Ninety-five Crenadactylus specimens were sampled for genetic analysis. Allozyme profiles were successfully scored for 94 individuals and a representative subset of these (N = 53) were sequenced for the ND2 gene (Additional file 1, Table S3). Based on the results of mitochondrial and allozyme analysis, we obtained nuclear data (RAG-1) for exemplars of the ten most divergent lineages of Crenadactylus. For dating analyses we also incorporated published C-mos data for three representative deep lineages spanning crown Crenadactylus . Outgroups (Additional file 1, Table S4) were selected from published diplodactylid, carphodactylid, pygopodid, gekkonid and sphaerodactylid sequences on GenBank .
Allozyme analyses of liver homogenates were undertaken on cellulose acetate gels according to established procedures . The final allozyme dataset (Additional file 1, Table S1) consisted of 94 Crenadactylus genotyped at 42 putative loci. The following enzymes displayed banding patterns of sufficient activity and resolution to permit allozymic interpretation: ACON, ACP, ACYC, ADH, AK, DIA, ENOL, EST, FDP, FUM, GAPD, GLO, GOT, GPD, GPI, GSR, IDH, LAP, LDH, MDH, MPI, NDPK, NTAK, PEPA, PEPB, PGAM, 6PGD, PGK, PGM, SOD, SORDH, TPI, and UGPP. Details of enzyme/locus abbreviations, enzyme commission numbers, electrophoretic conditions, and stain recipes are presented elsewhere . Allozymes were labelled alphabetically and multiple loci, where present, were labelled numerically in order of increasing electrophoretic mobility (e.g. Acp
b; Acon-1 <Acon-2).
The genetic affinities of individuals were explored using 'stepwise' Principal Co-ordinates Analysis (PCO), implemented on a pairwise matrix of Rogers' genetic distances. The rationale and methodological details of stepwise PCO are detailed elsewhere . Scatterplots of PCO scores in the first two dimensions were assessed for the presence of discrete clusters of individuals which were diagnosable from all other clusters by the presence of multiple fixed differences (i.e. loci at which the two groups shared no alleles). Separate rounds of PCO were then undertaken individually on these primary groups to assess whether any group harboured additional subgroups which were also diagnosable by multiple fixed differences. Having identified groups of individuals diagnosable from one another by multiple fixed differences, two between-taxon estimates of genetic similarity were calculated: (1) percentage fixed differences (%FD; 1), allowing a cumulative 10% tolerance for any shared alleles, and (2) Nei's unbiased Distance.
DNA laboratory protocols and phylogenetic analyses
DNA extraction and amplification protocols for ND2 and nuclear loci (RAG-1, C-mos) follow those outlined elsewhere [4, 12, 28]. Newly-obtained PCR products for this study were sequenced by the Australian Genome Research Facility in Adelaide using an AB3730 DNA Analyzer (Applied Biosystems) and Big Dye chemistry. New sequences were aligned and compared to pre-existing datasets, and translated to check for substitutions leading to stop codons or frameshifts using standard procedures [4, 12, 28]. Maximum Parsimony (PAUP* vb80) , Bayesian Inference (MrBayes v3.1.2)  and Maximum Likelihood (RaxML v7.0.4)  were used to estimate phylogenetic relationships.
The final ND2 alignment consisted of 828 sites. All sequences could be translated into protein with no evidence of misplaced stop codons. Within the genus Crenadactylus 380 sites were invariable, 32 were variable but not parsimony informative, and 416 were variable and parsimony informative. The final complete nuclear alignment consisted of 2253 sites (1740 RAG-1 and 513 C-mos) of which 88 sites were variable and 28 were parsimony informative within Crenadactylus.
We performed both individual and combined analyses for the mitochondrial and nuclear data. The mitochondrial data were partitioned into first, second and third base pair positions as previous studies using the same gene region and many of the same taxa have demonstrated this significantly improves likelihood . The Akaike information criteria in MrModeltest  found the GTR+I+G model to have the highest likelihood for all partitions. For our nuclear alignment we did not partition by gene, (see justification given elsewhere ) and compared likelihood and topology for three partitioning strategies (unpartitioned; by codon; 1st with 2nds, 3rds separate). Whereas all strategies returned the same topology, likelihood support for the two partition (1st with 2nds, 3rds separate) strategy was highest. Based on the Akaike Information Criterion we used the GTR+I+G model for 1st and 2nd sites, and the GTR+G model for 3rd sites. Combined mitochondrial and nuclear analyses were partitioned by gene, but otherwise partitioned as per the non-combined analysis. As phylogenetic inference has been shown to be robust to such missing data, especially if it is evenly distributed across divergent lineages , the combined dataset included some individuals for which nuclear sequence data were unavailable.
Final Bayesian analyses were run for 5 million generations × 4 chains (one cold and three heated) sampling every 200 generations, with a burn-in of 20% (5,000 trees), leaving 20,000 trees for posterior analysis. In all Bayesian analyses, comparison of parallel runs showed posterior probability convergence (standard deviation <0.01) and likelihood equilibrium, were reached within the burn-in phase. The Maximum Likelihood tree was calculated using the -f d search function in RaxML v7.0.4 and Maximum Likelihood bootstrap support values were calculated using the -f i search function for one thousand replicates. We experimented with both simple and complicated models and found that topology, branch lengths and support values were effectively identical. Maximum Parsimony analyses were performed using heuristic searches with 100 random additions of sequences to identify most parsimonious trees. Bootstrap support values for nodes in MP trees were calculated using 100 pseudo replicates.
Divergences dates were estimated using Bayesian dating in BEAST v.1.4 . Dating analyses were performed on three sets of alignment data; RAG-1 nuclear data only (nuc), nuclear and mtDNA data combined (comb), and nuc and mtDNA combined with 3rd positions removed from the mtDNA dataset (comb reduced). Mitochondrial data were not analysed alone as the combination of old calibrations and high levels of saturation at this locus would generate significant overestimation of dates . Comparisons between these different analyses focused on variation in both actual and relative date estimates , for A) Pygopodoidea, B) Carphodactylidae, C) Pygopodidae, D) Diplodactyidae, E) core Australian Diplodactylidae (as used by Oliver and Sanders ), F) crown Crenadactylus, and (G-J) major geographically isolated clades within Crenadactylus (Table 3, Figure 2A).
Relaxed clock uncorrelated lognormal and GTR+I+G models were applied to all partitions and analyses. Nuclear only dating analyses were run unpartitioned, whereas combined analyses were partitioned into nuclear and mitochondrial data. After multiple initial runs to optimise parameters and priors, final BEAST analyses were run for 10,000,000 generations sampling every 1000 generations using the Yule speciation prior. Adequate sampling and likelihood stability was assessed using TRACER . Two thousand trees (20%) were discarded as burn in. All BEAST runs reached independence and showed no evidence of autocorrelation for all relevant parameters (e.g. branch lengths, topology and clade posteriors).
We used secondary calibrations from two independent studies [11, 12] as broad secondary priors; basal divergences among diplodactyloids (mean 71.5 mya, 95% CI 50-90 mya, normal distribution) and a uniform prior at the root of our tree (all geckos 80-150 mya). The latter prior was primarily inserted to provide a broad constraint to ensure analyses never converged on unrealistic dates, and was not meant to explicitly reflect current estimates for the age of this radiation. We experimented with incorporation of a potential calibration within crown Pygopodidae, but while this fossil is clearly a pygopod, its position within the extant radiation is uncertain and it thus does not constrain dates very tightly , and its incorporation had negligible effect on date estimates, both within the Pygopodidae and amongst other clades (results not shown).
As an independent check of our inferred date estimates, we estimated rates of mitochondrial evolution within Crenadactylus using posterior age estimates from the nuclear and two different combined analyses. A reduced mitochondrial dataset was calibrated with normal priors reflecting the posterior age estimates for the genus, and the mean and range of rates of variation were then estimated using BEAST with settings outlined above.