Climate change in Europe between 90 and 50 kyr BP and Neanderthal territorial habitability

supported in a given area within natural resource limits. Carrying capacity values were attributed to each biome type using the correspondence table 3.2 published by [18] (see also S3 File). The anomaly method we use to obtain climate and biome reconstructions allows us to con- duct all our analysis on the same 0.5˚ x 0.5˚ grid as the modern observational and topographic datasets we use in producing our environmental variables. To account for variations in sea level, and thus the area of land available at the coast, coastlines were adjusted by applying approximate global mean sea level changes characteristic for each period, and were created by extrapolating beyond the coastline with the average value of the nearest 3 existing inland grid points, estimated from [6]. Where these lowered coastlines did not exactly match those used in the climate model simulations, climate variables were extrapolated to the coast using the near- est 3 valid grid cells and these extrapolations used in the biome reconstruction. Maxent The nature of our data, i.e. the presence of certain fossils, requires us to use a presence-only model to reconstruct the Neanderthals’ ecological niches. In fact, archaeological excavations inform us about where Neanderthal was present (presence data), but we have no certainty about where it was absent (absence data) (in contrast to, for example, systematic biological sur- veys that produce presence-absence data). We chose to use Maximum Entropy Modelling (Maxent), specifically Maxent v3.4.4 [19].

The aim of Maxent [20,21] is to estimate a target probability distribution by finding the proba- bility distribution of maximum entropy (i.e., that is most spread out, or closest to uniform), subject to a set of constraints (expressed in terms of simple functions of the environmental var- iables, called features; as for the predictors in usual modelling) that represent our incomplete information about the target distribution. This method has proven its effectiveness: 1. in presence-only data, indeed it really requires only presence data, and not pseudo-absence (i.e. absence data are locations randomly generated in the background, that’s to say sam- pling universe where environmental predictors are informed). This suppress the problem of “contaminated controls” [22,23], in which background data is treated as absence data, even though it is contaminated with presences; 2. in avoiding over-fitting by using l1-regularization [19]; 3. with small sample sizes [24,25]. Given our small number of fossil sites and to avoid overfitting our data with an overly com- plex model, only linear, quadratic and hinge (like linear, but it is constant below a threshold) features were used to model the constraint link between the environmental variables and pres- ence of Neanderthal. As the prevalence (proportion of occupied sites) of Neanderthals is typi- cally unknown, Maxent’s default prevalence was not modified (value set by the software designers: Prevalence = 0.5). Likewise, for regularization and background settings, defaults

(values set by the software designers: Regularization multiplier = 1; Maximum number of background points = 10,000, randomly sampled from covariate grids) were used. Results are shown in the cloglog output format, an estimate of the probability that the Neanderthals would have been able to live at a location, given the environmental variables at this location [26], which is classically called suitability in Species Distribution Modelling (SDM). Sites with small cloglog values are predicted to be unsuitable or only marginally suitable for the species. The constructed maps reflect these estimated probabilities, from red indicating high probabil- ity of suitable conditions for the Neanderthals, to blue indicating a low one. PLOS ONE Neanderthal habitability PLOS ONE | https://doi.org/10.1371/journal.pone.0308690 February 26, 2025 4 / 17