Abstract
Elite military units use an in-depth assessment and selection (A&S) process to acquire the most qualified candidates. A unique challenge is to objectively evaluate the human dimension of attributes like leadership, resilience, and initiative in candidates. The A&S process requires significant time and resources to execute. The specific A&S studied for this research is eight weeks long and has a high logistical demand between supplies, personnel, and facilities. Effective screening of candidates prior to the A&S saves resources and selecting the best candidates enables the unit to better conduct highly specialized missions. Improving the system will reap dividends for the military.
Most studies about military A&S have used small data sets, used descriptive statistics for analysis, and focused on identifying predictors of candidate success. This research was broader in scope. We used 11,885 candidate records taken over a five-year period with 89 total features that included administrative, performance, and psychological data on each candidate. We applied a robust data science approach involving feature engineering, feature selection, optimized predictive models, and data subsets analysis to extract meaningful information from the data. Our objective for this research was to evaluate the utility of applying data science techniques to a specific military A&S data set with the goal of improving the holistic A&S system.
We applied ten classification models to a variety of feature, candidate, and feature engineering data subset combinations created using data science techniques. Using all candidates, the best model performance yielded a kappa score of 52 and 77% accuracy. Candidate non-selection prediction accuracy (86% Negative Predictive Value) was higher than candidate selection (68% Positive Predictive Value). The strongest predictors of candidate success were performance features, followed by administrative, and lastly psychological features. Although prediction accuracy was modest (<90%), we discovered utility in applying data science techniques to the A&S data. We extracted valuable insights from the data, found features highly predictive of candidate non-selection, and learned methods to modify the existing data to improve predictive capability. In conclusion, this research 1) validates the importance of an A&S to observe the human dimension of candidates and 2) proposes recommendations to add value to the holistic A&S system.
