Concept Architecture
Concept
Theoretically, DFBETA is a regression diagnostic that quantifies the influence of an individual observation on a single regression coefficient. It measures the change in an estimated parameter when a particular observation is omitted from the analysis. DFBETA exists to identify influential observations that may disproportionately affect model estimates and subsequent inference. In health economics, it is used when developing regression models for costs, utilities, resource utilisation and risk prediction.
Mathematically, DFBETA is defined as the difference between the regression coefficient estimated using the full dataset and the coefficient estimated after deleting a single observation. Standardised versions (DFBETAS) further divide this difference by the estimated standard error, allowing comparisons across coefficients. Large absolute values indicate observations with substantial influence on the corresponding regression coefficient.
In practice, DFBETA is calculated following regression model estimation by sequentially deleting each observation and re-estimating the model. Analysts examine the magnitude of DFBETA values alongside other influence diagnostics to determine whether influential observations reflect genuine variation, data errors or model misspecification. Health economists commonly use DFBETA during model development and validation before incorporating regression results into economic evaluations.
Purpose
Used to identify observations that exert substantial influence on individual regression coefficients, supporting regression diagnostics, model validation and robust parameter estimation in health economic analyses.
Mathematical Formulae
Primary Formula
For observation i and coefficient j,
DFBETA?? = ??? ? ??????
where:
- ??? = regression coefficient estimated using the complete dataset
- ?????? = regression coefficient estimated after deleting observation i
Supporting Formulae
Standardised DFBETAS:
DFBETAS?? = (??? ? ??????) / �Var(??????)
A commonly used screening criterion is
|DFBETAS| > 2/�n
where n is the sample size.
Related Mathematical Methods
- Ordinary least squares regression
- Influence diagnostics
- Leave-one-out analysis
- Cook's Distance
- Leverage
- Studentised residuals
Example
A health economist fits a linear regression model predicting annual healthcare costs from age, sex and disease severity using data from 400 patients.
The estimated coefficient for disease severity is 1,250 using the full dataset. After removing patient 143, the coefficient increases to 1,330.
DFBETA = 1250 ? 1330 = ?80
The change indicates that patient 143 has a measurable influence on the estimated effect of disease severity. Additional influence diagnostics are examined before deciding whether the observation represents a recording error or legitimate clinical variation.
Excel Implementation
| Function | Example Formula | Health Economics Application |
|---|---|---|
| INDEX | =INDEX(B:B,MATCH(ID,E:E,0)) | Retrieves regression coefficients from repeated model estimates. |
| MATCH | =MATCH(ID,E:E,0) | Locates the omitted observation. |
| ABS | =ABS(B2-C2) | Calculates the absolute influence of deleting an observation. |
| SQRT | =2/SQRT(A1) | Calculates the commonly used DFBETAS screening threshold, where A1 contains sample size. |
VBA (Optional)
Automate leave-one-out regression estimation and generate DFBETA values and influence reports for every observation.
Sources
- Belsley DA, Kuh E, Welsch RE. Regression Diagnostics: Identifying Influential Data and Sources of Collinearity. Wiley.
- Cook RD, Weisberg S. Residuals and Influence in Regression. Chapman & Hall.
- Draper NR, Smith H. Applied Regression Analysis. 3rd ed.
- Harrell FE. Regression Modeling Strategies. 2nd ed.
- ISPOR Good Practice Reports on statistical modelling and validation.
Related Concepts (2)
Frequently Asked Questions (6)
What is DFBETA?
A regression diagnostic measuring how much a coefficient would change if a particular observation were removed, identifying disproportionately influential data points.
Source: Belsley, Kuh & Welsch 1980
Why does a single observation sometimes distort a regression?
In a regression, one unusual data point can pull an estimated coefficient noticeably toward itself, so that the fitted relationship reflects that point more than the rest of the data. DFBETA quantifies this by showing how much each coefficient would shift if a given observation were dropped, flagging points whose removal changes the result substantially. A large value warns that a conclusion rests heavily on one observation, which merits checking for error or special circumstances. Belsley and colleagues (1980) introduced such diagnostics.
Source: Belsley et al. 1980
How is DFBETA calculated?
DFBETA is calculated for each observation and each coefficient by comparing the coefficient estimated from the full data with the estimate obtained when that observation is omitted, expressing the difference, often standardised. A DFBETA value shows the change in the coefficient attributable to the observation. Computing DFBETAs across observations reveals which ones most affect each coefficient, so unusually large values flag observations whose removal would notably alter the fitted model, indicating high influence.
Source: Belsley, Kuh & Welsch 1980
What does DFBETA reveal about a regression?
DFBETA reveals which individual observations disproportionately influence the estimated coefficients of a regression, so that a few influential points driving the results can be identified. This matters because a regression's conclusions should not rest heavily on a small number of observations, and influential points may reflect errors, outliers, or genuinely unusual cases. By quantifying each observation's effect on the coefficients, DFBETA helps assess the robustness of the regression and whether particular data points warrant scrutiny.
Source: Belsley, Kuh & Welsch 1980
How does DFBETA help handle influential observations?
Influential observations flagged by DFBETA are examined to determine whether they are errors, outliers, or legitimate but unusual cases. Errors may be corrected or removed, while genuine but influential points prompt caution about how much the results depend on them. Analysts may report results with and without such observations to show their effect, or use methods less sensitive to influence. The aim is not automatically to delete influential points but to understand and address their disproportionate effect on the conclusions.
Source: Belsley, Kuh & Welsch 1980
Why are regression diagnostics like DFBETA important?
Regression diagnostics like DFBETA are important because a regression can be unduly influenced by a few observations, so that its conclusions rest on them rather than on the data as a whole, which may be misleading if those points are erroneous or atypical. Diagnostics reveal such influence, along with outliers and other problems, allowing the reliability of the regression to be assessed. Checking influence supports robust conclusions, ensuring that estimated relationships are not artefacts of a small number of influential data points.
Source: Belsley, Kuh & Welsch 1980
Trust Record
Verified by Dr Darrin Baines
British health economist
Professional identity: darrinbaines.org
Verification date: 14 Oct 2025
Content version: 1.0.0
Canonical Identity
- Persistent URI
- https://healtheconomics.wiki/concept/dfbeta
- Term code
- HE-EM-MV-018
Stable URI · Machine-readable · Resolvable · CC BY 4.0