A Guide to Model Building

Model Building

The process of creating simplified representations of reality to understand complex relationships, predict outcomes, or prescribe actions.

Family of Models

A broader grouping of possible model forms, such as linear models, non-parametric regression, and regression trees.

Form of the Model

The specific structure of the model, including choice of predictors and relationships between predictors and the response.

Fit

The process of adjusting a model to data using various methods, commonly including least squares.

Assumptions in Models

Conditions that models assume about data, such as linearity and normality in linear regression.

Problem Definition Phase

The initial phase in model building where the decision-maker identifies and defines the problem.

Design Phase

The process of developing, testing, and validating a model, including determining variables and relationships.

Choice Phase

The phase in model building where alternatives are compared and the best solution is selected.

Implementation Phase

The stage of model building where the selected solution is put to work.

Static vs Dynamic Models

Static models capture a snapshot in time, while dynamic models represent systems evolving over time.

Quality of Fit

The assessment of how well a model fits data, often using metrics like R² and RMSE.

Sensitivity Analysis

A method to assess the importance of input variables by analyzing the impact on model output.

Model Validation

The process of testing and validating a model to ensure its reliability and that it meets assumptions.

Akaike Information Criterion (AIC)

A metric used for model comparison that assesses the trade-off between model fit and complexity.

Explanation versus Prediction

The distinction where models for explanation prioritize interpretability, and those for prediction prioritize minimizing error.

Decision Support Systems (DSS)

Systems that use models to help describe real-world decision-making situations.

Overfitting

A modeling issue where a model performs well on training data but poorly on unseen data due to excessive complexity.