What is a Multi-Scale Analysis? Implications for Modeling Presence/Absence of Bird Species Kathryn M. 1 Georgitis , Alix I. 1 Gitelman , 1Department Don L. 1 Stevens , Nick 2 Danz , and JoAnn of Statistics Oregon State University Corvallis, Oregon 2 Natural Resources Research Institute University of Minnesota-Duluth Duluth, Minnesota 2 Hanowski R82-9096-01 We used logistic regression to explore the association of Pine Warbler with landscape composition variables at each spatial extent separately. 500 (m) Lowland conifer, Pine and Oak-Pine, Spruce-Fir, Spruce-Fir*Pine and Oak-Pine 1000 (m) Pine and Oak-Pine, Spruce-Fir, Spruce-Fir*Pine and Oak-Pine 5000 (m) Pine and Oak-Pine, foresta, foresta*Spruce-Fir, Spruce-Fir * Response Variable: presence/absence of Pine Warbler (PIWA) in the Western Great Lakes -Methods described in Howe et al 1998 * Explanatory Variables: percentage of each land cover type within 4 different buffer widths i.e. spatial extents (Figure 1) Figure 1: Diagram of G.I.S. buffering scheme used to create explanatory variables from TM Imagery. Table 1: Correlations at each spatial extent for subset of explanatory variables used in model selection procedure. Spatial Extent 100 m 500 m 1000 m 5000 m 1000m 500m a:The forest variable is an indicator for stands located in the Chequamegon national forest. 100 m 500 m 1000 m 5000 m 100m Drawbacks with Option 1: - 210 possible models for each spatial extent - Multi-collinearity of explanatory variables - Does not account for possible relationships between spatial extents Pine and Oak-Pine Lowland Conifer -0.233 -0.280 -0.269 -0.161 Pine and Oak-Pine Aspen-Birch -0.464 -0.585 -0.623 -0.531 Table 3: Explanatory variables included in logistic regression model for estimated log odds of Pine Warbler using variables from the 100m, 500m, and 1000m spatial extent in model selection procedure. Spatial Extent 100m 500m 1000m 100m*1000m Explanatory Variables Selected Using BIC criteria Aspen-Birch, Northern Hardwoods and Oaks, Pine and Oak-Pine, Spruce-Fir None Spruce-Fir Pine and Oak-Pine*Spruce-Fir Drawbacks with Option 2: - minimum of 230 possible models - Multi-collinearity of explanatory variables within and between spatial extents As spatial extent increases the correlation between land cover types changes (Table 1). This will affect the inclusion and exclusion of variables in the final model (see option 1). We created new explanatory variables to investigate if the presence of PIWA is associated with a change in land cover variables from the 100m to the 500m spatial extent. Table 4: Results for Proportional and Difference logistic regression models. X500m is the % of land cover type X at the 500m spatial extent, and X100m is the % at the 100m spatial extent. Model Proportional Difference New Variable X 500m/X100m X 500m- X100m Explanatory Variables Aspen-Birch, Pine and Oak-Pine Pine and Oak-Pine Drawbacks with Option 3: - Model selection problem - Correlation structure? - Is this a biologically meaningful function of the variables? F i g u r e 3 : P e r c e n t a g e o f S p r u c e F i r v e r s u s S p a t i a l E x t e n t C h e q u a m e g o n F o r e s t C h i p p e w a F o r e s t S t . C r o i x F o r e s t 0 1 0 0 0 2 0 0 0 3 0 0 0 4 0 0 0 5 0 0 0 C h e q u a m e g o n F o r e s t C h i p p e w a F o r e s t S t . C r o i x F o r e s t Another Alternative: Graphical Models Pine and Oak-Pine500 0 1 0 0 0 2 0 0 0 S p a t i a l E x t e n t ( m ) 3 0 0 0 4 0 0 0 S p a t i a l E x t e n t ( m ) Parameter Superior Forest Stand 3 Stand 1 Stand 2 Stand 1 Stand 6 Stand 2 Chequamegon Forest Stand 1 Stand 3 Chippewa Forest Stand 5 Stand 5 Stand 3 Stand 4 Mean -1.164 -0.0813 0.0026 0.00364 We are investigating the use of Graphical Models to address multi-scale questions in ecological studies. The conditional modeling approach allows for specification of correlation between variables at multiple scales. Also, this approach can be used to explore how the presence/absence of a species is related to vegetation characteristics at multiple scales, similar to Bayesian Hierarchical models. E s t i m a t e d O d d s o f P r e s e n c e o f P I W A w i t h 1 0 % I n c r e a s e i n P i n e a n d O a k P i n e 95% Wald Confidence Interval for C h e q u a m e g o n C h i p p e w a S t . C r o i x S u p e r i o r C h e q u a m e g o n C h i p p e w a S t . C r o i x S u p e r i o r Stand 4 Lowland Conifer100 Diagram 3: Schematic of Graphical Model for presence/absence of PIWA. For example, Pine and OakPine500 is the % of Pine and Oak-Pine at the 500m spatial extent. Lines without arrows represent correlation between two variables. Lines with arrows represent a directional association. 95% Confidence Interval -1.719, -0.633 -0.5654, 0.1523 -0.009755, 0.02342 -0.01925, 0.04036 E s t i m a t e d O d d s o f P r e s e n c e o f P I W A g i v e n 1 0 % I n c r e a s e i n L o w l a n d C o n i f e r 95% Wald Confidence Interval for Pine and Oak-Pine100 Presence/Absence of PIWA Results Model 1 Bayesian Hierarchical Models are a useful method for multi-scale analysis because they can be used to explore relationships between processes operating at different scales in ecological systems (e.g. Monleon et al 2002). A typical hierarchical structure is displayed in diagram 2. St.Croix Forest Lowland Conifer500 5 0 0 0 Figure 2 and Figure 3 display the % of Pine and Oak-Pine and % of Spruce-fir for nine stands versus spatial extent. Notice the stands located in Chippewa forest have very different %’s of Pine and Oak-Pine (and Spruce-Fir) at the 100m scale but at the 5000m scale they have similar values. These figures raise the question perhaps it is the change in the % of Pine and Oak-Pine from one spatial extent to another that is what is important when trying to model the presence of Pine Warbler (see option 3). Option 4: Analysis Using Bayesian Hierarchical Models - Multi-collinearity of explanatory variables within and between spatial extents * affects inclusion/exclusion of variables * inflates variance of coefficients * unclear interpretation of coefficients - What form of the explanatory variables are biologically meaningful? Possible statistical methods that can be used to address the question of interest: Option 1: Perform an analysis for each spatial extent separately Option 2: Use the variables from all spatial extents in one analysis Option 3: Use summary variables that combine information from multiple spatial extents Option 4: Use Bayesian Hierarchical Models Stand 4 Option 3: Analysis Using “New” Explanatory Variables F i g u r e 2 : P e r c e n t a g e o f P i n e a n d O a k P i n e v e r s u s S p a t i a l E x t e n t Pine and Oak-Pine Spruce-Fir -0.311 0.026 0.110 0.209 Lowland Conifer Spruce-Fir -0.068 -0.137 -0.263 -0.633 Option 2: Analysis Using all the Explanatory Variables We used all the variables from the 100m, 500m, and 1000m spatial extent in the logistic regression model selection procedure. - Variable selection procedures do not account for number of possible explanatory variables (Fred Ramsey, personal communication). P e r c n t a g o f S p r u c e F i 0 10 20 30 40 50 60 Table 2: Significant explanatory variables included in logistic regression model for estimated log odds of Pine Warbler at each spatial extent. Spatial Extent Explanatory Variables Selected Using BIC Criteria 100 (m) Lowland conifer, Pine and Oak-Pine Discussion: Problems with Options 1-4: The Data: P e r c n t a g o f P i e n d O a k P i e 0 10 20 30 40 50 60 Option 1: Analysis at Each Spatial Extent Separately What statistical methods can be used to address multi-scale questions such as: How is the presence of Pine Warbler related to vegetation characteristics at different spatial scales? Stand 2 Stand 2 Stand 1 Stand 3 Diagram 2: Diagram of hierarchical structure present in a forest ecosystem. Stands are nested within the forests. Likelihood model for the observations- 0 . 0 0 . 5 1 . 0 1 . 5 2 . 0 9 5 % W a l d C o n f i d e n c e I n t e r v a l s f o r Yi.j~ Bernoulli(pi,j) Logit(pi,j) = ,j +,j X1,i,j + 2,j X2,i,j where Yi,j =1 if PIWA was present in the ith stand located in the jth forest and 0 otherwise, and where X1,i,j is the percentage of Lowland Conifer at the 100m spatial extent for the ith stand in the jth forest X2,i,j is the percentage of Pine and Oak-Pine at the 100m spatial extent for the ith stand in the jth forest For the next level of the hierarchy we used: Model 1- ~N(mj,t) Model 2- ,~N(mj,t) mj = 0 +Z1,j +2Z2,j where Z1,j is the mean percentage of Pine and Oak-Pine for the jth forest Z2,j is the mean percentage of Lowland conifer for the jth forest 0 . 0 0 . 5 1 . 0 1 . 5 2 . 0 2 . 5 3 . 0 9 5 % W a l d C o n f i d e n c e I n t e r v a l s f o r Results Model 2 95% Wald C.I. for 9 5 % W a l d C . I . f o r E s t i m a t e d I n t e r c e p t f o r e a h F o r e s t c Parameter Mean -0.06529 0.02004 -0.4674 -0.02388 -0.05129 95% Confidence Interval -0.1361, -0.01759 0.008479, 0.03178 -5.35, 3.006 -0.1865, 0.2031 -0.4004, 0.3651 C h e q u a m e g o n C h i p p e w a S t . C r o i x S u p e r i o r Acknowledgements: I would like to thank Jerry Niemi for the opportunity to work on an interesting ecological problem. Also, Fred Ramsey who elucidated the problem with model selection criteria. 3 Drawbacks with Option 4: References: Howe, R. W., G. J. Niemi, S. J. Lewis, and D. A. Welsh. 1998. A standard method for monitoring songbird populations in the Great Lakes region. Loon 70:188-197. Monleon, V. A.I. Gitelman, and A.N. Gray. 2002. Multi-scale relationships between coarse woody debris and presence/absence of Western Hemlock in the Oregon Coast Range. In Case Studies in Bayesian Statistics Vol. VI Gatsonis et al (Eds.). Springer-Verlag, New York, Inc. 2 1 0 9 5 % W a l d C o n f i d e n c e I n t e r v a l s f o r - Forest level covariates not independent measurements - Concentric circle sampling design not a “true” nested hierarchy - Multi-collinearity of explanatory variables within same spatial extent 1 Funding/Disclaimer: The research described in this presentation has been funded by the U.S. Environmental Protection Agency through the STAR Cooperative Agreement CR82-9096-01 Program on Designs and Models for Aquatic Resource Surveys at Oregon State University. It has not been subjected to the Agency's review and therefore does not necessarily reflect the views of the Agency, and no official endorsement should be inferred.