testing markers: linkage equilibrium and hardy–weinberg equilibrium
In 242 tests, there were 17 deviations from Hardy–Weinberg equilibrium, 15 of which were due to homozygote excess at seven loci. After sequential Bonferroni corrections (Rice 1989), four deviations remained: loci 4K, 4D, 7F, in Ovens, Koondrook and Gunbower forests, respectively, in 2005 (homozygote excess), and locus 4K in Chiltern forest (heterozygote excess) in 2004. All individuals produced ≥ 1 band at all loci (i.e. there were no candidate null homozygotes). In parentage analyses, there were no mismatches consistent with null alleles at the loci showing homozygous excess between pairs of individuals that otherwise matched (see Lada et al. 2007a). Null alleles if present must therefore be at very low frequency.
We found four significant linkage disequilibria (see Lada et al. 2007a): 7F7K in both years and 4K7H in Chiltern (2004); and 1A7F in Ovens (2004). It seemed reasonable to retain all loci given the transitory and uncommon nature of linkage disequilibria and the amount of information lost if loci 7F and 4K were removed.
patterns of historic gene flow and population dynamics
Thirtytwo mtDNA sequence haplotypes were identified (biggest difference = 2%), 24 of which were unique to a region (Fig. 2; Supporting Information Appendix S4). Seven of these were from Chiltern, which shared only three of its 10 haplotypes with other populations: haplotype 23 in Millewa–Barmah, haplotype 19 in Millewa–Barmah and Rushworth–Reedy and, the very common and internal on the network and thus, probably ancestral, haplotype 2, found in all regions (Fig. 2). Other than the latter, all haplotypes in the Warby–Ovens region were unique (Fig. 2). Campbells–Guttrum–Koondrook–Gunbower had seven unique haplotypes but shared haplotypes 10 and 7 with Millewa–Barmah, and haplotypes 14 and 21 with the Reedy–Rushworth region (Fig. 2). Haplotype 3, which differed from the ancestral haplotype 2 by only 1 bp, was the second most common haplotype in the study area, but was not found in Warby–Ovens and Chiltern (Fig. 2). Overall, substantial retention of the ancestral haplotype was coupled with the presence of unique haplotypes in all regions, suggesting occurrence of relatively recent mutations and restrictions to femalemediated gene flow among regions. Possession of mostly unique haplotypes in the Chiltern and Warby–Ovens regions suggests that these populations may have been isolated from other regions for the longest time (Fig. 2; Supporting Information Appendix S4). With zero migration, the divergence time between Chiltern and Warby–Ovens was estimated at T = 1 = 0·5N_{e}, which was relatively longer than the estimated T = 0·2 = 0·1N_{e} between Chiltern and Millewa–Barmah.
gene flow and landscape features: farmland, vegetation and rivers
Structure analyses of the 2005 genotypic data suggested K = 8 (Fig. 3) as the most likely number of clusters. However, K = 7 produced similar regional structuring, and thus, for simplicity, is presented here (Fig. 1). Five regions each had strong membership with a different, single cluster (population q values, denoting proportional cluster membership, ranging from 0·61–0·87): (i) Guttrum–Campbells; (ii) Ovens–Warby; (iii) Reedy–Rushworth; (iv) Barmah–Millewa; and (v) Chiltern (Fig. 1). Low similarity in genetic cluster membership among these five regions suggested lack of genetic connectivity among them (Fig. 1), but connectivity typically was evident between forests within each region. Gunbower–Koondrook was genetically highly differentiated from all other regions, including the nearby Guttrum–Campbells. Structure results were unaltered for K between seven and 10, in the sense that all new clusters were absorbed into Gunbower–Koondrook with uniform q values, indicating that these clusters were artificial.
Microsatellite pairwise F_{ST} values for 2005 were significantly > 0 between all forests. Due to small sample sizes, five forests were not tested in 2004 (Fig. 4; Supporting Information Table S1). The tree of pairwise F_{ST} values (Fig. 4) shows that the Chiltern population may be genetically more similar to Barmah (146 km away) than to Ovens (29 km). Pairwise F_{ST} values were low (0·015–0·029) between local populations with the exception of Guttrum–Gunbower (0·045) and Campbells–Koondrook (0·060).
Transformed pairwise F_{ST} values among 15 sampling units (Fig. 5; Supporting Information Table S2) were significantly correlated with each of transformed Euclidean distances, original LCP costs, and original LCP lengths (Mantel tests, Table 3). LCP cost was significantly correlated with genetic distance corrected for Euclidean distance but Euclidean distance was not correlated with genetic distance conditioned on LCP cost (partial Mantel tests, Table 4). Thus, LCP cost explained additional variation not explained by Euclidean distance. The significant correlation between genetic distance and LCP cost remained after controlling for LCP length, but not vice versa (Table 4). LCP distance was marginally correlated with genetic distance corrected for Euclidean distance, and there was no correlation between Euclidean distance and genetic distance corrected for LCP distance (Table 4). Thus, variation in genetic distance is correlated with LCP cost > LCP distance > Euclidean distance.
Table 3. Correlation coefficients (from Mantel tests) in 2005 among transformed variables: pairwise F_{ST} values (F_{ST}), Euclidean distances, original LCP cost (LCP cost) and original LCP length (LCP length). All correlations were significant, (P < 0·001)  F_{ST}  Euclidean  LCP cost 

Euclidean  0·674   
LCP cost  0·705  0·990  
LCP length  0·684  0·994  0·995 
Table 4. Partial Mantel tests in 2005 among transformed variables: pairwise F_{ST} values (F_{ST}), Euclidean distances, original LCP cost (LCP cost) and original LCP length (LCP length) Correlation between  Conditioned on  r correlation coefficient  P value 


LCP cost and F_{ST}  Euclidean distance  0·357*  0·002 
Euclidean distance and F_{ST}  LCP cost  –0·233  0·952 
LCP cost and F_{ST}  LCP length  0·321*  0·010 
LCP length and F_{ST}  LCP cost  –0·229  0·958 
LCP length and F_{ST}  Euclidean distance  0·177  0·106 
Euclidean distance and F_{ST}  LCP length  –0·081  0·681 
Similar results were obtained for all cost layers in which frictional cost of novegetation > cost of vegetation. Correlation between LCP costs (e.g. Table S3) and genetic distances increased with cost for novegetation, and was greatest for farm100 (r = 0·747, P < 0·001), which was significantly greater than for either farm10 or farm2. LCP costs from both the original and farm5 layers explained genetic distances equally well, and thus, at this spatial scale, cost of moving through farmland seems important, unlike that of roads and rivers. For the counterintuitive layer (cost of vegetation > novegetation), after we controlled for Euclidean distances, there was no correlation between LCP costs and genetic distances.
In the GESTE analysis at the largest spatial scale, no model seemed superior to the others when Euclidean distance and farm100 cost of LCP were tested simultaneously (Table 1, Fig. 5). Models 1 (constant), 3 (constant + cost of LCP) and 7 (constant + cost of LCP + Euclidean distance) had the highest posterior probabilities (Table 1). Similar results were obtained when length of LCP and Euclidean distance were tested (models 1, 3, 5 and 9 favoured, results not shown). When each factor was tested separately, model 3 (constant plus factor) was best in each case (Table 1).
In the only large continuous forest in which such tests were possible (Gunbower), transformed genetic distances were correlated significantly with each of transformed Euclidean distances (r = 0·938, P < 0·010), LCP lengths (r = 0·935, P < 0·012) and LCP costs (r = 0·934, P < 0·003), indicating isolation by distance.