A comparative analysis of predictve data mining techniques
- 1. Journal of Vegetation Science 15: 285-292, 2004 © IAVS; Opulus Press Uppsala. - Comparison of statistical methods commonly used in predictive modelling - 285 Comparison of statistical methods commonly used in predictive modelling Muñoz, Jesús1* & Felicísimo, Ángel M.2 1Real Jardín Botánico, Plaza de Murillo 2, 28014 Madrid, Spain; 2Departamento de Expresión Gráfica, Universidad de Extremadura, 10071 Cáceres, Spain; E-mail amfeli@unex.es *Corresponding author; E-mail jmunoz@ma-rjb.csic.es Abstract. Logistic Multiple Regression, Principal Compo- Introduction nent Regression and Classification and Regression Tree Analy- sis (CART), commonly used in ecological modelling using Modelling studies have employed different statisti- GIS, are compared with a relatively new statistical technique, cal techniques to unravel the complexity of interactions Multivariate Adaptive Regression Splines (MARS), to test between distributions and environmental factors. Those their accuracy, reliability, implementation within GIS and ease of use. All were applied to the same two data sets, include Generalized Linear Models (GLM; Guisan et al. covering a wide range of conditions common in predictive 1998), especially Logistic Multiple Regression (LMR; modelling, namely geographical range, scale, nature of the Narumalani et al. 1997; Felicísimo et al. 2002); Gener- predictors and sampling method. alized Additive Models (GAM; Yee & Mitchell 1991) We ran two series of analyses to verify if model validation and Classification and Regression Trees (CART, also by an independent data set was required or cross-validation on known as Regression Tree Analysis, RTA; Moore et al. a learning data set sufficed. Results show that validation by 1991; Iverson & Prasad 1998). Recently, Guisan & independent data sets is needed. Model accuracy was evalu- Zimmermann (2000) made a comprehensive review of ated using the area under Receiver Operating Characteristics predictive modelling and noticed the lack of compara- curve (AUC). This measure was used because it summarizes performance across all possible thresholds, and is independent tive studies in which more than two statistical tech- of balance between classes. niques were applied to the same data set. MARS and Regression Tree Analysis achieved the best Most classical papers on predictive modelling are prediction success, although the CART model was difficult to based on methodologies that assume a Gaussian relation use for cartographic purposes due to the high model complex- between response and predictors, and also that the con- ity. tribution to the response from the interactions among predictors is uniform across their range of values. Both Keywords: Classification and Regression Tree; Fagus; assumptions are unwarranted in most cases (Austin & Grimmia; Logistic regression; Multivariate Adaptive Regres- Cunningham 1981; Austin et al. 1990, 1994). Neverthe- sion Splines; Regression Tree Analysis. less, LMR with a quadratic function to represent Gaussian responses has often implied high predictive success. Further problems associated with classical regression Abbreviations: AUC = Area under the ROC curve; CART = Classification and Regression Trees; FN = False negative; FP = analysis arise when many predictors are used. Such False positive; GAM = Generalized Additive Model; GIS = increase in the number of predictors implies an increase Geographic Information System; GLM = Generalized Linear greater than exponential in the number of possible re- Model; LMR = Logistic Multiple Regression; MARS = Multi- gression structures, and the almost inevitable problem variate Adaptive Regression Splines; NDVI = Normalized with multicollinearity. Difference Vegetation Index; PCR = Principal Components To sidestep these problems, analysts impose strong Regression; ROC = Receiver Operating Characteristics. model assumptions, forcing the variables to act globally over the response by limiting or eliminating local changes in response or interactions. This strategy can no longer be justified if suitable predictors are likely to act and interact differently on the response variable across their range of values. The search for a model that handles the above problems lead us to experiment with a relatively new statistical technique employed in data-mining strat- egies in fields such as chemical engineering, marketing
- 2. 286 Muñoz, J. & Felicísimo, Á.M. campaigns or weather forecasting: multivariate adap- (http://edcdaac.usgs.gov/gtopo30/gtopo30.html). Cli- tive regression splines (MARS; Friedman 1991). matic data span the period 1961-1991 and were obtained The aim of this paper is to compare, using the same from the Data Distribution Centre (DDC) of the Inter- data sets, two techniques commonly used in modelling governmental Panel on Climate Change (IPCC) at its species/communities distributions (LMR and CART) web site, http://ipcc-ddc.cru.uea.ac.uk/. Each observa- and a relatively new technique (MARS). We test their tion corresponds to the monthly 30-yr mean of the efficiency, accuracy, reliability, implementation within following variables: ground frost frequency (‘frost’, GIS and ease of use in ecological modelling studies. units: days·10), maximum temperature (‘tmax’, ∞C·10), minimum temperature (‘tmin’, ∞C·10), precipitation (‘prec’, 10·mm day–1 (=10 L.m–2.d–1)), radiation (‘rad’, Material and Methods W.m–2) and wet day frequency (‘wet’, days·10). Cli- matic grids have a horizontal grid spacing of 0.5∞ (ca. 55 Test data sets km at the Equator and 32 km at 55∞ S). NDVI was derived from the SPOT satellite remote sensing imagery The two data sets employed were selected to cover a (http://www.vgt.vito.be). We used four data sets, one per variety of characteristics. They are considered repre- year season, extracted from a ten day global synthesis sentative of typical ecological data sets, thus allowing compiled from daily syntheses over the previous ten results to be extrapolated to other studies. The data sets days at a resolution of 1 km (VGT-S10 SPOT products). cover different geographical ranges (continental vs re- NDVI was selected for its influence on factors such as gional), spatial resolutions (coarse vs high), nature of albedo and heat exchange, as well as being a good the predictors (direct vs indirect) and sampling methods descriptor of vegetation phenologic status or a surrogate (no sampling strategy vs stratified sampling). for vegetation classes. The Grimmia data set (1285 cases; 419 present, 866 The Fagus data set (103 181 cases; ca. 50% each of absent) represents the distribution of species of the moss presences and absences) was selected to represent high genus Grimmia in Latin America, from Mexico to Cape spatial resolution at a regional scale. The dependent Hornos (Fig. 3). Grimmia was recently revised for Latin variable was the presence/absence of Fagus sylvatica America and its taxonomy is well known worldwide oligotrophic forest in the La Liébana region (Cantabria (Muñoz 1999; Muñoz & Pando 2000). It is a genus Province, NW Spain). A random sampling was per- typical of bare rocks in cool or cold regions, in formed in the GIS over a digitized vegetation map intertropical regions only present at high altitudes created at the Earth Sciences Department of the Univer- (Grimmia austrofunalis, G. navicularis and G. longi- sity of Cantabria (Spain). Equal numbers of samples rostris recorded at 5300 m a.s.l. in the Andes). The were taken from the target forest type and the remaining dependent variable was presence/absence of plants of vegetation classes. Predictor variables were derived from this genus at a given locality. The Grimmia data set was the digital elevation model generated from topographic selected as an example to study model response at a maps, all of them have a horizontal grid spacing of 50 m. continental scale. It was also selected because it is Altitude (constructed using Delaunay’s triangulation representative of many environmental studies: pres- algorithm), slope (derived using Sobel’s operator; Horn ences are based on museum collections, not on a sam- 1981), potential insolation (constructed as a function of pling design and therefore suffer from biased sampling. the sun’s trajectory for standard date periods, which Absences, not recorded in museums, were generated estimates the amount of time that each point receives with a random grid in the GIS, and were approximately direct sun radiation with 20 minutes temporal resolution; double the number of presences. Such proportion was Fernández-Cepedal & Felicísimo 1987) and distance from chosen because with an equal number of absences/ the sea were used to estimate the oceanic-continental presences over such a large area, the model will tend to gradient given that other climatic data were not available. underestimate real absence. Randomly generated ab- Complete details of this data set can be found in Felicísimo sences will inevitably increase the number of false et al. (2002). Contrary to the Grimmia data set, all predic- negatives, but arising from the nature of the data this tors except radiation must be considered indirect, which problem has no solution and will equally affect all theoretically generate more local models. statistical methods tested. Predictor variables were of Software included the GIS ArcInfo and ArcView 3.2 three different types: altitude, climatic and normalized (ESRI Inc., http://www.esri.com/) and the packages SPSS difference vegetation index (NDVI). Apart from alti- 10.0 for LMR and Principal Component Regression, tude, they are all direct gradients, which can generate CART 4.0 (http://www.salford-systems.com) for Regres- more general models (Guisan & Zimmermann 2000). sion Trees and MARS 2.0 (http://www.salford-systems. Gridded altitude data were obtained from GTOPO30 com/) for multivariate adaptive regression splines.
- 3. - Comparison of statistical methods commonly used in predictive modelling - 287 Modelling methods the generated models can be extremely complex and difficult to interpret. This is exemplified by the work on Methods used in predictive modelling consist of two Australian forests by Moore et al. (1991), generating a main types: global parametric and local non-parametric. tree with 510 nodes for just ten predictors. In the present Global parametric models use a strategy of global study, the optimal tree obtained for the Fagus data set variable selection. Each variable enters the model as ‘a (103 181 cases) has 1726 terminal nodes! Such com- whole’ to explain its contribution to the response. This plexity makes the tree impossible to interpret, whereas strategy is clearly inappropriate when the hypothesis is in many studies interpretability is a key issue. Moreo- that variables interact in a non-homogeneous way across ver, implementation of such a tree within GIS is un- their range of values. However, global techniques are workable. Prediction maps are often a required outcome still appropriate for small data sets where the analyst is of modelling, and this shortcoming affects the use of forced to use parametric modelling techniques because CART when complexity grows beyond a reasonable all points will influence almost every aspect of the limit. model. As an example of global parametric model we Multivariate Adaptive Regression Splines (MARS; have used LMR. Although widely used (e.g. Augustin et Friedman 1991) is a relatively novel technique that al. 1996; van de Rijt et al. 1996; Narumalani et al. 1997; combines classical linear regression, mathematical con- Guisan et al. 1998; Mladenoff et al. 1999; Felicísimo et struction of splines and binary recursive partitioning to al. 2002), GLM and LMR have several important disad- produce a local model where relationships between vantages. Ecologists frequently assume a unimodal and response and predictors are either linear or non-linear. symmetric response to gradients, which real life obsti- To do this, MARS approximates the underlying func- nately tends to refute (Austin & Smith 1989; Yee & tion through a set of adaptive piecewise linear regres- Mitchell 1991; also see Rydgren et al. in press). Such sions termed basis functions (BF) For example, the multi-modal or skewed distributions are sometimes dealt basis functions and the final function from the Grimmia with using high-order polynomial functions, but this model with 2nd. order interactions include mean number strategy heavily increases the risk of over-fitting – find- of frost days during April·10 (FROST04), mean number ing patterns that only apply to the training data – creat- of frost days during August·10 (FROST08) and mean ing models that work almost perfectly with original data precipitation per day in April·10 (PREC04) (Fig. 3; but have poor predictive ability with new data. below) are: Secondly, in GLM the relationships between re- BF2 = max (0, 3318 – elevation); sponse and predictors are assumed to be linear, when BF3 = max (0, FROST04 – 67) * BF2; real-world effects are generally more complex. BF4 = max (0, 67 – FROST04) * BF2; Our hypothesis is that in modelling organism/com- BF6 = max (0, 6 – PREC04); munities distributions, response is related to predictor BF7 = max (0, FROST08 - 85) * BF2; variables in a non-linear and local fashion. Local non- parametric models are suitable under such a hypothesis Y = 0.913 + 0.363496E-05 * BF3 – 0.449372E-05 * BF4 – as they use a strategy of local variable selection and 0.072 * BF6 – 0.271023E-05 * BF7 (1) reduction, and are flexible enough to allow non-linear relationships. From this type we have tested CART and Changes in slope of those basis functions occur at MARS. points called knots (values 3318, 67 and 6 in the above Classification and Regression Trees (CART; Breiman examples). The regression line is thus allowed to bend at et al. 1984) is a rule based method that generates a the knots, which mark the end of one region of data and binary tree through binary recursive partitioning, a the beginning of another with different behaviour of the process that splits a node based on yes/no answers about function (Fig. 1). Like splits in CART, knots are estab- the values of the predictors. Each split is based on a lished in a forward/backward stepwise way. A model single variable. Some variables may be used many times which clearly overfits the data is produced first. In while others may not be used at all. The rule generated at subsequent steps, knots that contribute least to the effi- each step maximizes the class purity within each of the ciency of the model are discarded by backwards pruning two resulting subsets. Each subset is split further based steps. The best model is selected via cross-validation, a on entirely different relationships. CART builds an over- process that applies a penalty to each term (knot) added grown tree based on the node purity criterion that is later to the model to keep low complexity values. pruned back via cross-validation to avoid over-fitting. Another key improvement of MARS over global The main drawback of CART models, when used to parametric models is the way it deals with interactions. predict organism distributions, is that with more than In local modelling, interactions can no longer be treated just a handful of predictor variables or cases to classify, as global. MARS considers interactions not between the
- 4. 288 Muñoz, J. & Felicísimo, Á.M. Fig. 1. Example of MARS functions showing the relationship between response and a single predictor: A. Effect of Frost12 (mean number of frost days during December·10) on the response. This almost flat-bottom function with three very different slope regions is difficult to model under the assumption of a linear relationship, because the positive and negative sloping sections cancel each other out. B. The basis function has a negative slope from sea level to 1157 m a.s.l., then ascends tol 3318 m a.s.l. where it reaches a semi-plateau value. original predictors, but between the sub-regions of every Most statistical techniques perform poorly with high basis function generated. A particular sub-region of a dimensionality, a problem that in predictive modelling given basis function can interact with a particular sub- literature is usually circumvented by limiting the number region of another basis function, but other regions of of variables employed through a priori selection, which these basis functions might display none or a different is not always biologically warranted, or by eliminating interaction pattern. Examples are shown in Fig. 2, which the interactions from the model, as in Yee & Mitchell illustrate the contribution to the response by interactions (1991). Some models for acceptable selection of predic- specified in Table 1. tors have been proposed, including the use of CART to identify the interactions, then fit GLM or GAM based Interactions among predictors on those variables (Guisan et al. 2002). To compare the performance of each technique with Common regression analysis rapidly becomes unre- regard to interactions, all statistical analyses were run liable when dimensionality becomes high, a phenom- with no interactions (i.e. main effects) and with interac- enon known as the ‘curse of dimensionality’ (Hastie & tions up to 5th order in MARS (when the results start to Tibshirani 1990). An immediate consequence of the deteriorate). The same interactions that produced the ‘curse of dimensionality’ is the severe limitation in the best model in MARS were then used to run a new LMR number of usable variables in a given analysis. Suppose analysis to compare against MARS performance. we consider two regions for each predictor. With two predictors the number of regions to be considered will Dealing with multicollinearity be four; with three predictors they will be eight (23) etc. In the present study, and limiting the number of regions Multicollinearity occurs when one or more variables to 2 – a number we do not know a priori –, the number are exact or near exact linear functions of other variables of regions will be 32 with Fagus (five predictors), but in the data set. This is a common problem associated 15·1022 with the Grimmia data set (77 predictors) which with organism/communities modelling (Brown 1994; is simply not feasible. De Veaux & Ungar 1994). Assessment of model per- formance when multicollinearity may be an issue was done by including the following analyses (all LMR are Table 1. MARS sample output for the Grimmia data set with forward conditional stepwise LMR with P-to-enter = 2nd. order interactions. See p. 3 for explanation of variables. 0.05 and P-to-remove = 0.1, no polynomial functions Basis function Coefficient Variable Parent Knot used due to the high number of predictors): LMR with original variables; LMR with Varimax orthogonalized 0 0.913 3 3.63E-006 FROST04 altitude 67 PCA factors (Principal Component Regression); LMR 4 – 4.49E-006 FROST04 altitude 67 using as variables the basis functions generated in the 6 – 0.072 PREC04 6 MARS main effects (i.e. no-interactions) analysis (five 7 – 2.71E-006 FROST08 altitude 85 basis functions in Grimmia data set and ten in Fagus
- 5. - Comparison of statistical methods commonly used in predictive modelling - 289 Fig. 2. Example of MARS functions ob- tained from the Grimmia data set with 2nd. order interactions among predictors, show- ing the contribution to the response from such interactions: A. Altitude and Frost04 (mean number of frost days during April·10) only interact below 3318 m a.s.l., but not above this elevation. B. Altitude only inter- acts with Frost08 (mean number of frost days during August·10) below 3318 m a.s.l. and then only when there are less than 21 frost days. data set); CART with original variables; CART with the ‘percentage of class-0 observations’ and are therefore same PCA factors as above; CART with the same MARS independent of the balance between the two classes. basis functions as above and MARS with the original Hanley & McNeil (1983) have shown that when dealing variables. with a single scoring model, the AUC is equal to the We used two LMR projection methods to avoid empirically observed probability of a class-1 observation multicollinearity problems: principal component regres- attaining a higher score than a class-0 observation. They sion (PCR), already used in predictive modelling (see have also shown that the AUC is actually equivalent to references in Guisan & Zimmermann 2000) and LMR the normalized Mann-Whitney two-sample statistic, which using as variables the basis functions generated in the makes it equivalent to the Wilcoxon statistic. MARS main effects analysis. PCR uses the projections AUC is a measure of model accuracy, but it does not of the predictor variables onto a subset of PCs in place of provide a rule for the classification of cases (Fielding & the predictor variables themselves. Since the PCs are Bell 1997). The final decision about which threshold linearly uncorrelated (Varimax rotation), there are no should be selected in a particular study depends upon its multicollinearities between the projection coefficients. objectives. The relative importance of FP (False posi- To our knowledge, MARS basis functions has not previ- tives, Type I) and FN (False negatives, Type II) error ously been combined with LMR to circumvent rates must be individually considered in each study, and multicollinearity. this decision is independent of model accuracy. If the purpose of the study is to identify sites where we need to Evaluation of the models be certain that an organism will be found, we must select the threshold that minimizes FP error rates. Contrarily, Assessment of model performance was done by a if the aim is conservation of the same organism, the method that was independent of any threshold: the area threshold must be chosen to minimize FN error rates. under the Receiver Operating Characteristic (ROC) Finally, and in order to have a reliable estimate of the curve, commonly termed AUC. The ROC curve is rec- prediction power of each model, we use two approaches: ommended for comparing two class classifiers, as it 1. Tenfold cross-validation: The original data set is does not merely summarize performance at a single randomly divided into ten mutually exclusive subsets arbitrarily selected decision threshold, but across all (the folds) of about equal size. A subset is removed and possible decision thresholds (Fielding & Bell 1997). It the remaining cases are used to generate a model. This plots the sensitivity (i.e. true positives) vs (1– specificity) model is afterwards applied to the removed section and (i.e. false positives). An ideal classifier hugs the left side its empirical error calculated. The process is repeated and top side of the graph and the area under the curve with the remaining nine subsets and the mean empirical equals one. A random classifier should achieve ca. 0.5. error is used as the final estimate of the total error of the Whilst most analyses use 0.5 as the decision threshold to model. We report results obtained by applying the opti- consider a case as present or absent, this value is arbi- mal models to the entire (i.e. learning) data sets. trary and it does not necessarily give a more accurate 2. Use of independent data sets for training and model (Fielding & Bell 1997; Manel et al. 1999). AUC evaluation: Both the Grimmia and Fagus data sets were eliminates this problem by considering the performance split randomly into one training and one evaluation data of the model across all possible threshold values. Moreo- set, containing approximately 70% and 30% of the total ver, ROC curves are invariant under changing distribu- cases, respectively. The training data sets were used to tions of the binary classes (presence/absence), as they generate the models which were then tested with the actually plot the ‘percentage of class-1 observations’ vs independent evaluation data sets.
- 6. 290 Muñoz, J. & Felicísimo, Á.M. Table 2. Results with the Grimmia data set (best-fit model in bold). LMR = Logistic Multiple Regression; PCA = Principal Component Analysis; CART = Classification and Regression Trees; AUC = Area under ROC curve; MARS = Multiple Adaptive Regression Splines; CI = Confidence interval (95%). Evaluation by ten-fold Evaluation by cross-validation independent data set Model AUC CI AUC CI LMR with original variables 0.960 0.951-0.970 0.953 0.933-0.972 LMR with original variables, same interactions as MARS best model 0.963 0.954-0.972 0.951 0.931-0.971 LMR with PCA factors 0.892 0.874-0.910 0.879 0.844-0.914 LMR with MARS BFs 0.939 0.926-0.951 0.942 0.920-0.963 CART with original variables 0.970 0.960-0.980 0.912 0.876-0.947 CART with PCA factors 0.950 0.937-0.962 0.845 0.801-0.889 CART with MARS BFs 0.956 0.945-0.967 0.927 0.898-0.957 MARS with original variables, no interactions 0.977 0.970-0.984 0.931 0.899-0.964 MARS with PCA factors 0.933 0.919-0.946 0.897 0.865-0.928 MARS with original variables 2nd. order interactions 0.957 0.947-0.967 0.954 0.930-0.978 MARS with original variables 3rd. order interactions 0.980 0.974-0.987 0.945 0.915-0.974 MARS with original variables 4th. order interactions 0.982 0.976-0.988 0.939 0.908-0.969 Results biological relevance. In the particular case of the Grimmia data set, Fig. 2 shows that between ‘Elevation’ and Model performance ‘Frost04’ (Fig. 2A) or ‘Frost08’ (Fig. 2B) there is inter- action only below 3318 m a.s.l. and not above this Tables 2 and 3 present the results of the different altitude, where in both cases Contribution = 1. Fig. 2A statistical techniques applied to the Grimmia and Fagus should be interpreted as that in South America, below data sets, respectively. For both data sets differences 3318 m a.s.l., contribution of the interaction to the between evaluation via 10-fold cross-validation and model (Eq. 1) increases linearly above 6.7 frost days in independent data are evident. Confidence intervals rarely April (basis function BF3, units are days·10), according overlap between the two approaches, confirming the to the idea of Grimmia as a genus typical of cool or cold results of Manel et al. (1999) that evaluation via inde- regions. Similarly, Fig. 2B shows that contribution to pendent test samples is needed. The following discus- the model (Eq. 1) equals one below 3318 m a.s.l. only sion therefore applies to the evaluation by the independ- when frost days in August are less than 8.5 (basis ent data set. function BF7, units are days·10). As BF7 enters nega- With the Grimmia data set the best performance was tively in the model (Eq. 1, coefficient – 0.271023E-05), achieved by the MARS with 2nd. order interactions it means that probability of finding Grimmia increases model (Fig. 3), with AUC = 0.954, although LMR when the number of frost days in August increases. results were very similar. Improvement of MARS over LMR is evident for Fagus, with an increase from 0.778 Multicollinearity to 0.909 in AUC. In this data set, however, the best results were obtained with CART (AUC = 0.946). Un- Tables 2 and 3 show that the projection methods fortunately, the complexity of the tree which generates used to circumvent multicollinearity, which a priori such results, with 1726 terminal nodes, makes this model should outperform the methods using the original pre- inadequate for cartographic purposes, a frustration al- dictors (De Veaux & Ungar 1994), always reduced ready pointed out by Guisan & Zimmermann (2000). predictive power. This was an unexpected result, be- The MARS model, very similar in performance to the cause the data sets included a large number of predictors CART model, can then be considered as a surrogate that might intuitively be considered highly correlated. mapping technique (Fig. 4). The reason of this apparent contradiction is that interac- tions among predictors, demonstrated in the previous Interactions among predictors section, are masked when using projected variables (De Veaux & Ungar 1994). The best results were achieved with models allow- ing for higher order interactions. MARS was better than CART in that the models generated were easier to interpret, and interactions between predictors could be more easily understood and explained in terms of their
- 7. - Comparison of statistical methods commonly used in predictive modelling - 291 Table 3. Results with the Fagus data set (best-fit model in bold). LMR = Logistic Multiple Regression; PCA = Principal Component Analysis; CART = Classification and Regression Trees; AUC = Area under ROC curve; MARS = Multiple Adaptive Regression Splines; CI = Confidence interval (95%). a Model used to generate the map shown in Fig. 4. Evaluation by ten-fold Evaluation by cross-validation independent data set Model AUC CI AUC CI LMR with original variables 0.793 0.790-0.796 0.778 0.772-0.783 LMR with PCA factors 0.755 0.752-0.758 0.757 0.751-0.762 LMR with MARS BFs 0.898 0.896-0.900 0.897 0.894-0.901 CART with original variables 0.979 0.978-0.980 0.946 0.943-0.949 CART with PCA factors 0.805 0.802-0.808 0.794 0.789-0.799 CART with MARS BFs 0.976 0.975-0.977 0.946 0.943-0.949 MARS with original variables 0.896 0.894-0.898 0.909a 0.906-0.912 MARS with PCA factors 0.792 0.790-0.795 0.789 0.784-0.794 MARS with original variables 2nd. order interactions 0.907 0.905-0.909 0.906 0.903-0.912 Discussion LMR has been successfully used in predictive mod- elling to predict distributions of species and communi- Our results show that distributions of species and ties, despite its drawbacks such as its inability to deal communities can be better defined if we accept that they with skewed or multi-modal responses. CART, which may follow multivariate non-linear mappings. At present may produce better numerical predictions on new data, there are many sources of predictors which may interact generates complex models that can lead to no or spurious or correlate with each other in an unknown fashion, model interpretations, and is occasionally of no use for especially if we are considering remote sensing data. cartographic purposes. One goal of predictive model- Simplifications, both in the number of predictors em- ling is to generate ‘potential habitat distribution maps’. ployed or assuming linear relationships with the re- According to Guisan & Zimmermann (2000), such maps sponse, may render misleading models that perform are cartographic representations of (1) the probability of poorly on real world data. Our results support the view occurrence; (2) the most probable abundance; (3) the that global parametric methods are inferior in such predicted occurrence; or (4) the most probable entity. modelling studies. Cartographic implementation is therefore crucial. Fig. 3. Estimated probability of presence of plants of the genus Grimmia in South America according to MARS with 2nd. Fig. 4. Predicted map representing the probability of occur- order interactions model. Inset shows in detail the northern tip rence of oligotrophic forest with Fagus sylvatica in La Liébana of the Andean range. (NW Spain).
- 8. 292 Muñoz, J. & Felicísimo, Á.M. Contrary to all of these disadvantages, MARS is Guisan, A., Theurillat, J.-P. & Kienast, F. 1998. Predicting the better suited to model situations that include a high potential distribution of plant species in an alpine environ- number of variables, non-linearity, multicollinearity and/ ment. J. Veg. Sci. 9: 65-74. or a high degree of interaction among predictors. MARS Guisan, A., Edwards Jr., T.C.& Hastie, T. 2002. Generalized linear and generalized additive models in studies of species has been shown to perform as well as and more consist- distributions: setting the scene. Ecol. Model. 157: 89-100. ently than other methods with two very different data Hanley, J.A. & McNeil, B.J. 1983. A method of comparing the sets. Moreover, it is extremely easy to implement within areas under Receiver Operating Characteristic curves de- GIS, and we conclude that it can be seen as an alterna- rived from the same cases. Radiology 148: 839-843. tive that extends the use of Generalized Methods and Hastie, T.J. & Tibshirani, R.J. 1990. Generalized additive avoids unsupported inferences derived from uncritical models. Chapman & Hall, London, UK. use of standard procedures. Horn, B.K.P. 1981. Hill-shading and the reflectance map. Proc. Inst. Electr. Electron. Engin. 69: 14-47. Iverson, L.R. & Prasad, A.M. 1998. Predicting abundance of References 80 tree species following climate change in the eastern United States. Ecol. Monogr. 68: 465-485. Augustin, N.H., Mugglestone, M.A. & Buckland, S.T. 1996. Manel, S., Dias, J.-M. & Ormerod, S.J. 1999. Comparing An autologistic model for the spatial distribution of wild- discriminant analysis, neural networks and logistic regres- life. J. Appl. Ecol. 33: 339-347. sion for predicting species distributions: a case study with Austin, M.P. & Cunningham, R.B. 1981. Observational analy- a Himalayan river bird. Ecol. Model. 120: 337-347. sis of environmental gradients. Proc. Ecol. Soc. Aust. 11: Mladenoff, D.J., Sickley, T.R. & Wydeven, A.P. 1999. Test- 109-119. ing a predictive landscape model of favorable gray wolf Austin, M.P. & Smith, T.M. 1989. A new model for the habitat: logistic regression models vs. new field data. continuum concept. Vegetatio 83: 35-47. Ecol. Appl. 9: 37-44. Austin, M.P., Nicholls, A.O. & Margules, C.R. 1990. Meas- Moore, D.M., Lee, B.G. & Davey, S.M. 1991. A new method urement of the realized qualitative niche: environmental for predicting vegetation distributions using decision tree niche of five Eucalyptus species. Ecol. Monogr. 60: 161- analysis in a geographic information system. Environ. 177. Manage. 15: 59-71. Austin, M.P., Nicholls, A.O., Doherty, M.D. & Meyers, J.A. Muñoz, J. 1999. A revision of Grimmia (Musci, Grimmiaceae) 1994. Determining species response functions to an envi- in the Americas. 1: Latin America. Ann. Mo. Bot. Gard. ronmental gradient by means of a beta-function. J. Veg. 86: 118-191. Sci. 5: 215-228. Muñoz, J. & Pando, F. 2000. A world synopsis of the genus Breiman, L., Friedman, F., Olshen, R. & Stone, C. 1984. Grimmia (Musci, Grimmiaceae). Monogr. Syst. Bot. Mo. Classification and regression trees. Wadsworth, Pacific Bot. Gard. 83: 1-133. Grove, CA, US. Narumalani, S., Jensen, J.R., Althausen, J.D., Burkhalter, S. & Brown, D.G. 1994. Predicting vegetation types at treeline Mackey, H.E. 1997. Aquatic macrophyte modeling using using topography and biophysical disturbance variables. GIS and logistic multiple regression. Photogramm. Engin. J. Veg. Sci. 5: 641-656. Remote Sens. 63: 41-49. De Veaux, R.D. & Ungar, L.H. 1994. Multicollinearity: A tale Rydgren, K., Økland, R.H. & Økland, T. 2003. Species re- of two nonparametric regressions. In: P. Cheeseman, P. & sponse curves along environmental gradients. A case study Oldford, R.W. (eds.) Selecting models from data: AI and from SE Norwegian swamp forests. J. Veg. Sci. 14: 869- Statistics IV, pp. 293-302. Springer-Verlag. New York, 880. NY, US. van de Rijt, C.W.C.J., Hazelhoff, L. & Blom, C.W.P.M. 1996. Felicísimo, A.M., Francés, E., Fernández, J.M., González- Vegetation zonation in a former tidal area: A vegetation- Díez, A. & Varas, J. 2002. Modeling the potential distribu- type response model based on DCA and logistic regres- tion of forests with a GIS. Photogramm. Engin. Remote sion using GIS. J. Veg. Sci. 7: 505-518. Sens. 68: 455-461. Yee, T.W. & Mitchell, N.D. 1991. Generalized additive mod- Fernández-Cepedal, G. & Felicísimo, A.M. 1987. Método de els in plant ecology. J. Veg. Sci. 2: 587-602. cálculo de la radiación solar incidente en áreas con apantallamiento topográfico. Rev. Biol. Univ. Oviedo 5: Received 8 October 2002; 109-119. Accepted 17 November 2003. Fielding, A.H. & Bell, J.F. 1997. A review of methods for the Co-ordinating Editor: R.H. Økland. assessment of prediction errors in conservation presence/ absence models. Environ. Conserv. 24: 38-49. Friedman, J.H. 1991. Multivariate adaptive regression splines. Ann. Statist. 19: 1-141. Guisan, A. & Zimmermann, N.E. 2000. Predictive habitat distribution models in ecology. Ecol. Model. 135: 147- 186.