Long considered a consequence of anthropogenic agropastoralism, the origin of Madagascar’s central highland grassland is hotly disputed. Arguments that ancient endemic grasses formed grassland maintained by extinct grazers and fire have been persuasive. Consequent calls to repeal fire-suppression legislation, burn protected areas, and accept pastoralism as the ‘salvation’ of endemic grasses mount, even as the IUCN declares 98% of lemurs face extinction through fire-driven deforestation. By analysing grass data from contemporary studies, and assessing endemic vertebrate habitat and feeding guilds, we find that although the grassland potentially dates from the Miocene, it is inhospitable to endemic vertebrates, and lacks obligate grazers. Endemic grasses are absent from dominant grassland assemblages, yet not from woodland and forest assemblages. There is compelling evidence that humans entered a highland dominated by woodland and forest, and burned it; by 1000 CE grass pollens eclipsed tree pollens, reminiscent of prevailing fire-induced transformation of African miombo woodland to grassland. Endemic grasses are survivors from vanished woody habitats where grassy patches were likely small and ephemeral, precluding adaptive radiation by endemic vertebrates to form grazing guilds. Today forests, relic tapia woodland and outcompeted endemic grasses progressively retreat in a burning grassland dominated by non-endemic, grazing-adapted grasses and cattle.

            Vertebrate taxa adapted to grassland

Using the ‘Habitat and ecology in detail’ data field from IUCN species assessments, we compiled a database pairing 839 taxa with habitat (forest or grassland adaptation, electronic supplementary material, S1). We included every endemic Madagascan mammal radiating from a single colonisation event (i.e., lemurs, euplerids, tenrecs and rodents), and all reptiles and birds. We did the same for Australian endemic mammals.  We then compared number of species in these groups adapted to either grassland, forest or either habitat, between the two countries, designating species as either “grassland” (grassland-limited) or “not” (forest or either habitat), using a Wilcoxon rank sum test with continuity correction. Regional endemics (co-occurring on the Mascarenes, Tasmania and New Guinea) were excluded, as were predominantly aquatic birds, reptiles, and monotremes, groups neither forest nor grassland-adapted.

Comparing Madagascan and East Gondwanan herbivore guilds

Given Africa separated from East Gondwana 155 mya, India from Australia 130 mya, and Madagascar from India 88 mya, we compiled a database for the three established East Gondwana tropical grasslands, Australia, East Africa, and the Indian Terai-Duar, and compared herbivore guilds with a Madagascan suite comprising the only extant large herbivore (Lemur catta), and subfossils present in the highlands when humans arrived (25 species, ca. 2500 bp [4]; electronic supplementary material, S2). Guilds were divided into predominant browsers and folivores, obligate grazers, and mixed feeders, informed by a literature review of subfossil isotopic δ¹³C content [1,17], which differentiates C₄ grass diets from C₃ woody diets. To assess for an obligate Madagascan grazer-guild, we compared grazers to non-grazers across guilds and regions using a χ2 test, with Fisher’s exact test.

          Relative representation of endemic grass species under fire and grazing

We assessed and interpreted results from contemporary studies [7,9] to highlight the percentage of narrow endemic grass species in the highland relative to other ecoregions, and evaluated proportion of endemic grasses in dominant suites across highland vegetation types (electronic supplementary material, S3a,b). Next, using grasses from a fire-adapted functional group established by Solofondranohatra et al. [10], we compared number of sites occupied by endemic and non-endemic grasses to asses relative representation of endemics amongst dominant assemblages (electronic supplementary material, S3c). Lastly, we analysed data from 16 highland sites collected in December 2013 by Vorontsova et al. [7] to asses impact of fire, grazing and altitude on proportion of endemic grass species within sites. Grazing and fire represent the prevailing highland disturbance, impacting grasses in tandem, so we combined these two components to establish a variable that encapsulates the disturbance regime, by adding grazing (presence/absence) as recorded by Vorontsova et al. [7] and the frequency of fire over the three years preceding the data collection (2011-2013), using NASA Moderate Resolution Imaging Spectroradiometer data [19] (electronic supplementary material, S3d). If endemic grasses evolved in an ecosystem influenced by fire and grazing, they should be tolerant of these disturbances, and well-represented relative to non-endemics. We used a generalised linear model with number of endemic grass species versus non-endemics per site as the response variable, and altitude and disturbance regime (i.e., grazing index + fire frequency) as explanatory variables, binomial family with probit link. We validated the model visually, assessing residual fit for homogeneity, normality, and influential outliers [20].