850 JULIAN D. OLDEN ET AL. Ecology, Vol. 89, No. 3
<br />altered physical habitat conditions, predation by non-
<br />native fishes, competition with nonnative fishes, or other
<br />biological interactions, e.g., hybridization or disease).
<br />From the questionnaire results we used a majority rule
<br />to place each species into one of five categories that
<br />combined perceived level of extinction risk and primary
<br />source of threat over their lifetime (Appendix B): (1)
<br />none/low risk; (2) moderate risk-environmental degra-
<br />dation (first three sources listed above); (3) moderate
<br />risk-species invasions (last three sources listed above),
<br />(4) high risk-environmental degradation, and (5) high
<br />risk-species invasions. We assessed inter-respondent
<br />reliability by comparing the actual assignment to
<br />random assignments of the species to the five extinction
<br />level-source categories according to Fleiss' kappa (1971).
<br />Fleiss' kappa is interpreted as the extent to which the
<br />observed amount of agreement among raters exceeds
<br />what would be expected if all raters made their ratings
<br />completely randomly (complete agreement then x = 1,
<br />no agreement then x < 0). For the 22 fish species, we
<br />found that x = 0.41 (representing a mean concordance
<br />among participants of 66.8%); a value significantly
<br />greater than chance (P = 0.038). This indicates good
<br />inter-respondent consistency for assigning species to one
<br />of the five extinction level-source categories.
<br />Statistical analyses
<br />We used classification and regression trees (CART;
<br />Breiman et al. 1984) to model species' rarity (expressed
<br />as total kilometers of stream reaches occupied), fre-
<br />quency of local extirpation (probability between 0 and
<br />1), and perceived level and source of extinction risk (five
<br />nominal categories) as functions of the 10 biological
<br />traits and two indices of phylogenetic relatedness.
<br />CART is particularly powerful for ecological analyses
<br />because it allows the modeling of nonlinear, nonadditive
<br />relationships among mixed variable types, it is invariant
<br />to monotonic transformations of the data that are often
<br />required prior to using traditional methods, and it
<br />facilitates the examination of intercorrelated variables in
<br />the final model (De'ath and Fabricius 2000). Thus, the
<br />CART methodology is well suited for analyzing trait
<br />synergisms among inherently correlated biological traits
<br />that are both continuous and categorical.
<br />The CART methodology uses a recursive partitioning
<br />algorithm to repeatedly partition the data set according
<br />to the explanatory variables (i.e., biological traits) into a
<br />nested series of mutually exclusive groups, each as
<br />homogeneous as possible with respect to the response
<br />variable (i.e., rarity, extirpation, extinction). The
<br />branching topology of the resulting decision tree reveal
<br />nonadditive (or synergistic) trait effects, and the primary
<br />splits represent the most important predictor traits as
<br />well as indicating the best competitive traits (called
<br />surrogate splits) that show similar classification power.
<br />Therefore, the CART methodology has the favorable
<br />characteristic of allowing the simultaneous examination
<br />of biological traits (e.g., age and length at maturity) that
<br />may have similar predictive roles in the final model. We
<br />used the Gini impurity criterion to determine the
<br />optimal variable splits (minimum parent node size: n =
<br />5; minimal terminal node size: n = 2), and we determined
<br />the optimal size of the decision tree by constructing a
<br />series of cross-validated trees and selecting the smallest
<br />tree based on the one-standard-error rule (De'ath and
<br />Fabricius 2000). Cohen's x coefficient of agreement
<br />(classification tree) and Pearson's moment correlation
<br />coefficient (regression tree) were used to assess the
<br />predictive performance of the decision trees compared to
<br />random expectations (Fielding and Bell 1997). Analyses
<br />were conducted using CART 5.0 (Salford Systems, San
<br />Diego, California, USA).
<br />RESULTS
<br />Rarity
<br />Species rarity varied as a function of ecological and
<br />life-history traits (R = 0.72, P = 0.0002, Fig. 1). The
<br />branching sequence of the regression tree indicates that
<br />trophic specialization and reproductive strategy were the
<br />most important trait predictors of species' present-day
<br />range size (inversely related to rarity). Species with an
<br />herbivore/detritivore feeding behavior--considered a
<br />generalist trophic guild that consumes living and dead
<br />plant matter in benthic habitats-were three times more
<br />widespread compared to the other species (Fig. 1, node
<br />A). Of the species in the remaining trophic guilds (n =
<br />18), two groups had high rarity: those providing a low
<br />degree of parental care (surrogate split: non-guarders
<br />spawning on open substrates, i.e., low energetic contri-
<br />butions to reproduction with respect to guarding and
<br />nest building; node B), and those having high parental
<br />care coupled with low diet breadth (node Q. In contrast,
<br />species characterized by relatively high parental care and
<br />high diet breadth occupied considerably larger ranges
<br />(node D).
<br />Because the regression tree was optimized to separate
<br />species into relatively homogenous groups with respect
<br />to their level of rarity, it is useful to examine species that
<br />contribute the greatest to terminal node impurity (i.e.,
<br />model error). By comparing predicted to actual species
<br />range sizes (Appendix C), we found that the range sizes
<br />of Catostomus clarkii (desert sucker, node A) and Gila
<br />robusta (roundtail chub, node B) were underestimated
<br />by the model (i.e., according to the model these species
<br />should exhibit greater rarity), whereas the range sizes of
<br />C. latipinnis (flannelmouth sucker, node A), Oncorhyn-
<br />chus gilae apache (Gila trout, node C) and Poeciliopsis
<br />occidentalis (Gila topminnow, node D) were overesti-
<br />mated by the model (i.e., according to the model these
<br />species should exhibit lower rarity).
<br />Frequency of extirpation
<br />Species frequency of local extirpation was highly
<br />predictable as a function of multiple, interacting
<br />biological traits (R = 0.87, P < 0.0001, Fig. 2). Fish
<br />species with the highest frequency of local extirpation
|