Exponentiated Predictions as Geometric Means and Medians

What Does an Exponentiated Log Prediction Represent?

A regression model for $\ln(Y)$ predicts on the log scale. To express that prediction in the units of $Y$, we exponentiate it. The result is easy to mislabel: in general, it is not automatically the conditional arithmetic mean of $Y$. Its direct interpretation is tied to the mean of the log outcome, and therefore to a geometric mean.

Suppose $Y$ is a positive outcome and $X$ denotes the predictor or predictors. Write the log-outcome model as

\ln(Y)=m(X)+\varepsilon, \qquad E(\varepsilon\mid X)=0.

Here $m(X)$ describes the conditional mean of $\ln(Y)$, and $\varepsilon$ is the difference between an observation’s log outcome and that conditional mean. The zero conditional-mean assumption implies $E[\ln(Y)\mid X=x]=m(x)$. A fitted regression supplies an estimate $\hat m(x)$ of that log-scale quantity.

Definition: The conditional geometric mean of a positive outcome at $X=x$ is $\exp\{E[\ln(Y)\mid X=x]\}$. Under the model above, it is $\exp\{m(x)\}$. The exponentiated fitted value $\exp\{\hat m(x)\}$ estimates this conditional geometric mean.

This is the conditional version of the geometric mean familiar from a set of positive numbers: take their logarithms, average those logarithms, and exponentiate the result. In a regression model, the averaging is over the conditional distribution of outcomes at a specified predictor value. The prediction describes a multiplicative center on the outcome scale, not a prediction that every individual outcome will equal that value.

From a Fitted Value to an Outcome-Scale Prediction

Exponentiation reverses the natural logarithm. If a fitted model gives $\hat m(x)=c$, the back-transformed prediction is $e^c$. Units require care: if $Y$ was recorded in dollars, $e^c$ is in dollars; if it was recorded in hundreds of dollars, $e^c$ is in hundreds of dollars. Changing the outcome’s measurement scale shifts its logarithm, so the scale used to fit and interpret the model must be stated.

The model also requires $Y>0$, since the natural logarithm is defined only for positive values. A log-outcome regression is appropriate only when the outcome can be expressed as positive measurements on the scale used in the model. The regression form for $m(x)$ must be suitable for the predictor values being interpreted, and the conditional mean-zero assumption is what connects that fitted log-scale relationship to the conditional mean of $\ln(Y)$.

Worked Example: Back-Transforming a Fitted Log Prediction

Consider a hypothetical model for weekly sales $Y$, recorded in units of $100, using advertising expenditure $x$, measured in thousands of dollars:

\widehat{\ln(Y)}=1.20+0.10x.

At $x=4$, the fitted log value is

\hat m(4)=1.20+0.10(4)=1.20+0.40=1.60.

Exponentiating gives

\exp\{\hat m(4)\}=e^{1.60}\approx 4.953.

Because $Y$ is measured in units of $100, this is a prediction of about $4.953$ units of $100, or about $495.30. Provided the model’s conditional-mean assumption is appropriate, this estimates the conditional geometric mean of weekly sales at that advertising level. By itself, the calculation does not establish that the conditional arithmetic mean is $495.30.

The outcome is positive, as required for the logarithm, and the interpretation is restricted to the stated predictor value and measurement scale. The numerical calculation does not verify that the chosen log-linear relationship or the model assumptions are suitable; those are modeling questions rather than consequences of exponentiation.

When Is It Also a Conditional Median?

A stronger interpretation is available if the log-scale errors are symmetric around zero at each predictor value. Symmetry means that positive and negative deviations from zero balance in the conditional error distribution; for example, values equally far above and below zero have matching probabilities. A distribution symmetric around zero has zero as a median. Because exponentiation is strictly increasing, it preserves the ordering of values and carries a median on the log scale to a median on the original scale.

Result: If $\ln(Y)=m(x)+\varepsilon$ at $X=x$, and the conditional distribution of $\varepsilon$ is symmetric around zero, then $\exp\{m(x)\}$ is a conditional median of $Y$. Thus $\exp\{\hat m(x)\}$ estimates that median when the fitted model adequately estimates $m(x)$ and the symmetry condition is appropriate.

Symmetry is sufficient for this median conclusion; it is not required for the geometric-mean interpretation. The geometric-mean result follows from the conditional mean of $\ln(Y)$, whereas the median result depends on the location of the conditional median of the errors. A conditional mean of zero alone does not require the error median to be zero.

Worked Example: Symmetric Log Errors Give a Median

Suppose that at a particular predictor value, a hypothetical positive outcome $Y$, measured in milligrams per litre, has the log-scale form

\ln(Y)=\ln(50)+\varepsilon,

where $\varepsilon$ takes the values $-0.20,\ 0,\ 0.20$, each with probability $1/3$. These errors are symmetric around zero and have mean zero. The conditional mean log outcome is therefore $\ln(50)$, so the conditional geometric mean is

\exp\{E[\ln(Y)\mid X=x]\} =\exp\{\ln(50)\} =50\text{ mg/L}.

The three possible outcome values are $50e^{-0.20}$, $50$, and $50e^{0.20}$, approximately $40.94$, $50.00$, and $61.07$ mg/L. The middle value is the median, so the conditional median is $50$ mg/L as well.

The arithmetic mean of those three outcomes is different:

\frac{50e^{-0.20}+50+50e^{0.20}}{3} = 50\left(\frac{0.8187+1+1.2214}{3}\right) \approx 50.67\text{ mg/L}.

This example checks the symmetry and mean-zero conditions directly from the specified error values. It shows why the exponentiated log prediction can describe both a geometric mean and a median while not equalling the arithmetic mean.

Mean-Zero Errors Do Not Guarantee a Median

It is important not to infer symmetry merely from the assumption $E(\varepsilon\mid X)=0$. That condition locates the conditional mean of the log errors at zero. A skewed distribution can have mean zero while its median is elsewhere. In that case, exponentiating $m(x)$ still gives the conditional geometric mean under the stated model, but it need not give the conditional median.

Worked Example: Mean Zero Without Symmetry

At one predictor value, suppose a hypothetical log-outcome model is

\ln(Y)=\ln(10)+\varepsilon,

where the errors $-0.20,\ 0.10,\ 0.10$ are equally likely. Their mean is

E(\varepsilon\mid X=x) =\frac{-0.20+0.10+0.10}{3}=0.

Thus the conditional mean log outcome is $\ln(10)$, and the conditional geometric mean is $e^{\ln(10)}=10$ units. However, these errors are not symmetric around zero: the middle error is $0.10$, not zero. The corresponding outcome values are approximately $10e^{-0.20}=8.187$, $10e^{0.10}=11.05$, and $10e^{0.10}=11.05$ units. Their median is $11.05$ units, not $10$.

The example verifies the zero-mean condition but not the symmetry condition. It therefore supports the geometric-mean interpretation, but not the claim that the exponentiated fitted log value is a conditional median.

Why the Distinction Matters

The geometric mean, median, and arithmetic mean summarize a conditional outcome distribution in different ways. An exponentiated fitted log value has a clear interpretation through the mean of the log outcome. Under symmetric errors around zero, it also estimates the conditional median. Neither statement makes it automatically an estimate of the arithmetic mean. On the original scale, the arithmetic mean depends on the distribution of the errors after exponentiation as well as on the fitted log-scale value.

When reporting a prediction, name the target explicitly. For example, “the estimated conditional geometric mean is about $495 per week” communicates what follows from the log-scale mean model. Calling that value “the predicted average weekly sales” may lead readers to understand an arithmetic mean, which is not justified by exponentiation alone. If a project requires an arithmetic-mean prediction, it needs a method that accounts for the log-scale error distribution; the back-transformation by itself does not supply that adjustment.

The median interpretation likewise requires a stated condition. In practice, a symmetric-looking distribution of fitted residuals may be a useful diagnostic, but it cannot prove that errors are conditionally symmetric at every predictor value. The conclusion should match the assumptions supported by the model and the data, rather than follow automatically from using a logarithm.

Common Mistake

Pitfall: “The model predicts $\ln(Y)=c$, so $e^c$ is the expected value of $Y$.” This reasoning treats exponentiation as if it preserved averages. It does not. Under the conditional mean-zero log-error model, $e^c$ estimates the conditional geometric mean. It is also a conditional median when the log errors are symmetric around zero. Do not call it the arithmetic mean unless an appropriate additional prediction method justifies that target.

A careful report therefore identifies the outcome scale and the summary being predicted, states the relevant model conditions, and gives units. The distinction is not merely terminology: geometric means and medians describe multiplicative or middle behavior, while arithmetic means describe totals averaged on the original scale.

Check Your Understanding

Answer each question in terms of the log-outcome model and the scale of the outcome.

If $\hat m(x)=2.00$, what is the exponentiated fitted value, and what conditional summary does it estimate under a correctly specified mean-zero log-error model?
What additional condition lets the exponentiated fitted log value estimate a conditional median?
Why does a conditional mean-zero error assumption alone not guarantee that the exponentiated prediction is a conditional median?
A positive outcome is measured in hundreds of dollars, and its exponentiated fitted value is $3.2$. Give the prediction in the recorded units and in dollars.
Can an exponentiated fitted log value automatically be described as the conditional arithmetic mean? Explain briefly.