South Arabia’s prehistoric monument landscape shows social resilience to climate change

PLOS One | https://doi.org/10.1371/journal.pone.0323544 May 28, 2025 9 / 24 ridge and LASSO, allowing both variable selection and handling of collinear predictors. Both LASSO and elastic net can be used implicitly for variable selection [92]. A second parameter, λ, controls the overall penalty strength. As λ increases, more coefficients shrink towards zero, further refining the model. All variables were standardized to have a mean of zero and a variance of one before analysis to ensure comparable model coefficients. We performed K-fold cross validation in the R ‘glmnetUtils’ package to determine optimal values for α and λ, using deviance as the performance metric. The model with the lowest multinomial deviance identified α = 0.4 (elastic net regression) (S2 Fig). We report coefficients from the model with the largest λ within one standard-error of the minimum λ value (λ = 0.049, log λ = -3.023) (S3 Fig). This follows the one-standard-error rule [88], the default in glmnet, which accounts for estimation uncertainty in risk curves. We also used regularized multiple linear regression to examine how environmental factors influence total stone volume, providing insight into the relationship between environmental conditions and labor input. If monument builders selected locations to communicate something about their significance, we would expect greater labor investment in environmen- tally important areas. As before, we log-transformed stone volume for normality and standardized all variables. K-fold cross-validation identi-

fied α = 0.1 (elastic net regression) based on the model with the lowest mean-squared error (S4 Fig). We compare results from both the minimum λ and one-standard error λ models (S5 Fig). Since the one-standard error model with no variables showed similar accuracy and error to the minimum λ model, the latter’s results should be interpreted with caution. We repeated this analysis using only data of HCTs and triliths, the two most common monument types. Collocation analysis Using our regional chronological sequence of monument types, we also tested spatiotemporal relationships between mon- ument types. To communicate something about historical place and builders’ attachments to it, monument builders may have decided to build where prior monuments existed. Clustering would suggest that builders were reinforcing or adjusting prior perceptions of landscape through monument placement. We used collocation analysis to test whether the locations of preexisting monument types influenced the placement of subsequently constructed monuments [93]. This technique determines the degree to which a subsequent monument type is collocated with or isolated from the preceding monument type, relative to other monument types, in the neighboring 20 monument locations. In addition, we tested for spatiotemporal relationships between the monument types. We calculated the spatial mean center and standard distance of the monument types to observe general distributional trends of the different monument

types. The mean center is the geographic center, calculated by averaging the monument coordinates. Standard distance measures the degree to which the monuments are concentrated or dispersed around the mean center. We used a dis- tance of one standard deviation from the mean. We can visually compare these measures to observe geographic overlap or dispersion of monument types. Technological and labor data analysis We tested whether the technological and labor demands of monument construction followed a trend with aridification. We used the total stone volume of a monument as an indicator of labor input [66,94,95]. Using ANOVA with Tukey post- hoc tests, we compared total stone volumes across different monument types - platforms, SCABS, HCT, triliths, and boat graves -which represent distinct chronological periods. Since labor can be provided simultaneously by many or sequentially by few, we test whether earlier monuments also have the largest stones, a proxy for simultaneous labor. If monuments with larger total stone volumes also include the largest stones, it would support prior research suggesting that large groups convened episodically during the wetter Early Holocene to simultaneously invest labor for monument building [29,53].