Interpreting Models After Transformation
What you'll learn: How to correctly interpret regression coefficients and predictions when you've transformed variables like log, square root, or Box-Cox.
The Core Challenge
When you transform variables to fix violations of regression assumptions (like non-linearity or heteroscedasticity), your model lives in a new scale. A coefficient no longer means "a one-unit increase in X leads to β₁ change in Y"—because your X or Y (or both) aren't in their original units anymore.
Understanding Transformed Coefficients
Log transformation of Y: If you modeled log(Y) = β₀ + β₁X, the coefficient β₁ represents the proportional change in Y. Specifically, a one-unit increase in X multiplies Y by approximately exp(β₁). For small β₁, this is roughly a (β₁ × 100)% change.
Log transformation of X: If you modeled Y = β₀ + β₁log(X), then β₁ represents the change in Y when X is multiplied by some factor (like doubling).
Both transformed: With log(Y) = β₀ + β₁log(X), β₁ becomes an elasticity—the percent change in Y per 1% change in X.
Back-Transformation Considerations
To make predictions in original units, you back-transform (e.g., exp(ŷ) for log-transformed Y). Critical caveat: Due to Jensen's inequality, exp(E[log(Y)]) ≠ E[Y]. The naive back-transformation gives you the median prediction, not the mean. Bias correction factors may be needed for unbiased mean predictions.
Think of it like translating languages: you can convert words back, but idioms don't always translate literally.
Key Takeaway: After transformation, coefficients describe relationships in the transformed scale; always clarify whether you're interpreting proportional changes, elasticities, or multiplicative effects, and remember that back-transformed predictions require careful interpretation—often representing medians rather than means.