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

Learn More →

Building better wildlife‐habitat models

Building better wildlife‐habitat models Wildlife-habitat models are an important tool in wildlife management today, and by far tbe majority of these predict aspects of species distribution (abundance or presenee) as a proxy measure of babitat quality. L'nfortmiately, few are tested on independent data., anii of those Ihat are, few show useful predictive skill. We demonstrate tbat six eritieal assumptions nnderlie distribution based wildlife-habitat models, all of wbieb must be valid for tbe model to predict habitat quality. We outline tbese assumptions in a meta-model, and discuss metbods for tbeir validation. Even wbere all six assumptions sbow a bi}>b level of validity, tbere is still a strong likelibood that tbe model will not prediet habitat qnality. However, tbe meta-model does suggest babitat quality can be predicted more accurately if distributional data are ignored, and variables more indicative of babitat quality are modelled instead. Wildlile-habitat models appear regularly in the piiblished literature, and most use distributional data (abundance or presence) as their dependent variable. Ecologists develop these models Tor at least two purposes, to explain wildlife distributions and predict habitat use (Patton 1992, Morrison et al. 1992), but in all cases assume that habitat quality is indexed by distribution (Van Home 1983. Schamberger and O'Neil 1986. Hobbs and Hanley 1990). so that model output prediets habitat quality. Superticially. many distributional models appear to predict well, generally showing exeellent tit to the data from which they were derived (the training set). Unfortunately, models are optimised for their training set, so goodness-ot-tit to this data is an overly optimistic estimate of predictive skill outside the training set (Mostclier and Tukey 1977. Picard and Cook 1984). Despite this, an alarmingly large number of even the most recent studies tbcus only on the model's fit to the training set (Brown and Nicholls 1993, Sarre 1995, Pausas et al. 1995. Sarre et al. 1995, St-Georges et al. 1995, Poizat and Pont 1996. Abensperg-Trauii et al. 1996. Beyer et al. 1996. Smith et al. 1996. Welsh et al. ECOORAPHY 22:2 (19991 1996. Pedlar et al. 1997). Since this is an optimistically biased measure of predictive skill, the number of published models with genuine predictive skill is probably far fewer than a quick perusal of the literature suggests. Only a relatively small proportion of models are tested on data not used for model development (a test set), and among these, many do not fit the test set well (e.g. Rotenberry 1986. Lindeninayer et al. 1994, Fielding and Haworth 1995, Lindenmayer et al. 1995). Fitting models to a test set is a more useful test of predictive skill, though such tests are still data dependent (Oreskes et al. 1994), so even if a model fits one test set well, it will not necessarily fit others equally. Modellers need to understand the predictive skill of the models they develop, but current approaches provide limited guidance. In this paper we propose a meta-model which describes distribution based wildlifehabitat models, and discuss its three most important implications for building models that predict better. Firstly, we identify six critical assumptions which must all be valid for a model to predict habitat quality. Secondly, we discuss how each of these assumptions can be validated. Finally, and most importantly, we demonstrate why even if all critical assumptions are valid, a distribution based model is still unlikely to predict habitat quality, and argue that more direct measures of habitat quality should be modelled in preference to distributional data. The critical assumptions Superficially, distribution based models have a simple logic; habitat determines species distribution, so habitat characteristics predict wildlife distribution (Fig. lA). This conceptual model relies on two critical assumptions: that wildlife distributions are inlluenced by habi219 tat, and that predicted distributions are adequately modelled from habitat data. Unfortunately, factors such as predation and climale also inlUience species distribution (Schambcrger and O'Neil 1986). weakening the first assumption. The second assumption is invalid if data are modelled inappropriately, for example, when a nonlinear relationship is modelled linearly, or a nonadditive process is modelled addilively. Even if habitat does aeeount for substantial variation in species distribution, and appropriate modelling techniques are used, successful modelling further depends on how both habitat and distribution are quantified (Fig. IB). Important habitat variables can be overlooked during sampling, and their perceived significance ean depend on how they are quantified (i.e. intensiveness. extensiveness. definition). Similarly, the strength of the habitat-distribution relationship also depends on how distribution is measured (e.g. presence, absolute abundance, relative abundance, spoor, nest presence, etc.). As a consequence, both habitat and distribution must be measured adequately for predicted distribution and measured distribution to be adequately linked. Because the measurement of habitat quality underlies all distribution based wildlife-habitat models (Van Home 1983, Hobbs and Hanley 1990). two final assumptions are required to complete the meta-model (Fig. !C). A model might predict measured distribution very well, but if measured distribution does not equate with habitat quality, the model has no clear interpretation in terms of target speeies welfare. Thus, it is assumed that measured distribution equates with habitat quality. Habitat quality is an ambiguous expression, most widely defined in terms of a habitat's capacity to support a species (Maurer 1986, Brennan et al. 1986. Paradis and Croset 1995). Important aspects of habitat quality include reproductive success, survivorship and population density (Van Horne 1983), and the exact way each of these ultimately contribute to the habitat quality metric will determine the strength of the iink between measured and actual habitat quality. Thus, the final assumption in the meta-model requires that habitat quality is adequately measured. The validity of these latter two assumptions will not directly affect the ability of a distribution based model to predict measured distribution, but without a link between measured distribution and habitat quality, a model has no clear use, so these assumptions are important to a model's biological validity. It is important to note that the six critical assumptions outlined above are linearly related. Consequently, each assumption has equal importance, since any invalid assumption, or subset of assumptions, breaks the link between predicted distribution and measured habitat quality, regardless of the validity of any other assumption. Models that accurately predict habitat quality are therefore those for which all critical assumptions are valid, and assumption validation is a critical step in the process of identifying models with rehable predictive ability. Validating tiie criticai assumptions Assumption I is tested by fitting the model to the parent data, or even better, to independent data (Table 1). If the fit is good, then the validity ofthis assumption is partially supported. However, this strategy does not test assumption I in isolation from assumptions 2 through 4, since all four must be valid for the model to fit any data. This strategy is therefore inconclusive in the case of a negative result, unless it is used with more direct tests of assumptions 2-4. Independent data are strongly preferred over the parent data for this test because it provides a more realistic estimate of the Predicted distribution Habitat Dislribution Predicted distribution Measured habitat Habitat Distribution r y Measured distribution Predicted distribution .—^ r—V Measured habitat Habitat Distributioii Measured distribution Habitat quality Measured habitat quality Fig. I. Development of a meta-model explaining the underlying assutnptions of distribution based uildlifc-habitiit models. Bold nutnbers represent erilical model assumptions. 1: that habitat influetices distribution: 2; that predicted distribution is adequately modelled; 3: that habitat is adequately measured; 4: that distribution is adequately measured; 5: that measured distribution equates with habitat qtiality; 6; that habitat quality is adequately measured. b C O G K A P H Y 2^.2 Table 1. Validation strategies for distribution based wildlifehabitat rnodels. Validation strategics Assumption t: That habitat influences distribution. Fit the tiiodel to parent daia. Fit the model to independent data. Manipulate habitat and observe distributional changes. Assumption 2: That predicted distribution is adequately modelled. Check the model's underlying statistical assumptions. Apply several alternative modelling strategics. j^ssurnption 3: That habitat is adequately measured. Use standardised measurement procedures. Test measurement accuracy and precision. Assumption 4: That distribution is adequately measured. Use standardised measurement procedures. I Test measurement accuracy and precision. Assumption 5: That measured distribution equates with habitat quality. Test the relationship between habitat quality and measured distribution. Assumption 6; That habitat quality is adequately measured. Use standardised measurement procedures. Test measurement accuracy, precision, and scale adequacy. model's predictive power. A tiiorc direct alternative for testing assumption I is by experimental manipulation of habitat. F'or example, in a model which identifies grass biomass as an important indicator of herbivore abundance, grass biomass could be experimentally manipulated and the subsequent change in herbivore abundance measured to test assumption I. This test is more direct becatise it does not involve assumption 2. so where herbivore abundance does not follow grass biomass. there could be a weakness in assumptions 1, 3 or 4. but not 2. Although the experimental approach is more direct, like all field experiments, its application is limited by the financial, logistic and ethical eosts of habitat manipulation (Diamond 1983), and this largely explains why experitncntal manipulation has played such a limited role in model testing. Assumption 2 can be tested at least two ways (Table 1). Firstly, the underlying statistical assumptions of the modelling procedure can be examined and tested. These validations vary according to the statistical tools used in the modelling, but in most cases useful literature exists on appropriate methods (e.g. McCullagh and Nelder 1996). However, even if statistical assumptions are valid, it is always possible that an alternative modelling method will fit the data better. For example, if herbivore abundance is linearly related to grass biomass. then modelling the relationship in a regression tree model may be less efficient than in a linear regression. A second means of validating assumption 2 therefore is through using several modelling procedures. Checking statistical assumptions and applying several E C O G R A P l t Y 2 2 2 (1999) modelling methods should be viewed as complementary rather than alternative validations, because they perform slightly different functions; the first tests the internal validity of a given model, whereas the latter reduces the likelihood that a suboptimal alternative is used. Assumptions 3 and 4 are both also tested in at least two ways: by assessing the precision and accuracy of habitat and distribution measurements, and by sampling with standardised methods (Table 1). Assessment of precision and accuracy will vary according lo the measures in question, but will generally require comparison of the sampling methods used with alternative methods. There is a good, although diffuse, literature on many appropriate methods (e.g. Caughley 1977, McDonald el al. 1990). Standardised sampling methods are useful if they have been previously validated for the conditions of the modelling study, and in these cases, their use negates the need to assess precision and aeeuracy. Assumption 5. like assumption I. cannot be lested in isolation from other assumptions. It can be partially validated by testing the significance of the association between measured distribution and measured habitat quality. If the relationship is significant, there are grounds to accept its validity. However, a non-significant association ean result because assumption 5 and/or assumption 6 is invalid. By contrast, assumption 6 can be validated separately, so the results ofthis validation will help clarify the testing of assumption 5. Assumption 6 is tested in terms of the accuracy and precision of the habitat quality metric, along the same lines as assumptions 3 and 4. The use of standardised habitat quality measures (e.g. Van Horne 1983) is probably not feasible at present, given the lack of studies providing non-distributional habitat quality measures; however, the development of such measures is an outcome that the modelling community should move toward in lhe longer term. A better alternative The likelihood of building a distribution based wildlifehabitat model that predicts habitat quality outside the training data should be maximised where all critical assumptions are valid, and we have outlined appropriate procedures above. Unibrtunateiy, most assumptions will, at best, be partially rather than completely validated. Consequently, even if all assumptions are judged sufficiently valid, their eumulative validity may not be sufficient to produce a tuodel of acceptable predictive skill. For example, in a model in which each assumed link is 80% valid (such that SOV^t of the variation in distribution is accounted for by habitat, SO'Vn of the variation in habitat is captured in habitat measurements, and so on), predicted dislribution will account 221 for about only 2h"/- (O.S'') ot" [he variation in hiibitiit quality. R-squarcd and equivalent v;ilues uround t).8 are quite high in many biological contexts, and suggest model reliability. Unfortunately the underlying assumptions of modelling arc linearly related (Fig. IC). so model reliability quickly degenerates even when all assumptions are relatively sound. This severely reduces ihe likelihood of building a distribution-based model that predicts habitat quality adequately. Figure IC does, however, also suggest an alteriialive modelling approach that is more likely to accurately predict habitat quality. If habitat quality rather than distribution is modelled, the number of critical assumptions in the modelling process is reduced by iwo (Fig. 2). This approach will greatly reduce the work involved in validation of critical assumptions. More importantly. il should also increase the cumulative validity of the critical assumptions in any model: using the example of WVn validity again, predicted habitat quality will account for ca •X\% (O.S"*) of the variation in habitat quality using the approach in Fiy. 2. The idea of modelling habitat quality rather than distribution echoes Van Horne (1983). who argued strongly against the use of distributional data as an indicator of habitat quality. Our work reinforces hers, and further shows that even where distribution does relate to habitat quality, the sheer number of assutnptions in a distributional model make it unlikely that habitat quality will be predicted effectively. The way forward is through direct models of habitat quality rather than abundance or presence, ihough there are still significant hurdles to dear. For example, there will be debate about how to actually measure habitat quality more directly. Current definitions of habitat quality are ambiguous at best, and as a result, indices as diverse as rcproducti\e success (Beyer et al. 1996), survivorship (Paradis and Croset 1995). and physiological condition (Virgl and Messier 1993). have all been suggested. More work is required to determine which (combination) of these is the optimal hahitat quality index for any speeies. and even then, some indices may be impractical lo measure across many sites. In spite of these difficulties, modellers face a clear choice. They can continue tnodelling under a distribution-based paradigm, wherein the modelled dependent variable is relatively easy to measure, but the model is unlikely lo actually predict habitat quality. Alternatively, they can rise to the challenge of measuring and modelling habitat quality more directly, and ultimately de\elop models much more likely to predict the true value of a habitat. .•icknfwk-tii^t'nicitr.\ - This work was funded by tlie Univ. of Queensland's School of Natural and Rural Systems Maiiagemeul and by ttie Queensland Dept of linviromnent. We would also like to thank M. J. Page and B. J. Sullivan for eommenlini; on drafts of the manuscript. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Ecography Wiley

Building better wildlife‐habitat models

Beutel, T. S.; Beeton, R. J. S.; Baxter, G. S.
Ecography , Volume 22 (2) – Apr 1, 1999
Download PDF
6 pages

Loading next page...
 
/lp/wiley/building-better-wildlife-habitat-models-UOMySxF6ts

References (36)

Publisher
Wiley
Copyright
Copyright © 1999 Wiley Subscription Services, Inc., A Wiley Company
ISSN
0906-7590
eISSN
1600-0587
DOI
10.1111/j.1600-0587.1999.tb00471.x
Publisher site
See Article on Publisher Site

Abstract

Wildlife-habitat models are an important tool in wildlife management today, and by far tbe majority of these predict aspects of species distribution (abundance or presenee) as a proxy measure of babitat quality. L'nfortmiately, few are tested on independent data., anii of those Ihat are, few show useful predictive skill. We demonstrate tbat six eritieal assumptions nnderlie distribution based wildlife-habitat models, all of wbieb must be valid for tbe model to predict habitat quality. We outline tbese assumptions in a meta-model, and discuss metbods for tbeir validation. Even wbere all six assumptions sbow a bi}>b level of validity, tbere is still a strong likelibood that tbe model will not prediet habitat qnality. However, tbe meta-model does suggest babitat quality can be predicted more accurately if distributional data are ignored, and variables more indicative of babitat quality are modelled instead. Wildlile-habitat models appear regularly in the piiblished literature, and most use distributional data (abundance or presence) as their dependent variable. Ecologists develop these models Tor at least two purposes, to explain wildlife distributions and predict habitat use (Patton 1992, Morrison et al. 1992), but in all cases assume that habitat quality is indexed by distribution (Van Home 1983. Schamberger and O'Neil 1986. Hobbs and Hanley 1990). so that model output prediets habitat quality. Superticially. many distributional models appear to predict well, generally showing exeellent tit to the data from which they were derived (the training set). Unfortunately, models are optimised for their training set, so goodness-ot-tit to this data is an overly optimistic estimate of predictive skill outside the training set (Mostclier and Tukey 1977. Picard and Cook 1984). Despite this, an alarmingly large number of even the most recent studies tbcus only on the model's fit to the training set (Brown and Nicholls 1993, Sarre 1995, Pausas et al. 1995. Sarre et al. 1995, St-Georges et al. 1995, Poizat and Pont 1996. Abensperg-Trauii et al. 1996. Beyer et al. 1996. Smith et al. 1996. Welsh et al. ECOORAPHY 22:2 (19991 1996. Pedlar et al. 1997). Since this is an optimistically biased measure of predictive skill, the number of published models with genuine predictive skill is probably far fewer than a quick perusal of the literature suggests. Only a relatively small proportion of models are tested on data not used for model development (a test set), and among these, many do not fit the test set well (e.g. Rotenberry 1986. Lindeninayer et al. 1994, Fielding and Haworth 1995, Lindenmayer et al. 1995). Fitting models to a test set is a more useful test of predictive skill, though such tests are still data dependent (Oreskes et al. 1994), so even if a model fits one test set well, it will not necessarily fit others equally. Modellers need to understand the predictive skill of the models they develop, but current approaches provide limited guidance. In this paper we propose a meta-model which describes distribution based wildlifehabitat models, and discuss its three most important implications for building models that predict better. Firstly, we identify six critical assumptions which must all be valid for a model to predict habitat quality. Secondly, we discuss how each of these assumptions can be validated. Finally, and most importantly, we demonstrate why even if all critical assumptions are valid, a distribution based model is still unlikely to predict habitat quality, and argue that more direct measures of habitat quality should be modelled in preference to distributional data. The critical assumptions Superficially, distribution based models have a simple logic; habitat determines species distribution, so habitat characteristics predict wildlife distribution (Fig. lA). This conceptual model relies on two critical assumptions: that wildlife distributions are inlluenced by habi219 tat, and that predicted distributions are adequately modelled from habitat data. Unfortunately, factors such as predation and climale also inlUience species distribution (Schambcrger and O'Neil 1986). weakening the first assumption. The second assumption is invalid if data are modelled inappropriately, for example, when a nonlinear relationship is modelled linearly, or a nonadditive process is modelled addilively. Even if habitat does aeeount for substantial variation in species distribution, and appropriate modelling techniques are used, successful modelling further depends on how both habitat and distribution are quantified (Fig. IB). Important habitat variables can be overlooked during sampling, and their perceived significance ean depend on how they are quantified (i.e. intensiveness. extensiveness. definition). Similarly, the strength of the habitat-distribution relationship also depends on how distribution is measured (e.g. presence, absolute abundance, relative abundance, spoor, nest presence, etc.). As a consequence, both habitat and distribution must be measured adequately for predicted distribution and measured distribution to be adequately linked. Because the measurement of habitat quality underlies all distribution based wildlife-habitat models (Van Home 1983, Hobbs and Hanley 1990). two final assumptions are required to complete the meta-model (Fig. !C). A model might predict measured distribution very well, but if measured distribution does not equate with habitat quality, the model has no clear interpretation in terms of target speeies welfare. Thus, it is assumed that measured distribution equates with habitat quality. Habitat quality is an ambiguous expression, most widely defined in terms of a habitat's capacity to support a species (Maurer 1986, Brennan et al. 1986. Paradis and Croset 1995). Important aspects of habitat quality include reproductive success, survivorship and population density (Van Horne 1983), and the exact way each of these ultimately contribute to the habitat quality metric will determine the strength of the iink between measured and actual habitat quality. Thus, the final assumption in the meta-model requires that habitat quality is adequately measured. The validity of these latter two assumptions will not directly affect the ability of a distribution based model to predict measured distribution, but without a link between measured distribution and habitat quality, a model has no clear use, so these assumptions are important to a model's biological validity. It is important to note that the six critical assumptions outlined above are linearly related. Consequently, each assumption has equal importance, since any invalid assumption, or subset of assumptions, breaks the link between predicted distribution and measured habitat quality, regardless of the validity of any other assumption. Models that accurately predict habitat quality are therefore those for which all critical assumptions are valid, and assumption validation is a critical step in the process of identifying models with rehable predictive ability. Validating tiie criticai assumptions Assumption I is tested by fitting the model to the parent data, or even better, to independent data (Table 1). If the fit is good, then the validity ofthis assumption is partially supported. However, this strategy does not test assumption I in isolation from assumptions 2 through 4, since all four must be valid for the model to fit any data. This strategy is therefore inconclusive in the case of a negative result, unless it is used with more direct tests of assumptions 2-4. Independent data are strongly preferred over the parent data for this test because it provides a more realistic estimate of the Predicted distribution Habitat Dislribution Predicted distribution Measured habitat Habitat Distribution r y Measured distribution Predicted distribution .—^ r—V Measured habitat Habitat Distributioii Measured distribution Habitat quality Measured habitat quality Fig. I. Development of a meta-model explaining the underlying assutnptions of distribution based uildlifc-habitiit models. Bold nutnbers represent erilical model assumptions. 1: that habitat influetices distribution: 2; that predicted distribution is adequately modelled; 3: that habitat is adequately measured; 4: that distribution is adequately measured; 5: that measured distribution equates with habitat qtiality; 6; that habitat quality is adequately measured. b C O G K A P H Y 2^.2 Table 1. Validation strategies for distribution based wildlifehabitat rnodels. Validation strategics Assumption t: That habitat influences distribution. Fit the tiiodel to parent daia. Fit the model to independent data. Manipulate habitat and observe distributional changes. Assumption 2: That predicted distribution is adequately modelled. Check the model's underlying statistical assumptions. Apply several alternative modelling strategics. j^ssurnption 3: That habitat is adequately measured. Use standardised measurement procedures. Test measurement accuracy and precision. Assumption 4: That distribution is adequately measured. Use standardised measurement procedures. I Test measurement accuracy and precision. Assumption 5: That measured distribution equates with habitat quality. Test the relationship between habitat quality and measured distribution. Assumption 6; That habitat quality is adequately measured. Use standardised measurement procedures. Test measurement accuracy, precision, and scale adequacy. model's predictive power. A tiiorc direct alternative for testing assumption I is by experimental manipulation of habitat. F'or example, in a model which identifies grass biomass as an important indicator of herbivore abundance, grass biomass could be experimentally manipulated and the subsequent change in herbivore abundance measured to test assumption I. This test is more direct becatise it does not involve assumption 2. so where herbivore abundance does not follow grass biomass. there could be a weakness in assumptions 1, 3 or 4. but not 2. Although the experimental approach is more direct, like all field experiments, its application is limited by the financial, logistic and ethical eosts of habitat manipulation (Diamond 1983), and this largely explains why experitncntal manipulation has played such a limited role in model testing. Assumption 2 can be tested at least two ways (Table 1). Firstly, the underlying statistical assumptions of the modelling procedure can be examined and tested. These validations vary according to the statistical tools used in the modelling, but in most cases useful literature exists on appropriate methods (e.g. McCullagh and Nelder 1996). However, even if statistical assumptions are valid, it is always possible that an alternative modelling method will fit the data better. For example, if herbivore abundance is linearly related to grass biomass. then modelling the relationship in a regression tree model may be less efficient than in a linear regression. A second means of validating assumption 2 therefore is through using several modelling procedures. Checking statistical assumptions and applying several E C O G R A P l t Y 2 2 2 (1999) modelling methods should be viewed as complementary rather than alternative validations, because they perform slightly different functions; the first tests the internal validity of a given model, whereas the latter reduces the likelihood that a suboptimal alternative is used. Assumptions 3 and 4 are both also tested in at least two ways: by assessing the precision and accuracy of habitat and distribution measurements, and by sampling with standardised methods (Table 1). Assessment of precision and accuracy will vary according lo the measures in question, but will generally require comparison of the sampling methods used with alternative methods. There is a good, although diffuse, literature on many appropriate methods (e.g. Caughley 1977, McDonald el al. 1990). Standardised sampling methods are useful if they have been previously validated for the conditions of the modelling study, and in these cases, their use negates the need to assess precision and aeeuracy. Assumption 5. like assumption I. cannot be lested in isolation from other assumptions. It can be partially validated by testing the significance of the association between measured distribution and measured habitat quality. If the relationship is significant, there are grounds to accept its validity. However, a non-significant association ean result because assumption 5 and/or assumption 6 is invalid. By contrast, assumption 6 can be validated separately, so the results ofthis validation will help clarify the testing of assumption 5. Assumption 6 is tested in terms of the accuracy and precision of the habitat quality metric, along the same lines as assumptions 3 and 4. The use of standardised habitat quality measures (e.g. Van Horne 1983) is probably not feasible at present, given the lack of studies providing non-distributional habitat quality measures; however, the development of such measures is an outcome that the modelling community should move toward in lhe longer term. A better alternative The likelihood of building a distribution based wildlifehabitat model that predicts habitat quality outside the training data should be maximised where all critical assumptions are valid, and we have outlined appropriate procedures above. Unibrtunateiy, most assumptions will, at best, be partially rather than completely validated. Consequently, even if all assumptions are judged sufficiently valid, their eumulative validity may not be sufficient to produce a tuodel of acceptable predictive skill. For example, in a model in which each assumed link is 80% valid (such that SOV^t of the variation in distribution is accounted for by habitat, SO'Vn of the variation in habitat is captured in habitat measurements, and so on), predicted dislribution will account 221 for about only 2h"/- (O.S'') ot" [he variation in hiibitiit quality. R-squarcd and equivalent v;ilues uround t).8 are quite high in many biological contexts, and suggest model reliability. Unfortunately the underlying assumptions of modelling arc linearly related (Fig. IC). so model reliability quickly degenerates even when all assumptions are relatively sound. This severely reduces ihe likelihood of building a distribution-based model that predicts habitat quality adequately. Figure IC does, however, also suggest an alteriialive modelling approach that is more likely to accurately predict habitat quality. If habitat quality rather than distribution is modelled, the number of critical assumptions in the modelling process is reduced by iwo (Fig. 2). This approach will greatly reduce the work involved in validation of critical assumptions. More importantly. il should also increase the cumulative validity of the critical assumptions in any model: using the example of WVn validity again, predicted habitat quality will account for ca •X\% (O.S"*) of the variation in habitat quality using the approach in Fiy. 2. The idea of modelling habitat quality rather than distribution echoes Van Horne (1983). who argued strongly against the use of distributional data as an indicator of habitat quality. Our work reinforces hers, and further shows that even where distribution does relate to habitat quality, the sheer number of assutnptions in a distributional model make it unlikely that habitat quality will be predicted effectively. The way forward is through direct models of habitat quality rather than abundance or presence, ihough there are still significant hurdles to dear. For example, there will be debate about how to actually measure habitat quality more directly. Current definitions of habitat quality are ambiguous at best, and as a result, indices as diverse as rcproducti\e success (Beyer et al. 1996), survivorship (Paradis and Croset 1995). and physiological condition (Virgl and Messier 1993). have all been suggested. More work is required to determine which (combination) of these is the optimal hahitat quality index for any speeies. and even then, some indices may be impractical lo measure across many sites. In spite of these difficulties, modellers face a clear choice. They can continue tnodelling under a distribution-based paradigm, wherein the modelled dependent variable is relatively easy to measure, but the model is unlikely lo actually predict habitat quality. Alternatively, they can rise to the challenge of measuring and modelling habitat quality more directly, and ultimately de\elop models much more likely to predict the true value of a habitat. .•icknfwk-tii^t'nicitr.\ - This work was funded by tlie Univ. of Queensland's School of Natural and Rural Systems Maiiagemeul and by ttie Queensland Dept of linviromnent. We would also like to thank M. J. Page and B. J. Sullivan for eommenlini; on drafts of the manuscript.

Journal

Ecography – Wiley

Published: Apr 1, 1999

There are no references for this article.