mismodeling

Mismodeling is the mismatch between a model's assumptions, structure, or inputs and the real system it seeks to represent. It occurs when the chosen functional form, variable set, or distributional assumptions fail to capture essential relationships, dynamics, or heterogeneity, leading to biased estimates, poor predictions, or misguided decisions.

Causes include model misspecification (omitted variables, incorrect functional form, wrong error structure), data issues (measurement error, sampling bias), and model-class limitations (insufficient complexity, or overly flexible models that capture noise). In econometrics and statistics, mismodeling often yields biased estimators and wrong standard errors due to omitted variable bias, endogeneity, or incorrect assumptions about independence and homoscedasticity. In machine learning, mismodeling can arise from distribution shift between training and deployment, target leakage, feature engineering errors, or a mismatch between the model class and the true data-generating process.

Consequences include reduced predictive accuracy, biased inference, miscalibrated risk, and poor decision support. Examples include assuming linear relationships in nonlinear regimes, ignoring interactions, assuming normal errors when heavy tails exist, or extrapolating beyond the range of observed data.

Detection and mitigation involve diagnostic checks such as residual analysis, goodness-of-fit tests, calibration curves, backtesting, cross-validation, and out-of-sample evaluation. Mitigation strategies include revising the model specification, incorporating additional variables, using robust or Bayesian methods, performing sensitivity and scenario analyses, correcting for measurement error, and aligning data and model to domain knowledge. Ongoing validation and uncertainty quantification help manage mismodeling risk.