A comparison of methods for modeling soundscape dimensions based on different datasetsa).

Soundscape studies vary considerably in study design, statistical methods, and model fit metrics used. Due to this confounding of data and methods, it is difficult to assess the suitability of statistical modelling techniques used in the literature. Therefore, five different methods and two performance metrics were applied to three existing soundscape datasets to model soundscape Pleasantness and Eventfulness based on seven acoustic and three sociodemographic predictors. Datasets differed in soundscape type (urban outdoor vs indoor), experimental setting (field- vs lab-based), size, and study design (site- vs person-centered). The fixed-effects and mixed-effects methods ranged from linear to nonlinear regression based on advanced machine learning approaches. Results showed that models performed better for Eventfulness than for Pleasantness in most cases, while performance as measured by the out-of-sample R2 was dependent on the total variance of the target, especially in both field studies with imbalanced targets and groups. Nonlinear methods consistently outperformed linear regression, with random forest and extreme gradient boosting performing particularly well, while the performance levels of all nonlinear methods remained comparable. Mixed-effects models provided a more generalized, albeit slightly smaller prediction performance when tested on unknown groups. Finally, this study motivates the use of cross-validation with special splitting for analyzing small imbalanced datasets.