Interpreting Log-Log Models as Elasticities

Why a Log-Log Model Is Useful

Some relationships are easier to describe in proportional terms than in the original measurement units. For example, a change of $100 may be substantial for a small budget but minor for a large one. If the question concerns how a percentage change in one quantity relates to a percentage change in another, a model with both variables logged gives a direct interpretation: its slope is an elasticity.

The log-log interpretation is especially useful because the slope has no units. It describes relative changes rather than changes measured in dollars, kilograms, or other particular scales. It does not, by itself, show that changing the predictor causes the outcome to change.

The Log-Log Model and Elasticity

Definition: A log-log regression model for positive predictor $X$ and positive outcome $Y$ is $$ \log Y_i=\beta_0+\beta_1\log X_i+\varepsilon_i. $$ Here $\beta_0$ is the intercept, $\beta_1$ is the slope, and $\varepsilon_i$ is the error for observation $i$. The model describes a linear relationship between the conditional mean of $\log Y$ and $\log X$, with $E(\varepsilon_i\mid X_i)=0$. The predictor and outcome must be positive for their logarithms to be defined.

The slope $\beta_1$ is the change in fitted $\log Y$ associated with a one-unit increase in $\log X$. A one-unit change on a log scale is a multiplicative change on the original scale, so the slope can be expressed as the rate of change in log outcome per rate of change in log predictor:

\frac{d\log \widehat{Y}}{d\log X}=\beta_1.

This rate is called the elasticity of $Y$ with respect to $X$. A slope of $0.7$, for example, means that a 1% increase in $X$ is associated with approximately a 0.7% increase in $Y$, for a sufficiently small change. The word “approximately” matters: the slope gives an exact relationship between changes in logarithms, while translating finite changes back to percentages requires exponentiation.

For interpretation, the model assumptions are that $X$ and $Y$ are positive, that the conditional mean of $\log Y$ is linear in $\log X$ over the range being interpreted, and that the conditional errors have mean zero. For statistical inference, observations are also commonly assumed independent under the sampling design. These conditions support a model-based association; a causal interpretation requires additional design or assumptions and does not follow from the regression slope alone.

From a Predictor Change to an Outcome Change

Suppose the predictor changes from $x$ to $x(1+g)$, where $g$ is the proportional change. For instance, a 10% increase has $g=0.10$, and a 20% decrease has $g=-0.20$. Subtract the fitted log outcomes at the two predictor values:

\begin{aligned} \Delta\log \widehat{Y} &=\beta_1\log\{x(1+g)\}-\beta_1\log x\\ &=\beta_1\log(1+g). \end{aligned}

Exponentiating this difference gives the ratio of the exponentiated fitted log outcomes at the new and original predictor values. Therefore, the exact proportional change on that model-implied scale is:

\frac{\exp\{\widehat{\log Y}\text{ at }x(1+g)\}} {\exp\{\widehat{\log Y}\text{ at }x\}} =(1+g)^{\beta_1}, \qquad \text{exact percent change}=100\left[(1+g)^{\beta_1}-1\right]\%.

The intercept cancels from this comparison. The expression applies when $1+g>0$, so the new predictor value remains positive. It describes a ratio implied by the fitted log-scale relationship; it should not automatically be read as a change in the arithmetic mean of $Y$.

For small changes, $\log(1+g)$ is approximately $g$, and $\exp(z)-1$ is approximately $z$ when $z$ is small. Thus the exact percentage change is approximately $100\beta_1g\%$. In words, multiply the predictor’s percentage change by the elasticity to obtain an approximate outcome percentage change.

Worked Example: A 10% Increase With Positive Elasticity

Suppose a hypothetical analysis of positive measurements gives the fitted model $\widehat{\log Y}=1.10+0.80\log X$. Assume the log-linear relationship is suitable over the predictor range considered, the observations are independent, and the model’s conditional errors have mean zero. What outcome change does the model imply when $X$ increases by 10%?

The elasticity is $0.80$. The quick approximation is $0.80\times 10\%=8.0\%$ higher $Y$. To calculate the exact change, set $g=0.10$ and $\beta_1=0.80$:

100\left[(1+0.10)^{0.80}-1\right] =100\left[1.10^{0.80}-1\right] =100\left[\exp\{0.80\log(1.10)\}-1\right] \approx 7.92\%.

Thus, the exact model-implied change is about 7.92% higher on the exponentiated fitted log scale, close to the 8.0% approximation. The predictor and outcome are positive, and the interpretation is restricted to the range where the assumed log-linear relationship is suitable. This association does not establish that increasing $X$ would cause $Y$ to rise.

Direction, Size, and the Limits of the Approximation

A positive elasticity means the fitted log outcome increases as the predictor increases; a negative elasticity means it decreases. The magnitude describes responsiveness. For example, an elasticity of $-0.4$ indicates a less-than-proportional decrease in the fitted outcome for a small increase in the predictor. An elasticity greater than 1 in magnitude indicates a more-than-proportional response for small changes.

For a finite change, the exact formula is preferable when the predictor shift is not small. The approximation uses a local rate, while the exact calculation accounts for the nonlinear conversion from log changes to percentage changes. The difference may be noticeable for a large predictor change or a large absolute elasticity.

Worked Example: A 50% Decrease With Negative Elasticity

A hypothetical fitted log-log model has slope $\hat{\beta}_1=-0.60$. Assume the predictor and outcome are positive, the log-linear association is suitable across the change being considered, and the observations and model errors meet the conditions described above. What change does the model imply when the predictor is reduced by 50%?

Here $g=-0.50$, so the new predictor is $0.50$ times its original value. The small-change approximation would be $(-0.60)(-50\%)=+30\%$. For the exact calculation:

\begin{aligned} \text{exact percent change} &=100\left[(1-0.50)^{-0.60}-1\right]\\ &=100\left[0.50^{-0.60}-1\right]\\ &=100\left[\exp\{-0.60\log(0.50)\}-1\right]\\ &\approx 100(1.5157-1)\\ &\approx 51.57\%. \end{aligned}

The model implies an increase of about 51.6% on the exponentiated fitted log scale when the predictor is halved. The sign is positive because the elasticity is negative and the predictor decreases. The 30% approximation is not close here: a 50% change is too large to treat the local approximation as a precise calculation.

Worked Example: Comparing a Small Increase With a Doubling

Suppose a hypothetical fitted model has elasticity $\hat{\beta}_1=0.35$. Assume both variables are positive and that the model’s log-linear relationship and error conditions are appropriate over the predictor range in question. Compare the model-implied outcome changes for a 4% predictor increase and for a doubling of the predictor.

For the 4% increase, the approximation is $0.35\times4\%=1.40\%$. The exact value is:

100\left[1.04^{0.35}-1\right] =100\left[\exp\{0.35\log(1.04)\}-1\right] \approx 1.38\%.

The approximation is close for this small shift. A doubling means $g=1$, not $g=0.50$. The approximation would suggest $0.35\times100\%=35\%$, but the exact calculation is:

100\left[(1+1)^{0.35}-1\right] =100\left[2^{0.35}-1\right] \approx 27.46\%.

Thus, doubling $X$ implies about a 27.5% increase on the exponentiated fitted log scale, rather than the 35% suggested by the small-change rule. A doubling is a large change, so the exact calculation is the more appropriate report.

Units and Interpretation

The elasticity is unchanged if either variable is expressed in a different positive unit. If $X$ is converted to a scale $X^*=cX$ for a positive constant $c$, then $\log X^*=\log X+\log c$. This adds a constant to the predictor’s log values, shifting the intercept in the fitted log-log relationship but not changing its slope. Rescaling $Y$ likewise adds a constant to $\log Y$, changing the intercept but not the slope. Consequently, the elasticity has no measurement units, even though the variables themselves do.

This unit invariance does not make the choice of variables or model irrelevant. The predictor and outcome must be positive, and the log-log relationship must be appropriate for the range being interpreted. Also, the model describes an association between $X$ and $Y$; it does not rule out other factors related to both variables.

Common Mistake

Wrong reasoning: “The coefficient is 0.80, so a 10% increase in $X$ produces exactly an 8% increase in $Y$.” This treats the small-change approximation as an exact finite-change rule.
Correction: An elasticity of 0.80 gives an approximate 8% increase for a 10% predictor increase. The exact model-implied change is $100[1.10^{0.80}-1]\approx7.92\%$. For larger changes, calculate the exact percentage using $100[(1+g)^{\beta_1}-1]$.

A second interpretive caution is to avoid translating an observational regression into a causal statement. “A 10% increase in $X$ is associated with an approximately 8% increase in $Y$” reflects the model’s fitted association. “Increasing $X$ will increase $Y$” makes a causal claim that the slope alone cannot justify.

Key takeaway: In a log-log model, the slope is the elasticity: the change in log outcome per change in log predictor. Multiply the predictor’s percentage change by the slope for a small-change approximation; use $100[(1+g)^{\beta_1}-1]\%$ for the exact model-implied change from a finite proportional shift.

Check Your Understanding

Use the elasticity interpretation and exact-change formula to answer these questions.

A log-log model has slope $0.55$. What is the approximate outcome percentage change associated with a 2% increase in the predictor?
For the same slope, write the exact formula for the model-implied outcome percentage change when the predictor increases by 20%. Is the approximation likely to be as accurate as it is for a 2% increase?
A model has elasticity $-0.30$. What direction of outcome change is associated with a small increase in the predictor?
Why does a coefficient of $0.80$ not mean that a 10% predictor increase implies exactly an 8% outcome increase?
What happens to the elasticity when the predictor is converted from dollars to thousands of dollars, and why?