Interpreting Log-Level Models

Why a Logged Outcome Changes the Interpretation

In an ordinary level-level regression, a slope describes an additive change in the outcome for a one-unit change in the predictor. When the outcome is logged instead, the predictor is still untransformed, but its linear association is with the logarithm of the outcome. Translating that association back to the outcome scale gives a multiplicative comparison: a specified predictor change corresponds to multiplying the model’s fitted outcome-scale value by a particular factor.

That factor can be converted into an exact percentage change. The distinction matters because, for example, a model coefficient of 0.04 does not mean an exact 4% increase for a one-unit predictor increase. Four percent is a useful approximation when the coefficient is small; the exact percentage follows from exponentiating the coefficient.

The Log-Level Model

Definition: A log-level regression model has a logged outcome and an untransformed predictor. In a simple model, write $$ \ln(Y_i)=\beta_0+\beta_1x_i+\varepsilon_i, $$ where $Y_i$ is a positive outcome, $x_i$ is the predictor in its original units, $\beta_0$ is the intercept, $\beta_1$ is the slope, and $\varepsilon_i$ is the error on the log scale. For interpreting the fitted log relationship, assume that the conditional mean error is zero, so $E(\varepsilon_i\mid x_i)=0$, and that the mean of $\ln(Y_i)$ is linear in $x_i$ over the range being interpreted.

Because the logarithm is used, the outcome must be positive. The usual convention here is the natural logarithm, $\ln$. The predictor is not logged: if $x$ is measured in years, a one-unit increase means one year; if it is measured in kilometres, a one-unit increase means one kilometre. State those units whenever interpreting the slope.

The coefficient $\beta_1$ describes the change in the fitted log outcome for a one-unit increase in $x$. More generally, if the predictor changes from $x_0$ to $x_1$, let $\Delta x=x_1-x_0$. The difference between the two fitted log outcomes is $\beta_1\Delta x$. Exponentiating this difference gives the ratio of the corresponding exponentiated fitted log values:

\frac{\exp(\beta_0+\beta_1x_1)} {\exp(\beta_0+\beta_1x_0)} =\exp\!\left(\beta_1(x_1-x_0)\right) =\exp(\beta_1\Delta x).

Thus, $\exp(\beta_1\Delta x)$ is the model-implied multiplicative factor for that predictor change. The intercept cancels from the ratio. To express the factor as an exact percentage change, subtract 1 and multiply by 100:

\text{Exact percentage change} =100\left[\exp(\beta_1\Delta x)-1\right]\%.

A positive value means an increase; a negative value means a decrease. This calculation compares exponentiated fitted log values. It does not, by itself, establish a percentage change in the arithmetic conditional mean of $Y$. Nor does a regression association alone establish a causal effect.

Worked Interpretations

Worked Example: A One-Year Increase in Age

Suppose a hypothetical regression for a positive monthly expense $Y$, measured in dollars, is $\ln(Y_i)=\beta_0+0.025x_i+\varepsilon_i$, where $x_i$ is age in years. Assume the log-linear mean relationship is appropriate over the ages being compared and the conditional mean error is zero. Compare the model-implied outcome-scale values for people whose ages differ by four years.

The response is positive, as required for taking its logarithm, and the predictor remains age measured in years. Here $\Delta x=4$, so the fitted log outcome changes by $0.025(4)=0.100$. The multiplicative factor is $\exp(0.100)\approx1.1052$.

The exact percentage change is $100[\exp(0.025(4))-1]\%=100(1.1052-1)\%\approx10.52\%$. Thus, the model-implied exponentiated fitted value for a four-year increase is about 1.105 times as large, or about 10.5% higher. This is an association described by the fitted model, not evidence that aging causes the expense to change.

For comparison, the small-change approximation gives $100(0.025)(4)=10\%$. It is close here but not exact; the exponentiation is needed for the exact percentage.

Worked Example: A Negative Slope and a Decrease

Consider a hypothetical model for a positive laboratory measurement $Y$ with an untransformed exposure index $x$: $\ln(Y_i)=\beta_0-0.12x_i+\varepsilon_i$. Suppose the fitted log relationship is appropriate across the three-unit comparison, and $E(\varepsilon_i\mid x_i)=0$. What model-implied percentage change corresponds to increasing the exposure index by 3 units?

The predictor change is $\Delta x=3$, and the fitted log outcome changes by $(-0.12)(3)=-0.36$. Therefore the outcome-scale factor is $\exp(-0.36)\approx0.6977$.

Converting the factor to a percentage gives $100[\exp(-0.36)-1]\%=100(0.6977-1)\%\approx-30.23\%$. The model-implied exponentiated fitted value is therefore about 0.698 times its comparison value, a decrease of about 30.2% for a 3-unit increase in the index.

It would be inaccurate to report an exact 36% decrease by multiplying $-0.12$ by 3 and converting directly to a percentage. That product is the change on the log scale; exponentiating it produces the exact outcome-scale factor. The outcome is positive and the comparison is within the stated range of model use, but the association should not be described as causal without an appropriate causal design and assumptions.

Worked Example: A Predictor Measured in Tens

Suppose a hypothetical model for a positive response $Y$ uses a predictor $x$ measured in tens of kilometres: $\ln(Y_i)=\beta_0+0.06x_i+\varepsilon_i$. Assume the outcome is positive, the log-linear relationship is suitable for the comparison, and the conditional mean error is zero. Find the exact percentage change associated with an increase of 15 kilometres.

Because one unit of $x$ represents 10 kilometres, an increase of 15 kilometres is $\Delta x=1.5$, not 15. The change in the fitted log outcome is $0.06(1.5)=0.09$. The outcome-scale factor is $\exp(0.09)\approx1.0942$.

The exact percentage change is $100[\exp(0.06(1.5))-1]\%=100(1.0942-1)\%\approx9.42\%$. The model therefore implies that the exponentiated fitted value is about 9.42% higher for a 15-kilometre increase. The predictor’s scale is essential: treating the 15 kilometres as 15 units would incorrectly calculate $\exp(0.06(15))$.

Why the Exponentiation Gives the Exact Change

The conversion follows directly from the model, rather than from a special rule to memorize. At predictor value $x_0$, the fitted log outcome is $\beta_0+\beta_1x_0$; at $x_1$, it is $\beta_0+\beta_1x_1$. Subtracting the first fitted value from the second removes the intercept:

(\beta_0+\beta_1x_1)-(\beta_0+\beta_1x_0) =\beta_1(x_1-x_0) =\beta_1\Delta x.

Since the logarithm is the inverse of the exponential function, exponentiating a difference of logarithms gives the ratio of the corresponding exponentiated values. If $m_0=\exp(\beta_0+\beta_1x_0)$ and $m_1=\exp(\beta_0+\beta_1x_1)$, then

\ln(m_1)-\ln(m_0)=\beta_1\Delta x \quad\Longrightarrow\quad \ln\!\left(\frac{m_1}{m_0}\right)=\beta_1\Delta x \quad\Longrightarrow\quad \frac{m_1}{m_0}=\exp(\beta_1\Delta x).

The definition of percentage change is $100(m_1-m_0)/m_0$. Rewriting it as $100(m_1/m_0-1)$ and substituting the ratio gives $100[\exp(\beta_1\Delta x)-1]\%$, as claimed. This derivation concerns the exponentiated fitted log values; it does not require interpreting those values as arithmetic conditional means of the original outcome.

Approximations, Units, and Scope

For a small value of $\beta_1\Delta x$, the exponential function is approximately $1+\beta_1\Delta x$. Consequently, $100[\exp(\beta_1\Delta x)-1]\%$ is approximately $100\beta_1\Delta x\%$. This is a convenient estimate, not the exact answer. It becomes less reliable as the absolute value of $\beta_1\Delta x$ grows. For an exact interpretation, retain the exponential.

The same coefficient also produces different percentage comparisons for different predictor changes. A one-unit increase has factor $\exp(\beta_1)$, while a five-unit increase has factor $\exp(5\beta_1)$. A decrease of the same size has the reciprocal factor: changing $x$ by $-\Delta x$ gives $\exp(-\beta_1\Delta x)=1/\exp(\beta_1\Delta x)$. Percentage increases and decreases are therefore not symmetric. For example, multiplying by 1.10 is a 10% increase, while reversing that comparison multiplies by $1/1.10$, which is about a 9.09% decrease.

Changing the units of the predictor changes the numerical slope: a coefficient per kilometre differs from one per 10 kilometres. The product $\beta_1\Delta x$, and hence the comparison for the same physical change, remains the same when the units are converted consistently. Changing the units used to record the positive outcome shifts the log scale by a constant and affects the intercept, but not the slope-based ratio for a predictor change.

Common Mistakes

Reading the slope as an exact percentage. A slope of 0.04 is a 0.04-unit change in fitted log outcome per predictor unit. For one unit, the exact percentage is $100[\exp(0.04)-1]\%\approx4.08\%$, not exactly 4%.
Using the wrong predictor increment. First express $\Delta x$ in the units used by the model. If one model unit is 10 kilometres, a 15-kilometre increase means $\Delta x=1.5$.
Forgetting the minus one. $\exp(\beta_1\Delta x)$ is a ratio or multiplier. Subtract 1 and multiply by 100 to obtain a percentage change.
Calling the association causal. A fitted regression coefficient describes a model-based association. A causal claim needs suitable study design and assumptions beyond the log-level form.
Claiming an arithmetic-mean change from the exponentiated fitted log values. The calculation here compares exponentiated fitted log values. It does not by itself give an arithmetic-mean comparison on the original outcome scale.

Key takeaway: For a predictor change $\Delta x$, a log-level slope $\beta_1$ gives the model-implied outcome-scale factor $\exp(\beta_1\Delta x)$. The exact percentage change is $100[\exp(\beta_1\Delta x)-1]\%$; state the predictor units and describe the result as a model-based association.

Check Your Understanding

Use the exponential conversion for exact percentage changes, and keep the predictor’s units explicit.

A fitted log-level slope is $0.03$ per year. What is the exact percentage change in the exponentiated fitted value for a 2-year increase?
A model has slope $-0.08$ per index unit. Calculate the exact percentage change for a 4-unit increase and state whether it is an increase or decrease.
A predictor is measured in 5-kilogram units, and its slope is $0.02$ per model unit. What is $\Delta x$ for a 15-kilogram increase, and what expression gives the exact percentage change?
Explain why $\exp(\beta_1\Delta x)$ is not itself a percentage change.
What additional claim would be unsupported if a researcher described a model-based association as proof that changing the predictor causes the outcome to change?

What You'll Learn