Orchidaceae is the largest and one of the ecologically and morphologically most diverse families of flowering plants . Several ages have been proposed for the origin of modern orchid lineages (i.e., their crown age), ranging from ~26 million years (Ma) , ~40 Ma , ~80 Ma  to as much as ~110 Ma . A correct time estimation is essential for our understanding of the mechanisms underlying the diversification of orchids, and could contribute to discern between alternative hypotheses of diversification - such as significant increases in speciation rates temporally correlated to climatic changes, tectonic events, or radiation of pollinators.
Many parameters have been identified to affect divergence time estimates in phylogenies, including taxon sampling, reliability, number and placement of internal calibration points, and dating method [6–16]. Until recently, molecular dating of the Orchidaceae has been challenging due the complete absence of reliable orchid fossils. The finding of a 15-20 Ma fossil of an extinct stingless bee (Proplebeia dominicana), covered with pollinia from an orchid species belonging to the subtribe Goodyerinae, finally allowed for temporal calibration of the family . Using this fossil as an internal calibration point, and departing from a phylogenetic tree obtained from the analysis of plastid DNA sequences (matK and rbcL), Ramirez et al.  estimated the origin of Orchidaceae at 76-84 Ma. These results supported an 'ancient' origin of orchids in the Late Cretaceous.
Although the study by Ramirez et al.  unquestionably constituted a milestone in orchid research, the large discrepancies in age estimates obtained in the last decade - some 80 Ma between the youngest  and oldest  crown ages - suggests that the matter is probably not completely settled. In a recent study in the family Begoniaceae, Goodall-Copestake et al. found that the second largest source of variance in age estimates (after availability and placement of internal calibration points) was derived from the choice of dating method employed, a result consistent with previous evaluations of empirical data [10, 11]. In particular, recent developments in molecular dating techniques have called into question the assumptions and algorithms implemented in Non-parametric Rate Smoothing [NPRS; ] and Penalized Likelihood [PL; ] - the two methods employed by Ramirez et al. . Whereas NPRS has been largely abandoned in favour of its successor PL [see discussion in ], both implemented in the software r8s , PL competes today in popularity with Bayesian dating  implemented in the software BEAST .
PL and BEAST operate in very different ways: i) PL requires a fixed phylogram as input, whereas BEAST samples topologies simultaneously as it calculates divergence times under a MCMC analysis, and allows the choice of several different priors and models; ii) PL assumes autocorrelation of rates within the phylogeny (i.e. that mutational rates are inherited, resulting in closely related taxa exhibiting similar evolutionary rates), whereas BEAST allows branches to vary in evolutionary rate; iii) in PL, nodes can be calibrated to be either fixed to a certain age, or constrained by a maximum or a minimal bound; whereas in BEAST, several additional alternatives are available for calibrating a node, because such calibrations represent age priors drawn from distributions of various shapes (e.g., normal, lognormal, exponential, or uniform). The methodological and conceptual differences between r8s, BEAST, and some other methods available today for molecular dating have been reviewed by several authors [12, 14, 22–24].
Although the methodology and assumptions implemented in each molecular dating method can be readily compared, our knowledge of how time estimates are influenced by the choice of method is still poor. For instance, Goodall-Copestake et al. obtained younger ages in the Begoniaceae using PL than using NPRS, but as these authors noted the inverse situation was found by Clement et al. [25, 26] on the same taxonomic group. According to Goodall-Copestake , this surprising discrepancy was probably caused by differences in density of sampled taxa and calibration points. Similarly, it may be very difficult to predict differences in age estimates using PL and BEAST: in the study by Goodall-Copestake , PL produced considerably younger ages than BEAST, whilst the opposite situation was found within family Caryophyllaceae . These results exemplify the potential influence of methodology on age estimations.
In this study we aim at reassessing the temporal origin and diversification of Orchidaceae, using the Bayesian uncorrelated relaxed molecular clock approach implemented in BEAST. In addition to choosing a different dating method, we conduct a new analysis on an expanded taxon sampling by adding two internal calibration points in the orchid phylogeny. We base these calibrations on fossil leaves described subsequent to the study by Ramirez et al.  from Early Miocene deposits of New Zealand, which were confidently assigned to genera Dendrobium and Earina . Then, to explore how the high-level age estimates obtained here may affect the evolutionary interpretation of a species-level orchid clade, we date the origin and diversification of the rare South American orchid genus Hoffmannseggella.
Hoffmannseggella belongs to the Epidendroideae, the largest subfamily within Orchidaceae, which comprises over half of all orchid species . The subfamily has been divided into 'lower' and 'higher' Epidendroids  and this latter clade includes the monophyletic subtribe Laeliinae, where Hoffmannseggella is nested . The genus is endemic to Brazil, where it is confined to the High Altitude Rocky Complexes (Brazilian Campos Rupestres and Campos de Altitude) of Minas Gerais, Rio de Janeiro, Espírito Santos and Bahia states. It comprises exclusively rupicolous species, i.e. growing among rocks . Adding to the 32 different species recognized by Chiron and Castro Neto , several new species have recently been described and today Hoffmannseggella comprises 41 species [31, 33–39]. Half of these are "micro-endemic" - known from a single natural population, and some only from the type collection. We have been able to obtain or generate sequences for all but a single species, thus reaching a 98% complete species sampling.