The process of comparing and choosing between competing mathematical or statistical models based on criteria such as predictive performance, complexity, parsimony, information criteria, or cross-validation scores.

Principle

Model selection balances goodness-of-fit against model complexity to avoid overfitting, using formal criteria (AIC, BIC), predictive validation (cross-validation), or decision-theoretic metrics that weigh expected predictive loss and model interpretability.

Demonstration

Choosing between linear and polynomial regressions by cross-validated prediction error, selecting the number of components in a mixture model by BIC, or deciding between competing mechanistic epidemiological models based on out-of-sample forecasts and parsimony.

Misapplication

Selecting models solely on in-sample fit or on a single p-value without penalizing complexity, or conducting selection using the same data later used for inference or testing (data leakage), which inflates apparent performance and leads to poor generalization.

Consequence

Appropriate model selection yields models that generalize better to new data, simplifies interpretation, and supports reliable decisions; it also clarifies model uncertainty and may prompt model averaging when no single model is decisively superior.

Reversal

Model averaging or ensemble approaches: instead of choosing a single model, combine multiple models weighted by performance or posterior probability to hedge against selection uncertainty and often improve predictive accuracy.

Boundary

Focuses on structural choices between competing model forms and complexity levels; excludes parameter estimation within a fixed model (though both interact) and calibration aimed at tuning a chosen model's predictive performance rather than changing its structure.

Semantic Tension

Tension arises between selection and averaging: selection commits to one structure and simplifies interpretation, while averaging acknowledges model uncertainty and can yield better predictive performance but at the cost of interpretability and operational simplicity.

Synthesis

Model selection is the principled comparison and choice among alternative models using penalties for complexity and measures of predictive performance, providing a defensible structural decision while acknowledging residual model uncertainty and the option of averaging.