We created a data matrix of 279 characters for 63 sauropod terminal taxa (seven outgroups and 56 putative ingroup titanosauriforms) using MESQUITE (Maddison & Maddison, 2011). Our outgroup taxa are Shunosaurus, Omeisaurus, Mamenchisaurus, Camarasaurus,Nigersaurus, Apatosaurus, and Diplodocus. These taxa were selected as they represent relatively complete and well-known genera, including a basal macronarian (the more inclusive clade including Titanosauriformes), three representatives of Diplodocoidea (the sister group to Macronaria), and three non-neosauropods; they also span three continents. The ingroup taxa include some unnamed forms that potentially represent distinct taxa and/or unusual character state combinations. Many ingroup taxa are extremely incomplete; however, excluding them merely on the basis of their completeness is problematic as they might preserve important data and unique character combinations that could have a significant impact on our results (see also discussion on incorporation of incomplete specimens in Kearney & Clark, 2003; Carrano, Benson & Sampson, 2012; and Wiens & Tiu, 2012). Table 4 lists all taxa included in the analysis, as well as our basis (i.e. references and/or personal observations) for coding taxa. Many have never been incorporated into a phylogenetic analysis prior to this study. Coding for Brachiosaurus altithorax was based only on the type specimen FMNH P25107 (Riggs, 1903). We included FMNH PR 977 [originally described as ‘Pleurocoelus sp.’ (Gallup, 1989)] in our coding of Cedarosaurus, following its referral by D'Emic (2013). The ‘Cloverly titanosauriform’ is restricted to the associated skeleton YPM 5449 (Ostrom, 1970). Although we did not disagree a priori with a recent referral of this material and Paluxysaurus to Sauroposeidon (D'Emic & Foreman, 2012; D'Emic, 2013), we retained these three taxa as separate OTUs to test this hypothesis (note that the material comprising Sauroposeidon and the ‘Cloverly titanosauriform’ both overlap with that of Paluxysaurus). The ‘French Bothriospondylus’ refers only to the individual described by Lapparent (1943). Material included in Janenschia is restricted to the type and all elements from Quarry P (see Bonaparte et al., 2000), although we also included a pubis [Janensch, 1961: pl. 19, fig. 4 (HMN MB.R.2090.2)] and ischium (HMN MB.R.2090.4: P. D. Mannion, pers. observ., 2011) that were recovered from site B [the same locality as the type material (Bonaparte et al., 2000) and distinguishable from Tornieria: P. D. Mannion, pers. observ., 2011]. We excluded the isolated manus recovered from site Nr. 5 as there are no shared autapomorphies between this and the only remaining overlapping elements from Quarry P. Unlike the recent analysis of Carballido et al. (2011b), we also did not include the caudal series HMN MB.R.2091.1−30, as there is currently no basis for its referral to Janenschia; instead it was treated here as a separate OTU. Lapparentosaurus consists only of the material discussed by Ogier (1975; see Mannion, 2010). We excluded the tentatively referred tooth in our coding of Xianshanosaurus, as this element was not found in association with the holotype skeleton (Lü et al., 2009a). Although the focus of this study was the relationships of basal titanosauriforms, we included a number of relatively complete and widely accepted derived titanosaurs (Alamosaurus, Malawisaurus, Opisthocoelicaudia, Rapetosaurus, and Saltasaurus) to help understand character evolution at the base of Titanosauria. Furthermore, these taxa have been included in most previous sauropod analyses, thereby enabling direct comparisons with our results, and several are clade specifiers. We followed previous authors by including referred remains in our Alamosaurus OTU (e.g. Upchurch et al., 2004a; D'Emic, 2012, 2013; see Table 4), although note that some of these referrals are currently based on non-overlapping material (see D'Emic, Wilson & Williamson, 2011). Supplementary information on scoring was incorporated from Wilson (2002), Upchurch et al. (2004a), and D'Emic (2012), and data on the slenderness index (SI) values of teeth were augmented from the supporting information provided by Chure et al. (2010).

The majority of characters used in this analysis were derived from Upchurch (1995, 1998), Salgado et al. (1997), Wilson & Sereno (1998), and Wilson (2002). These were revised and modified, including removal of problematic gaps between plesiomorphic and derived character states (see also Mannion et al., 2012 for a similar treatment of diplodocoid characters). This core data set was supplemented with characters from subsequent phylogenetic analyses (i.e. Upchurch et al., 2004a; Curry Rogers, 2005; Canudo et al., 2008; Remes et al., 2009; Chure et al., 2010; Ksepka & Norell, 2010; Carballido et al., 2011b; Whitlock, 2011b; D'Emic, 2012, 2013; Mannion et al., 2012), new characters based on descriptions and revisions (e.g. Wedel, Cifelli & Sanders, 2000a; Apesteguía, 2005a, b; Bonaparte, González Riga & Apesteguía, 2006; Rose, 2007; Taylor, 2009; Wilson & Upchurch, 2009; D'Emic et al., 2011; Mannion, 2011; Mannion & Calvo,2011), and entirely novel characters presented here, based on personal observations and an extensive review of the literature. We did not exclude characters based on a priori assumptions about their level of homoplasy. Where possible, we attempted to quantify, or at least more precisely define, characters and state boundaries to remove ambiguity (see similar attempts by Harris, 2006). The complete character list, including references, as well as our data matrices, is provided in Appendices 1–3. The MESQUITE versions of the data matrices are also presented as Supporting Information Appendices S1–S3 and their TNT equivalents are available from the first or second authors on request.

Reductive (or ‘contingent’) coding was used here, rather than absence coding. Although problems exist with the former method, simulation studies suggest that it is the best solution when there is no logical interpretation of a character for a given taxon (Strong & Lipscomb, 1999; Brazeau, 2011). For example, our character number 127 (C127) relates to the absence or presence of postzygapophyseal epipophyses on cervical vertebrae, whereas C128 relates to the posterior extent of these epipophyses. However, a taxon scored as zero for C127 (i.e. that lacks epipophyses) cannot be scored for either the basal state (epipophyses do not extend beyond the postzygapophyses) or derived state (epipophyses extend beyond the postzygapophyses) for C128. In absence coding, our taxon would be scored as ‘0’ for C128, designating it as the basal condition without any actual evidence for this decision, whereas in reductive coding we scored this character as a ‘?’.

There are two versions of the data matrix, known here as the ‘Lusotitan standard discrete matrix’ (LSDM) and the ‘Lusotitan continuous + discrete matrix’ (LCDM). These matrices differ in the way they deal with quantitative characters, such as the relative proportions of skeletal elements. In both matrices, characters 1–74 represent quantitative characters, whereas characters 75–279 score discrete variation (mostly binary characters such as the absence/presence of a feature). Thus, the numbers referring to characters in the character list (seeAppendix 1) apply to both the LSDM and LCDM, even though the treatment of characters 1–74 is different in the two matrices. In the LSDM, the quantitative characters have been ‘discretized’ by dividing the observed variation in the ratio between two measurements into two or more discrete states (with state boundaries determined from previous studies and/or based on our outgroup taxa). For example, C2 in the character list (see Appendix 1) is: ‘External naris, greatest diameter to greatest diameter of orbit ratio: greater than 1.0 (0); 1.0 or less (1)’. This is the standard method for dealing with quantitative morphological characters and it has been used by virtually all previous phylogenetic analyses of dinosaur relationships (although see, for example, Maidment et al., 2008). We applied this standard approach in order to make our results more easily comparable with previous analyses of sauropod relationships, all of which have discretized their quantitative characters. However, there are several problems with the treatment of quantitative characters in this way, one of which is that the discretization of continuous variation is somewhat arbitrary (Wiens, 2001). In many cases, the relative proportions of two skeletal elements vary in a continuous fashion across the known taxa: that is, variation does not fall neatly into two or more separate clusters with gaps between them. This means that different systematists might find evidence to support alternative tree topologies because they have defined the boundaries between states at different points in a character's continuous variation. For example, Wilson & Sereno (1998) noted that macronarian sauropods have a derived condition in which the ratio of the length of the longest metacarpal to the length of the radius is 0.45 or higher (C52 in our study). Consequently, taxa such as Apatosaurus louisae and Mamenchisaurus youngi, with metacarpal to radius length ratios of approximately 0.40 (Table 5), are scored with the plesiomorphic state. However, metacarpal to radius length ratios in sauropods cover a wide range of values (Table 5), with no clear gaps in this variation that might be used to produce a more objective division into plesiomorphic and derived states. For example, it would be equally legitimate to define the division between states 0 and 1 as a ratio of 0.4 (as implemented by Upchurch et al., 2004a, and in this study), so that Apatosaurus and Mamenchisaurus would be scored with state 1. A second drawback with discretization of quantitative characters is that it might fail to capture some of the potentially phylogenetically informative signal in the data. For example, Table 5 suggests that additional derived states in longest metacarpal to radius length ratios could be recognized: titanosaurs such as Aeolosaurus and Argyrosaurus have very high values of 0.53–0.6, whereas other taxa have values in the range of 0.45–0.50 (e.g. Camarasaurus and Rapetosaurus). At present, this variation is obscured by assigning state 1 to all taxa with a ratio of 0.4/0.45 or higher (see Wiens, 2001 for further discussion of this issue). These problems can be circumvented by representing quantitative characters by continuous data: that is, by using the ratios between two parameters as the character states themselves, as proposed by Goloboff, Mattoni & Quinteros (2006) and as can be implemented in TNT (Goloboff, Farris & Nixon, 2008). For example, in the case of C2 in our data set, the LCDM scored Shunosaurus with state 1.26, Diplodocus with state 0.71, and Giraffatitan with state 1.15. This means that relatively larger changes between character states (i.e. the values of the ratio) required by a given tree topology will cost more (in terms of parsimony) than a relatively smaller shift. In effect, treatment of quantitative characters as continuous data means that the cost of evolutionary transformation of a character (on a given tree topology) is proportional to the required change in the value of the character states (Goloboff et al., 2006). Thus, treating quantitative characters as continuous data eliminates the need to define arbitrary state boundaries in order to discretize the observed variation (Wiens, 2001; Goloboff et al., 2006), which otherwise imposes artificial and potentially subjective divisions onto a quantitative data series that might result in the a priori biasing of an analysis (Maidment et al., 2008). A further benefit is that future analyses do not need to modify the state boundary, or reverse the polarity, when using the same character for revised data sets or analyses focused at different taxonomic levels (e.g. an analysis examining all sauropods, rather than just titanosauriforms). However, continuous coding sometimes means that less information can be gleaned from a description or personal examination of a specimen. For example, if an author states that one dimension of an element is greater than another (e.g. ‘wider than tall’) but does not provide the actual value, or if an element is incomplete such that its exact dimensions are unknown but it is clear that one value is greater than the other, then the character cannot be coded for the continuous data matrix but often can for the discrete matrix (in the current study, 30 vs. 32% of the quantitative character data matrix could be scored for the continuous and discrete data sets, respectively). Additionally, the use of continuous characters makes it more difficult to identify synapomorphies because characters represent a spectrum of morphological variation. Different but analogous treatments of quantitative characters have been implemented in a small number of other palaeontological studies (e.g. Maidment et al., 2008; Angielczyk & Rubidge, 2010; Ketchum & Benson, 2010; Kammerer, Angielczyk & Fröbisch, 2011), and the method used here has been implemented in palaeontological studies of mammalian and trilobite phylogeny (Prevosti, 2010; Hopkins, 2011), and in one preliminary study of sauropodomorphs (Upchurch, 2009).

One important issue raised by the treatment of quantitative characters as continuous data is that of ‘scaling’ (Goloboff et al., 2006). The range of values covered by each quantitative character is likely to vary, resulting in characters having different weights. For example, suppose that character x has state values that range from 0.5 to 0.8, whereas character y has values that range from 0.5 to 3.5. The range of values covered by x is 0.3 whereas that for y is 3.0. This means that the weight of character y is ten times greater than that assigned to x: all things being equal, y is more likely to have a stronger influence on final tree topology than x (see Goloboff et al., 2006 for further discussion and examples). Most systematists who apply parsimony have preferred to commence their analyses with the assumption that all characters are equally weighted (see also our LSDM analysis). Thus, the treatment of quantitative characters as continuous data means that we need a method that deals with the scaling issue. One such approach is ‘implied weights’, (Goloboff, 1993; Goloboff et al.,2006). This approach weights each character in proportion to its fit to a given tree topology, and is calculated as follows:

where k is a constant (defined by the user) and h is a measure of the homoplasy of the character (i.e. the number of steps required by the character on a given tree minus the minimum number of steps required by that character when it displays no homoplasy). Thus, implied weights provide a means for identifying and down-weighting homoplastic characters and, unlike other methods (such as rescaled weighting), is implemented during a tree search rather than afterwards (Goloboff et al., 2006). One byproduct of implied weighting is that homoplasy in a relatively scaled-up quantitative character (e.g. y above) is penalized more strongly than the same amount of homoplasy in a relatively scaled-down one (e.g. x above). The result of applying implied weights is that the differential weighting of quantitative characters is reduced, at least partially, and it has been recommended as a method for dealing with the scaling issues created by the treatment of quantitative characters as continuous data (Goloboff et al., 2006). Here, therefore, we analyse the LCDM using implied weighting with the default k-value of 3 in TNT (Goloboff et al., 2006), although we note that different k-values can produce alternative topologies (Goloboff, 1993).

All tree searches, identification of ‘wild card’ taxa, and tests of topological robustness were carried out in TNT (Goloboff et al., 2008). Character mapping and tree drawing were carried out in MESQUITE (Maddison & Maddison, 2011). PAUP 4.0 (Swofford, 2002) was used to implement Templeton's tests.

In all analyses of the LSDM, characters 11, 14, 15, 27, 104, 122, 147, 148, 177, 205, and 259 were treated as ordered multistate characters. Safe taxonomic reduction was applied to this data set using the program TAXEQ3 (Wilkinson, 1995), but no redundant taxa were identified. The full LSDM was then analysed using the ‘Stabilise Consensus’ option in the ‘New Technology Search’ in TNT vs. 1.1 (Goloboff et al., 2008). In these analyses, searches were carried out using sectorial searches, drift, and tree fusing, with the consensus stabilized five times. This yielded 142 trees of length 1070 steps. In order to search for additional topologies, these 142 trees were used as the starting trees for a ‘Traditional Search’ using tree bisection-reconstruction (TBR). This produced 24 192 most parsimonious trees (MPTs) of length 1069 steps [consistency index (CI) = 0.275, retention index (RI) = 0.532, rescaled consistency index (RCI) = 0.147]. The strict consensus tree is shown in Figure 20. The pruned trees option in TNT was then used to identify the most unstable OTUs in the MPTs. This analysis indicated that the greatest increase in strict consensus tree resolution could be achieved through the a posteriori deletion of Australodocus and Malarguesaurus, resulting in the strict reduced consensus tree (see Wilkinson, 1994) shown in Figure 21.

The agreement subtree (i.e. the largest fully resolved topology common to all MPTs) was then calculated in TNT, requiring the a posteriori pruning of a further ten OTUs (see Fig. 22). These OTUs are: Abydosaurus, Angolatitan, Astrophocaudia, Cedarosaurus, Chubutisaurus, ‘Cloverly titanosauriform’, Europasaurus, Fukuititan, Fusuisaurus, and Ligabuesaurus. The relationships of each of the 12 pruned OTUs were then investigated individually by sequentially deleting 11 of them from the 24 192 MPTs and then constructing a strict reduced consensus tree for each. These studies indicated that: (1) Abydosaurus and Cedarosaurus are derived brachiosaurids, forming a polytomy with Giraffatitan and Venenosaurus; (2) Australodocus is a member of Titanosauria, but detailed relationships are difficult to assess; (3) Angolatitan, Chubutisaurus, ‘Cloverly titanosauriform’, and Malarguesaurus are non-titanosaurian somphospondylans, more closely related to Titanosauria than the euhelopodid clade; (4) Astrophocaudia is a basal somphospondylan that occupies positions within Euhelopodidae in some of the MPTs, and basal to this clade in others; (5) Europasaurus is a basal brachiosaurid, in a polytomy with forms such as Brachiosaurus and the ‘French Bothriospondylus’; (6) Fukuititan and Fusuisaurus are very unstable non-titanosaurian titanosauriforms that occupy positions within Brachiosauridae and basal Somphospondyli in the MPTs; and (6) Ligabuesaurus is either a basal titanosaur or the sister-taxon to Titanosauria, in a trichotomy with Andesaurus and a clade of all other titanosaurs. Based on these results, a reduced strict consensus cladogram (Fig. 23) was generated via the a priori pruning of the five least stable OTUs (Astrophocaudia, Australodocus, Fukuititan, Fusuisaurus, and Malarguesaurus).

The support for the relationships produced by the LSDM was evaluated using symmetric resampling. Bremer supports were not calculated because collection of suboptimal trees indicated that the limit of 99 999 topologies was reached when trees of just one extra step were retained (i.e. trees with lengths of 1069 and 1070 steps). This means that Bremer supports could not be evaluated because of the limits on memory in TNT.

Symmetric resampling is similar to bootstrapping and jack-knifing, but the probability of down-weighting a given character is equal to the probability of up-weighting it (Goloboff et al., 2003). Following Goloboff et al. (2003), symmetric resampling was used here to generate the relative, rather than absolute, frequencies of groups of taxa in the trees produced by multiple replicate analyses. This was because these authors demonstrated that absolute frequencies (i.e. the number of times a clade occurs, divided by the total number of resampled replicate trees) often under- or over-estimate support. Relative frequency (termed the ‘GC value’ by Goloboff et al., 2003) is defined as the frequency of a given group of taxa minus the frequency of the most common contradictory group. GC values can thus vary from 1 to −1, where 1 indicates maximum support, 0 indicates indifferent support, and −1 indicates maximum contradiction. Symmetric resampling was applied to the LSDM using 5000 replicates in TNT. All tree searches were carried out using ‘Traditional Search’ with TBR. The resulting GC values are shown in Figure 24A (note that in all Figures, GC values are multiplied by 100). As noted by Wilkinson (1996), support values can be lowered by OTU instability. For example, a clade containing five OTUs (e.g. A to E) can appear weakly supported, but this can arise from the combination of a strongly supported clade of four OTUs (e.g. A–D) and the instability of one OTU (e.g. E). In order to investigate this phenomenon in the LSDM trees, the 12 OTUs excluded from the agreement subtree (Fig. 22) were pruned from the replicate trees generated by symmetric resampling, and the GC values were recalculated (see Fig. 24B).

We also analysed the LSDM using implied weighting in TNT with a k-value of 3 (see ‘Data sets’ above). These analyses are here referred to as LSDM_iw. The purpose of these analyses was to enable determination of the relative contributions of implied weighting and the alternative treatments of continuous data to the LSDM and LCDM results. Differences between the LSDM and LSDM_iw topologies indicate the impact of applying implied weighting, whereas differences between the LSDM_iw and LCDM indicate the effects of treating continuous data as such.

The LSDM_iw was analysed using the same protocols as the LSDM (i.e. an initial ‘New Technology Search’ in which the consensus was stabilized five times, followed by the use of the resulting MPTs as the starting trees for two consecutive rounds of TBR). These analyses produced 45 MPTs of length 111.22035 steps. The strict consensus cladogram of these 45 MPTs is shown in Figure 25 and the agreement subtree, following the a posteriori pruning of eight OTUs (Abydosaurus, Angolatitan, Baotianmansaurus, Brachiosaurus,Cedarosaurus, Europasaurus, ‘Pelorosaurus’ becklesii, and Xianshanosaurus), is shown in Figure 26.

In all analyses described below, characters 104, 122, 147, 148, 177, 205, and 259 were ordered. The LCDM was analysed in TNT using the ‘New Technology Search’ followed by ‘Traditional Search’ with TBR, as described above for the LSDM. However, for the LCDM, implied weights were used with a k-value of 3, as was implemented for our LSDM_iw. These searches yielded three MPTs of length 102.79682 steps. The non-integer step length reflects the presence of 74 quantitative characters in which states are expressed directly as continuous data, combined with the down-weighting of homoplastic characters produced by the application of implied weights. The latter is also responsible for the fact that tree length is substantially lower than the total number of characters. The strict consensus tree of the three MPTs is shown in Figure 27.

Symmetric resampling and Bremer support were applied to the LCDM. The GC values produced by symmetric resampling (5000 replicates; see above for methodology) are shown in Figure 28A. Bremer support values were generated in TNT by applying the ‘New Traditional Search’ using TBR [with implied weights (k = 3)] and collecting suboptimal topologies. In the case of the LCDM, only relative Bremer supports were calculated because it is not clear how additional fractional steps should be interpreted in the context of absolute Bremer supports. Relative Bremer support values (termed the relative fit difference, RFD) are calculated as follows:

where F is the sum of the fits of characters that fit the MPTs better than suboptimal trees, and C is the sum of the fits of characters that fit suboptimal trees better than MPTs (Goloboff & Farris, 2001). RFD therefore varies from 0 to 1, so that a value of 0 indicates that a clade has no support, and a value of 1 indicates that the clade is entirely uncontradicted (Goloboff & Farris, 2001). For example, if twice as many binary characters support a given node as contradict it, the RFD value is 0.5. The RFD values for the LCDM are shown in Figure 28B (as before, the GC and RFD values are multiplied by 100 in the figures).

Lusotitan

Other brachiosaurids

Note on the taxonomy of Brachiosaurus

Somphospondyli is defined as the most inclusive clade that includes Saltasaurus loricatus but excludes Brachiosaurus altithorax (Wilson & Sereno, 1998; Upchurch et al., 2004a), and within this Titanosauria is defined as the least inclusive clade that includes Andesaurus delgadoi and Saltasaurus loricatus (Wilson & Upchurch, 2003) (see Table 6). Our analyses produced notably different topologies for Somphospondyli: whereas our LSDM recovered a large paraphyletic array of basal somphospondylans leading to a relatively ‘traditional’ Titanosauria, the LSDM_iw and LCDM resulted in either one or zero basal somphospondylans, respectively, and an extremely diverse Titanosauria. In our LSDM_iw and LCDM, Titanosauria is composed of an ‘andesauroid’ clade that is the sister taxon to a clade of titanosaurs containing Lithostrotia (the least inclusive clade containing Malawisaurus dixeyi and Saltasaurus loricatus; Wilson & Upchurch,2003; Upchurch et al., 2004a). This split at the base of Titanosauria into an andesauroid and derived titanosaur clade has only previously been recovered in the analyses of Canudo et al. (2008) and Carballido et al. (2011b). ‘Andesauridae’ was originally erected as a family comprising Andesaurus, Argentinosaurus, and Epachthosaurus (Bonaparte & Coria, 1993), but all subsequent phylogenetic analyses have recovered this assemblage as paraphyletic (e.g. Salgado et al., 1997; Sanz et al., 1999; Curry Rogers, 2005; Calvo et al., 2007), and several authors have recommended the disuse of this name (Salgado et al., 1997; Salgado, 2003; Wilson & Upchurch, 2003; Mannion & Calvo, 2011). Following our recovery of a clade containing Andesaurus and other taxa, we refer to this as Andesauroidea, following Salgado (2003; see Table 6 for the phylogenetic definition). The sister clade to Andesauroidea in our analysis would traditionally be called Titanosauroidea (Upchurch, 1995; Salgado, 2003); however, Titanosaurus is indeterminate and thus its co-ordinated rank-taxa must be abandoned (Wilson & Upchurch, 2003). As a result of Andesaurus being recovered as the sole representative of Andesauroidea in our LSDM, we refrain from providing a new name for it sister clade, referring to it as ‘Titanosauroidea’.

Dongbeititan was placed as the basal-most member of Somphospondyli in our LSDM, but as a non-titanosauriform macronarian in our LCDM. Wang et al. (2007) described Dongbeititan, from the Barremian of China, as a non-titanosaurian somphospondylan, Mannion & Calvo (2011) noted possible titanosaurian affinities, and D'Emic (2012) indicated a placement within Somphospondyli. Its variable position here might reflect the highly deformed nature of the material (R. N. Barnes, pers. observ., 2011).

The LSDM recovered a euhelopodid clade of East Asian basal somphospondylans, forming a sister clade with all other somphospondylans. Euhelopodidae (see Table 6 for phylogenetic definition) comprises five (Erketu, Euhelopus, Phuwiangosaurus,Qiaowanlong, and Tangvayosaurus) of the six taxa recovered in this clade by D'Emic (2012), as well as two additional taxa not included in the latter analysis (Gobititan and Liubangosaurus). Our composition primarily differs in that Daxiatitan was recovered as a derived titanosaur in the LSDM. However, our LSDM_iw and LCDM produced notably different results in the placements of these taxa: all were recovered in the titanosauroid clade, but form a paraphyletic array, rather than a monophyletic clade.

Liubangosaurus was recovered as the basal-most euhelopodid in our LSDM, but as a saltasaurid in our LSDM_iw and LCDM. Known from five middle-posterior dorsal vertebrae from the Early Cretaceous of China, Liubangosaurus was described as a basal eusauropod by Mo, Xu & Buffetaut (2010), recovered in a polytomy with Macronaria and Diplodocoidea at the base of Neosauropoda. If correctly identified, the presence of such a basal taxon would be highly unusual: the Cretaceous sauropod fauna of China is otherwise entirely composed of titanosauriforms (Wilson, 2005a; Wilson & Upchurch, 2009), or at least derived macronarians (based on the placement of Dongbeititan: see above). A putative diplodocid (Upchurch & Mannion, 2009) might represent an additional non-titanosauriform presence in the Cretaceous of China, although Whitlock, D'Emic & Wilson (2011) have argued for a probable titanosauriform identification of this specimen. With the exception of Spanish occurrences whose stratigraphical age cannot be constrained more precisely than Tithonian–middle Berriasian (e.g. Royo-Torres et al., 2006), there is no other evidence for the survival of non-neosauropods across the Jurassic/Cretaceous (J/K) boundary in any part of the world (Mannion et al., 2011b). This is consistent with the results from both of our analyses, which suggest that Liubangosaurus is a somphospondylan, rather than a late surviving Cretaceous non-neosauropod as proposed by Mo et al. (2010).

The phylogenetic affinity of the Chinese taxon Euhelopus has a complex history, with the most recent thorough treatment regarding it as a basal somphospondylan (Wilson & Upchurch, 2009), although it was placed as a basal macronarian by Royo-Torres (2009) and Carballido et al. (2011b). Its role as a clade specifier for Euhelopodidae (D'Emic, 2012) makes it a member of this clade by default, but in our LSDM_iw and LCDM, Daxiatitan is the only other euhelopodid. Daxiatitan is an Early Cretaceous Chinese taxon that was recovered as a somphospondylan or basal titanosaur by You et al. (2008). However, the exclusion of Andesaurus from this, and many other analyses (e.g. Wilson, 2002; Suteethorn et al., 2010; Zaher et al., 2011), despite its importance as a clade specifier for Titanosauria (see above and Table 6), means that the placement of the titanosaurian node has often been arbitrarily determined. Here we found support for a slightly more derived placement within Titanosauria. Erketu, from the early Late Cretaceous of Mongolia [Ksepka & Norell, 2006, 2010 (see the latter reference regarding this younger age, rather than the Early Cretaceous age originally proposed)] has previously been included in three phylogenetic analyses, in which it was recovered as a non-titanosaurian somphospondylan (Ksepka & Norell, 2006, 2010), and an euhelopodid (D'Emic, 2012). Whereas it was recovered as the sister taxon to Euhelopus in our LSDM, Erketu was placed as a derived titanosauroid in our LSDM_iw and LCDM, forming a clade with Mongolosaurus + Rapetosaurus.

The Early Cretaceous Chinese taxon Qiaowanlong was originally described as a brachiosaurid by You & Li (2009), but was recovered as a basal somphospondylan in the analysis of Ksepka & Norell (2010), with these authors demonstrating the wider distribution of the features used by You & Li (2009) to support a brachiosaurid position (see also Mannion, 2011). Here, Qiaowanlong formed a sister-taxon relationship with Phuwiangosaurus in our LSDM_iw and LCDM, rather than clustering with Erketu + Euhelopus as in our LSDM. The Thai taxon Phuwiangosaurus has been regarded as either a basal titanosaur (e.g. Upchurch, 1998; Curry Rogers, 2005; Canudo et al., 2008; Carballido et al., 2011a; Zaher et al., 2011) or non-titanosaurian somphospondylan (although see earlier discussion regarding the exclusion of Andesaurus from many analyses) in nearly all previous analyses (e.g. Rose, 2007; González Riga et al., 2009; Wilson & Upchurch, 2009; Suteethorn et al., 2010; Carballido et al., 2011b; D'Emic, 2012, 2013). Royo-Torres (2009) provided the only strongly contrasting viewpoint, recovering Phuwiangosaurus as a brachiosaurid, but its somphospondylan affinities appear secure.

Tangvayosaurus, from the Aptian–Albian of Laos, was originally described as a basal titanosaur, although Allain et al. (1999: 610) noted that this was only supported by one feature (‘pubis significantly longer than ischium’). A basal titanosaurian position was also supported in the analysis of Zaher et al. (2011; although see above regarding the exclusion of Andesaurus). Several authors have listedTangvayosaurus as a non-titanosaurian titanosauriform (e.g. Wilson, 2005a; Wilson & Upchurch, 2009; Mannion & Calvo, 2011), whereas Royo-Torres (2009) placed it as the sister taxon to Titanosauriformes, and Suteethorn et al. (2010) recovered it as a non-titanosaurian somphospondylan. Here, Tangvayosaurus was placed as the sister taxon to Fusuisaurus [forming a clade with Gobititan + ‘Huanghetitan’ruyangensis (see below)] in our LSDM_iw and LCDM, rather than as the sister taxon to Phuwiangosaurus as in our LSDM (see also D'Emic,2012). You et al. (2003) described Gobititan, from the Albian of China, as a basal titanosaur, but this position has been doubted by several subsequent workers, who have suggested a more basal position within Titanosauriformes (Wilson & Upchurch, 2009; Mannion & Calvo,2011; Mannion & Otero, 2012), including listing it as a somphospondylan (D'Emic, 2012). Euhelopodid or titanosauroid affinities are supported based on our analyses. The Early Cretaceous Chinese taxon Fusuisaurus was considered a titanosauriform of uncertain affinities by Mo et al. (2006) and according to our LSDM (see above). This instability most likely relates to the incompleteness and brief description of the specimen [character completeness metric (CCM) = 9%; Mannion & Upchurch, 2010b], and we regard it as a titanosauriform of uncertain placement, with probable somphospondylan affinities based on our LSDM_iw and LCDM.

The late Early Cretaceous North American genus Brontomerus (Taylor, Wedel & Cifelli, 2011) was recovered as the most basal member of the non-euhelopodid somphospondylan clade in our LSDM, and occupies a similar position within the titanosauroid clade of our LSDM_iw and LCDM. Brontomerus was described by Taylor et al. (2011) as a camarasauromorph (i.e. basal macronarian), with some evidence for a somphospondylan position, as supported here, although D'Emic (2012) regarded it as an indeterminate titanosauriform.Astrophocaudia was recovered as the sister taxon to Brontomerus in our LSDM_iw and LCDM, and as a basal somphospondylan in our LSDM. Astrophocaudia is known from the early Albian of North America (Langston, 1974; D'Emic, 2013) and has been regarded as either a brachiosaurid or titanosaur by previous workers (see D'Emic, 2013 for a review). However, a recent revision of this material and its first incorporation into a phylogenetic analysis (D'Emic, 2013) positioned it as a nontitanosaurian somphospondylan. Thus, it seems that the somphospondylan affinities of Brontomerus and Astrophocaudia are secure, but their exact placement within this clade cannot currently be ascertained.

Our LSDM recovered a basal somphospondylan clade comprising ((Sonorosaurus + Tastavinsaurus) + (Diamantinasaurus + ‘Pelorosaurus’ becklesii)), whereas none of these taxa clustered together in our other analyses. In our LSDM_iw and LCDM,Sonorosaurus is regarded as a brachiosaurid (see above), and Tastavinsaurus is recovered as a derived andesauroid. Tastavinsaurus is known from two individuals from the early Aptian of Spain (Canudo et al., 2008; Royo-Torres, 2009; Royo-Torres et al., 2012). Originally recovered as a non-titanosaurian somphospondylan (Canudo et al., 2008; see also D'Emic, 2012), the analyses of Carballido et al. (2011a, b) placed it as a non-titanosauriform macronarian. Royo-Torres et al. (2012) recovered it as the sister taxon to Cedarosaurus in all of their analyses, forming a clade with Venenosaurus that these authors called Laurasiformes (see also Royo-Torres, 2009, who foundAragosaurus and Sonorasaurus to be additional members of this clade). However, the position of Laurasiformes varied in Royo-Torres et al. (2012), being recovered as a basal somphospondylan clade in the analyses based on the data matrix of Wilson (2002), and as a brachiosaurid clade in those based on Upchurch et al. (2004a), whereas in Royo-Torres (2009) this clade was placed outside Titanosauriformes. Here, we found no evidence to support the monophyletic clustering of these taxa as Laurasiformes (see also D'Emic,2012): Tastavinsaurus is a somphospondylan, Cedarosaurus and Venenosaurus are brachiosaurids (see above), Sonorasaurus might represent a brachiosaurid or somphospondylan (see above), and Aragosaurus is placed outside Titanosauriformes (see below). The LSDM_iw placed ‘Pelorosaurus’ becklesii within the titanosauroid clade, whereas the LCDM analysis recovered it as a basal member of Andesauroidea. This taxon is known from an upper forelimb and skin impression from the Berriasian–Valanginian of England. It has a complicated taxonomic history (Taylor & Naish, 2007; Upchurch, Mannion & Barrett, 2011b), but has been regarded as a basal titanosaur by some authors (Upchurch, 1995; Upchurch et al., 2004a, 2011b; Mannion & Calvo, 2011). Although known only from limited material, ‘Pelorosaurus’ becklesii represents a distinct and diagnostic somphospondylan, requiring a new generic name (Upchurch et al., 2011b), although its titanosaurian affinities cannot currently be confirmed (see also D'Emic, 2012). Diamantinasaurus, from the early Late Cretaceous of Australia (Hocknull et al., 2009; see Tucker et al., 2013 regarding this younger age, rather than the late Albian age traditionally ascribed to the Winton Formation), has previously been recovered as a lithostrotian by Hocknull et al., (2009) and Zaher et al. (2011). It was placed as the sister taxon to Lithostrotia in our LSDM_iw and LCDM analyses.

The late Early Cretaceous North American taxa Sauroposeidon and Paluxysaurus were both originally described as brachiosaurids (Wedel et al., 2000a, b; Rose, 2007), with this position supported in subsequent analyses incorporating Paluxysaurus (Ksepka & Norell,2010; Carballido et al., 2011b). However, D'Emic (2013) referred Paluxysaurus to Sauroposeidon and recovered this combined OTU close to the base of Somphospondyli (see also D'Emic, 2012). The two were recovered as sister taxa in all of our analyses, supporting D'Emic's synonymization, although they were recovered as basal somphospondylans in our LSDM (albeit in a more derived position than in D'Emic,2012), and andesauroids in our LSDM_iw and LCDM. The contemporaneous ‘Cloverly titanosauriform’ has not previously been included in an analysis by itself, but was originally described as a titanosaur by Ostrom (1970) and referred to Sauroposeidon by D'Emic & Foreman (2012). In our LSDM the ‘Cloverly titanosauriform’ was recovered in a trichotomy with Sauroposeidon + Paluxysaurus and a clade leading to Titanosauria, whereas it was placed as a distantly related andesauroid in the LSDM_iw and LCDM analyses. As such, we find only equivocal support for its referral to Sauroposeidon.

Wintonotitan, from the early Late Cretaceous of Australia has previously been recovered as a non-titanosaurian somphospondylan (Hocknull et al., 2009; Carballido et al., 2011a), a non-titanosauriform macronarian (Carballido et al., 2011b), and as a basal titanosaur [asAustrosaurus (Upchurch et al., 2004a) – see below]. Hocknull et al. (2009) also noted several morphological features (not included in their phylogenetic analysis) that might indicate a position within Titanosauria. The early Late Cretaceous Chinese sauropod Ruyangosauruswas originally described as a basal titanosaur (‘Andesauridae’) by Lü et al. (2009b), but Mannion & Calvo (2011) argued that it showed no clear titanosaurian features and should be considered a non-titanosaurian somphospondylan. In our LSDM, Wintonotitan andRuyangosaurus form a clade of derived non-titanosaurian somphospondylans, whereas they are andesauroids in our LSDM_iw and LCDM results.

Huanghetitan liujiaxiaensis, from the Early Cretaceous of China, was recovered as a non-titanosaurian somphospondylan by You et al. (2006) and Hocknull et al. (2009). ‘Huanghetitan’ ruyangensis is an early Late Cretaceous Chinese taxon described as a second species of Huanghetitan, with Huanghetitanidae erected as a monogeneric family (Lü et al., 2007). However, we found that the two Huanghetitanspecies did not cluster together in our analyses: they are successive outgroups to Titanosauria in our LSDM, whereas in our LSDM_iw and LCDM the type species is an andesauroid and the referred species a titanosauroid. As such, we do not support the monophyly of Huanghetitanidae and, pending revision of the two species, suggest that the diagnostic ‘Huanghetitan’ ruyangensis (Henan Geological Museum specimen 41HIII-0001: P. D. Mannion & P. Upchurch, pers. observ., 2012) is in need of a new generic name.

In our LSDM_iw and LCDM analyses, several taxa not yet mentioned were recovered within the andesauroid clade. The Turonian-aged African genus Angolatitan was recovered as a basal somphospondylan in its only previous inclusion in a phylogenetic analysis (Mateuset al., 2011), although D'Emic (2012) suggested lithostrotian affinities. Our LSDM supports the non-titanosaurian somphospondylan placement of Mateus et al. (2011), whereas it was recovered as an andesauroid in our other analyses. Although the middle Cretaceous Argentinean genus Chubutisaurus was included within Brachiosauridae by McIntosh (1990), it has been recovered as a non-titanosaurian somphospondylan by most other authors, usually closely related to the titanosaur radiation (Salgado et al., 1997; Bonaparte et al., 2006; Calvo et al., 2007; González Riga et al., 2009; Carballido et al., 2011a; D'Emic, 2012), but was placed outside Titanosauriformes by Carballido et al. (2011b). In all our analyses, Chubutisaurus clusters closely with Angolatitan. The Late Cretaceous Chinese taxonBaotianmansaurus was originally described as a titanosauriform with probable somphospondylan affinities by Zhang et al. (2009), and D'Emic (2012) indicated euhelopodid affinities. In its first incorporation into a phylogenetic analysis, Baotianmansaurus was recovered as a titanosaur in all analyses: in our LSDM it was placed within Saltasauridae, as the sister taxon to Opisthocoelicaudia, whereas it was recovered as an andesauroid in our LSDM_iw and LCDM. Ligabuesaurus, from the Aptian–Albian of Argentina (Bonaparte et al., 2006), was recovered in a trichotomy with the middle Cretaceous Argentinean taxon Andesaurus (Calvo & Bonaparte, 1991; Mannion & Calvo, 2011) and a clade comprising all other titanosaurs in our LSDM, and was recovered as the sister taxon of Andesaurus in our LSDM_iw and LCDM analyses, in a deeply nested position within the andesauroid clade. Ligabuesaurus has been recovered as a non-titanosaurian somphospondylan by most authors (Bonaparte et al., 2006; González Riga et al., 2009; Carballido et al., 2011a), and was placed as the most basal member of this clade in the analysis of D'Emic (2012). However, Carballido et al. (2011b) found it to form a clade withAndesaurus, a position partly supported here.

Australodocus was recovered as a titanosaur in all of our analyses, forming a polytomy with titanosaurs more derived than Andesaurus in our LSDM, and placed as a basal titanosauroid in our LSDM_iw and LCDM. This Tithonian-aged taxon from the Tendaguru Formation of Tanzania was originally described as a new genus of diplodocid (Remes, 2007), but revision by Whitlock (2011a) demonstrated its titanosauriform affinities (see also D'Emic, 2012), a position supported by phylogenetic analysis (Whitlock, 2011a, b; Mannion et al., 2012). These previous analyses were focused on the inter-relationships of diplodocoids and thus Brachiosaurus (a composite OTU ofBrachiosaurus and Giraffatitan) and Sauroposeidon (Whitlock, 2011a) were the only other titanosauriforms included, limiting further determination of the position of Australodocus. However, Whitlock (2011a) recovered a clade of Australodocus + Brachiosaurus, to the exclusion of Sauroposeidon, indicating a closer relationship with Brachiosauridae for Australodocus. The recovered position in our analyses demonstrates the spatiotemporally contemporaneous existence of both brachiosaurids (i.e. Giraffatitan) and somphospondylans in the Tendaguru Formation, and provides the earliest body fossil evidence for titanosaurs (see ‘Jurassic titanosaurs?’ below).

Fukuititan, from the Barremian of Japan, was regarded as a titanosauriform of uncertain affinities by Azuma & Shibata (2010), and in our LSDM. In our LSDM_iw and LCDM it was recovered as a basal titanosauroid. This taxon shares one potentially unusual feature with the putative titanosaur Borealosaurus from the Late Cretaceous of China (You et al., 2004): both species possess opisthocoelous middle-posterior caudal centra. Future studies incorporating both taxa should assess whether this feature is a synapomorphy or convergently acquired, with the latter presumably the case in the non-neosauropod eusauropod Turiasaurus from the latest Jurassic–earliest Cretaceous of Spain (Royo-Torres et al., 2006).

Malarguesaurus, from the early Late Cretaceous of Argentina (González Riga et al., 2009) was placed in a basal position within Titanosauria by Carballido et al. (2011a), whereas González Riga et al. (2009) recovered Malarguesaurus as a non-titanosaurian somphospondylan. D'Emic (2012) listed it as a macronarian. Here it was recovered as a non-titanosaurian somphospondylan of uncertain affinities in our LSDM, but as a titanosauroid in our other analyses.

All remaining taxa were recovered within Titanosauria in all analyses, and we do not discuss in detail the positions of the ‘higher titanosaurs’ (i.e. Alamosaurus, Malawisaurus, Opisthocoelicaudia, Rapetosaurus, and Saltasaurus) because they are not the focus of this study. Xianshanosaurus is known from the early Late Cretaceous of China and was considered a neosauropod of uncertain affinities by Lü et al. (2009a). Mannion & Calvo (2011) suggested possible titanosaurian affinities, and a lithostrotian identification was postulated by D'Emic (2012). Xianshanosaurus + Daxiatitan (see above) is the sister clade to Lithostrotia in our LSDM, and the former taxon forms a trichotomy with Diamantinasaurus (see above) and Lithostrotia in our LSDM_iw and LCDM. In our LSDM trees, Lithostrotia is composed of two clades: Saltasauridae, and a clade of Malawisaurus + (Mongolosaurus + Rapetosaurus). However, Mongolosaurus + Rapetosaurus was recovered outside of Lithostrotia in our LSDM_iw and LCDM analyses. Mongolosaurus, from the Aptian–Albian of China, has been considered as a diplodocid, euhelopodid, and titanosaur in the past (see Barrett et al., 2002; Mannion, 2011; D'Emic, 2012), although its only prior inclusion in a phylogenetic analysis demonstrated probable titanosaurian affinities, with a possible derived position (Mannion,2011; see also Wilson, 2005a), as supported here.

In all our analyses, Saltasaurus + (Alamosaurus + Jiangshanosaurus) formed a clade. Jiangshanosaurus, from the Albian of China, has previously been regarded as a titanosaur (Tang et al., 2001b; Upchurch et al., 2004a; Wilson, 2005a; Wilson & Upchurch, 2009), and D'Emic (2012) suggested saltasaurid affinities, but this is the first time that this has been tested and received support from phylogenetic analysis. Its position as the sister taxon to Alamosaurus has interesting biogeographical implications (see ‘East Asian Cretaceous sauropods’ below). The sister clade to this grouping includes Opisthocoelicaudia and Dongyangosaurus in all three analyses, withBaotianmansaurus (see above) and Liubangosaurus (see above) comprising the third member of this clade in our LSDM and LSDM_iw/LCDM analyses, respectively. The early Late Cretaceous Chinese taxon Dongyangosaurus was regarded as a titanosaur by Lü et al. (2008), and euhelopodid by D'Emic (2012), but this is its first inclusion in a phylogenetic analysis and the first time that a saltasaurid identification has been proposed.

A number of taxa included in our analysis have previously been identified as titanosauriforms or macronarians, but are here recovered in more basal positions, either as non-titanosauriforms or placed outside Neosauropoda in all analyses. We discuss each of these taxa below.

HMN MB.R.2091.1−30 is a series of 30 articulated caudal vertebrae from the Tithonian Tendaguru Formation of Tanzania that was originally referred to Janenschia (see below) and considered a titanosaur by Janensch (1929), on the basis of the strong procoely of the first ten caudal vertebrae (see also Jacobs et al., 1993; Salgado & Calvo, 1997). Bonaparte et al. (2000) questioned its titanosaurian affinities (see also McIntosh, 1990) as a result of the wider distribution of caudal procoely in sauropods (see ‘Implications for phylogenetic characters’ below), and removed it from Janenschia based on the lack of overlapping elements. Royo-Torres & Cobos (2009) also suggested that it might represent a non-neosauropod turiasaur. Here, we recovered it as the sister taxon to Mamenchisaurus in all analyses (see Table 6 for the phylogenetic definition of Mamenchisauridae). Previous authors have found evidence for clustering of some Middle–Late Jurassic North African and European sauropod taxa with Chinese forms [Atlasaurus and Jobaria with Bellusaurus (Upchurchet al., 2004a); Cetiosauriscus and Losillasaurus with Mamenchisaurus and Omeisaurus (Rauhut et al., 2005; Harris, 2006; Nair & Salisbury, 2012); see also discussion of similarities between the Middle Jurassic North African genus Spinophorosaurus and Asian mamenchisaurids in Remes et al., 2009], although this is the first time that a taxon from the southern part of Africa has been linked with these Asian species. Most evidence indicates that East Asia was separated from the rest of Pangaea during the Late Jurassic (see Barrettet al., 2002; Upchurch, Hunn & Norman, 2002; and Wilson & Upchurch, 2009 for reviews and ‘East Asian Cretaceous sauropods’ below), which might indicate an earlier and more widespread diversification of mamenchisaurids prior to this geographical isolation. One caveat is that the focus of character and taxon sampling in this study was on basal titanosauriforms; thus, to test fully the position of HMN MB.R.2091.1−30, consideration of other non-neosauropods is required [e.g. Jobaria (Sereno et al., 1999), Spinophorosaurus (Remes et al., 2009), and turiasaurs (Royo-Torres et al., 2006)]. However, we note that at least one feature [an approximate 200% anteroposterior length increase of middle caudal neural spines from their base to their summit (C199)] appears to be uniquely shared by HMN MB.R.2091.1−30 and the mamenchisaurids Mamenchisaurus [Mamenchisaurus hochuanensis (Young & Zhao, 1972: pl. 3; IVPP type: P. Upchurch, pers. observ., 2010) and Mamenchisaurus youngi (Ouyang & Ye, 2002: pl. I5)] and Chuanjiesaurus (Sekiya, 2011: figs 32–35) amongst Sauropoda.

Lapparentosaurus madagascariensis, from the Bathonian (Middle Jurassic) of Madagascar, has a complex taxonomic history (Mannion,2010). Originally described as ‘Bothriospondylus madagascariensis’ by Lydekker (1895), with additional material referred by Thevenin (1907) and Ogier (1975), Bonaparte (1986) created the new combination Lapparentosaurus madagascariensis to include all of this material. However, in the same paper, Bonaparte (1986) also erected a new genus and species – Lapparentosaurus madagascariensis – with the material described by Ogier (1975) elected as the holotype. As such, the latter name is a homonym. Mannion (2010) demonstrated that the type material of ‘Bothriospondylus madagascariensis’ represents a nondiagnostic non-neosauropod eusauropod, and thus Bonaparte's (1986) first usage of Lapparentosaurus madagascariensis (as a new combination) is a nomen dubium. Consequently, the name Lapparentosaurus is not available for the Ogier (1975) material and, following International Commission on Zoological Nomenclature (ICZN) rules, this taxon requires a new generic name. However, the name Lapparentosaurus has been associated solely with the Ogier (1975) material in subsequent taxonomic works (e.g. McIntosh, 1990; Upchurch, 1995, 1998; Upchurch et al., 2004a; Läng & Goussard, 2007; Mannion, 2010; this work), including the only previous inclusions of this taxon in cladistic analyses (Upchurch, 1998; Upchurch et al., 2004a). As such, we recommend submitting a case to the ICZN for the conservation of the nameLapparentosaurus madagascariensis in its current usage to assist nomenclatural stability (see the recent example of Cetiosaurus; Upchurch & Martin, 2003; Upchurch, Martin & Taylor, 2009), and we follow this proposal in the following discussion. Bonaparte (1986) placed Lapparentosaurus as a basal member of the Cetiosauridae, a family of non-neosauropod eusauropods, whereas McIntosh (1990) and Upchurch (1995) both suggested brachiosaurid affinities. In the first phylogenetic analysis to include Lapparentosaurus, Upchurch (1998) found support for a brachiosaurid position, although in an updated analysis its position was modified to Titanosauriformesincertae sedis (Upchurch et al., 2004a). Here, we found no evidence to support a placement within Titanosauriformes forLapparentosaurus; instead, it was recovered as a non-neosauropod eusauropod in all analyses. Upchurch (1995) noted that the lack of lateral pneumatic fossae in the cervical and dorsal vertebrae might support the basal position proposed by Bonaparte (1986); however, the cervical centra do possess shallow lateral fossae (e.g. MNHN MAA 5, 13, 53, 90, and 172: P. D. Mannion, pers. observ., 2011), as is also the case in the anterior-most dorsal vertebra preserved (MNHN MAA 201: P. D. Mannion, pers. observ., 2011). One note of caution is that Lapparentosaurus is known only from juvenile material (Ogier, 1975), and Upchurch (1995) mentioned the possibility that some of these plesiomorphic features might instead reflect ontogeny. However, a partial skeleton of an adult (MNHN MAJ 289) found at a locality close to that of Lapparentosaurus, also displays a number of plesiomorphic features suggesting a non-titanosauriform position (Läng & Goussard, 2007). As with the juvenile material, the anterior-most dorsal vertebrae of this adult specimen possess a lateral pneumatic fossa, with this feature absent from more posterior dorsal centra (MNHN MAJ 289: P. D. Mannion, pers. observ., 2011). This shared characteristic, coupled with a lack of clear anatomical differences, provides tentative evidence that MNHN MAJ 289 might be referable toLapparentosaurus (although see Läng & Goussard, 2007), and suggests that this is probably not an ontogenetic feature in the juvenile material. In conclusion, a non-neosauropod position for Lapparentosaurus appears to be well supported, but its exact placement will have to await future analyses focused on the inter-relationships of basal eusauropods.

Atlasaurus, from the Bathonian–Callovian (Middle Jurassic) of Morocco, was initially described as ‘brachiosaurid-like’ (Monbaron, Russell & Taquet, 1999); if this identification is correct, this would represent the oldest known member of Neosauropoda. However, the taxon was recovered as a non-titanosauriform macronarian by Upchurch et al. (2004a), and as a non-neosauropod eusauropod in all other analyses (Table 7), including all of our trees. Although its brachiosaurid affinities seem unlikely (see also D'Emic, 2012), Atlasaurus has received only a very brief description (Monbaron et al., 1999), and a revision of this taxon will be vital in determining its precise phylogenetic affinities.

Tendaguria, from the Tithonian-aged Tendaguru Formation of Tanzania, was regarded as Sauropoda incertae sedis by Bonaparte et al. (2000) and Upchurch et al. (2004a), with the former study erecting the monospecific family Tendaguriidae. McIntosh (1990) merely noted that the material comprising Tendaguria was unique (at the time the two holotypic dorsal vertebrae were assigned to Janenschia), and Mannion & Calvo (2011) suggested possible titanosaurian affinities. The only previous analysis to include Tendaguria recovered it as either a non-neosauropod or as a non-titanosauriform macronarian (Carballido et al., 2011b). All of our analyses placed Tendaguriaoutside of Macronaria: in our LSDM and LCDM it was recovered as the sister taxon to Neosauropoda, whereas it was placed as a basal diplodocoid in our LSDM_iw analysis. Clearly, Tendaguria appears to be positioned close to the base of Neosauropoda: revision of the material, and its inclusion in a phylogenetic analysis focused on basal eusauropod relationships, should help to determine its exact placement.

Janenschia has a complex taxonomic history (Bonaparte et al., 2000), but has generally been regarded as a basal titanosaur by most workers, even after the exclusion of the referred tail sequence (HMN MB.R.2091.1−30 – see above) from the genus (e.g. Janensch, 1929; Upchurch, 1995; Wilson & Sereno, 1998; Upchurch et al., 2004a; Curry Rogers, 2005; Mannion & Calvo, 2011). Consequently, Janenschiahas been considered to represent the earliest body fossil evidence for Titanosauria (Upchurch, 1995; Upchurch et al., 2004a). However, Bonaparte et al. (2000) disagreed with a titanosaurian position for Janenschia, and instead suggested close affinities with the basal macronarian Camarasaurus. Royo-Torres & Cobos (2009) also suggested that at least some of the material included in Janenschiamight be referable to the non-neosauropod group Turiasauria, although this is yet to be tested through phylogenetic analysis. Carballidoet al. (2011b) recovered Janenschia in a non-titanosauriform macronarian position, and D'Emic (2012) noted a lack of evidence supporting somphospondylan affinities. Here, we support the placement in the analysis of Carballido et al. (2011b), with Janenschia positioned at or close to the base of Macronaria, forming a sister-taxon relationship with Tehuelchesaurus in our LSDM and LSDM_iw. Tehuelchesaurus, from the Late Jurassic of Argentina (Rich et al., 1999; Rauhut et al., 2005), was initially regarded as a cetiosaurid by Rich et al. (1999), and a non-neosauropod eusauropod position was supported by the analysis of Upchurch et al. (2004a). However, Rauhut et al. (2005) recovered it in a polytomy with other macronarians, and the analyses of Carballido et al. (2011a, b) and D'Emic (2012) agree with a nontitanosauriform macronarian placement.

Aragosaurus, from the Early Cretaceous of Spain (Sanz et al., 1987), was placed as a non-titanosauriform macronarian in all our analyses. Sanz et al. (1987) regarded it as a ‘camarasaurid’, and Upchurch et al. (2004a) referred it to Eusauropoda incertae sedis, whereas Barco, Canudo & Cuenca-Bescós (2006), Mannion & Calvo (2011), Canudo et al. (2012), and D'Emic (2012) all suggested that it represented a basal titanosauriform. However, its only previous inclusion in a phylogenetic analysis also placed it as a non-titanosauriform macronarian, albeit within ‘Laurasiaformes’ (Royo-Torres, 2009; see above).

In our LSDM analysis, the latest Jurassic–earliest Cretaceous taxon Galveosaurus is recovered as the sister taxon to Titanosauriformes, and as the sister taxon to Dongbeititan + (Lusotitan + Titanosauriformes) in our LCDM. Galveosaurus was originally described as a cetiosaurid (Sánchez-Hernández, 2005), and its position as a non-neosauropod eusauropod (in a turiasaur clade) was supported by Royo-Torres et al. (2006) and Royo-Torres & Upchurch (2012). D'Emic (2012) suggested a basal titanosauriform placement, whereas Barco et al. (2006) and Carballido et al. (2011a, b) recovered it in the same position as the current study.

Quantitative characters comprise 27% of our data matrix and thus clearly play an important role in the elucidation of titanosauriform relationships. Our use of discretized and continuous characters for quantitative morphological data produced notable differences to tree topology in our two parallel sets of analyses (LSDM and LCDM). The positions of most taxa are only subtly affected, with the composition of Titanosauria the most disparate between the two analyses. In particular, Titanosauria is considerably more diverse in our LCDM than LSDM trees, with the latter largely resembling previous analyses, whereas numerous taxa almost universally regarded as basal somphospondylans or basal titanosauriforms (see Table 7 for a summary of previous assignments) have been included within Titanosauria in our LCDM. The latter might be largely caused by pushing Andesaurus, a clade specifier for Titanosauria, lower in the LCDM trees. Clearly, numerous placements of taxa in at least one of our agreement subtrees are incorrect.

By comparing the LSDM with the LSDM_iw and LCDM, it is clear that the use of implied weights has a greater effect on the data set and resultant trees than the treatment of quantitative characters as continuous data. Although implied weighting, by itself (i.e. the LSDM_iw), produced MPTs that still lie in a set of trees that are not statistically significantly distinguishable from those produced by the LSDM, the treatment of continuous data as such pushes the resulting MPTs into the significant difference range (see ‘Templeton's tests’ above). The majority of the differences in the topology of the LCDM, in comparison to that of the LSDM, were also recovered in the LSDM_iw. As such, only a few alternative placements between the trees can be explained by the application of continuous rather than discrete data. These are: (1) the placement of Dongbeititan and Lusotitan outside of Titanosauriformes in the LCDM, rather than within Somphospondyli and Brachiosauridae, respectively, in the LSDM; (2) the basal position of Venenosaurus within Brachiosauridae in the LCDM, in contrast to its derived placement in the LSDM; and (3) the placement of the ‘French Bothriospondylus’ as a derived brachiosaurid in the LCDM, rather than a basal member of this clade as in the LSDM. All other differences in topology seem to result from the effect of character weighting, suggesting that the continuous data had only a relatively small impact on tree topology.

Numerous character states used in this analysis have previously been regarded as well-supported synapomorphies of certain titanosauriform clades, but have been shown here to be highly homoplastic and/or to characterize more inclusive clades. Here we discuss a small number of exemplar characters and detail their more complex distributions. Lists of synapomorphies for the main macronarian clades recovered from the LSDM and LCDM are provided in Appendices 4 and 5, respectively.

A lateral pneumatic foramen (‘pleurocoel’) situated within a fossa in dorsal vertebrae (C145) has been considered a characteristic of titanosaurs by previous authors (Bonaparte & Coria, 1993; Upchurch et al., 2004a). However, it is present in a number of non-titanosaurian somphospondylans based on the LSDM analysis, as well as being convergently present in the brachiosaurid Cedarosaurus (Tidwell et al.,1999: fig. 2; DMNH 39045: P. D. Mannion, pers. observ., 2008).

Procoely in anterior caudal vertebrae (C27) has traditionally been regarded as a derived titanosaur feature, with this convergently present in Mamenchisaurus and flagellicaudatans (McIntosh, 1990; Jacobs et al., 1993; Upchurch, 1995, 1998; Salgado et al., 1997; Wilson, 2002). However, mild procoely is present in the anterior caudal vertebrae of the basal titanosaur Andesaurus (Mannion & Calvo, 2011), as well as the somphospondylan Tastavinsaurus (Canudo et al., 2008), and the non-titanosaurian Dongbeititan possesses strongly procoelous anterior caudal vertebrae (Wang et al., 2007). Furthermore, a number of probable non-neosauropods also display strongly procoelous anterior caudal vertebrae [i.e. Bellusaurus (Dong, 1990), Chuanjiesaurus (Sekiya, 2011), HMN MB.R.2091.1−30 (Bonaparte et al., 2000; this study), and Losillasaurus (Casanovas, Santafe & Sanz, 2001)]. Similarly, the extension of procoely into middle-posterior caudal vertebrae (C184) has been considered a feature of lithostrotian titanosaurs (Jacobs et al., 1993; Upchurch, 1995, 1998; Salgado et al.,1997; Wilson, 2002). Although its uninterrupted presence still remains a feature of derived titanosaurs (C177), the possession of some procoelous middle caudal vertebrae is much more widespread amongst titanosauriforms. Some middle caudal vertebrae referred to the Late Jurassic brachiosaurid Giraffatitan (Janensch, 1950: pl. 3, fig. 22; HMN MBR specimens: P. D. Mannion, pers. observ., 2011) display the procoelous condition (D'Emic, 2012, 2013), and a number of taxa of uncertain taxonomic affinities within Somphospondyli (this study) also display this morphology [e.g. Brontomerus (Taylor et al., 2011), Gobititan (You et al., 2003), Malarguesaurus (González Riga et al.,2009), Astrophocaudia (D'Emic, 2013), and Tastavinsaurus (Royo-Torres, 2009)].

Salgado et al. (1997) noted that titanosaurian pubes are markedly longer than their ischia, whereas in other sauropods these two elements are subequal. However, a number of taxa recovered as non-titanosaurian somphospondylans in our LSDM trees have very low (< 0.8) ischium to pubis length ratios (C60) [e.g. Qiaowanlong (You & Li, 2009) and Tastavinsaurus (Canudo et al., 2008)], suggesting that this feature characterizes a more inclusive clade.

The presence of a lateral bulge and medially deflected proximal femur (C255) has long been regarded as a synapomorphy of Titanosauriformes (McIntosh, 1990; Calvo & Salgado, 1995; Salgado et al., 1997); however, this feature has previously been poorly defined and in its current state categorizes most eusauropods to some extent (see Salgado et al., 1997: fig. 10). Here we have modified the character so that it describes whether the proximolateral margin, above the lateral bulge, is level with/lateral (plesiomorphic state), or medial (derived state) to the lateral margin of the distal half of the shaft, based on Royo-Torres (2009) and Royo-Torres et al. (2012). As such, the derived feature characterizes macronarians, other than Camarasaurus, and is lost in some derived titanosaurs (e.g.Opisthocoelicaudia, Rapetosaurus, Saltasaurus, and Xianshanosaurus), as well as the more basal somphospondylan Tastavinsaurus.

The presence of camellae in presacral vertebrae has generally been regarded as either a titanosauriform synapomorphy, or convergently acquired by brachiosaurids and derived titanosaurs (Wedel, 2003), along with the Chinese non-neosauropod Mamenchisaurus (Young & Zhao, 1972; Russell & Zheng, 1993; Ouyang & Ye, 2002). However, presacral camellate vertebrae are now recognized as characterizing most somphospondylans, rather than just derived titanosaurs (e.g. Carballido et al., 2011b; D'Emic, 2012). The non-titanosauriform macronarian Galveosaurus also possesses camellate cervical (Barco et al., 2006; Barco, 2009) and anterior-most dorsal vertebrae (C115) [note that the ‘posterior cervical’ in Barco et al. (2006: fig. 2) is much more likely to represent an anterior dorsal vertebra]. Along with the absence of camellae in more posterior dorsal vertebrae and thoracic ribs of Galveosaurus (Barco et al., 2006; Barco, 2009), this suggests that a camellate internal tissue structure within the anterior portion of the presacral series is synapomorphic of Galveosaurus + Titanosauriformes (or a slightly more inclusive clade), with its extension into the full dorsal series (C141) [including thoracic ribs (C170)] characterizing Titanosauriformes (see also discussion in Carballido et al., 2011b). Moreover, this interpretation is supported by the position of Tehuelchesaurus as a non-titanosauriform macronarian (see also Carballido et al., 2011b; D'Emic, 2012), because its dorsal vertebrae lack camellae, although no cervical vertebrae are preserved (Carballido et al., 2011b). Interestingly, a similar ‘front-to-back’ pattern is also observed in the evolution of theropod dinosaur pneumaticity (Benson et al., 2012).

Extension of camellae into the sacrum (C172) is restricted to Somphospondyli and its continuation into the ilium (C249) characterizes a more exclusive grouping within this clade. Caudal pneumaticity, in the form of camellae (C176), has generally been thought to be restricted to a small clade of latest Cretaceous South American titanosaurs, the Saltasaurinae (Wilson, 2002; Powell, 2003; Wedel, 2003). However, this internal tissue structure is also present in the caudal neural arches of the latest Cretaceous North American saltasauridAlamosaurus (Fronimos, 2011), the early Late Cretaceous Chinese derived titanosaur Xianshanosaurus (Henan Geological Museum specimen KLR-07–62-6: P. D. Mannion & P. Upchurch, pers. observ., 2012), and the Aptian-aged African basal lithostrotian Malawisaurus, based on referred (Wedel, 2009) and holotypic material (SAM 7405: P. D. Mannion, pers. observ., 2008). It is also present in the probable Aptian-aged Chinese specimen PMU R263 (Wiman, 1929; Whitlock et al., 2011), which was described as a diplodocid by Upchurch & Mannion (2009), but reinterpreted as a titanosauriform by Whitlock et al. (2011). As our data sets do not include many ‘traditional’ derived titanosaurs, it is difficult to determine whether the possession of camellate caudal vertebrae truly optimizes as a lithostrotian (or slightly more inclusive titanosaurian clade) character, or if this feature is convergently acquired by these taxa and Saltasaurinae; nevertheless, it clearly has a more widespread distribution amongst Titanosauria than previously realised. Currently, saltasaurines represent the only known taxa in which camellae invade the distal portions of the tail, as well as the pectoral girdle (Cerda, Salgado & Powell, 2012b).

The presence of titanosaurs in the Jurassic was originally founded upon a titanosaurian interpretation of remains attributed to the Tithonian-aged Tendaguru genus Janenschia (e.g. Janensch, 1929; Jacobs et al., 1993; Upchurch, 1995; Salgado & Calvo, 1997; Wilson & Sereno, 1998). Day et al. (2002, 2004) presented evidence for earlier occurrences of titanosaurs in the form of Middle Jurassic ‘wide-gauge’ trackways. However, Janenschia and all previously referred remains, have been shown to lie outside of Titanosauria by a number of authors (Bonaparte et al., 2000; Carballido et al., 2011b; this study), and a number of studies have expressed doubt as to whether wide-gauge trackways are produced solely by titanosaurs (e.g. Henderson, 2006; Santos, Moratalla & Royo-Torres, 2009). As such, here we review additional specimens and discuss further lines of evidence for the presence of Jurassic titanosaurs.

A poorly preserved and incomplete middle caudal centrum was originally misidentified as a left metatarsal V by Lapparent & Zbyszewski (1957: 38 and pl. 25, fig. 89) (Fig. 30). These authors referred this isolated element from the Late Jurassic of Castanheira, Portugal, toLourinhasaurus (‘Apatosaurus’) alenquerensis, despite the lack of overlapping material with that taxon. The element (MG 4799) preserves only the right half of a strongly procoelous centrum (figured in lateral view in Lapparent & Zbyszewski, 1957: pl. 25, fig. 89). Its lateral surface is dorsoventrally convex and anteroposteriorly concave, and the condylar convexity merges smoothly into the lateral surface of the main body of the centrum, lacking a distinct rim (Fig. 30). In lateral view the dorsal and ventral margins are concave, and the internal tissue structure appears to be solid. The anterior end is dorsoventrally taller than the posterior end. It is clear that the neural arch was restricted to the anterior half of the centrum (excluding the condylar ball). Although it is not possible to accurately determine the elongation index (EI) value, it is less than 1.5. Procoely in middle caudal centra has traditionally been regarded as a feature restricted to derived titanosaurs (Jacobs et al., 1993; Upchurch, 1995, 1998; Wilson, 2002), i.e. lithostrotians (Upchurch et al., 2004a). As such, this caudal vertebra would represent the earliest occurrence of a lithostrotian titanosaur, extending this lineage from the earliest known example in the Barremian (Early Cretaceous) of the UK [‘Titanosaurus (Iuticosaurus) valdensis’: Wilson & Upchurch, 2003; Mannion et al., 2011b] back into the Kimmeridgian–Tithonian (Late Jurassic). However, as noted above, procoelous middle caudal centra are now known to be more widespread amongst Titanosauriformes, casting doubt on the lithostrotian, or even titanosaurian, affinities of MG 4799 and ‘Titanosaurus valdensis’. This leaves three possible conclusions regarding the taxonomic affinities of MG 4799 (and ‘Titanosaurus valdensis’): (1) it represents an early lithostrotian titanosaur; (2) it represents an individual of the temporally contemporaneous Giraffatitan brancai (or, less likely, one of the Cretaceous taxa mentioned above with procoelous middle caudal centra); or (3) it represents an indeterminate titanosauriform. Based on our current understanding of titanosauriforms, as well as subtle morphological differences with Giraffatitan, we consider it preferable to follow the third option.

Although Janenschia, Tendaguria, and HMN MB.R.2091.1−30 are all excluded from Titanosauria in the current study (see also Carballidoet al., 2011b), our recovery of Australodocus as a basal titanosaur maintains the presence of pre-Cretaceous body fossil evidence for this clade of sauropods. This pre-Cretaceous origin is also partially supported by trackways from the middle Bathonian (late Middle Jurassic) Ardley tracksite in England (Day et al., 2002, 2004). At this site, ‘narrow’ and ‘wide’ gauge sauropod trackways are found on the same bedding plane, with the former preserving manual claw impressions, whereas these are absent in the wide-gauge tracks. Wide-gauge tracks have been attributed to titanosaurs based on a number of anatomical specializations (Wilson & Carrano, 1999; Mannion & Upchurch, 2010a), although there is some evidence to suggest that other sauropods might also have produced such tracks (Henderson,2006; Santos et al., 2009). Although it remains possible that non-titanosaurs, including some basal titanosauriforms (Wilson & Carrano,1999; Mannion & Upchurch, 2010a), might also have produced wide-gauge trackways, only derived titanosaurs (based on this analysis) entirely lack manual claws (Salgado et al., 1997; Wilson & Sereno, 1998; Day et al., 2002). As such, the absence of claw impressions in the wide-gauge tracks at the Ardley tracksite, when combined with their presence in the spatiotemporally contemporaneous narrow-gauge tracks, suggests that this absence is unlikely to merely represent a taphonomic artefact, and instead supports the view that these wide-gauge tracks were produced by titanosaurs (Day et al., 2002). One final note of caution reflects the dearth of preserved titanosauriform manüs: only five members of this clade (in our analysis) preserve the relevant manual material. Thus, currently, we do not know how reduced the manual claws were in most titanosauriform taxa, although it should be noted that the claw of the somphospondylanDiamantinasaurus is fairly prominent (Hocknull et al., 2009).

A number of putative basal titanosauriform taxa were not included in our analyses as a result of their fragmentary nature. In light of our revised understanding of titanosauriforms, we discuss the taxonomic affinities of these body fossils and ichnofossils (in approximate stratigraphical order), including indeterminate material that would potentially extend stratigraphical or geographical ranges of clades if correctly identified.

Early–Middle Jurassic

Late Jurassic

Early Cretaceous

Late Cretaceous

We created phylogenetic diversity estimates (PDEs) for titanosauriforms from our LSDM agreement subtree and LCDM strict consensus tree using time-calibrated phylogenies (Figs 22, 27), with ghost ranges or lineages reflecting gaps in the fossil record (see Mannion et al.,2011b, and references therein). In general the two PDEs match one another closely, although diversity is always higher for the LCDM PDE as a result of no taxa being pruned from this tree. Both PDEs indicate that late Oxfordian titanosauriform diversity is notably higher than any assessment based only on observed diversity (i.e. a single taxon, the ‘French Bothriospondylus’). This suggests that the Oxfordian was not as depauperate a time interval as indicated by previous analyses (e.g. Upchurch & Barrett, 2005; Mannion et al., 2011b), suggesting that this stage is especially poorly sampled and/or that some Late Jurassic outcrops might be inaccurately dated (Mannion et al., 2011b). The stratigraphical age and phylogenetic placement of the ‘French Bothriospondylus’ results in the extension of numerous brachiosaurid lineages back into the late Oxfordian, as well as the somphospondylan lineage. Currently there are no body fossil occurrences that suggest an earlier origin for titanosauriforms, but the trackway record suggests that the clade had already originated and diversified in the late Middle Jurassic [the middle Bathonian Ardley tracksite (Day et al., 2002); see above]. Unambiguous examples of titanosauriforms are known from the Late Jurassic of Africa, Europe, North America, and probably South America and Pakistan (see above). Their absence in Jurassic rocks in Antarctica and Australasia most likely reflects a sampling bias, although their Asian absence might be a genuine feature (see below). Thus, this near-global pattern, coupled with putative Middle Jurassic titanosauriform occurrences from Europe, suggests that Titanosauriformes most likely originated no later than the Middle Jurassic, i.e. prior to the initiation of Pangaean fragmentation (Day et al.,2002; Upchurch et al., 2002).

Whereas the LSDM PDE maintains an approximately constant diversity throughout the Late Jurassic (Fig. 22), there is a Tithonian peak in the LCDM PDE (Fig. 27). This is mainly a result of the placement of Australodocus as a titanosaur, which extends three additional titanosaur lineages back into the Tithonian in the LCDM PDE (note that Australodocus was pruned a posteriori from the LSDM agreement subtree, although was recovered as a member of Titanosauria). The near absence of titanosaurs from the Jurassic body fossil record most likely indicates that they were a rare component of Middle–Late Jurassic terrestrial faunas, but it is also possible that they displayed a different environmental preference (inland terrestrial) to other sauropods (Mannion & Upchurch, 2010a; see also Falkingham, Bates & Mannion, 2012). The newly recovered position of the coastal taxon Janenschia outside of Titanosauria strengthens the results of Mannion & Upchurch's (2010a) environmental analyses, although other modifications in terms of the composition of Titanosauria might result in additional changes. It is interesting to observe that rebbachisaurid diplodocoids also show a comparable and contemporaneous ghost lineage (Upchurch & Barrett, 2005; Sereno et al., 2007) to that of titanosaurs, and Mannion & Upchurch (2010a) noted that rebbachisaurids display some evidence for the same inland terrestrial environmental association as that recovered for titanosaurs. Thus, it is plausible that some combination of low diversity and abundance, coupled with a rarity in preservation of suitable environments, might account for the paucity of titanosaurs and absence of rebbachisaurids in the Jurassic, and their consequent parallel ghost lineages.

There is no decline in titanosauriform diversity across the J/K boundary (Figs 22, 27), which supports the view that the sauropod ‘diversity crash’ at this boundary was mainly restricted to diplodocids and non-neosauropods (Upchurch & Barrett, 2005; Barrett, McGowan & Page,2009; Mannion et al., 2011b; Upchurch et al., 2011a; Upchurch & Mannion, 2012). In fact, diversity is roughly constant (LCDM) or even increased (LSDM) across the boundary, although the latter is largely an artefact of the exclusion of Australodocus, whose omission prevents the extension of several somphospondylan lineages into the Tithonian, thereby creating a time lag for this initial diversity peak. Although overall titanosauriform diversity does not decrease over the J/K boundary, brachiosaurid diversity outside of North America appears to have declined sharply or even reduced to zero in the Early Cretaceous. However, this might merely reflect the absence of suitably aged terrestrial outcrops for much of this interval, a view that is supported by putative brachiosaurid occurrences in the Early Cretaceous of Lebanon (Buffetaut et al., 2006). The absence of brachiosaurids in the Cretaceous of East Asia, once reconnected to Europe (see below), provides weak support that the clade did genuinely go extinct in Europe and thus never had the opportunity to disperse into East Asia (Ksepka & Norell, 2010), but it is also possible that ecological barriers prevented their dispersal.

Ghost ranges indicate the existence of several unsampled somphospondylan lineages in the earliest Cretaceous (Berriasian–Hauterivian); however, ‘Pelorosaurus’ becklesii is the only body fossil evidence from this interval, again indicative of the lack of terrestrial outcrops at this time. A prominent increase in titanosauriform diversity can be detected in the Barremian, reflecting the radiation of somphospondylans and marking the earliest known putative lithostrotian (NHMUK R5333; see also D'Emic, 2012). The Aptian–Albian represents a peak in titanosauriform diversity (Figs 22, 27), with derived titanosaurs (lithostrotians, or taxa closely related to this radiation) known from north-western (O'Leary et al., 2004) and south-eastern Africa (Malawisaurus: Jacobs et al., 1993; Gomani, 2005), northern (Mongolosaurus: Mannion, 2011) and south-eastern Asia (Jiangshanosaurus: Tang et al., 2001b), and South America (Tapuisaurus: Zaheret al., 2011). This wide distribution of derived forms suggests an earlier origin and dispersal for titanosaurs, and supports the view that poor sampling of the earliest Cretaceous (and possible environmental biases in the Late Jurassic) obscures their early evolution (see also D'Emic, 2012).

Titanosauriform diversity suffered a severe drop in species numbers at, or around, the Albian/Cenomanian boundary in the LCDM PDE (40% decrease; Fig. 27), although this is less pronounced in the LSDM PDE (23% decrease; Fig. 22). However, in both cases, this is followed by a further gradual decline, representing a faunal turnover whereby basal titanosauriforms (comprising brachiosaurids, basal somphospondylans, and andesauroids) were replaced by derived ‘titanosauroids’ (N.B. the low diversity throughout the remainder of the Late Cretaceous is an artefact generated by the exclusion of most ‘definite’ titanosaurs from our study). This transition to a sauropod fauna dominated by ‘titanosauroids’ (especially lithostrotians) is consistent with previous work that has noted a middle Cretaceous faunal turnover in dinosaurs (e.g. Coria & Salgado, 2005), squamates (e.g. Evans, 2003), crocodylomorphs (e.g. Martin & Delfino, 2010), and other groups, and it seems that this transition occurred in a spatiotemporally staggered fashion (Benson et al., 2013). Brachiosaurids appear to have become extinct at, or close to this boundary (see the stratigraphical range uncertainty of the putative brachiosauridSonorasaurus: Table 4), and there are no unambiguous occurrences of this clade in post-Cenomanian outcrops. A few late surviving basal titanosauriforms are known from the early Late Cretaceous of Asia (e.g. ‘Huanghetitan’ ruyangensis and Ruyangosaurus), South America (Malarguesaurus), and Africa (Angolatitan). Post-Cenomanian, sauropods are absent from Europe and North America until the latest Cretaceous [with the exception of a late Turonian–early Coniacian Croatian tracksite (Mezga et al., 2006)], although whether this ‘hiatus’ reflects regional extinctions, poor sampling, or a combination of the two remains controversial (Lucas & Hunt, 1989; Le Loeuff,1993; Williamson & Weil, 2008; D'Emic, Wilson & Thompson, 2010; Mannion & Upchurch, 2011; D'Emic & Foreman, 2012; see below).

Although a poor fossil record might partly account for low somphospondylan diversity prior to the ‘middle’ Cretaceous, in comparison to the relatively high, early burst in Late Jurassic diversity of its sister clade Brachiosauridae, there is a clearer imbalance in tree shape when comparing the overall topology of Titanosauriformes: whereas the total number of known brachiosaurid species is fewer than ten, the Somphospondyli clade diverged into approximately 80 known species (including derived titanosaurs not included in this study). This asymmetry in tree shape and in the timing of diversification indicates variation in speciation and/or extinction rates in the two titanosauriform clades (see Mooers & Heard, 1997 for a discussion of tree shape). Although the causes of this difference are not yet understood and it is beyond the scope of this study to fully explore this issue, ecological differences and geographical factors are usually proposed as the most likely explanations (Schluter, 2000; Pigot et al., 2010). Pigot et al. (2010) showed through simulations that diversification rates slowed under scenarios of stable geographical ranges, and increased under conditions of faster rates of range expansion. It is therefore possible that the eventual global distribution of somphospondylans accounts for their disproportionately higher diversity, with brachiosaurids apparently never reaching Asia (see also Ksepka & Norell, 2010; Mannion, 2011), Australasia (this study), or (probably) South America (this study), and unambiguous Cretaceous occurrences restricted to North America (D'Emic, 2012, 2013; this study).

The Cretaceous of East Asia has played an important role in discussions of sauropod evolutionary history and biogeography. Prior to the Cretaceous, there is currently no unambiguous evidence for the presence of neosauropods in East Asia (Upchurch et al., 2004a; Wilson,2005a; this study), whereas a diverse array of Late Jurassic representatives of this clade is known from all other continents preserving appropriately aged terrestrial sediments (i.e. Africa, Europe, and the Americas; Upchurch et al., 2004a; Weishampel et al., 2004; Mannion & Calvo, 2011; Whitlock, 2011b; D'Emic, 2012; Mannion et al., 2012). Although it is possible that their absence might be a sampling bias (Barrett et al., 2002; Wilson & Upchurch, 2009; for example, they might have been low in diversity and abundance, or lived in different, unpreserved/undersampled environments), Middle–Late Jurassic East Asian non-neosauropods are extremely diverse and abundant (Upchurch et al., 2004a), leading to the view that the absence of pre-Cretaceous neosauropods might be a genuine pattern, resulting from the isolation of East Asia from the rest of Pangaea during the Middle Jurassic (Russell, 1993; Upchurch, 1995; see discussions in Barrettet al., 2002, Upchurch et al., 2002, and Wilson & Upchurch, 2009). Poor dating obfuscates the exact end of this isolation (or earliest evidence of East Asian neosauropods, should a sampling bias explanation turn out to be correct), but the East Asian basal titanosauriforms Dongbeititan and Fukuititan are known from Barremian deposits (Wang et al., 2007; Azuma & Shibata, 2010), and earlier macronarian occurrences are known from the ?Berriasian–?Valanginian of Japan (Barrett et al., 2002; Saegusa & Tomida, 2011).

A Middle Jurassic–Early Cretaceous East Asian isolation has also been used to support a monophyletic and endemic basal eusauropod clade (Euhelopodidae), comprising the Jurassic taxa Shunosaurus, Omeisaurus, and Mamenchisaurus, and Early Cretaceous Euhelopus(Upchurch, 1995, 1998). Whereas the monophyly of this group of taxa has broken down, with Euhelopus now regarded as a somphospondylan (e.g. Wilson & Sereno, 1998; Wilson, 2002; Wilson & Upchurch, 2009; D'Emic, 2012; this study), the idea of an euhelopodid clade of titanosauriforms closely related to Euhelopus has remained (Canudo et al., 2002; Wilson & Upchurch, 2009; see also Ksepka & Norell, 2010), and has recently been formally defined (D'Emic, 2012; see Table 6 for phylogenetic definition). Six Early–middle Cretaceous East Asian species (Daxiatitan, Erketu, Euhelopus, Phuwiangosaurus, Qiaowanlong, and Tangvayosaurus) were recovered as a monophyletic Euhelopodidae by D'Emic (2012), with a number of contemporaneous or slightly stratigraphically younger taxa (Baotianmansaurus, Dongyangosaurus, Huabeisaurus, and Mongolosaurus) tentatively assigned to this group based on recovered synapomorphies (D'Emic, 2012: table 9). The affinities of several other East Asian taxa (Dongbeititan, Fukuititan, Fusuisaurus, Gobititan,Huanghetitan, ‘Huanghetitan’ ruyangensis, Jiutaisaurus, Pukyongosaurus, and Ruyangosaurus) remained ambiguous within Macronaria, whereas a number of primarily Late Cretaceous taxa (Jiangshanosaurus, Nemegtosaurus, Opisthocoelicaudia, Qingxiusaurus,Sonidosaurus, and Xianshanosaurus) were regarded as derived titanosaurs (D'Emic, 2012).

Here, we found support for an euhelopodid clade in our LSDM analysis, comprising five of the six taxa recovered by D'Emic (2012), as well as two additional taxa (Gobititan and Liubangosaurus; see Fig. 22). Our Euhelopodidae composition primarily differs in that Daxiatitan is recovered as a derived titanosaur. Although the cervical vertebrae of Daxiatitan share the presence of ventrolaterally deflected parapophyses, a posteriorly extensive intrapostzygapophyseal lamina, and prong-like epipophyses with euhelopodids, Daxiatitan appears to differ from this clade in a number of ways: its cervical neural spines are unbifurcated, its middle-posterior dorsal neural spines are gently bifid, and its anterior caudal vertebrae are strongly procoelous (You et al., 2008). However, in our LSDM_iw and LCDM analyses (Figs 26, 27), Euhelopodidae is composed only of Euhelopus and Daxiatitan. Currently, regardless of its composition, euhelopodid taxa are known only from East Asia and thus might represent an endemic clade; however, at least one putative occurrence has been described from outside this region. Canudo et al. (2002) described spatulate teeth from the Barremian of Spain that possess a lingual boss, a feature currently known only in Euhelopus and isolated teeth from the Early Cretaceous of China and Thailand (Buffetaut et al., 2002; Barrett & Wang, 2007; Wilson & Upchurch, 2009). Although extension of a clade from East Asia to western Europe based solely on teeth means that this referral must remain tentative, the Early–middle Cretaceous European sauropod fossil record is dominated by fragmentary and poorly preserved specimens (e.g. Upchurch et al., 2011b), and so any absence of unambiguous euhelopodid remains might purely reflect a sampling bias.

In addition to a clade of euhelopodids, all other Cretaceous East Asian taxa also represent titanosauriforms (and probably are members of Somphospondyli; although see the non-titanosauriform position of Dongbeititan in our LCDM), but appear to correspond to an array of lineages. Some cluster with other East Asian taxa in all analyses, although the membership of these clades often varies. Several taxa are extremely fragmentary and it is difficult to fully reconcile their phylogenetic positions, but there is evidence for derived titanosaurs in the Aptian–Albian, with Mongolosaurus recovered as a lithostrotian (or placed just outside that clade), and the caudal vertebra PMU R263 (Upchurch & Mannion, 2009) also displays evidence for belonging to this clade (Whitlock et al., 2011; this study). A lithostrotian identification of a contemporaneous embryo from Mongolia (Grellet-Tinner et al., 2011) has been refuted and we follow D'Emic (2012) in regarding it as an indeterminate member of Macronaria. The Albian-aged Jiangshanosaurus is placed as a saltasaurid (see also Tang et al., 2001b; D'Emic, 2012), and as the sister taxon to the latest Cretaceous North American sauropod Alamosaurus in all our analyses (Figs 22, 26, 27). This potentially has interesting implications for our understanding of the ‘sauropod hiatus’ (see above) in that it would support an Asian, rather than South American, origin for the Alamosaurus lineage. Whereas an Asian origin has been proposed by other authors (e.g. Wilson & Sereno, 1998), and can be inferred from previous titanosaurian phylogenetic analyses (see discussion in Mannion & Upchurch, 2011), this has always been based on the nesting of Alamosaurus with the latest Cretaceous Mongolian taxonOpisthocoelicaudia. Thus, the placement of Jiangshanosaurus is novel and would imply an approximately 30 million year ghost range for the Alamosaurus lineage. Although we cannot determine when the lineage leading to Alamosaurus dispersed to North America, dinosaur dispersal between Asia and North America appeared to be possible throughout the middle–Late Cretaceous (D'Emic et al., 2010, and references therein), and this ghost range is coeval with the duration of the North American ‘sauropod hiatus’. As such, the Alamosauruslineage might be: (1) unsampled purely in East Asia, followed by a latest Cretaceous dispersal to North America, or (2) dispersal might have occurred earlier and the lineage is unsampled in North America until the latest Cretaceous. Teasing these two scenarios apart will have to await increased sampling of Late Cretaceous sediments from both regions. Finally, one important caveat to this is our taxon sampling: the focus of our analysis was on basal titanosauriforms. Although all North American titanosauriform taxa are included, and most East Asian taxa (including Opisthocoelicaudia), the vast majority of South American titanosaurs were not incorporated into this study (32 valid, or potentially valid genera omitted; Mannion & Otero, 2012: table 8). As such, a sister-taxon relationship between a South American taxon and Alamosaurus or Jiangshanosaurus would contradict, or at least complicate, an Asian origin. Increased taxonomic sampling of titanosaurs will be crucial to resolving this issue, as well as determining the ancestry of latest Cretaceous European titanosaurs (see Pereda-Suberbiola, 2009; Garcia et al., 2010; Mannion & Upchurch, 2011; Grellet-Tinner et al., 2012).

The Late Cretaceous sauropod faunas of China seem to be comprised mainly of titanosaurs, or derived somphospondylans (LSDM), withErketu (and possibly Huabeisaurus; D'Emic, 2012) potentially representing a late-surviving euhelopodid (LSDM only). Dongyangosaurus[and possibly Baotianmansaurus (LSDM) and Liubangosaurus (LSDM_iw and LCDM)] appears to represent a titanosaur with close affinities to the latest Cretaceous Mongolian taxon Opisthocoelicaudia, suggesting a clade of East Asian opisthocoelicaudine saltasaurids, andXianshanosaurus is placed just outside of Lithostrotia (Figs 22, 27). Derived titanosaurian affinities for four further Late Cretaceous Asian taxa that were not included in this analysis (Nemegtosaurus, Qingxiusaurus, Quaesitosaurus, and Sonidosaurus) seem well supported (Curry Rogers, 2005; Wilson, 2005a; Xu et al., 2006; Mo et al., 2008; Wilson & Upchurch, 2009; Ksepka & Norell, 2010; Mannion & Calvo,2011; D'Emic, 2012), although their relationships to other titanosaurs from East Asia and elsewhere will have to await their incorporation into new phylogenetic analyses.

Lastly, it has been proposed that titanosaurs (You et al., 2003), or a more inclusive clade of titanosauriforms (Xu et al., 2006), might have originated in East Asia. Dongbeititan is recovered as the most basal somphospondylan in our LSDM (and LSDM_iw) and lies just outside Titanosauriformes in our LCDM (Figs 22, 26, 27), which might give some support to this hypothesis. However, the presence of a titanosaur in the Tithonian (Late Jurassic) of Tanzania, the extension of the somphospondylan lineage back to at least the late Oxfordian (as a consequence of the age of the earliest known brachiosaurid), and the existence of possible titanosaurian tracks from the late Middle Jurassic, all indicate that titanosaur (or somphospondylan) origins remain enigmatic, with poor sampling of key time intervals/geographical regions shrouding their early evolutionary history.