Kuletz and others 



Chapter 15 



Inland Habitat Suitability in Southcentral Alaska 



sites were surveyed from shore, eight from boats, and nine 

 from upland sites. At Afognak Island, 76 dawn watch sites 

 were arbitrarily selected with efforts to sample equally 

 throughout the north and southwest parcels. Two sites were 

 surveyed from shore and 74 upland. 



Sites were not randomly located within the entire spill 

 zone. Therefore, our statistical results apply directly only to 

 the sampled sites, and caution should be used when making 

 inferences about other areas. Application of results to the 

 entire area is based on the assumption (supported by our 

 observations) that the study sites were representative of 

 habitat types throughout the spill zone. 



Because epiphyte cover and platforms were not recorded 

 at Naked Island, we used Naked Island data for preliminary 

 analyses, but not for the final multivariate analyses. For 

 analyses, we used detections over land <200 m from the 

 observer because it produced stronger relationships with 

 predictor variables in preliminary analysis of portions of the 

 data set. Data from boat- and shore-based surveys were 

 combined with upland survey data because these data are 

 highly correlated (Marks and others, in press). Data from all 

 areas were grouped because preliminary analyses indicated 

 that within-site trends were similar to trends exhibited for all 

 sites combined. 



Multiple Regression Analyses ofMurrelet Activity Levels 



We used multiple regression analyses to examine the 

 continuum of murrelet activity levels relative to independent 

 variables, to examine the interactive effects of those variables, 

 and to describe the amount of variation explained by the 

 model. Although season and weather affect inland activity 

 level, we incorporated all these variables into the model 

 rather than attempting to develop standardization factors. 

 Our initial set of 19 predictor variables were factors known 

 or suspected to be associated with high levels of activity or 

 nesting of Marbled Murrelets, based on previously conducted 

 analyses (Kuletz and others, in press; Marks and others, in 

 press; Naslund and others, in press), and on univariate statistics 

 across the four study areas. We used Kruskal-Wallis 

 nonparametric analysis of variance to test categorical variables 

 for significant effects on the number of detections. We 

 calculated Pearson correlation coefficients between continuous 

 variables and the number of detections, and between each 

 pair of continuous variables. To control for colinearity, only 

 one of a pair of variables with r > 0.80, whichever had the 

 strongest correlation with the number of detections, was 

 included in the same regression analysis. 



Because categorical and continuous variables were 

 included in the multiple regression model, we used a General 

 Linear Model procedure (SAS Institute 1988) to examine 

 variation in murrelet activity levels. We transformed the 

 number of detections by using natural logarithms and the 

 percent data (canopy cover, forest cover, alder cover, and 

 slope) by using square roots to stabilize residuals. We ran 

 our initial regression model with all sites, and included all 

 significant (P < 0.05) categorical variables and those 



continuous variables which were measured across all four 

 study areas. We ran a second regression model for the three 

 areas for which variables more directly related to Marbled 

 Murrelet nest site selection (epiphyte cover and platforms 

 per tree) were estimated. For this model we included all 

 variables in the initial regression and epiphyte cover, which 

 was highly correlated with platforms per tree. We reduced 

 the model to include t probabilities for parameter estimates 

 where P < 0.25 in the original model. This criterion was 

 selected because our objective was to include all variables 

 that explained variation in murrelet activity. Standardized 

 parameters (parameter estimates divided by their standard 

 error) were used to determine the relative importance of 

 variables included in the models. 



Discriminant Analyses ofMurrelet Occupancy 



We used univariate tests and stepwise logistic regression 

 to identify variables that could predict the probability of 

 detecting occupied behavior at a survey site. This analysis 

 included a test of how well the logistic model performed in 

 classifying individual observations. For all four areas 

 combined, we tested frequencies of classes of categorical 

 variables for differences between occupied sites and sites 

 of unknown status by using chi-square; and for differences 

 in rank sums of continuous variables between occupied 

 and unknown status sites by using the Wilcoxon 2-Sample 

 Test (procedure NPAR1WAY; SAS Institute 1988). 

 Significant variables (P < 0.05) in these tests were entered 

 into a stepwise logistic regression model (procedure 

 LOGISTIC; SAS Institute 1990; Naked Island group 

 excluded). Inclusion and retention of variables in the 

 stepwise logistic analysis were allowed at P < 0.10. We 

 included platforms per tree in the model because it performed 

 marginally better than one including epiphyte cover. 

 Standardized parameter estimates were estimated by dividing 

 the parameter estimate by the ratio of the standard deviation 

 of the underlying distribution to the sample standard 

 deviation of the explanatory variable (SAS Institute 1990), 

 and were used to determine the relative importance of 

 variables in the model. The classification error rate was 

 calculated using a jackknife approach to reduce the bias of 

 classifying the same data from which the classification 

 criterion was derived (SAS Institute 1990). 



Results 



Marbled Murrelet Activity Levels 



Activity of Marbled Murrelets differed by study area 

 (P = 0.018), with the greatest level of activity occurring at 

 Afognak Island, the least at Naked Island, and intermediate 

 levels in western Prince William Sound and Kenai Fjords 

 National Park (table 1). Activity was greater during late 

 summer than during spring and early summer (table 1). 

 Activity was greater when the cloud ceiling was low than 

 when there was a high ceiling or clear conditions (table 1). 

 Activity was also greater at survey sites located at the heads 



144 



USDA Forest Service Gen. Tech. Rep. PSW-152. 1995. 



