Log Transformations in Regression: Linear-Log, Log-Linear, and Log-Log Interpretation (Worked Examples)

There are four common combinations of linear and log forms in regression: **Linear-linear (Y on X):** β represents the change in Y for a 1-unit increase in X. **Log-linear (log Y on X)**: β represents approximately a 100β percent change in Y for a 1-unit increase in X. Exact: 100(e^β − 1) percent for larger β. **Linear-log (Y on log X)**: β represents the change in Y for a 1 percent increase in X (specifically: β/100 change in Y for a 1% increase in X). Or: β represents the change in Y for a doubling of X if you used log base 2. **Log-log (log Y on log X)**: β IS the elasticity — a 1 percent increase in X produces a β percent increase in Y. This is the most common interpretation in economics where price elasticity, income elasticity, and demand elasticity all use this form.

Log transformation serves several purposes: **Linearize curved relationships.** Many natural relationships are multiplicative or exponential rather than additive. Log-transforming converts y = a·b^x to log(y) = log(a) + x·log(b) — a linear form regression can fit cleanly. **Stabilize variance (homoscedasticity).** Many variables have variance that grows with the mean (income, sales, populations, prices). Log transformation often produces residuals with constant variance, satisfying the OLS homoscedasticity assumption. **Make distributions more symmetric.** Right-skewed variables (income, time-to-event, market caps) become roughly symmetric after log transformation. This stabilizes coefficient estimation and makes confidence intervals more reliable. **Convert to elasticity (log-log).** Economists love log-log because the coefficient IS the elasticity — a unit-free measure that's comparable across studies and contexts. **When NOT to log-transform:** - Variables that take zero or negative values (log undefined). Workarounds: log(x + 1) for non-negative integers, or use a different transformation. - Variables already roughly symmetric and homoscedastic. - When the original units are meaningful for interpretation (loss of "increase by 1 dollar" interpretation matters). **Diagnostic checks:** - Plot residuals vs fitted: fan-shape suggests need for log transformation of Y - Plot Y vs X: curved relationship suggests log transformation may help - Q-Q plot of residuals: heavy right tail suggests log Y

A wage regression: log(Wage) = 0.85 + 0.12 × Education + ε, where Education is years of schooling. **Interpretation of 0.12:** approximately a 12% increase in wages for each additional year of education. More precisely: 100(e^0.12 − 1) = 100(1.1275 − 1) = 12.75% increase per year. The approximation 12% vs exact 12.75% — the gap is small for coefficients under 0.1 and grows with larger coefficients. For β = 0.30, the approximation gives 30% but the exact is 100(e^0.30 − 1) = 35%. Always use the exact formula for larger coefficients. **Worked numbers**: someone with 12 years of education earns: e^(0.85 + 0.12 × 12) = e^(0.85 + 1.44) = e^2.29 = $9.87 (in whatever monetary unit). Someone with 16 years (4 more years): e^(0.85 + 0.12 × 16) = e^(0.85 + 1.92) = e^2.77 = $15.96. Ratio: 15.96 / 9.87 = 1.617 — a 61.7% increase for 4 more years of education. Per year: 61.7^(1/4) − 1 ≈ 12.7% per year, matching the exact formula. This is why log-linear is the workhorse for percent-effect interpretation in economics, finance, and biostatistics.

A health regression: BMI = 28.5 + (-1.5) × log(Income) + ε, where Income is in dollars. **Interpretation of -1.5:** for a 1 percent increase in income, BMI decreases by approximately 0.015 units (= 1.5/100). For a doubling of income (100% increase), BMI decreases by 1.5 × log(2) = 1.5 × 0.693 = 1.04 BMI units. For a 10% income increase, BMI decreases by approximately 1.5 × log(1.10) = 1.5 × 0.0953 = 0.143 BMI units. **Why this form is useful**: when Y is in natural units (BMI, test score, height, count) but X is heavily skewed (income, population, price), the linear-log form preserves Y's natural interpretation while handling X's skewness. The "1 percent change in X produces β/100 change in Y" interpretation makes intuitive sense. **Worked numbers**: someone with $30,000 income: log(30000) = 10.31. BMI = 28.5 − 1.5 × 10.31 = 28.5 − 15.47 = 13.03. (Note: this is illustrative — real BMI doesn't go this low; this is just to show the math.) Someone with $60,000 (doubled): log(60000) = 11.0. BMI = 28.5 − 1.5 × 11.0 = 28.5 − 16.5 = 12.0. Difference: 1.03 BMI units, matching our prediction of 1.04 for a doubling.

A demand regression: log(Quantity) = 5.2 + (-1.8) × log(Price) + ε. **Interpretation of -1.8:** the coefficient IS the elasticity. A 1 percent increase in price produces a 1.8 percent decrease in quantity demanded. Demand is "elastic" because |elasticity| > 1. **Worked numbers**: at price $10, quantity = e^(5.2 + (-1.8) × log(10)) = e^(5.2 − 4.14) = e^1.06 = 2.89. At price $11 (10% increase), quantity = e^(5.2 + (-1.8) × log(11)) = e^(5.2 − 4.32) = e^0.88 = 2.41. Change: 2.41/2.89 − 1 = -16.6%, close to the predicted -18% for a 10% price increase (the difference reflects the discreteness of a 10% change in this approximation). **Elasticity classifications**: - |elasticity| > 1: elastic (price changes produce more-than-proportional quantity changes) - |elasticity| < 1: inelastic (price changes produce less-than-proportional quantity changes) - |elasticity| = 1: unit elastic - elasticity = 0: perfectly inelastic (insulin, life-saving medications without substitutes) - elasticity = -∞: perfectly elastic (commodity markets at the price-taker margin) **Common elasticities in economics**: - Price elasticity of demand (negative for normal goods) - Income elasticity of demand (positive for normal goods, negative for inferior goods) - Cross-price elasticity (positive for substitutes, negative for complements) - Wage elasticity of labor supply Log-log regressions are the standard estimation tool for these elasticities in empirical economics.

Suppose you're modeling sales (Y) as a function of advertising spend (X). Both are positive. Three plausible models: 1. **Linear-linear**: Sales = α + β·Adv. β interpretation: each additional dollar of advertising increases sales by β dollars. Diminishing returns are not modeled. 2. **Log-linear**: log(Sales) = α + β·Adv. β interpretation: each additional dollar of advertising produces a β·100% increase in sales (approximately). Implies a constant percent return regardless of current spend level. 3. **Log-log**: log(Sales) = α + β·log(Adv). β interpretation: a 1% increase in advertising produces a β% increase in sales. Models diminishing returns naturally because doubling adv only doubles log(adv) by log(2). Which to choose: - If you believe the relationship is roughly linear (each ad dollar produces similar incremental sales): linear-linear. - If you believe percent effects are constant (each ad dollar produces a constant percent boost): log-linear. Sometimes called "exponential growth" model. - If you believe elasticity is constant and there are diminishing returns: log-log. Common for advertising, where doubling spend rarely doubles sales. Diagnostic approach: fit all three, compare R², residual plots, and prediction error on a holdout set. The best model isn't necessarily the one with highest R² — it's the one that satisfies regression assumptions and produces meaningful, stable coefficients.

Snap a photo of the regression problem and StatsIQ identifies whether the model is linear-linear, log-linear, linear-log, or log-log, applies the correct interpretation to each coefficient, runs the percent-change calculation, and shows the comparison between approximate and exact formulas where relevant. For applied work, StatsIQ also discusses when log transformation is appropriate based on residual patterns and variable distributions, and walks through diagnostic checks like Q-Q plots and residuals vs fitted.

**Pitfall 1: Logging zero values.** log(0) is negative infinity. If your variable has zeros, you can't log it directly. Common workarounds: log(x+1) for non-negative integers; censor zeros and report separately; use a Tobit or Heckman selection model if zeros represent a meaningful state. **Pitfall 2: Confusing the approximation with the exact.** For β under 0.1, the approximation 100β ≈ percent change is accurate. For β = 0.50, the exact is 100(e^0.50 − 1) = 64.9%, not 50%. Don't use the approximation for large coefficients. **Pitfall 3: Comparing R² across linear and log models directly.** Different scales mean R² values aren't directly comparable. Use AIC, BIC, or out-of-sample MSE on the original scale. **Pitfall 4: Mis-stating the interpretation.** "A 1-unit increase in log(X) produces β change in Y" is correct but unhelpful — readers want the original-units interpretation. Translate: "A 1% increase in X produces β/100 change in Y" or "A doubling of X produces β·log(2) change in Y." **Pitfall 5: Forgetting the back-transformation bias.** If you log-transform Y, fit the regression, then exponentiate predictions back to the original scale, the predictions are biased downward. Use the smearing estimator (Duan 1983) or assume log-normal residuals to correct: predicted Y = exp(predicted log Y) × exp(σ²/2), where σ² is the residual variance. This content is for educational purposes only and does not constitute statistical advice.