Interpreting Level-Log Models

Why a Logged Predictor Changes the Interpretation

A one-unit increase can mean very different things depending on where the predictor starts. For example, a 10-unit increase from 10 to 20 doubles a predictor, while the same increase from 100 to 110 raises it by only 10%. When the predictor is logged, the model describes outcome changes in relation to proportional changes in that predictor. This is useful when relative changes are more meaningful than fixed-unit changes.

A level-log model keeps the outcome in its original units and uses the logarithm of a positive predictor. Its slope does not give the outcome change for a one-unit increase in the predictor. Instead, it gives the outcome change associated with a one-unit increase in the log of the predictor, which corresponds to multiplying the predictor by a factor of $e$. Derivatives describe the local change, and a log-ratio calculation gives the exact change between any two predictor values.

The Model and Its Assumptions

Definition: A level-log model has a quantitative outcome $Y$ in its original units and a positive predictor $X$ entered as a logarithm. Write its conditional mean as $$ E(Y\mid X=x)=\beta_0+\beta_1\ln\left(\frac{x}{x_{\mathrm{ref}}}\right), \qquad x>0, $$ where $x_{\mathrm{ref}}>0$ is a fixed reference value in the same units as $X$. Equivalently, the model for observation $i$ is $$ Y_i=\beta_0+\beta_1\ln\left(\frac{X_i}{x_{\mathrm{ref}}}\right)+\varepsilon_i, $$ where $E(\varepsilon_i\mid X_i)=0$. The coefficient $\beta_1$ is measured in outcome units per unit of log ratio.

The ratio $x/x_{\mathrm{ref}}$ is dimensionless, and the reference value makes the intercept easy to interpret: when $x=x_{\mathrm{ref}}$, the log term is zero and the model gives $E(Y\mid X=x_{\mathrm{ref}})=\beta_0$. Choosing a different reference value changes the intercept's reference point, but not the slope or the predicted difference between two predictor values.

The interpretation relies on the conditional mean being adequately represented as linear in $\ln(x/x_{\mathrm{ref}})$ across the predictor range of interest. The predictor must be positive, and the observations used to estimate the model must provide reasonable support for the comparisons being made. The condition $E(\varepsilon_i\mid X_i)=0$ means the model's errors average to zero at each predictor value. Independent observations are needed for standard regression uncertainty calculations; additional assumptions may be required for particular inference procedures. These conditions do not, by themselves, show that changing $X$ causes a change in $Y$.

From the Slope to a Proportional Change

Let $m(x)=E(Y\mid X=x)$. Differentiating the conditional mean with respect to $x$ gives its instantaneous rate of change in outcome units per unit of $X$. The chain rule yields

\frac{dm(x)}{dx} = \beta_1\frac{d}{dx}\ln\left(\frac{x}{x_{\mathrm{ref}}}\right) = \frac{\beta_1}{x}.

Thus the marginal change depends on the starting value $x$. For a positive slope, the predicted increase per additional predictor unit gets smaller as $x$ grows; for a negative slope, the predicted decrease per additional predictor unit gets smaller in magnitude as $x$ grows. The derivative is a local rate, not necessarily the exact change over a finite interval.

For an exact comparison, take two positive predictor values, $x_1$ and $x_2$, and subtract their conditional means. The intercept and reference value cancel:

m(x_2)-m(x_1) = \beta_1\left[ \ln\left(\frac{x_2}{x_{\mathrm{ref}}}\right) - \ln\left(\frac{x_1}{x_{\mathrm{ref}}}\right) \right] = \beta_1\ln\left(\frac{x_2}{x_1}\right).

If $x_2$ is a proportion $r$ greater than $x_1$, then $x_2=(1+r)x_1$, so the exact change is $\beta_1\ln(1+r)$. Here $r=0.10$ means a 10% increase, and $r=-0.10$ means a 10% decrease. The multiplier must be positive, so a proportional decrease must be less than 100%.

Theorem: In a level-log model, changing $X$ from $x_1$ to $x_2$ changes the model's conditional mean by exactly $\beta_1\ln(x_2/x_1)$ outcome units. For a proportional increase of $r$, the exact change is $\beta_1\ln(1+r)$. For a small proportional change, $\ln(1+r)\approx r$, so the approximate change is $\beta_1 r$ outcome units.

The approximation follows from the derivative of $\ln(1+r)$ at $r=0$, which is 1. In other words, near zero, $\ln(1+r)$ changes approximately as fast as $r$. The approximation is useful for small percentages, but the exact logarithmic calculation is straightforward and should be used when the change is not small.

Worked Example: A 50% Increase in a Predictor

Suppose a hypothetical analysis of positive monthly advertising expenditure $X$, measured in hundreds of dollars, and sales $Y$, measured in thousands of dollars, gives the fitted conditional mean $ \hat{m}(x)=20+6\ln(x/10) $. The hypothetical observations cover expenditures from 10 to 40 hundred dollars, the log-linear mean pattern is considered adequate over that range, and the comparison below lies within it. Interpret the fitted change when expenditure goes from 20 to 30 hundred dollars.

The multiplier is $30/20=1.5$, so expenditure rises by 50%. The exact difference in fitted mean sales is

\hat{m}(30)-\hat{m}(20) = 6\ln\left(\frac{30}{20}\right) = 6\ln(1.5) = 6(0.4055) = 2.433.

The model therefore predicts an increase of about 2.433 thousand dollars in mean sales for this change in expenditure. As a check, the fitted values are $\hat{m}(20)=20+6\ln(2)=24.159$ thousand dollars and $\hat{m}(30)=20+6\ln(3)=26.592$ thousand dollars; their difference is approximately 2.433 thousand dollars. This is a model-based association, not evidence by itself that increasing advertising causes sales to rise.

Worked Example: A Small Percentage Increase

Suppose a hypothetical fitted model relates a positive training measure $X$, in hours per week, to a performance score $Y$, in points: $ \hat{m}(x)=70+4\ln(x/5) $. Assume the hypothetical data support this log-linear mean pattern from 5 to 20 hours per week. Find the exact and approximate fitted outcome changes for a 1% increase from 10 hours per week.

A 1% increase takes $X$ from 10 to $10(1.01)=10.1$ hours per week. The exact change is

\hat{m}(10.1)-\hat{m}(10) = 4\ln\left(\frac{10.1}{10}\right) = 4\ln(1.01) = 4(0.009950) = 0.03980\text{ points}.

Using the small-change approximation, $r=0.01$, so the approximate change is $4(0.01)=0.04000$ points. The approximation is close to the exact result because 1% is a small proportional change. The units are points, not points per hour: this is the predicted outcome difference for the specified percentage increase.

The derivative at 10 hours per week is $4/10=0.4$ points per hour. Multiplying that local rate by the 0.1-hour increase gives another approximation, $0.4(0.1)=0.04000$ points. The exact log calculation remains $0.03980$ points.

Doubling, Halving, and Other Multipliers

A fixed multiplicative change has the same fitted outcome difference wherever it starts. For example, doubling from $x$ to $2x$ always changes the model's conditional mean by $\beta_1\ln(2)$. This holds because the ratio of the new predictor value to the old one is always 2. Likewise, a halving has the same magnitude with the opposite sign, since $\ln(1/2)=-\ln(2)$.

This does not mean every equal-sized absolute change has the same effect. Moving from 10 to 20 is a doubling, whereas moving from 20 to 30 is only a 50% increase. The model predicts different outcome changes for these comparisons because their ratios differ.

Worked Example: Interpreting a Doubling and a 10% Decrease

Suppose a hypothetical model for a positive pollutant concentration $X$, in micrograms per cubic meter, and a respiratory symptom score $Y$, in points, is $ \hat{m}(x)=12+3\ln(x/5) $. Assume the hypothetical observations range from 2 to 20 micrograms per cubic meter, the log-linear mean pattern is adequate in that range, and the comparisons are within the observed range.

For a doubling from 5 to 10 micrograms per cubic meter, the fitted change is

\hat{m}(10)-\hat{m}(5) = 3\ln\left(\frac{10}{5}\right) = 3\ln(2) = 3(0.6931) = 2.079\text{ points}.

Thus a doubling is associated with a fitted increase of about 2.079 points in the mean symptom score. The fitted means confirm the result: $\hat{m}(5)=12$ points and $\hat{m}(10)=14.079$ points.

For a 10% decrease from 10 to 9 micrograms per cubic meter, $r=-0.10$ and the multiplier is $0.90$. The exact fitted change is

\hat{m}(9)-\hat{m}(10) = 3\ln\left(\frac{9}{10}\right) = 3\ln(0.90) = 3(-0.1054) = -0.3161\text{ points}.

The model predicts a decrease of about 0.3161 points in the mean score for this 10% reduction. The estimate describes an association under the stated model; it does not establish that reducing concentration would cause the score to change.

What the Coefficient Does and Does Not Say

The slope $\beta_1$ is not the predicted outcome change for a one-unit increase in $X$. Its direct interpretation is the change in conditional mean for a one-unit increase in $\ln(X/x_{\mathrm{ref}})$. Equivalently, multiplying $X$ by $e$ changes the conditional mean by $\beta_1$ outcome units, because $\ln(e)=1$. For more familiar changes, calculate $\beta_1\ln(1+r)$ for a proportional increase $r$, or $\beta_1\ln(x_2/x_1)$ for two specified values.

A 1% increase is approximately associated with a change of $0.01\beta_1$ outcome units, but this is an approximation. The exact change is $\beta_1\ln(1.01)$. Similarly, for a 10% increase the exact multiplier on $\beta_1$ is $\ln(1.10)=0.09531$, not exactly 0.10. A negative coefficient reverses the direction of each predicted change but does not alter the formulas.

Because the formula depends on a ratio, converting $X$ to another unit does not alter the proportional comparisons. If both $x_1$ and $x_2$ are expressed in a new unit, their ratio remains $x_2/x_1$. The reference value helps define the intercept, while the exact change between predictor values depends only on their ratio.

Common Mistake: Treating the Slope as a Per-Unit Effect

A common error is to say, “The coefficient is 4, so one additional hour of $X$ increases the outcome by 4 units.” This applies the slope as though the model were linear in $X$, but the model is linear in $\ln(X)$. In a level-log model, the per-unit marginal change is $\beta_1/x$, so it varies with the starting value of $X$.

The correction is to specify a meaningful comparison. For a 20% increase, use $\beta_1\ln(1.20)$; for a move from $x_1$ to $x_2$, use $\beta_1\ln(x_2/x_1)$. Also check that both values are positive and that the comparison is supported by the data. A correctly calculated fitted difference remains an association unless the study design and assumptions justify a causal interpretation.

Key takeaway: A level-log model translates proportional changes in a positive predictor into additive changes in the outcome. The exact change from $x_1$ to $x_2$ is $\beta_1\ln(x_2/x_1)$; the derivative $\beta_1/x$ describes the local change per predictor unit at $x$.

Check Your Understanding

Answer each question using the level-log formulas and include outcome units in your interpretations.

In a model with slope $\beta_1=-5$, what is the exact fitted outcome change for a 20% increase in $X$? Write the formula and describe the direction.
Explain why a doubling from 10 to 20 and a doubling from 50 to 100 have the same fitted outcome difference in a level-log model.
For a level-log model with slope 8, find the exact change for a decrease from $X=40$ to $X=36$. What proportional decrease does this represent?
A student says that a slope of 2 means each one-unit increase in $X$ raises the fitted outcome by 2 units. Explain why this interpretation is incorrect.
For a slope of 6, compare the exact change for a 1% increase with the small-change approximation. State both calculations.

What You'll Learn