DFFITS: Influence on Fitted Values
What you'll learn: How to measure how much removing a single observation changes its own predicted value, scaled appropriately for comparison.
The Core Idea
After learning about Cook's Distance, which measures overall influence on the entire model, DFFITS (pronounced "dee-fits") zooms in on a more specific question: "How much does the predicted value for observation i change when we remove observation i from the dataset?"
Think of it like testing your weight on a bathroom scale. If you step off and the scale suddenly reads very differently for "where you were standing," something's wrong—you were exerting unusual influence on that measurement. DFFITS quantifies this effect for regression predictions.
How DFFITS Works
DFFITS compares:
- The fitted value with observation i included: ŷᵢ
- The fitted value without observation i: ŷᵢ₍₋ᵢ₎
The difference (ŷᵢ - ŷᵢ₍₋ᵢ₎) is then scaled by its standard error to make values comparable across different observations and datasets.
The formula structure (without diving deep into notation):
DFFITS = (Change in fitted value) / (Standard error of prediction)
Interpreting DFFITS
Larger absolute values indicate more influence. Common cutoffs:
- |DFFITS| > 2√(p/n) suggests high influence (where p = parameters, n = sample size)
- For simple regression with n=100: threshold ≈ 0.28
Unlike Cook's Distance (which measures global influence), DFFITS focuses specifically on how well that point predicts itself.
Key Takeaway: DFFITS measures how much an observation's own fitted value changes when it's removed from the model, scaled by standard error—perfect for identifying observations that "pull" their prediction toward themselves.