Interpreting Log Revenue Models for Growing Firms

Why Model Revenue on a Log Scale?

Revenue among firms can vary greatly: a small change in predictors may correspond to a modest dollar difference for one firm and a much larger difference for another. A model for log revenue instead describes proportional changes. This can be useful when revenue is positive and its relationship with predictors is more naturally expressed as percentage differences than as constant dollar differences.

Logging the outcome changes how coefficients and predictions must be explained. A coefficient is not a dollar change in revenue. Exponentiating a fitted log value also does not, in general, give the predicted arithmetic mean revenue. Those distinctions matter when communicating what the model says.

The Log-Revenue Regression Model

Let $R_i$ denote the revenue of firm $i$, recorded as a positive number in millions of dollars. For example, $R_i=4$ represents revenue of $4 million. Define $Z_i=\ln(R_i)$, where $\ln$ is the natural logarithm. A multiple linear regression for log revenue can be written as

Z_i=\beta_0+\beta_1x_{i1}+\cdots+\beta_px_{ip}+\varepsilon_i.

Here $x_{ij}$ is predictor $j$ for firm $i$, $\beta_0,\ldots,\beta_p$ are population regression coefficients, and $\varepsilon_i$ is the error. The fitted value $\hat{\eta}_i=\hat{\beta}_0+\hat{\beta}_1x_{i1}+\cdots+\hat{\beta}_px_{ip}$ is on the log-revenue scale. The observed revenue must be positive for its logarithm to be defined.

Model assumptions: The conditional mean of log revenue is linear in the included predictors; errors have conditional mean zero and constant variance; and errors from different firms are independent. Normal errors are additionally assumed when using normal-theory inference or the lognormal mean correction derived below. These are assumptions about the model and data, not automatic consequences of taking logarithms. In particular, residual patterns should be assessed on the log scale.

A coefficient describes the change in the model's conditional mean of log revenue for a one-unit increase in its predictor, holding the other included predictors fixed. The coefficient itself is measured in log units. Exponentiating a coefficient or a coefficient contrast gives a multiplicative comparison on the revenue scale.

From Log-Scale Coefficients to Percentage Changes

Suppose one predictor $x_j$ increases by $\Delta$, while the other predictors stay fixed. The fitted log revenue changes by $\hat{\beta}_j\Delta$. Exponentiating this difference gives a revenue ratio of $\exp(\hat{\beta}_j\Delta)$. Thus the corresponding percentage change is

100\left[\exp(\hat{\beta}_j\Delta)-1\right]\%.

For a small value of $\hat{\beta}_j\Delta$, the approximation $\exp(u)-1\approx u$ gives $100\hat{\beta}_j\Delta\%$. The approximation is convenient for small effects, but it can differ noticeably from the exact percentage change when the log-scale contrast is larger. If several predictors change, use the sum of their coefficient changes in the exponent.

Worked Example: A One-Unit Increase in Marketing Spending

Suppose a hypothetical model is

\widehat{\ln(R)}=1.20+0.045A+0.18M,

where $A$ is firm age in years and $M$ is annual marketing spending measured in units of $100,000. Consider firms with the same age, but with $M$ differing by one unit—that is, by $100,000.

Check the model conditions relevant to interpretation. Revenue is positive, so $\ln(R)$ is defined. The fitted equation specifies a linear conditional mean in $A$ and $M$. For the usual regression interpretation and inference, the error assumptions stated above must be reasonable; this hypothetical calculation alone cannot establish that they hold.

The fitted log-revenue difference is $0.18(1)=0.18$. The approximate percentage difference is $100(0.18)=18\%$. The exact model-based percentage difference is

100[\exp(0.18)-1] =100(1.1972-1) \approx 19.72\%.

Thus the model associates an additional $100,000 in annual marketing spending with about 19.7% higher revenue, comparing firms of the same age. This is an adjusted association, not evidence by itself that increasing spending causes revenue to rise.

Comparing Firms with More Than One Predictor Difference

When two firms differ on several predictors, the fitted log-revenue difference is the sum of the relevant coefficient contrasts. Exponentiating that sum gives the fitted revenue ratio. The order in which the predictor differences are considered does not change the result: multiplication of the corresponding revenue ratios is equivalent to adding their log contrasts.

Worked Example: Combining Age and Marketing Differences

Use the fitted model $\widehat{\ln(R)}=1.20+0.045A+0.18M$ from the previous example. Compare a firm that is two years older and has marketing spending $200,000 higher with an otherwise comparable firm. Assume the predictor changes are within the range represented by the data and that the model's form is appropriate for this comparison.

The age difference contributes $0.045(2)=0.090$ log units. The marketing difference is two $100,000 units, contributing $0.18(2)=0.360$ log units. The total fitted log difference is

$$ 0.090+0.360=0.450. $$

The fitted revenue ratio is $\exp(0.450)\approx1.568$. Therefore the model-implied percentage difference is

100[\exp(0.450)-1] \approx100(1.568-1) =56.8\%.

Using the small-change approximation would give $100(0.450)=45.0\%$, which is substantially lower than the exact result. Under the model, the firm with both higher age and higher marketing spending has fitted revenue about 56.8% higher, conditional on the other model features. This comparison remains an association, not a causal effect.

Exponentiating a Fitted Value Is Not Always a Mean Prediction

For a firm with specified predictors, write the model as $Z=\eta+\varepsilon$, where $\eta$ is the conditional mean of log revenue. Then revenue is $R=\exp(\eta+\varepsilon)$. The exponentiated fitted value $\exp(\hat{\eta})$ is often useful as a typical or central prediction on the revenue scale. But because the exponential function is curved, it is generally not the arithmetic mean of revenue among firms with those predictor values.

Under normally distributed errors with constant variance $\sigma^2$, revenue conditional on the predictors has a lognormal distribution. The conditional median is $\exp(\eta)$, while the conditional arithmetic mean is larger. The distinction follows directly from the normal moment-generating function.

Theorem: If $Z=\ln(R)=\eta+\varepsilon$ and $\varepsilon\mid x\sim N(0,\sigma^2)$, then $R\mid x$ is lognormal, with conditional median $\exp(\eta)$ and conditional arithmetic mean $\exp(\eta+\sigma^2/2)$. The median and mean are different when $\sigma^2>0$.

To justify the mean result, use $R=\exp(\eta)\exp(\varepsilon)$. Since $\eta$ is fixed for the specified predictors,

E(R\mid x) =\exp(\eta)E[\exp(\varepsilon)\mid x].

For a normal random variable with mean zero and variance $\sigma^2$, $E[\exp(\varepsilon)]=\exp(\sigma^2/2)$. Substitution gives

E(R\mid x) =\exp(\eta)\exp(\sigma^2/2) =\exp(\eta+\sigma^2/2).

This is the lognormal correction. In practice, $\eta$ and $\sigma^2$ are unknown, so fitted values and the estimated residual variance are used. The resulting arithmetic-mean prediction relies on the normal, constant-variance error model being reasonable. If the distribution of errors changes with predictors, a single variance correction need not adequately describe mean revenue across firms.

Worked Example: Returning from a Fitted Log Value to Revenue

Suppose a hypothetical firm's fitted log revenue is based on the model

\widehat{\ln(R)} =1.80+(0.080)(2)-(0.012)(5),

where the predictors have values $x_1=2$ and $x_2=5$, and $R$ is in millions of dollars. Suppose the estimated conditional error variance on the log scale is $0.160$. For this calculation, assume the log errors are approximately normal with constant variance across the relevant firms.

First compute the fitted log value:

$$ 1.80+0.160-0.060=1.90. $$

Exponentiating gives

\exp(1.90)\approx6.686.

This is about $6.686 million. Under the normal-error model it is the fitted conditional median revenue, not the conditional arithmetic mean. The variance correction is $\hat{\sigma}^2/2=0.160/2=0.080$, so the estimated arithmetic mean is

\exp(1.90+0.080) =\exp(1.98) \approx7.243.

The corrected estimate is about $7.243 million. It exceeds the median-scale prediction because a right-skewed lognormal distribution has a mean above its median. The calculation uses the stated fitted log value and variance; without the correction, reporting $6.686 million as the predicted arithmetic mean would understate the model-based mean.

How to Communicate the Result

A clear report identifies the outcome scale, the predictor contrast, the comparison being made, and whether the number describes a log-scale change, a multiplicative ratio, a median, or a mean. State revenue units explicitly. For example, “The fitted median revenue is $6.686 million” and “The estimated conditional mean revenue is $7.243 million under the lognormal error model” describe different quantities.

For a predictor contrast $\Delta$, the exact ratio $\exp(\hat{\beta}_j\Delta)$ applies to the model's conditional log location. If the error distribution is the same across the compared predictor values, it also gives the ratio of conditional arithmetic means, because the common factor $E[\exp(\varepsilon)\mid x]$ cancels. With normal errors and constant variance, this condition holds. If the residual distribution or variance differs, the mean ratio may not equal the exponentiated log contrast.

Key takeaway: A log-outcome coefficient gives a change in conditional mean log revenue. Exponentiating a coefficient contrast gives a multiplicative comparison; convert that ratio to a percentage with $100[\exp(\text{contrast})-1]\%$. For an arithmetic-mean revenue prediction under normal, constant-variance log errors, include the variance correction.

Common Mistakes

Calling a log coefficient a dollar change. A coefficient such as $0.08$ is in log units, not millions of dollars. Exponentiate the relevant contrast to obtain a revenue ratio.
Using the approximation for a large contrast. The approximation $100u\%$ can differ substantially from the exact $100[\exp(u)-1]\%$. Calculate both when the distinction matters.
Confusing the exponentiated fitted value with the arithmetic mean. Under normal log errors, $\exp(\hat{\eta})$ is a median-scale prediction; the mean prediction includes the variance correction.
Interpreting association as causation. Adjusting for included predictors does not remove all possible confounding or establish that changing a predictor would change revenue.

Check Your Understanding

Use the definitions and calculations in this tutorial to answer the following.

A log-revenue coefficient is $0.06$ per $100,000 of spending. What is the approximate percentage difference for that increment, and what is the exact percentage difference?
Why does a fitted value of $\widehat{\ln(R)}=2$ not mean that fitted revenue is $2 million?
For a log contrast of $0.30$, calculate the fitted revenue ratio and exact percentage difference.
Under normal, constant-variance log errors with variance $0.25$, how does the conditional arithmetic mean relate to the conditional median?
What additional caution is needed before describing a predictor's adjusted association with revenue as a causal effect?