Common Transformations: Log, Square Root, Reciprocal
What you'll learn: How to apply three powerful mathematical transformations that fix common regression problems like curved relationships and unequal variance.
Why These Transformations Matter
When residual plots reveal non-linearity or heteroscedasticity (remember, that's non-constant variance), transforming your variables can "straighten out" relationships and stabilize variance—making linear regression assumptions hold.
The Three Workhorses
Logarithmic Transformation
Apply log(y) or log(x) when:
- Data spans several orders of magnitude (e.g., income, population)
- Variance increases with the mean
- Growth is exponential or multiplicative
Effect: Compresses large values more than small ones, pulling in the right tail of skewed distributions.
Square Root Transformation
Apply sqrt(y) or sqrt(x) when:
- You have count data (always non-negative)
- Variance increases with the mean, but less dramatically than log situations
- You want a gentler transformation than log
Effect: Moderately compresses larger values while preserving zero.
Reciprocal Transformation
Apply 1/y or 1/x when:
- Diminishing returns are present (effect weakens as X increases)
- You're modeling rates or speeds
- The relationship approaches an asymptote
Effect: Dramatically compresses large values; makes infinity become zero. Changes interpretation significantly.
Choosing the Right One
Think of these as strength levels: reciprocal is strongest, log is medium, square root is gentlest. Start with the mildest transformation that fixes your diagnostics. After transforming, always re-check residual plots and Q-Q plots to confirm improvement.
Key Takeaway: Logarithmic, square root, and reciprocal transformations are your toolkit for linearizing curved relationships and stabilizing variance—choose based on your data's behavior and check diagnostics afterward to verify the fix worked.