Log-Response Models and Multiplicative Effects

Why a Log Response Changes the Interpretation

A linear regression model for $Y$ describes additive changes in the response: a one-unit increase in a predictor changes the fitted response by a fixed amount. In some settings, the response instead changes proportionally. For example, the same increase in a predictor might correspond to a similar percentage change in expected claim severity at different starting levels. Modeling the logarithm of a positive response allows the regression to represent this multiplicative pattern.

The coefficient is still a change on the modeled scale, which is the log scale. Its original-scale meaning is obtained by exponentiating it. That conversion gives an exact percentage change; multiplying the coefficient by 100 gives only an approximation, which is useful when the coefficient is small.

Definition: A log-response linear regression model has the form $\log(Y_i)=\beta_0+\beta_1x_{i1}+\cdots+\beta_px_{ip}+\epsilon_i$, where $Y_i>0$, the $x_{ij}$ are predictors, and $\epsilon_i$ is an error term. Unless otherwise stated, $\log$ denotes the natural logarithm. The model is linear in the coefficients, even though its fitted relationship with $Y_i$ is generally nonlinear.

For the usual regression-based inference, the errors are assumed to have conditional mean zero, constant variance, and independence; normal errors are additionally assumed when exact small-sample $t$- and $F$-inference is used. Coefficient interpretation on the log scale requires a correctly specified conditional mean model. To compare original-scale responses using a coefficient, the compared cases must differ in the stated predictor change while other predictors are held fixed. The exponential ratio interpretation also assumes the error distribution is the same for both cases.

A log response requires positive observed responses. A zero or negative response cannot be logged without changing the analysis, and adding an arbitrary constant changes the interpretation. The transformation should therefore be justified by the data and modeling objective, not applied mechanically.

From a Log-Scale Coefficient to a Percentage Change

Suppose a predictor $x_j$ increases by $c$ units, with all other predictors held fixed. The fitted log response changes by $\beta_jc$. Exponentiating the two fitted log responses gives their original-scale ratio:

\frac{\exp(\widehat{\eta}_{\text{new}})} {\exp(\widehat{\eta}_{\text{old}})} = \exp(\beta_jc), \qquad \text{exact percentage change} = 100\{\exp(\beta_jc)-1\}\%.

Here $\widehat{\eta}$ denotes the fitted value of $\log(Y)$. The coefficient's sign determines the direction of the change: a positive coefficient gives a ratio greater than one, and a negative coefficient gives a ratio below one. The percentage change is relative to the original fitted level, not a fixed number of response units.

Key result: For a one-unit increase in $x_j$, the exact percentage change associated with its coefficient is $100\{\exp(\beta_j)-1\}\%$, holding the other predictors fixed. For a $c$-unit increase, use $100\{\exp(c\beta_j)-1\}\%$. The approximation $100c\beta_j\%$ is appropriate only when $|c\beta_j|$ is small.

The approximation follows from the expansion $\exp(z)=1+z+z^2/2+\cdots$. When $z$ is close to zero, the terms after $z$ are small, so $\exp(z)-1$ is close to $z$. Thus the approximate percentage change is $100z\%$, where $z=c\beta_j$. The approximation becomes less reliable as the absolute log-scale change increases, and it can be especially misleading for larger negative changes.

\exp(z)-1 \approx z \quad\Longrightarrow\quad 100\{\exp(c\beta_j)-1\}\% \approx 100c\beta_j\%.

Worked Interpretations

Worked Example: A Small Positive Coefficient

A hypothetical insurer models the positive amount of a covered repair, $Y$, using $\log(Y)=\beta_0+\beta_1x+\epsilon$, where $x$ is the vehicle's age in years. The estimated coefficient is $\widehat{\beta}_1=0.08$. Assume the model errors are independent $N(0,0.09)$, with variance $0.09$, and have the same distribution at the two vehicle ages being compared. No interest rate or mortality assumption is involved. Interpret a one-year increase in vehicle age, holding other predictors fixed.

Method. Calculate the exact multiplicative ratio as $\exp(\widehat{\beta}_1)$, then subtract one and express the result as a percentage.

Step 1: Exact ratio. The log response increases by $0.08$, so the ratio of the exponentiated fitted log responses is $\exp(0.08)=1.083287$.

Step 2: Exact percentage change. Subtract the baseline ratio of $1$: $1.083287-1=0.083287$. Converting to a percentage gives $100(0.083287)=8.3287\%$, or an estimated 8.33% increase, rounded to two decimal places.

Step 3: Approximation. The small-coefficient approximation gives $100(0.08)=8\%$. This is close to, but not equal to, the exact 8.33%.

Check and common mistake. Since the coefficient is positive, the ratio must exceed one and the percentage change must be positive; $1.083287$ satisfies this check. Reporting 0.08% confuses a decimal proportion with a percentage. The exact result is about 8.33%, not 0.08%.

Worked Example: A Negative Coefficient and a Multi-Unit Change

In a hypothetical model of positive administrative costs, $\log(Y)=\beta_0+\beta_1x+\epsilon$, $x$ is the number of weeks a process has been in operation. The estimated coefficient is $\widehat{\beta}_1=-0.15$ per week. Assume independent errors $N(0,0.04)$, with variance $0.04$, and the same error distribution at both process ages. No interest rate or mortality assumption applies. Estimate the percentage change associated with an increase of two weeks.

Method. Multiply the coefficient by the full two-unit change before exponentiating. This gives the exact change, rather than adding two one-week percentage changes.

Step 1: Log-scale change. The change is $c\widehat{\beta}_1=2(-0.15)=-0.30$.

Step 2: Original-scale ratio. The ratio is $\exp(-0.30)=0.740818$. It is below one, consistent with a decrease.

Step 3: Exact percentage change. Subtract one: $0.740818-1=-0.259182$. Multiplying by 100 gives $-25.9182\%$, or an estimated 25.92% decrease, rounded to two decimal places.

Step 4: Approximation. The approximation is $100(-0.30)=-30\%$. This is noticeably different from the exact decrease. The approximation uses the two-week log change, not the one-week coefficient alone.

Check and common mistake. A decrease of 25.92% leaves about 74.08% of the original fitted level, matching the ratio $0.740818$. A common error is to report $100(-0.15)=-15\%$ for the two-week change; that is the one-week approximation, not the requested two-week effect.

Worked Example: Combining a Continuous and an Indicator Predictor

A hypothetical model for positive claim-processing expense is $\log(Y)=\beta_0+\beta_1x+\beta_2D+\epsilon$, where $x$ is processing time in days and $D=1$ indicates use of a particular review method ($D=0$ otherwise). Suppose $\widehat{\beta}_1=0.05$ per day and $\widehat{\beta}_2=-0.10$. Assume independent errors $N(0,0.16)$, with variance $0.16$, and the same error distribution for the cases compared. No interest rate or mortality assumption applies. Compare an otherwise identical case with $x$ increasing by three days and $D$ changing from 0 to 1.

Method. Add the contributions to the fitted log response, then exponentiate the total change. This accounts for the changes jointly.

Step 1: Continuous-predictor contribution. A three-day increase contributes $3(0.05)=0.15$ to the fitted log response.

Step 2: Indicator contribution. Changing from $D=0$ to $D=1$ contributes $(-0.10)(1-0)=-0.10$.

Step 3: Total change and ratio. The total log-scale change is $0.15-0.10=0.05$. The fitted original-scale ratio is $\exp(0.05)=1.051271$.

Step 4: Percentage effect. The exact percentage change is $100(1.051271-1)=5.1271\%$, or a 5.13% increase, rounded to two decimal places. The approximation based on the combined change is $100(0.05)=5\%$.

Check and common mistake. The positive continuous-predictor contribution exceeds the negative indicator contribution, so the net effect should be a modest increase; the ratio $1.051271$ confirms this. Do not add separately computed percentages: the three-day effect is $\exp(0.15)-1$, about 16.18%, and the indicator effect is $\exp(-0.10)-1$, about -9.52%, but these percentages do not add to the combined percentage. The combined log-scale changes add; their multiplicative ratios multiply.

Why Exponentiating a Fitted Log Response Needs Care

The fitted value $\widehat{\eta}$ is on the log-response scale. Exponentiating it gives $\exp(\widehat{\eta})$, a value on the original scale, but that operation does not automatically make the result an estimate of the conditional arithmetic mean $E[Y\mid x]$. The distinction matters whenever a question asks for an average response or an expected dollar amount.

To see why, suppose the model is $\log(Y)=\eta(x)+\epsilon$, with $\epsilon\mid x\sim N(0,\sigma^2)$. Then $Y=\exp(\eta(x))\exp(\epsilon)$. The conditional median is $\exp(\eta(x))$, because the median of a mean-zero normal error is zero. The conditional mean, however, is $E[Y\mid x]=\exp(\eta(x))E[\exp(\epsilon)\mid x]$. For a normal error, $E[\exp(\epsilon)\mid x]=\exp(\sigma^2/2)$. Therefore:

\operatorname{Median}(Y\mid x)=\exp(\eta(x)), \qquad E[Y\mid x]=\exp\{\eta(x)+\sigma^2/2\}.

This is a consequence of the nonlinearity of the exponential function: in general, $E[\exp(\epsilon)]$ is not $\exp(E[\epsilon])$. If $\sigma^2$ is the same across cases, the common factor $\exp(\sigma^2/2)$ cancels in a ratio of conditional means. The coefficient-based percentage comparison can then describe that mean ratio as well as the ratio of exponentiated fitted log responses. If the error distribution or variance differs between cases, the correction need not cancel.

Interpretation caution: Exponentiating a fitted log response gives a log-scale location expressed on the original scale; under normal errors it is the conditional median, not the conditional mean. A mean prediction requires accounting for the error distribution. The coefficient's exact multiplicative interpretation is separate from the question of which original-scale summary a retransformed fitted value estimates.

Common Traps and Practical Checks

Using the approximation as if exact. A coefficient of $-0.30$ implies an exact decrease of $1-\exp(-0.30)$, about 25.92%, not 30%.
Forgetting the predictor units. A coefficient per day must be multiplied by the number of days in the change. If a predictor is recorded in hundreds of dollars, a one-unit change means $100, not $1.
Interpreting a coefficient as a response-unit change. In a log-response model, $\beta_j$ is a log-scale change. It is not an increase of $\beta_j$ dollars or other response units.
Ignoring the other predictors. A coefficient comparison holds the remaining predictors fixed. If several predictors change, add their contributions to the log-scale change before exponentiating.
Turning association into causation. A fitted coefficient describes a model-based association. It does not establish that changing a predictor would cause the response to change unless the study design and assumptions support a causal conclusion.
Calling a retransformed fit an expected value without justification. The conditional mean and the exponentiated fitted log response need not agree. Check which response summary is being requested and what is assumed about the errors.

Key takeaway: For a log-response model, a predictor change $c$ changes the fitted log response by $c\beta_j$ and multiplies the exponentiated fitted response by $\exp(c\beta_j)$. Convert that ratio to an exact percentage with $100\{\exp(c\beta_j)-1\}\%$. Treat $100c\beta_j\%$ as a small-change approximation, and distinguish an exponentiated log-scale fit from an original-scale conditional mean.

Check Your Understanding

For each problem, state the exact percentage effect and use the stated assumptions. Round final percentages to two decimal places.

A hypothetical model for positive inspection cost has a log-response coefficient of $0.04$ per inspection hour. The error is independent $N(0,0.01)$, with the same distribution at both cases; no interest rate or mortality assumption applies. What is the exact percentage change for a two-hour increase? (Answer: $100\{\exp(0.08)-1\}=8.33\%$ increase.)
A model for positive account servicing time has coefficient $-0.12$ for a binary indicator that equals one for a revised process. Errors are independent $N(0,0.04)$ and have the same distribution by process; no interest rate or mortality assumption applies. What is the exact percentage change for revised versus standard process? (Answer: $100\{\exp(-0.12)-1\}=-11.31\%$, a decrease.)
A log-response model has coefficient $0.06$ per month for account age. Errors are independent $N(0,0.09)$ with the same distribution for both cases; no interest rate or mortality assumption applies. Find the exact effect of a five-month increase and compare it with the approximation. (Answer: $100\{\exp(0.30)-1\}=34.99\%$ increase; approximation $30\%$.)
A hypothetical positive expense model has coefficients $0.03$ per service day and $-0.08$ for an indicator that changes from zero to one. Errors are independent $N(0,0.04)$ and have the same distribution across the cases; no interest rate or mortality assumption applies. For a two-day increase together with the indicator change, find the exact combined percentage effect. (Answer: total log change $2(0.03)-0.08=-0.02$; $100\{\exp(-0.02)-1\}=-1.98\%$, a decrease.)
Suppose $\log(Y)=\eta(x)+\epsilon$, with $\epsilon\mid x\sim N(0,0.25)$. If $\eta(x)=2$, give the conditional median and conditional mean of $Y$. (Answer: median $\exp(2)\approx 7.39$; mean $\exp(2+0.25/2)=\exp(2.125)\approx 8.37$.)

What You'll Learn