Not everything that matters is measured. A latent variable is something inferred from observed data that influences outcomes but is not directly seen. Intelligence, motivation, and disease severity all fall into this category.

Models use latent variables to explain patterns in observed data. Factor analysis groups correlated measurements into underlying factors. Variational autoencoders learn a latent space that captures the structure of images or text. Topic models infer themes from documents.

Common latent variable uses

• Dimensionality reduction
• Generative modelling
• Topic discovery in text
• Psychometric measurement
• Causal inference with unobserved confounders

Interpretation is tricky. A latent variable is a mathematical construct, not a proven entity. Two models may find different latent structures that explain the same data equally well. Assigning meaning to latent dimensions requires care.