The introduction and application of recursive partitioning methods in organizational science

recursive partitioning;classification and regression trees;random forests;machine learning;personnel selection

Jin, Jing

关键词: recursive partitioning;    classification and regression trees;    random forests;    machine learning;    personnel selection;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/46675/Jing_Jin.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Traditionally, multiple linear regression has been widely used in the field of organizational science for predictive modeling. Despite its pervasive use, the classical regression model falls short in several aspects, including the lack of flexibility in handling complex nonlinear relationships and the strict assumptions imposed by parametric approaches. To overcome these limitations, the current study examined an alternative, nonparametric recursive partitioning method – Classification and Regression Trees (CART), and its advanced successor, random forests. Results from two Monte Carlo simulations (Study 1 and 2) showed that random forests consistently produced comparable predictive accuracy as the traditional methods when the data was structured in a linear or simple additive model, yet exhibited substantially more accurate results when the data was structured in a complex nonlinear manner. CART outperformed the traditional methods for evaluating model fit (i.e., resubstituition accuracy), but was not as effective when generalizability was evaluated, except when the data was structured in a nonlinear tree-like pattern. Two empirical studies were also conducted to illustrate the application of the two recursive partitioning methods for predicting employee turnover (Study 3) and job performance (Study 4). Practical guidance is provided regarding how the feature selection procedure of random forests and a single decision tree constructed by CART could be combined to explore complex relationships within the data and better facilitate model interpretation. Limitations and implications for future research are also discussed.