An overriding assumption of this study is that disjunct distributions of today were formerly connected by intermediate populations (either instantaneously or over time). In other words, it is deemed impossible for species to "jump" dr ainages. This may conceivably be violated when several drainages connect during lowered sea-levels and for some reason(s) (i.e., a waterfall), a species is unable to colonize drainages intervening between separated ones. It may also be violated if speci es go over a drainage divide and then at another point cross back into a different drainage not adjacent the original source drainage.
Another assumption is that species as now defined represent monophyletic units. A recognized flaw in this is adjacent populations that may not be the most closely related (Platnick & Nelson, 1978); however, at some hierarchical level (between popu lation and genus), they must be monophyletic if present nomenclature is relatively accurate. Through studying additional species, one can predict where adjacent populations of widespread taxa may not be closely related.
An assumption of many studies is species with similar distributional patterns have similar histories. McDowall (1978, 1990) demonstrated problems with this when the roles of dispersal and ecology are ignored. However, similarity implies connection at some time over an organism's history, which alone is useful to recognize, although determining if changes in distributions occurred simultaneously or otherwise may be extremely difficult. The final assumption of all biogeographic studies is that negativ e data do not exist except where indicated by the fossil record. This is clearly invalid, but impossible to circumvent without a tardis.
Data Sources A>
Approximately 200 freshwater fishes are recorded in Australia. Of these, some are marine or estuarine vagrants, while others spend a portion of their life cycle in oceanic conditions and were excluded (Appendix II). My analysis was o f 167 species that cannot survive more than brief exposure to seawater during any stage of their life cycle (Appendix III).
Nomenclature of 156 described species followed Eschmeyer (1998), except for Oxyeleotris selheimi (see Appendix IV) and Hephaestus tulliensis (Allen & Pusey, in prep.), four species described since (Allen & Feinberg, 1998), and fam ily designations follow Allen (1989) for Galaxiodea and Petromyzontiformes, Kemp (1997b) for Dipnoi, and Mooi & Gill (1995) for Latidae. The percichthyid genus Edelia is subsumed within Nannoperca following Kuiter, Humphries & Arthi ngton (1996). Eleven undescribed species are included, seven of which were listed by Allen (1989); in addition I include Hypseleotris sp. C (Unmack, in press); Mogurnda sp. 2 (Glover, 1989); a new species and genus of percichthyid being des cribed by B. Pusey and associates; and Hypseleotris sp. D being described by H. Larson. Specific designations for several species in Bulloo drainage (BULL; Fig. 6) were problematic. The melanotaeniid was referred to Melanotaenia splendida, which it appears to most closely resemble. The Mogurnda, Ambassis, and Neosilurus "false-spined" species could not be assigned to species due to insufficient numbers of specimens and unusual character combinations. All three were e xcluded from analyses except for richness calculations. A recently discovered population of Craterocephalus in the Cooper Creek drainage (LEB) appears nearest C. stercusmuscarum (R. Wager, pers. comm.) and was included as such. The status of Craterocephalus in TORR was not addressed in a recent revision of the C. eyresii complex (Crowley & Ivantsoff, 1990), it is left as C. eyresii. Several cryptic species have been alluded to based upon genetic evidence within Tandanus tandanus (Musyl & Keenan, 1996; Jerry & Woodland, 1997), Macquaria ambigua (Musyl & Keenan, 1992), M. australasica (Dufty, 1986), Mogurnda adspersa, and M. mogurnda (M. Adams, pers. comm.). No formal systematic clarifications have appeared, hence they were treated as single species.
Museum specimens were the primary data sources except for TAS (Frankenberg, 1974; Allen, 1989; Chilcott & Humphries, 1996) and to a lesser extent QLD (Wager, 1993; Herbert et al., 1995; Pusey, Kennard & Arthington, pers. comm.) and WA (H utchins, 1981; Allen 1982, 1989; Allen & Leggett, 1990; Morgan, Gill & Potter, 1998). The following Australian museum collections were examined primarily by browsing their catalogs, AM, MoV, NTM, QM, SAM, and WAM. Several museums in the USA were also visited, AMNH, CAS, FMNH, USNM, and UMMZ. Specimens found therein were examined. In Australian museums, identifications were typically checked only if the record seemed unique or unusual for a given drainage based on primary literature or personal knowledge. Groups and species commonly misidentified, i.e., Plotosidae, Oxyeleotris lineolatus, O. selheimi, Ambassis spp., and Hypseleotris spp., all were examined. Additional records were obtained from primary and gray lit erature, or by personal communication with specialists for groups or regions. Several species, as follows, were assumed present based on either widespread occurrences, or by likely artificial gaps due to a lack of sampling. Scleropages jardinii i s assumed present in SGC; Pseudomugil tenellus is assumed present in ARCH; Ambassis agrammus is assumed present in NICH; Ambassis elongatus is assumed present in EGC; and Hypseleotris compressa is assumed present in EKIM, VOR, NICH, EGC, and ARCH. The inclusion of these assumptions were trivial and only resulted in minor changes in the positions of NICH and SGC when analyzing relationships among regions. Richness values are slightly higher as a consequence. Incorrect records and corrections of literature are in Appendix IV.
Presence/absence data were collected at river basin scale (hereafter referred to as a drainage), largely following the drainage designations in Australian Water Resources Council (1976), but with several minor boundary changes. MDB was the only major deviation, divided into three drainages, lower Murray River below Darling River confluence, and Murray and Darling rivers each above their confluence. Drainages were summarized into 31 regions (Fig. 6) for ease of analysis. Boundaries were chosen to max imize differences in presence/absence of fish taxa between adjacent regions. Certain boundaries were somewhat arbitrary; however, this is not expected to have significant effects since some regions with clinal changes between them lack distinct boundarie s. Furthermore, poorly sampled areas were problematic to analyze and judicious merging of drainages reduced this difficulty. These include inaccessible portions of the Kimberleys (WA), drainages between Victoria and Daly rivers (NT), Aboriginal lands in Arnhem Land (NT), drainages between Roper and Nicholson rivers (NT) and to a lesser extent between Nicholson and Mitchell rivers (QLD), between Murray River and Waterpark Creek, except for Burdekin River (QLD), and most of WP (WA, NT, SA). Minor alterat ions to regional boundaries or scale change some richness and endemism values, but have little effect on overall patterns.
Richness and Endemism
Presence/absence data were entered into the spreadsheet program Microsoft Excel 97 and richness and endemism computed. Richness is defined as the number of species recorded from a region, determined by summing species occurrences. En demism is defined as any species restricted to only one region. It was determined by summing the number of records for each species, eliminating all species with values > 1 (i.e., occurring in more than one region), then adding remaining records by re gion. Both absolute numbers, and percentage endemism were calculated.
Data were analyzed via Q-mode which measures the relationship between objects (regions) based on descriptors (taxa) (compared to R-mode which measures the relationship between descriptors based on objects). Similarity coefficients use binary data to measure association between objects. A favorable characteristic of similarity coefficients is their ability to exclude double-zeros, otherwise regions with low species richness would be grouped on the basis of shared absences rather than presences (Legendre & Legendre, 1983). Shi (1993) provided a review and classification of similarity coefficients. On the basis of his review, and availability in NTSYS (Rohlf, 1997), Dice's, Jaccard's, Kulczynski's #2, and Ochiai's coefficients wer e utilized. Formulae are given in Legendre & Legendre (1983) and Rohlf (1997). The most notable difference between these coefficients is the effect of sample-size is strongest in Kulczynski's #2, moderate in Ochiai's, and low in the remaining two (S hi, 1993). This is an important consideration when choosing a suitable index, as objects with few species will be more closely ranked where sample-size effect is highest. Whether a similarity coefficient is metric or non-metric also is important in deci ding which ordination technique(s) is appropriate. Only Jaccard's Coefficient is metric (Shi, 1993). Commands used in batch files for clustering and ordination analyses in NTSYS for Jaccard's Coefficient are provided in Appendix V.
All analyses involving similarity coefficients were conducted using the statistical program NTSYS 2.02i (Rohlf, 1997). UPGMA was used to cluster similarities. This technique provides an unweighted arithmetic average between each indi vidual object and other members of the cluster or between members of clusters as they merge (Legendre & Legendre, 1983). To test if data actually contain clusters (as clustering forces objects into clusters whether or not they exist) the cluster matr ix is compared to a cophenetic value matrix of the original data to produce a cophenetic correlation value. If the two matrices show the same clustering patterns they will produce high cophenetic correlation, indicating low distortion (Rohlf & Fisher , 1968). Values of r>0.9, 0.8<0.9, 0.7<0.8, and r<0.7 indicate a very good, good, poor, and very poor fit (Rohlf, 1997).
Problems with dendrogram clustering methods include loss of information once initial pairs are formed. Hence, it is valuable for describing relationships within pairs of objects and less accurate when determining major clusters of objects (Legendre &a mp; Legendre, 1983). To show the most distinct clusters, the consensus of complete- and single-link dendrograms were calculated. Clusters present are called ball clusters which are more similar to each other than any object within the cluster is to any object outside the cluster (Rohlf, 1997).
The advantage of ordination relative to clustering is all objects are compared together rather than primarily by individual pairs. In other words, the general relationships of objects are more accurately portrayed (Legendre & Lege ndre, 1983). This compliments clustering methods which find the closest relationships within pairs of objects. Many ordination techniques require data to be distributed in Euclidean space (Shi, 1993). Since both metric and non-metric similarity coeffic ients were applied, the ordination procedure of non-metric multidimensional scaling was used. It is suitable for showing ordered relationships of objects when either non-metric or metric data are analyzed relative to other ordination methods (i.e., princ ipal coordinate analysis) (Legendre & Legendre, 1983). Stress is the term applied to variance between the two ranked orders, distance and similarity (Shi, 1993), which provides an indication of distortion relative to the original data. Objects are a nalyzed initially in two dimensions, with the number of dimensions increased until they no longer decrease stress significantly (Kruskal, 1964; Rohlf, 1972). Since stress was decreased to a reasonable value (fair or better) and graphical presentation was restricted to two or three dimensions, data were not analyzed beyond three dimensions. Cophenetic correlations were calculated so distortion between ordination and clustering results could be compared. Outputs are presented in two dimensions for ease o f interpretation and because three dimensional plots did not significantly increase information content. A minimum spanning tree connecting regions is shown to indicate whether close pairs of regions in the plot are actually close, or distant if other di mensions are considered. Based on the similarity of several clustering results (see below) only Jaccard's and Kulczynski's #2 coefficients were analyzed through multidimensional scaling.
Parsimony Analysis of Ende micity
Parsimony analysis has a long history of use in systematics for reconstructing phylogenies based on ancestral and derived characteristics. Rosen (1984, 1985) introduced parsimony analysis for examining presence / absence data of taxa by areas to define regions with shared presence of taxa. This differs from the above methods in that clustering is based on individual characters within areas rather than gross similarity between areas. Endemicity here refers to regions with common grou pings of species. In order to avoid confusion with endemism (as defined above), I refer to this method as parsimony analysis.
Analysis was conducted using PAUP 4.01b (Swofford, 1998), consisting of heuristic searches with 500 random addition sequences. Characters were coded as Dollo reversed. This specifies it is more difficult to change from absence to presence (zero to on e) than presence to absence (one to zero). In other words, extinction is more likely than colonization. All non-informative characters were removed (i.e., taxa only found in one operational taxonomic unit (OTU), e.g., endemic, and OTUs with only one cha racter e.g., WP and TORR). All most parsimonious trees were retained; only the strict consensus of these is presented. Rooting was done artificially for ease of comparison at the separation on the unrooted tree that most closely matched that of the UPGM A trees. Assessment of branch support were made using bootstrapping in PAUP and by decay index (Bremer, 1994).
Drainage patterns during lowered sea-level were modeled using Spatial Analyst 1.1 and ArcView 3.1 based upon a bathymetric 30 arc second grid produced by the Australian Geological Survey Organisation.