Recommended Free Tools
Log models express the outcome, one or more predictors, or both on a logarithmic scale. The choice changes the relationship being modeled and the meaning of its coefficients: depending on the form, a slope describes a change in units, a percentage change, or an elasticity. Use a log model when that relationship makes sense for the subject—not just to make data look more normal or to improve a fit statistic.
What a log model changes
A regression with a logarithm is not merely the same model with different-looking numbers. Taking the log of a variable changes the scale on which the model describes its relationship with other variables. In his January 7, 2018 article, Sibashis Chakraborty frames the central question as why variables are logged in regression and whether they should be.
The answer depends on the pattern you want to represent and explain. A log of the response can model a percentage-type change in the original outcome; a log of a predictor can make its percentage changes correspond to changes in the outcome’s units. Logging both yields a percentage-to-percentage relationship. These interpretations are conditional on the other terms in the model and describe association, not causation.
Three common regression forms and their coefficients
Let β1 be the slope. These interpretations describe how the fitted outcome changes as a predictor changes, holding the model’s other terms constant.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →#1 Best Overall
| Form | Specification | Meaning of the slope |
|---|---|---|
| Level-log: log the predictor | Y = β0 + β1 ln(X) + u | A 1% increase in X corresponds approximately to a 0.01β1-unit change in Y. |
| Log-level: log the outcome | ln(Y) = β0 + β1X + u | A one-unit increase in X corresponds approximately to a 100β1% change in Y when the slope is small. The exact conversion is 100(exp(β1) − 1)%. |
| Log-log: log both | ln(Y) = β0 + β1 ln(X) + u | β1 is the elasticity: a 1% increase in X corresponds approximately to a β1% change in Y. |
The approximations are useful for small changes; for a large coefficient in a model with logged outcome, use the exact exponential conversion rather than treating 100β1% as exact. The Introduction to Econometrics with R explains these common specifications and interpretations.
When a logarithmic form makes sense
Choose the form by considering the relationship the subject matter suggests, then check whether the fitted model is adequate. The functional-form reasoning described by Chakraborty is a useful starting point:
- If a one-unit change in X is expected to correspond to a roughly constant percentage change in Y, consider modeling ln(Y) against X.
- If a percentage change in X is expected to correspond to a roughly constant unit change in Y, consider modeling Y against ln(X).
- If percentage changes in X and Y are expected to move together, a log-log model gives the slope an elasticity interpretation.
Taking logs can also turn a power relationship into a form that is linear in its parameters. A transformation may stabilize variance in some applications, but it does not do so automatically. Damodar N. Gujarati’s Basic Econometrics discusses log-linear forms alongside the assumptions about the disturbance term that still need attention after transformation.
What logging does not guarantee
A log transformation is not required merely because a predictor or outcome is non-normal. Ordinary least squares does not require predictors to be normally distributed; when normality is used for classical inference, the relevant assumption concerns the model errors. Logging the response changes the outcome scale and the residual structure, so it also changes what the model estimates.
Rank #3
Nor does logging automatically normalize data, remove outliers, or cure heteroskedasticity. It can reduce the influence of some large values or make variance more stable in a particular setting, but those effects must be checked rather than assumed. Evaluate residual diagnostics and predictive or inferential performance on the scale that matches the question you need to answer. Depending on the problem, robust regression, quantile regression, or methods such as MARS may be alternatives to a transformed ordinary least-squares model.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Check values before taking logs
The ordinary real logarithm is defined only for positive inputs. If X or Y includes zeros or negative values, do not silently add a constant and interpret the resulting coefficient as though it came from the original log model. A shift changes the transformation and therefore its interpretation; decide explicitly how such observations should be modeled.
Quick Recap
Rank #4
A practical way to choose a form
- State the relationship you expect. Decide whether the meaningful pattern is in absolute units, percentage changes, or an elasticity, and explain why the subject matter supports it.
- Write down the candidate equation. Mark whether the log applies to the outcome, predictor, or both. This makes the outcome scale and coefficient interpretation explicit before fitting.
- Check the data domain. Confirm that every variable to be logged is positive, or make an explicit, defensible modeling decision for nonpositive observations.
- Fit and diagnose the model on its specified scale. Examine residual behavior and whether the model’s assumptions are reasonable; do not assume the transformation fixes a problem.
- Compare models for the question you care about. Consider interpretability, substantive plausibility, diagnostics, and predictive or inferential goals. Do not choose a form solely because its R-squared is higher; Gujarati emphasizes theoretical basis and coefficient interpretation as well as fit.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
