VerifiedEvidence: highv1.0.0

Mean Squared Error

A measure of the average squared difference between predicted and actual values, penalising larger errors more heavily due to squaring.

Last reviewedDarrin Baines IP Ltd

Concept Architecture

Concept


Theoretically, Mean Squared Error (MSE) is a statistical measure of prediction error that quantifies the average squared difference between observed and predicted values. It combines both the variance and bias of prediction errors into a single metric and serves as one of the principal loss functions in regression, forecasting and statistical estimation. Mean Squared Error exists to evaluate predictive performance while assigning greater penalty to larger errors through the squaring operation.

Mathematically, Mean Squared Error is calculated by averaging the squared residuals between observed and predicted values. Because errors are squared before averaging, large prediction errors contribute disproportionately to the final value, making MSE particularly sensitive to outliers. In estimation theory, MSE is also defined as the expected squared difference between an estimator and the true parameter value and can be decomposed into variance and squared bias.

In practice, MSE is calculated to assess regression models, forecasting algorithms, machine learning models and simulation outputs. In health economics, it is used to evaluate models predicting healthcare costs, quality-adjusted life years, disease progression, resource utilisation and survival outcomes. Lower MSE values indicate better predictive accuracy and are commonly used to compare competing models during model development and validation.

Purpose


Used to quantify prediction error, compare predictive models, evaluate regression and forecasting performance and assess estimator accuracy in health economic analyses.


Mathematical Formulae

Primary Formula

Mean Squared Error:

MSE = (1/n) ? ????� (y? ? ??)�

where:

  • n = number of observations
  • y? = observed value
  • ?? = predicted value

Supporting Formulae

Residual:

e? = y? ? ??

Bias-variance decomposition:

MSE = Var(??) + Bias(??)�

Root Mean Squared Error:

RMSE = �MSE

Related Mathematical Methods

  • Root Mean Squared Error
  • Mean Absolute Error
  • Mean Absolute Percentage Error
  • Residual Analysis
  • Cross-Validation
  • Regression Diagnostics
  • Bias-Variance Trade-off

Example

A health economist develops a model to predict annual healthcare costs for five patients.

PatientObserved (�)Predicted (�)Squared Error (��)
14,8004,60040,000
26,3006,50040,000
35,7005,40090,000
47,1006,90040,000
55,1005,40090,000

Mean Squared Error:

MSE = (40,000 + 40,000 + 90,000 + 40,000 + 90,000) � 5

MSE = 60,000

The model therefore has an average squared prediction error of 60,000 (��). The corresponding Root Mean Squared Error is:

RMSE = �60,000 � �245


Excel Implementation

FunctionExample FormulaHealth Economics Application
POWER=POWER(B2-C2,2)Calculate the squared prediction error for each observation.
AVERAGE=AVERAGE(D2:D501)Calculate Mean Squared Error across all observations.
SQRT=SQRT(E2)Calculate Root Mean Squared Error from MSE.
SUM=SUM(D2:D501)Calculate the total squared prediction error.
COUNT=COUNT(D2:D501)Determine the number of observations contributing to MSE.

VBA (Optional)

A VBA routine can automatically calculate Mean Squared Error and Root Mean Squared Error for multiple predictive models and generate comparative model performance summaries.


Sources

  • Hastie T, Tibshirani R, Friedman J. The Elements of Statistical Learning. Springer.
  • James G, Witten D, Hastie T, Tibshirani R. An Introduction to Statistical Learning. Springer.
  • Bishop CM. Pattern Recognition and Machine Learning. Springer.
  • Montgomery DC, Peck EA, Vining GG. Introduction to Linear Regression Analysis. Wiley.
  • Briggs A, Claxton K, Sculpher M. Decision Modelling for Health Economic Evaluation. Oxford University Press.
  • ISPOR Good Practice Reports.

Library

Publications

1
  • Book

    Statistical Analysis of Cost-Effectiveness Data — Willan & Briggs, 1st Edition ed., 2006 (John Wiley & Sons)

    A synthesis of statistical methods for analysing cost-effectiveness data, including net-benefit regression, confidence intervals for the ICER, cost-effectiveness acceptability curves, and covariate adjustment. Part of the Wiley Statistics in Practice series.

Frequently Asked Questions (6)

  • What is the mean squared error?

    A measure of the average squared difference between predicted and actual values, penalising larger errors more heavily due to squaring.

    Source: Casella G, Berger RL. Statistical Inference. 2nd ed. Duxbury; 2002.

  • Why does the mean squared error weigh large misses so heavily?

    The mean squared error averages the squared differences between predictions and actual values, and squaring is what makes it weigh large misses so heavily: an error twice as big contributes four times as much. This emphasis suits situations where big mistakes are especially costly, and it has a convenient mathematical form that decomposes neatly into the model's bias and its variance. The trade-off is heightened sensitivity to outliers. Punishing large errors disproportionately is its defining feature. Hastie and colleagues (2009) discuss this measure.

    Source: Hastie et al. 2009

  • How is the mean squared error calculated?

    The mean squared error is calculated by taking the difference between each predicted and actual value, squaring it, summing the squared errors, and dividing by the number of observations, giving the average squared error. So the mean squared error is calculated as the average of the squared prediction errors, which measures error in squared units and emphasises larger deviations through the squaring, and taking its square root gives the root mean squared error, which returns the measure to the original scale of the data, making it more directly interpretable while retaining the emphasis on larger errors.

    Source: Casella & Berger 2002

  • Why does the mean squared error penalise large errors?

    The mean squared error penalises large errors because it squares each error before averaging, so an error twice as large contributes four times as much to the total, making large errors dominate the measure. This emphasis reflects a view that large errors are disproportionately undesirable. So the mean squared error penalises large errors through the squaring, which weights big deviations far more than small ones, making it sensitive to outliers and appropriate when large errors are especially costly, though this same property means a few extreme errors can dominate the measure, which is why the mean absolute error is preferred when such robustness is wanted.

    Source: Casella & Berger 2002

  • How does the mean squared error relate to bias and variance?

    The mean squared error of an estimator can be decomposed into the square of its bias plus its variance, so it captures both systematic error and variability in a single measure. This decomposition underlies the bias-variance tradeoff. So the mean squared error relates to bias and variance as their combination, expressing the total expected squared error as the sum of the squared bias and the variance, which is why minimising the mean squared error involves balancing bias against variance, and this makes the mean squared error a central criterion in estimation and prediction, linking accuracy to the tradeoff between fitting too simply and too flexibly.

    Source: Casella & Berger 2002

  • How is the mean squared error used?

    The mean squared error is used to assess and compare the accuracy of predictions or estimators, with lower values indicating better performance, and as a criterion to fit models, since methods such as least squares minimise it. It is common in regression and prediction. So the mean squared error is used both to evaluate how well predictions match actual values and as an objective to be minimised in fitting models, which makes it a foundational measure in statistics and machine learning, though its sensitivity to large errors and its squared units are considered, with the root mean squared error often reported for interpretability on the original scale.

    Source: Casella & Berger 2002

Trust Record

Verified by Dr Darrin Baines

British health economist

Professional identity: darrinbaines.org

Verification date: 18 Dec 2025

Content version: 1.0.0

Canonical Identity

Term code
HE-ES-SA-113

Stable URI · Machine-readable · Resolvable · CC BY 4.0