Interpreting Coefficients in Log-Transformed Models

Why Log-Scale Coefficients Need Translation

A regression coefficient always describes a change on the scale used in the fitted equation. When a logarithm appears in that equation, a slope may describe a multiplicative change rather than a fixed increase. The translation depends on which variable was logged: the response, the predictor, or both.

This tutorial uses the natural logarithm, written \(\ln\), and its inverse, the exponential function \(\exp\). Unless stated otherwise, \(X\) and \(Y\) are positive measurements, recorded numerically in specified units. Logarithms of measurements are shorthand for taking the logarithm of their numerical values in those units. The fitted coefficients describe a statistical relationship, not necessarily a causal effect.

Definition: A response-log model regresses \(\ln(Y)\) on \(X\). A predictor-log model regresses \(Y\) on \(\ln(X)\). A log-log model regresses \(\ln(Y)\) on \(\ln(X)\). A coefficient must be interpreted from the equation in which it appears, not from its numerical value alone.

Response-Log Models: A Multiplicative Response Change

A response-log model has the form

\ln(Y)=\beta_0+\beta_1X+\varepsilon

Here \(\beta_0+\beta_1X\) describes the conditional mean of \(\ln(Y)\), and \(\varepsilon\) represents deviations from that mean on the log scale. For the usual least-squares interpretation, the mean relationship in \(\ln(Y)\) should be reasonably linear over the observed predictor range, and the residual diagnostics should not show a systematic pattern that makes the linear model unsuitable. The response must be positive for its logarithm to be defined. These conditions concern the model on the transformed scale; they do not make the relationship automatically appropriate on the original scale.

If \(X\) increases by one unit, the fitted log response changes by \(\beta_1\). Exponentiating that change gives the multiplicative factor \(e^{\beta_1}\) for the modeled geometric-scale response. The exact percentage change in that response is \(100(e^{\beta_1}-1)\%\). When \(\beta_1\) is near zero, \(100\beta_1\%\) is a useful approximation, but it is not generally the exact percentage.

\begin{aligned} \Delta\ln(Y) &= \beta_1 \\ \text{response factor} &= e^{\beta_1} \\ \text{exact percent change} &= 100(e^{\beta_1}-1)\% \end{aligned}

The intercept \(\beta_0\) is the fitted log response at \(X=0\). Exponentiating it gives \(e^{\beta_0}\), the fitted geometric-scale response at that predictor value. This intercept is useful only if \(X=0\) is meaningful in the setting or at least is within a reasonable range for the model. Exponentiating a fitted mean of \(\ln(Y)\) does not, in general, give the arithmetic conditional mean of \(Y\).

Worked Example: Interpreting a Response-Log Slope

In a hypothetical analysis of a positive chemical concentration, let \(Y\) be concentration in milligrams per liter and \(X\) be storage time in days. The fitted response-log model is

\widehat{\ln(Y)}=1.20+0.080X.

Slope. One additional day corresponds to a fitted change of \(0.080\) in \(\ln(Y)\). On the concentration scale, the factor is

e^{0.080}=1.0833.

Thus, for each additional day, the fitted geometric-scale concentration is multiplied by about \(1.0833\), or increases by \(100(1.0833-1)=8.33\%\). The small-slope approximation is \(100(0.080)=8.0\%\), close but not exact.

Intercept. At \(X=0\) days, the fitted log concentration is \(1.20\). Exponentiating gives \(e^{1.20}=3.320\), so the fitted geometric-scale concentration at the start of storage is about \(3.320\) mg/L. This interpretation assumes that zero days is within the relevant setting. It does not establish that storage time causes the observed pattern.

Predictor-Log Models: A Fixed Response Change for a Predictor Factor

A predictor-log model has the form

Y=\beta_0+\beta_1\ln(X)+\varepsilon.

In this model, the conditional mean of \(Y\) is linear in \(\ln(X)\), not in \(X\). The predictor must be positive, and the linear mean relationship should be plausible across the observed range of \(\ln(X)\), as assessed using the model and its residuals. Unlike a response-log model, the response remains on its original scale.

If \(X\) is multiplied by a factor \(c\), where \(c>0\), then \(\ln(X)\) increases by \(\ln(c)\). Therefore the modeled mean response changes by \(\beta_1\ln(c)\) response units. For a doubling, the change is \(\beta_1\ln(2)\); for a tenfold increase, it is \(\beta_1\ln(10)\). The coefficient is not the response change for a one-unit increase in \(X\).

\ln(cX)-\ln(X)=\ln(c) \quad\Longrightarrow\quad \Delta\widehat{Y}=\beta_1\ln(c)

The intercept is the fitted response when \(\ln(X)=0\), which means \(X=1\) in the chosen measurement unit. It is not the response at \(X=0\), because \(\ln(0)\) is undefined.

Worked Example: Interpreting a Predictor-Log Slope

Suppose a hypothetical model relates a plant’s weekly growth, \(Y\) in centimeters, to light exposure \(X\), measured in hours per day:

\widehat{Y}=4.0+12.5\ln(X).

The slope \(12.5\) does not mean that one extra hour of light adds \(12.5\) cm of weekly growth. Instead, doubling the hours per day changes the fitted response by

12.5\ln(2)=12.5(0.6931)=8.664\text{ cm}.

So, wherever the model is appropriate, doubling light exposure corresponds to an increase of about \(8.664\) cm in fitted weekly growth. A 10% increase in light exposure corresponds to \(c=1.10\), giving

12.5\ln(1.10)=12.5(0.09531)=1.191\text{ cm}.

The intercept is \(4.0\) cm at \(X=1\) hour per day, since \(\ln(1)=0\). That interpretation is relevant only if \(X=1\) is a meaningful value in the setting.

Log-Log Models: Elasticity and Power-Law Effects

A log-log model has the form

\ln(Y)=\beta_0+\beta_1\ln(X)+\varepsilon.

Both measurements must be positive. The model describes a linear conditional mean relationship between \(\ln(Y)\) and \(\ln(X)\); its suitability should be checked on that transformed scale. A log-log slope is called an elasticity: it gives the approximate percentage change in the response associated with a 1% increase in the predictor, for small changes. The exact finite-change calculation follows from the equation.

If \(X\) is multiplied by \(c\), the fitted log response changes by \(\beta_1\ln(c)\). Exponentiating gives a response factor \(c^{\beta_1}\). Thus the exact percentage change for that predictor multiplication is \(100(c^{\beta_1}-1)\%\). When \(c\) is close to 1, this is approximately \(100\beta_1(c-1)\%\), where \(c-1\) is the proportional change in \(X\).

Derivation: Compare the fitted log responses at \(X\) and \(cX\). Their difference is

\begin{aligned} \widehat{\ln(Y)}\big|_{cX}-\widehat{\ln(Y)}\big|_X &=\beta_1\ln(cX)-\beta_1\ln(X)\\ &=\beta_1\ln(c). \end{aligned}

Exponentiating this difference gives the ratio \(c^{\beta_1}\) between the fitted geometric-scale responses. Equivalently, the fitted relationship has the power-law form \(e^{\beta_0}X^{\beta_1}\) on that scale.

The log-log intercept \(\beta_0\) is the fitted log response at \(X=1\) in the chosen unit. Its exponent \(e^{\beta_0}\) is the fitted geometric-scale response there. Changing the unit used to record \(X\) changes what numerical value \(X=1\) represents, so the intercept interpretation changes accordingly.

Worked Example: Interpreting a Log-Log Slope

A hypothetical model relates positive package mass \(Y\), in grams, to package volume \(X\), in cubic centimeters:

\widehat{\ln(Y)}=0.50+0.65\ln(X).

The slope \(0.65\) means that a 1% increase in volume is associated with approximately a 0.65% increase in the fitted geometric-scale mass, for small percentage changes. For a doubling, use the exact factor:

2^{0.65}=e^{0.65\ln(2)} =e^{0.65(0.6931)} =e^{0.4505} \approx 1.569.

The fitted geometric-scale mass is therefore multiplied by about \(1.569\), an exact increase of \(100(1.569-1)=56.9\%\) for a doubling of volume. The approximation from elasticity would be \(0.65(100\%)=65\%\), which is less accurate for such a large change.

At \(X=1\) cubic centimeter, the fitted log mass is \(0.50\), and the fitted geometric-scale mass is \(e^{0.50}=1.649\) g. This is the intercept interpretation at the stated unit value, not at zero volume.

Units, Comparisons, and Common Mistakes

The measurement units are part of the interpretation. A predictor-log slope describes the response change for a specified factor in \(X\), so doubling or a 10% increase is often clearer than saying “one unit.” In a response-log model, the predictor’s unit determines what a one-unit increase means. In a log-log model, the slope is unchanged by multiplying \(X\) or \(Y\) by a fixed unit-conversion constant, but the intercept’s meaning depends on those units.

If base-10 logarithms are used instead of natural logarithms, the same reasoning applies with powers of 10 in place of powers of \(e\). Always check the log base used in the fitted equation before translating a coefficient.

Model	Change being considered	Fitted response interpretation
\(\ln(Y)=\beta_0+\beta_1X\)	\(X\) increases by one unit	Multiply geometric-scale \(Y\) by \(e^{\beta_1}\)
\(Y=\beta_0+\beta_1\ln(X)\)	\(X\) is multiplied by \(c\)	Change fitted \(Y\) by \(\beta_1\ln(c)\) response units
\(\ln(Y)=\beta_0+\beta_1\ln(X)\)	\(X\) is multiplied by \(c\)	Multiply geometric-scale \(Y\) by \(c^{\beta_1}\)

A frequent mistake is to read a response-log coefficient of \(0.08\) as an exact 8% increase. The exact factor is \(e^{0.08}\), giving an 8.33% increase; 8% is only the small-coefficient approximation. Another mistake is to read a predictor-log coefficient as the response change for a one-unit increase in \(X\). Its interpretation requires a multiplicative change in \(X\), such as doubling. In a log-log model, calling the slope “the percentage change” without naming both the predictor change and the approximate or exact calculation also leaves the interpretation incomplete.

Finally, regression describes a fitted association. Even a clear multiplicative interpretation does not by itself show that changing \(X\) would cause \(Y\) to change. Study design and other evidence determine whether a causal conclusion is justified.

Key takeaway: First identify which variable or variables are logged. Then translate the coefficient on the scale used by the model: exponentiate a response-log slope for a response factor, multiply a predictor-log slope by \(\ln(c)\) for a predictor factor \(c\), and use \(c^{\beta_1}\) for a log-log response factor.

Check Your Understanding

For each question, state the relevant scale and include units or a percentage interpretation where appropriate.

A response-log model has slope \(0.03\). What is the exact percentage change in fitted geometric-scale response for a one-unit increase in the predictor, and what is the small-slope approximation?
In a predictor-log model, the slope is \(-6\) meters. What is the fitted response change when the predictor is multiplied by 3?
A log-log model has slope \(1.4\). What response factor corresponds to doubling the predictor? Is the response percentage change exactly 140%?
In a predictor-log model, what predictor value does the intercept describe, and why is it not zero?
Why does a regression coefficient’s multiplicative interpretation not, by itself, establish causation?