DFBETAS: Influence on Individual Coefficients
What you'll learn: How to measure how much each individual regression coefficient (slope and intercept) changes when you remove a specific observation from your model.
Why DFBETAS Matters
You've already learned about Cook's Distance (overall influence) and DFFITS (influence on predictions). But sometimes you need finer detail: which specific coefficient is being influenced by a particular observation?
DFBETAS answers exactly that question. It tells you how much β₀ (intercept) or β₁ (slope) changes—measured in standard errors—when you drop one observation and refit the model.
The Core Idea
Think of DFBETAS like testing each ingredient in a recipe. Cook's Distance tells you "this ingredient really changes the whole dish," but DFBETAS tells you "it specifically makes the dish saltier" or "it specifically makes it sweeter."
For each observation i and each coefficient j, DFBETAS measures:
DFBETAS_ij = (coefficient with all data - coefficient without observation i) / standard error
A large absolute value means that observation has strong influence on that specific coefficient.
Interpreting DFBETAS
- Rule of thumb: Values greater than 2/√n (where n = sample size) are considered influential
- You get one DFBETAS value per observation per coefficient (so two values per observation in simple regression: one for slope, one for intercept)
- Positive DFBETAS: removing the point decreases that coefficient
- Negative DFBETAS: removing the point increases that coefficient
Key Takeaway: DFBETAS reveals which observations disproportionately affect individual regression coefficients, helping you identify points that may be distorting your slope or intercept estimates specifically, not just predictions overall.