Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Prediction of bird community composition based on point‐occurrence data and inferential algorithms: a valuable tool in biodiversity assessments

Prediction of bird community composition based on point‐occurrence data and inferential... Introduction A crisis exists with regard to the survey and inventory of life on Earth ( Wilson, 1988 ). Given the unknown nature of much of biological diversity, development of rigorous approaches for evaluation and understanding of patterns of the distribution and diversity of that life for purposes of conservation becomes critical ( Davis , 1990 ; Vane‐Wright , 1991 ; Lomolino, 1994 ; Peyton, 1994 ; Williams , 1996 ; Prendergast , 1999 ; Peters & Myers, 1991–92 ). Considerable effort has focused recently on predicting species’ geographical distributions based on known occurrences ( Nix, 1986 ; Austin , 1990 ; Jones & Gladkov, 1999 ). Given the poor state of knowledge of the distribution and natural history of most species, these interpolation techniques must work even in situations of small numbers of known occurrence points ( Peterson , in press ), as is the case for most species ( Peterson , 1998 ). One particular such interpolation technique, the Genetic Algorithm for Rule‐set Prediction (GARP) ( Stockwell & Noble, 1992 ; Stockwell & Peters, 1999 ; Stockwell & Peterson, in press ) has seen extensive testing with diverse regions and taxa ( Peterson & Cohoon, 1999 ; Peterson , 1999 ; Peterson & Vieglais, 2001 ; Peterson, 2001; Peterson ., 2002 ; Peterson , in press , Stockwell & Peterson, in press ). A critical extension of these approaches, however, is prediction of presences and absences across local communities of species. Single‐species tests have proven successful in this challenge ( Peterson & Cohoon, 1999 ; Stockwell & Peters, 1999 ; Godown & Peterson, 2000 ; Peterson & Vieglais, 2001 ; Peterson, 2001; Peterson ., 2002 ; Peterson , in press , Stockwell & Peterson, in press ), although assumptions regarding historical effects on species’ distributions may be necessary ( Peterson , 1999 ). However, tests of community‐level predictions have not, to our knowledge, been attempted. We report the results of such a test here: we predict local communities of birds across the Balsas Basin of south‐western México one species at a time, and test predictions based on independent field inventories. If such a methodology is viable, it may prove invaluable in making future surveys and inventories more efficient, providing preliminary hypotheses of key areas for conservation, and prioritizing areas for detailed study. METHODS Distributional data were derived from specimens housed in 32 natural history museums ( Peterson , 1998 ) (see Acknowledgements) and summar‐ized in the Atlas of Mexican Bird Distributions (Navarro and Peterson, in prep.). All records were georeferenced via direct consultation of 1 : 100 000 and 1 : 50 000 topographic maps ( Peterson , 1998 ), and unique locality records accumulated in spreadsheets for each species. As an initial filter, locality records (12 308 records of a total of 354 species in the region 17°–20°N latitude and 97°30′–103°5′W longitude) were separated for analysis. Later, we identified 89 resident species for which ≥ 20 specimens were available (median 61, range 20–399 specimens; median 29, range 8–81 unique latitude‐longitude combinations); the range of these species is broadly within the Balsas Basin’s dry desert scrub vegetation, based on personal experience and published habitat descriptions ( Howell & Webb, 1995 ). Ecological/environmental base layers (elevation, annual mean precipitation, annual mean temperature and potential vegetation) were obtained from the Comisión Nacional para el Conocimiento y Uso de la Biodiversidad (http://www.conabio.gob.mx) at a resolution of 0.01° (approximately 1 km) grid cells, and the study region was clipped out for analysis; point‐occurrence data were rasterized to this same grid system. Prior studies have shown that all four environmental data sets contribute significantly to the predictions ( Peterson & Cohoon, 1999; Peterson, unpubl. data). Species’ ecological niches can be defined as the range of conditions under which they are able to maintain populations without immigration ( Grinnell, 1917 ), a definition more easily visualized in geographical dimensions than later views ( MacArthur, 1972 ). One approach for modelling these niches is GARP, a machine‐learning procedure based on a genetic algorithm ( Stockwell & Noble, 1992 ; Stockwell & Peters, 1999 ; Stockwell, 1999 ). Occurrence points are divided evenly into training and test data sets. GARP works in an iterative process of rule selection, evaluation, testing, and incorporation or rejection: a method is chosen from a set of possibilities (e.g. logistic regression and bioclimatic rules), applied to the training data, and a rule is developed or evolved. Predictive accuracy is then evaluated based on 1250 points resampled from the test data and 1250 points sampled randomly from the study region as a whole to provide pseudoabsences (note that these subsamples and tests are internal to GARP, and are not available to the user for model validation, making necessary additional subsampling prior to analysis). Rules (which take the form of IF … THEN statements) may evolve by a number of means that mimic DNA evolution: point mutations, deletions, crossing over, etc. The change in predictive accuracy from one iteration to the next is used to evaluate whether a particular rule should be incorporated into the model, and the algorithm runs either 2560 iterations or until convergence. When the modelling process ceases, the ecological niche model is projected back onto the landscape to predict the species’ geographical distribution in the form of a raster grid, in which each grid square is designated as present or absent. GARP has seen strong statistical support for its predictive abilities under diverse conditions ( Peterson & Cohoon, 1999 ; Peterson & Vieglais, 2001 ; Peterson, 2001; Peterson ., 2002 ; Peterson , in press , Stockwell & Peterson, in press ). We developed GARP models for all 89 species, exported resulting predictions as ASCII raster grids, and uploaded them as GIS layers into ArcView (version 3.2). In ArcView, the 89 species’ predictions were overlaid, permitting queries of expected presence and absence for each species at any point across the study region. To test these community predictions, we accumulated single‐site inventory results from across the Balsas Basin: in all, data from eight localities were available to us. The sample sites were chosen a priori to fulfil three requirements: (a) to represent the variety of environments — and likely biotic communities–across the Balsas Basin, (b) to be accessible to permit efficient surveys, and (c) to be located in areas that were safe for access. Three were based on previous undergraduate thesis studies ( Villasenor Gomez, 1985 ; Feria, 1997 ; Mendez R., 1997 ), one was based on an inventory developed as part of a field course (Peterson et al ., unpubl. data), and four were surveyed by Feria as part of this study ( Feria, 2001 ). In all cases, inventories consisted of regular surveys by experienced field observers, in which species were recorded by sight and by call; sampling intensity ranged from 4 days (Peterson et al . unpubl. data) to many months ( Villasenor Gomez, 1985 ). For the six inventories in which we were involved, accumulation curves were used to assess inventory completeness ( Peterson & Slade, 1998 ). For model validation, for each site, predicted and observed faunas were organized in 2 × 2 contingency tables of species predicted present and absent vs. species actually present and absent [= confusion matrices ( Fielding & Bell, 1997 )], and compared using chi‐square tests. RESULTS AND DISCUSSION We developed GARP models for all 89 bird species (11 effectively endemic to the region) in the dry desert scrub vegetation of the Balsas Basin (examples in Fig. 1 ). Although the robustness of these predictions was not tested individually, owing to relatively small sample sizes within the Balsas Basin, prior tests of this type based on well‐sampled species across Mexico analysed with the same data sets have indicated excellent predictive ability ( Peterson ., 2002 ). Here, we analysed all species that met minimal sampling criteria, including many for which sample sizes were quite small. Combining these models across species produced clear concentrations of species richness and endemism ( Fig. 2 ): both were focused in narrow rings around the deepest and driest parts of the basin (see Feria & Peterson, unpubl. data, for a quantitative analysis of conservation priorities). 1 Example maps showing modelled geographical distributions of Passerina leclancherii (top) and Cacicus melanicterus (bottom) in the portion of their distributions that falls in the Balsas Basin of south‐western Mexico (thick, grey line shows limits of study region). Gray shading indicates predicted presence, and crosses represent known occurrences (note that some are very close to others and not all are visible); circles indicate the eight test inventory sites. 2 Geographic patterns summed across 89 bird species in dry tropical forests in the Balsas Basin of south‐western Mexico: top, species richness (grey scale: 0 species = white, 1–9 species = lightest grey, 10–19 species = light grey, … 60–66 species = black); bottom, richness of 16 species endemic to Mexico ( Escalante‐Pliego , 1993 ; Peterson & Navarro‐Siguenza, 1999 ; Navarro‐Siguenza & Peterson, 2000 ) (grey scale: 0 species = white, … 16 species = black). We validated these hypotheses of species composition at eight localities across the Basin ( Fig. 1 ). Accumulation curves for the six sites surveyed by us were all asymptotic, or nearly so, suggesting that the inventories were close to complete (Feria & Peterson, unpubl. data; but see below). At each of the eight sites, presence and absence of species was predicted statistically significantly better than at random ( P < 0.05), with false‐negative error rates ranging 2.3–19.1% and false‐positive error rates of 8.5–40.5% ( Table 1 ). 1 Summary of test sites and predictive accuracy of community predictions at the eight test sites across the Balsas Basin, south‐western Mexico No. Locality and source Data source Latitude Longitude Species predicted Species observed 1 Zicuiran, Michoacán Villaseñor Gomez (1985) 18°57′ 101°54′ 74 50 2 Tacámbaro, Michoacán Méndez (1997) 19°14′ 101°28′ 71 70 3 Infiernillo, Michoacán Feria (2001) 18°40′ 101°52′ 43 36 4 Las Tinajas, Guerrero Feria (2001) 18°21′ 100°43′ 40 37 5 Bejucos, Mexico Feria (2001 ) 18°45′ 100°25′ 38 38 6 Valente Trujillo, Guerrero Feria (2001 ) 17°56′ 99°34′ 32 37 7 Huautla, Morelos Peterson & Feria (unpubl. data) 18°47′ 98°99′ 44 44 8 Tlancualpicán, Puebla Feria (1997) 18°26′ 98°42′ 77 43 Interestingly, error in predictions was principally in terms of predicting species present that were not detected in the onsite inventories (false‐positive error) ( Table 1 ). At five sites, false‐positive error exceeded false‐negative error, but only at one was the converse the case (two sites were tied between the two error components). This difference suggests non‐detections in inventories, rather than model failure, as a principal factor in producing prediction error ( Peterson & Slade, 1998 ). To understand patterns of model failure in greater depth, we inspected lists of species on which the error statistics were based. The species most commonly predicted present but not detected were low‐density (e.g. Campephilus guatemalensis ), vulnerable to hunting pressure (e.g. Philortyx fasciatus ), or difficult to detect and identify (e.g. Cynanthus sordidus ) species. In contrast, species unexpectedly present included species poorly sampled in the specimen data base (e.g. Cathartes aura ) and species associated with human disturbance (e.g. Carduelis psaltria ). Hence, although many species were mispredicted at the eight test sites, patterns of error distribution suggest that further refinement of models and inventory data can yield much‐improved prediction of species’ presences and absences. The ability to predict many species’ distributions simultaneously may seem a minor extension of the ability to predict single species. If one species can be modelled with sufficient accuracy, as GARP has been shown to do efficiently in several studies ( Peterson & Cohoon, 1999 ; Peterson , 1999 ; Peterson ., 2000 ; Peterson & Vieglais, 2001 ; Peterson, 2001; Peterson ., 2002 ), at least for birds, then extension to multiple species should be a logical consequence. This methodology may, of course, have different strengths at different spatial scales (Ortega‐Huerta & Peterson, unpubl. data). For instance, at regional or continental scales, predictive ability may be high, yet at local scales, where interaction effects may be more common, predictive ability may prove lower. Nevertheless, this demonstration — GARP community predictions corroborated with actual field inventory data — opens the way for several new capabilities for biodiversity surveys, inventories and analyses. Community composition can be assessed based on existing information across large regions before initiation of fieldwork, and key sites (hotspots of diversity and endemism, areas with unexpected species combinations, etc.) can be chosen for field study. These key sites can then be surveyed to produce more reliable onsite field data, but less critical or less interesting sites can be accorded lower priority. In this way, the always‐limited resources available for biodiversity survey and inventory ( Peterson , 1998 ) can be put to maximum use, and the information obtained from results can maximized. Although error in our predictions was non‐trivial, additional testing under diverse conditions combined with more robust inventory data may produce a methodology that has the necessary predictive ability. An additional theoretical implication also arises. To the extent that species composition can be predicted from individual, niche‐based models for each species, then autecology would emerge as the force that dominates the shaping of biological communities, as opposed to widespread interactions among species. Of course, this conclusion will depend largely on the spatial scale at which the question is asked, so this first result must await detailed analysis and testing. Nevertheless, our approach presents a new opportunity to attack this fundamental, but very difficult, ecological question. Obviously, the evidence remains equivocal, but parallel tests of community composition vs. species’ autecology in diverse regions for varied taxa would present a very useful insight. Acknowledgments We thank Dr A.G. Navarro S. for his kind assistance in many phases of this study. Museums kindly providing data include: American Museum of Natural History, New York; Academy of Natural Sciences, Philadelphia; Bell Museum of Natural History, Minneapolis; British Museum (Natural History), Tring; Carnegie Museum of Natural History, Pittsburg; Canadian Museums of Nature, Ottawa; Denver Museum of Natural History; Delaware Museum of Natural History; Fort Hays State University; Field Museum of Natural History, Chicago; Iowa State University, Ames; University of Kansas Natural History Museum, Lawrence; Los Angeles County Museum of Natural History; Natuurhistorische Museum, Amsterdam; Louisiana State University Museum of Zoology, Baton Rouge; Museum of Comparative Zoology, Harvard University, Cambridge; Moore Laboratory of Zoology, Occidental College, Los Angeles; Museum Nationale d’Histoire Naturelle, Paris; Museum of Vertebrate Zoology, Berkeley; Museo de Zoología, Facultad de Ciencias, Universidad Nacional Autónoma de México; University of Nebraska, Lincoln; Royal Ontario Museum, Toronto; San Diego Natural History Museum; Southwestern College, Kansas; Texas Cooperative Wildlife Collections; University of Arizona; University of British Columbia Museum of Zoology; University of California at Los Angeles; Universidad Michoacana de San Nicolás de Hidalgo; US National Museum of Natural History; Western Foundation of Vertebrate Zoology; and Peabody Museum, Yale University. This work was supported by the US National Science Foundation, and TPFA was supported by a graduate fellowship from the Consejo Nacional de Ciencia y Tecnología. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Diversity and Distributions Wiley

Prediction of bird community composition based on point‐occurrence data and inferential algorithms: a valuable tool in biodiversity assessments

Loading next page...
 
/lp/wiley/prediction-of-bird-community-composition-based-on-point-occurrence-FotIpX0AVs

References (38)

Publisher
Wiley
Copyright
Copyright © 2002 Wiley Subscription Services, Inc., A Wiley Company
ISSN
1366-9516
eISSN
1472-4642
DOI
10.1046/j.1472-4642.2002.00127.x
Publisher site
See Article on Publisher Site

Abstract

Introduction A crisis exists with regard to the survey and inventory of life on Earth ( Wilson, 1988 ). Given the unknown nature of much of biological diversity, development of rigorous approaches for evaluation and understanding of patterns of the distribution and diversity of that life for purposes of conservation becomes critical ( Davis , 1990 ; Vane‐Wright , 1991 ; Lomolino, 1994 ; Peyton, 1994 ; Williams , 1996 ; Prendergast , 1999 ; Peters & Myers, 1991–92 ). Considerable effort has focused recently on predicting species’ geographical distributions based on known occurrences ( Nix, 1986 ; Austin , 1990 ; Jones & Gladkov, 1999 ). Given the poor state of knowledge of the distribution and natural history of most species, these interpolation techniques must work even in situations of small numbers of known occurrence points ( Peterson , in press ), as is the case for most species ( Peterson , 1998 ). One particular such interpolation technique, the Genetic Algorithm for Rule‐set Prediction (GARP) ( Stockwell & Noble, 1992 ; Stockwell & Peters, 1999 ; Stockwell & Peterson, in press ) has seen extensive testing with diverse regions and taxa ( Peterson & Cohoon, 1999 ; Peterson , 1999 ; Peterson & Vieglais, 2001 ; Peterson, 2001; Peterson ., 2002 ; Peterson , in press , Stockwell & Peterson, in press ). A critical extension of these approaches, however, is prediction of presences and absences across local communities of species. Single‐species tests have proven successful in this challenge ( Peterson & Cohoon, 1999 ; Stockwell & Peters, 1999 ; Godown & Peterson, 2000 ; Peterson & Vieglais, 2001 ; Peterson, 2001; Peterson ., 2002 ; Peterson , in press , Stockwell & Peterson, in press ), although assumptions regarding historical effects on species’ distributions may be necessary ( Peterson , 1999 ). However, tests of community‐level predictions have not, to our knowledge, been attempted. We report the results of such a test here: we predict local communities of birds across the Balsas Basin of south‐western México one species at a time, and test predictions based on independent field inventories. If such a methodology is viable, it may prove invaluable in making future surveys and inventories more efficient, providing preliminary hypotheses of key areas for conservation, and prioritizing areas for detailed study. METHODS Distributional data were derived from specimens housed in 32 natural history museums ( Peterson , 1998 ) (see Acknowledgements) and summar‐ized in the Atlas of Mexican Bird Distributions (Navarro and Peterson, in prep.). All records were georeferenced via direct consultation of 1 : 100 000 and 1 : 50 000 topographic maps ( Peterson , 1998 ), and unique locality records accumulated in spreadsheets for each species. As an initial filter, locality records (12 308 records of a total of 354 species in the region 17°–20°N latitude and 97°30′–103°5′W longitude) were separated for analysis. Later, we identified 89 resident species for which ≥ 20 specimens were available (median 61, range 20–399 specimens; median 29, range 8–81 unique latitude‐longitude combinations); the range of these species is broadly within the Balsas Basin’s dry desert scrub vegetation, based on personal experience and published habitat descriptions ( Howell & Webb, 1995 ). Ecological/environmental base layers (elevation, annual mean precipitation, annual mean temperature and potential vegetation) were obtained from the Comisión Nacional para el Conocimiento y Uso de la Biodiversidad (http://www.conabio.gob.mx) at a resolution of 0.01° (approximately 1 km) grid cells, and the study region was clipped out for analysis; point‐occurrence data were rasterized to this same grid system. Prior studies have shown that all four environmental data sets contribute significantly to the predictions ( Peterson & Cohoon, 1999; Peterson, unpubl. data). Species’ ecological niches can be defined as the range of conditions under which they are able to maintain populations without immigration ( Grinnell, 1917 ), a definition more easily visualized in geographical dimensions than later views ( MacArthur, 1972 ). One approach for modelling these niches is GARP, a machine‐learning procedure based on a genetic algorithm ( Stockwell & Noble, 1992 ; Stockwell & Peters, 1999 ; Stockwell, 1999 ). Occurrence points are divided evenly into training and test data sets. GARP works in an iterative process of rule selection, evaluation, testing, and incorporation or rejection: a method is chosen from a set of possibilities (e.g. logistic regression and bioclimatic rules), applied to the training data, and a rule is developed or evolved. Predictive accuracy is then evaluated based on 1250 points resampled from the test data and 1250 points sampled randomly from the study region as a whole to provide pseudoabsences (note that these subsamples and tests are internal to GARP, and are not available to the user for model validation, making necessary additional subsampling prior to analysis). Rules (which take the form of IF … THEN statements) may evolve by a number of means that mimic DNA evolution: point mutations, deletions, crossing over, etc. The change in predictive accuracy from one iteration to the next is used to evaluate whether a particular rule should be incorporated into the model, and the algorithm runs either 2560 iterations or until convergence. When the modelling process ceases, the ecological niche model is projected back onto the landscape to predict the species’ geographical distribution in the form of a raster grid, in which each grid square is designated as present or absent. GARP has seen strong statistical support for its predictive abilities under diverse conditions ( Peterson & Cohoon, 1999 ; Peterson & Vieglais, 2001 ; Peterson, 2001; Peterson ., 2002 ; Peterson , in press , Stockwell & Peterson, in press ). We developed GARP models for all 89 species, exported resulting predictions as ASCII raster grids, and uploaded them as GIS layers into ArcView (version 3.2). In ArcView, the 89 species’ predictions were overlaid, permitting queries of expected presence and absence for each species at any point across the study region. To test these community predictions, we accumulated single‐site inventory results from across the Balsas Basin: in all, data from eight localities were available to us. The sample sites were chosen a priori to fulfil three requirements: (a) to represent the variety of environments — and likely biotic communities–across the Balsas Basin, (b) to be accessible to permit efficient surveys, and (c) to be located in areas that were safe for access. Three were based on previous undergraduate thesis studies ( Villasenor Gomez, 1985 ; Feria, 1997 ; Mendez R., 1997 ), one was based on an inventory developed as part of a field course (Peterson et al ., unpubl. data), and four were surveyed by Feria as part of this study ( Feria, 2001 ). In all cases, inventories consisted of regular surveys by experienced field observers, in which species were recorded by sight and by call; sampling intensity ranged from 4 days (Peterson et al . unpubl. data) to many months ( Villasenor Gomez, 1985 ). For the six inventories in which we were involved, accumulation curves were used to assess inventory completeness ( Peterson & Slade, 1998 ). For model validation, for each site, predicted and observed faunas were organized in 2 × 2 contingency tables of species predicted present and absent vs. species actually present and absent [= confusion matrices ( Fielding & Bell, 1997 )], and compared using chi‐square tests. RESULTS AND DISCUSSION We developed GARP models for all 89 bird species (11 effectively endemic to the region) in the dry desert scrub vegetation of the Balsas Basin (examples in Fig. 1 ). Although the robustness of these predictions was not tested individually, owing to relatively small sample sizes within the Balsas Basin, prior tests of this type based on well‐sampled species across Mexico analysed with the same data sets have indicated excellent predictive ability ( Peterson ., 2002 ). Here, we analysed all species that met minimal sampling criteria, including many for which sample sizes were quite small. Combining these models across species produced clear concentrations of species richness and endemism ( Fig. 2 ): both were focused in narrow rings around the deepest and driest parts of the basin (see Feria & Peterson, unpubl. data, for a quantitative analysis of conservation priorities). 1 Example maps showing modelled geographical distributions of Passerina leclancherii (top) and Cacicus melanicterus (bottom) in the portion of their distributions that falls in the Balsas Basin of south‐western Mexico (thick, grey line shows limits of study region). Gray shading indicates predicted presence, and crosses represent known occurrences (note that some are very close to others and not all are visible); circles indicate the eight test inventory sites. 2 Geographic patterns summed across 89 bird species in dry tropical forests in the Balsas Basin of south‐western Mexico: top, species richness (grey scale: 0 species = white, 1–9 species = lightest grey, 10–19 species = light grey, … 60–66 species = black); bottom, richness of 16 species endemic to Mexico ( Escalante‐Pliego , 1993 ; Peterson & Navarro‐Siguenza, 1999 ; Navarro‐Siguenza & Peterson, 2000 ) (grey scale: 0 species = white, … 16 species = black). We validated these hypotheses of species composition at eight localities across the Basin ( Fig. 1 ). Accumulation curves for the six sites surveyed by us were all asymptotic, or nearly so, suggesting that the inventories were close to complete (Feria & Peterson, unpubl. data; but see below). At each of the eight sites, presence and absence of species was predicted statistically significantly better than at random ( P < 0.05), with false‐negative error rates ranging 2.3–19.1% and false‐positive error rates of 8.5–40.5% ( Table 1 ). 1 Summary of test sites and predictive accuracy of community predictions at the eight test sites across the Balsas Basin, south‐western Mexico No. Locality and source Data source Latitude Longitude Species predicted Species observed 1 Zicuiran, Michoacán Villaseñor Gomez (1985) 18°57′ 101°54′ 74 50 2 Tacámbaro, Michoacán Méndez (1997) 19°14′ 101°28′ 71 70 3 Infiernillo, Michoacán Feria (2001) 18°40′ 101°52′ 43 36 4 Las Tinajas, Guerrero Feria (2001) 18°21′ 100°43′ 40 37 5 Bejucos, Mexico Feria (2001 ) 18°45′ 100°25′ 38 38 6 Valente Trujillo, Guerrero Feria (2001 ) 17°56′ 99°34′ 32 37 7 Huautla, Morelos Peterson & Feria (unpubl. data) 18°47′ 98°99′ 44 44 8 Tlancualpicán, Puebla Feria (1997) 18°26′ 98°42′ 77 43 Interestingly, error in predictions was principally in terms of predicting species present that were not detected in the onsite inventories (false‐positive error) ( Table 1 ). At five sites, false‐positive error exceeded false‐negative error, but only at one was the converse the case (two sites were tied between the two error components). This difference suggests non‐detections in inventories, rather than model failure, as a principal factor in producing prediction error ( Peterson & Slade, 1998 ). To understand patterns of model failure in greater depth, we inspected lists of species on which the error statistics were based. The species most commonly predicted present but not detected were low‐density (e.g. Campephilus guatemalensis ), vulnerable to hunting pressure (e.g. Philortyx fasciatus ), or difficult to detect and identify (e.g. Cynanthus sordidus ) species. In contrast, species unexpectedly present included species poorly sampled in the specimen data base (e.g. Cathartes aura ) and species associated with human disturbance (e.g. Carduelis psaltria ). Hence, although many species were mispredicted at the eight test sites, patterns of error distribution suggest that further refinement of models and inventory data can yield much‐improved prediction of species’ presences and absences. The ability to predict many species’ distributions simultaneously may seem a minor extension of the ability to predict single species. If one species can be modelled with sufficient accuracy, as GARP has been shown to do efficiently in several studies ( Peterson & Cohoon, 1999 ; Peterson , 1999 ; Peterson ., 2000 ; Peterson & Vieglais, 2001 ; Peterson, 2001; Peterson ., 2002 ), at least for birds, then extension to multiple species should be a logical consequence. This methodology may, of course, have different strengths at different spatial scales (Ortega‐Huerta & Peterson, unpubl. data). For instance, at regional or continental scales, predictive ability may be high, yet at local scales, where interaction effects may be more common, predictive ability may prove lower. Nevertheless, this demonstration — GARP community predictions corroborated with actual field inventory data — opens the way for several new capabilities for biodiversity surveys, inventories and analyses. Community composition can be assessed based on existing information across large regions before initiation of fieldwork, and key sites (hotspots of diversity and endemism, areas with unexpected species combinations, etc.) can be chosen for field study. These key sites can then be surveyed to produce more reliable onsite field data, but less critical or less interesting sites can be accorded lower priority. In this way, the always‐limited resources available for biodiversity survey and inventory ( Peterson , 1998 ) can be put to maximum use, and the information obtained from results can maximized. Although error in our predictions was non‐trivial, additional testing under diverse conditions combined with more robust inventory data may produce a methodology that has the necessary predictive ability. An additional theoretical implication also arises. To the extent that species composition can be predicted from individual, niche‐based models for each species, then autecology would emerge as the force that dominates the shaping of biological communities, as opposed to widespread interactions among species. Of course, this conclusion will depend largely on the spatial scale at which the question is asked, so this first result must await detailed analysis and testing. Nevertheless, our approach presents a new opportunity to attack this fundamental, but very difficult, ecological question. Obviously, the evidence remains equivocal, but parallel tests of community composition vs. species’ autecology in diverse regions for varied taxa would present a very useful insight. Acknowledgments We thank Dr A.G. Navarro S. for his kind assistance in many phases of this study. Museums kindly providing data include: American Museum of Natural History, New York; Academy of Natural Sciences, Philadelphia; Bell Museum of Natural History, Minneapolis; British Museum (Natural History), Tring; Carnegie Museum of Natural History, Pittsburg; Canadian Museums of Nature, Ottawa; Denver Museum of Natural History; Delaware Museum of Natural History; Fort Hays State University; Field Museum of Natural History, Chicago; Iowa State University, Ames; University of Kansas Natural History Museum, Lawrence; Los Angeles County Museum of Natural History; Natuurhistorische Museum, Amsterdam; Louisiana State University Museum of Zoology, Baton Rouge; Museum of Comparative Zoology, Harvard University, Cambridge; Moore Laboratory of Zoology, Occidental College, Los Angeles; Museum Nationale d’Histoire Naturelle, Paris; Museum of Vertebrate Zoology, Berkeley; Museo de Zoología, Facultad de Ciencias, Universidad Nacional Autónoma de México; University of Nebraska, Lincoln; Royal Ontario Museum, Toronto; San Diego Natural History Museum; Southwestern College, Kansas; Texas Cooperative Wildlife Collections; University of Arizona; University of British Columbia Museum of Zoology; University of California at Los Angeles; Universidad Michoacana de San Nicolás de Hidalgo; US National Museum of Natural History; Western Foundation of Vertebrate Zoology; and Peabody Museum, Yale University. This work was supported by the US National Science Foundation, and TPFA was supported by a graduate fellowship from the Consejo Nacional de Ciencia y Tecnología.

Journal

Diversity and DistributionsWiley

Published: Mar 1, 2002

There are no references for this article.