Construct validity (convergent and discriminant)

Grounded in peer-reviewed studies (Garbi, 2022; Garbi et al., 2024): Glimpse's explainables (traits like Agility, Perseverance) show convergent validity (different tasks measure the same trait) and discriminant validity (traits are distinct from each other and from General Mental Ability).

Predictive and concurrent validity

Glimpse models are rigorously validated using out-of-sample validation, which involves testing each model on independent data not used during training. This ensures that model performance reflects true predictive accuracy, rather than "memorizing" the training data.

By using this approach, Glimpse ensures robustness and reliability in real-world applications. In contrast, evaluating a model on the same data it was trained on can produce misleading results: the model may simply memorize patterns in that data rather than learn generalizable rules. This leads to a well-known issue called overfitting, where performance appears strong during model development but fails when applied to new candidates.

How Glimpse compares

  • Across use cases, Glimpse models achieve AUC-ROC values between 0.65 and 0.85, with corresponding correlations of r = 0.26 to 0.59, indicating strong and practical predictive value.
  • Big Five Personality Questionnaires typically achieve up to an equivalent of in-sample AUC of ~0.61.
  • General Mental Ability (GMA) typically reaches an equivalent of in-sample AUC ~0.66. Structured interviews perform slightly better, around in-sample AUC ~0.71, but are less scalable and are susceptible to bias.

Importantly, Glimpse remains predictive even in populations where GMA and interviews lose effectiveness, typically because candidates have already been pre-screened for high GMA, making them indistinguishable on that basis. This phenomenon, known as range restriction, limits the usefulness of traditional tools.