checklist. Data is entered on line using a specially adapted version of eBird, developed with BirdSource, a joint project of the Cornell University Lab of Ornithology and the National Audubon Society. The database is shared with land managers and researchers annually.
In the past, a lack of standardized protocols and the need to digitize extensive field data meant that much birders’ information remained largely inaccessible or at best difficult to use for region-wide studies. The standardized protocol of the BCN Survey transformed avian monitoring practices in the Chicago area. As this was one of the first projects to extensively collect birders’ sightings online, we hope that the analysis of the data sixteen years into the project will be of help to the many others around the country who are organizing similar projects.
The primary purpose for the collection of these data is to provide land managers with information that can help guide restoration practices. Its creators envisioned a second purpose: when the database grew large enough, it might be used to describe local population trends for species of concern. After amassing over 100,000 records and eight years of data, in 2004 BCN and Audubon-Chicago Region set out to investigate species trends in the region. This initial phase showed that the database is large enough to yield useful data for many species, and that it is possible to apply analysis methods that will yield credible trends.
In 2007, after amassing over 180,000 records, a more intensive analysis attempted to refine previous results and ask new questions. In 2013, after another five years had passed, a third analysis was initiated to update the trends and in an effort to continue to improve the accuracy and significance of the results. The results from this latest analysis are displayed here on the website. Today, the BCN ebird database contains over 3.5 million records including over 125,000 records following the breeding season point count protocol – well over the entire database of all records from 2004. The 2013 analysis was performed on the 125,000 breeding season point count records in the BCN Survey database.
POINT COUNT PROTOCOL
The BCN Survey has three protocols that are used by our monitors, but only the point count data were selected for this analysis. Point counts are conducted during the breeding season. Monitors are asked to conduct a minimum of two visits during June. They are instructed to start close to sunrise (between 5:15 a.m. and 5:20 a.m.), and complete counts by no later than 9:00 a.m. Points are located at least 150 meters apart. Point counts are five minutes in length and all birds seen or heard within a 75 meter radius of the point are recorded. Birds detected outside the radius are noted separately. It is requested that monitors walk the route in a different order on different days in order to maximize the number of birds detected during the hours when birds are most active.
After monitoring has been completed, data is entered on the BCN eBird website.
For more details regarding our survey methods, please visit the BCN Survey protocol page.
POINT SELECTION METHODS
The BCN Survey is designed to cover preserved land in the Chicago Wilderness region. Preserves were chosen based on preferences of available citizen science monitors, and thorough coverage of the preserve or a section of the preserve was attempted through point placement. Points were selected in conjunction with land managers and were selected in a variety of ways. Some were placed at intervals designed to get the best views or coverage of the habitat being surveyed. Others were randomly placed throughout the site, often using a grid method. Others were placed along walkable trails at intervals of 150 feet. In recent years, points have also been randomly assigned in habitat units. Points are almost exclusively located in natural areas that are usually owned by one of the county forest preserve districts, the state, or local municipalities. The Forest Preserve District of DuPage County point count data is also included in BCN eBird and this analysis. Their data are collected using similar protocols, but point counts last ten minutes. They collect data in two five minute intervals. For the sake of consistency, we used only the first five minute interval for our analysis.
The BCN Survey now contains sixteen years of point count data. Based on results from the first trends analysis in 2004, it was determined that there is little information content in the first two years' data and therefore more accurate trends are revealed when the data from the first two years are omitted (1997-1998). Data was limited during these initial "start-up" years of the BCN Survey, and a full level of participation from volunteers was not achieved until 1999 (see Figure 1).
Data collected from the BCN eBird website was imported into BirdSTATs, "an open source Microsoft Access database for the preparation and statistical analysis of bird counts data in a standardised way. The BirdSTATs tool is programmed to use and automatically run the program TRIM (Trends and Indices for Monitoring data) in batch mode to perform the statistical analysis for series of bird counts in the dataset."
"The BirdSTATs tool is developed at the request of the Pan European Common Bird Monitoring Scheme (PECBMS) by Bioland Informatie. Designing and programming of the tool is funded by the European Commission through British Royal Society for the Protection of Birds (RSPB)." The bulk of the analysis work is then completed in a statistical software package called TRIM (Trends & Indices for Monitoring Data), developed by Statistics Netherlands. Microsoft Access was also used to substantially decrease manual labor / work time and automate many tasks including the export of records (for each species), and to automate the creation of web pages for each species (including automatically inserted titles, names, data, linked images, and linked graphs).
The number of birds of each species detected at each point was calculated for each year. Records of adults and juveniles are not separated out by the BCN Survey. Species not recorded on a point count during a visit are assigned a count value of zero in order to distinguish between species absences during survey years and missing data during non-survey years. Data is limited to observations between June 1st and July 15th as a means of attempting to filter out some of the records pertaining to late migrants (common through end of May), post-breeding dispersal (July-Aug.), and other factors. Based on general consensus from our scientific advisors, the inclusion of July records provides better data for breeding species that arrive late in the season and still avoids the period when post-breeding dispersal becomes an issue. Because point counts are visited multiple times during the breeding season, summary statistics must be developed for each species during each year of the analysis. Mean abundances were used for this analysis. Although maximum abundance is fairly widely accepted, mean abundance is used more frequently among researchers. A number of extraneous variables are more associated with maximum abundances. Data could be inflated by the occasional presence of migrants, nomadic individuals, or birds that occasionally wander outside of their established territories. The number of visits and observer effort also have more significant effects on the results when using maximum abundance. Additional valuable information can be found in the recent publication: Betts, M., Simon, N.P.P., Nocera, J.J. 2005. Point count summary statistics differentially predict reproductive activity in bird-habitat relationship studies. Journal of Ornithology 146: 151-159.
A wide diversity of methods exists for analyzing data in order to reveal population trends of wildlife (Thomas 1996). For analysis of the BCN data, a log-linear analysis was conducted using TRIM, the comprehensive software package developed by Statistics Netherlands noted above (TRIM: Trends and Indices for Monitoring Data 2006) for the explicit purpose of analyzing wildlife population trend data collected by numerous individuals over large spatial scales (Fewster et al. 2000). The TRIM software accounts for the large variances associated with this type of data and models the autocorrelation. TRIM is used by the British Trust for Ornithology and has also been adopted by the European Bird Census Council and is being used throughout the European Union on a region-wide scale (European Bird Census Council).
Count data collected by volunteers often has missing values (Fewster et al. 2000). The TRIM program estimates missing values based on data from plots monitored during the same time frame. The algorithms in TRIM model the structure of the count data and, making assumptions about the structure of the data, impute the missing values. The objective is to estimate a model using the observed counts and then to use this model to predict the missing counts. When analyzing large, incomplete databases, simulation studies have shown this approach does a better job of describing trends in wildlife abundance than other alternative methods. Since it requires at least two points in a time series to estimate parameters, points with data from only one year are excluded from analyses (Pannekoek and van Strien 1996).
Annual population indices for each species are calculated using TRIM. This program uses a log-linear regression model with Poisson error terms for the analysis of time series of counts that contain missing observations (we estimated only linear trends in our analyses, although TRIM can also model some forms of non-linear trends).
Because the actual numbers of individuals are always unknown, time series are converted to index numbers with the base year set as the year the census began, or in this case the first year that the census had an adequate amount of data for analysis (1999). The base year is set at 1. The index values on the resulting graphs show how percentages have changed with respect to the base year and allow for easier comparison of changes for various species (see Figure 2).
The TRIM program also attempts to take into account the effects of overdispersion and serial correlation. "The usual (maximum likelihood) approach to estimation and testing procedures for count data are based on the assumption of independent Poisson distributions (or a multinomial distribution) for the counts. Such an assumption is likely to be violated for counts of animals because the variance is often larger than expected for a Poisson distribution (overdispersion), especially when they occur in colonies. Furthermore, the counts are often not independently distributed because the counts in a particular year will also depend on the counts in the year before (serial correlation). Therefore, TRIM uses statistical procedures for estimation and testing that take these two phenomena into account." (Pannekoek & van Strien 1996). Outputted errors will be higher for species that do not fit standard assumptions due to either of those reasons.
TRIM offers a variety of trend analysis methods. Various models were tested in 2007, and the Linear Trend model was used in 2007 and again in 2013 to analyze each species because trends across the entire monitoring period were desired and data for some species is too sparse to run the Time Effects model. It is also possible to test for specific site effects. With this type of analysis the investigator does not actually test for changes in population size at any particular site, but whether the number of individuals for a particular species are changing significantly differently among sites across all sites on average. Specifically, TRIM Model 2 was used: a model with a site-effect and log-linear time effect.
When calculating trends in TRIM, two indices are calculated in the program and displayed on graphs generated by the program: Imputed and Model-based. (See Figure 2.)
Imputed indices: Observed counts are used when available. Missing counts are predicted (more realistic course in time).
- Imputed indices use total counts for species for each year a point is monitored. For point / year combinations without species total counts (years when a particular point was not monitored), estimated total counts are based on other existing data are used.
- Only missing values are estimated.
Model-based indices: All counts are based on model predictions.
- Model indices always display estimated totals for all point / year combinations (including years when a particular point was monitored).
- All values are estimated
Two tests are used to determine the Goodness of Fit of each model: The Pearson’s chi-squared statistic ("Chi-square") and the likelihood ratio statistic ("Likelihood Ratio") tests (see Pannekoek and van Strien 1996, Sec. 2.5 for further details). The Chi-square p-values have been included with the graphs on the website.
If the model fails, the "model" line and estimated trends are significantly different from the data. For species that fail the Goodness of Fit test, the failure can sometimes be explained by significantly different trends in different habitat types indicating that a species is increasing in one habitat where it breeds while declining in another. In these instances, the Wald-Test for significance of habitat covariates has a significant p-value of less than 0.05, and the trends calculated for each of the individual habitats was closely observed in addition to the overall trend.
Of the 116 species with sufficient data to be analyzed using TRIM, 97 (83.6%) passed the goodness-of-fit test and fit the linear trend model (compared to 66.4% in the 2007 analysis).
General credibility ratings for the analysis were assigned based on relative degrees of deficiency in the data. These ratings aid the casual viewer in comparing the relative accuracy of trends and help highlight species trends that we feel have notable deficiencies and should be evaluated more cautiously. This concept is comparable to the rationale for the "regional credibility measures" that are now being used by the Breeding Bird Survey. Credibility ratings were assigned based on a broad interpretation of the ranges of error associated with the trends, coupled with goodness of fit and the number of records.