Why Transformed-Scale Interpretation Needs Care
A regression model may be fitted using a transformed response, a transformed predictor, or both. Its coefficients describe relationships on the scale used to fit the model. Interpreting them on the original measurement scale requires an additional step—and that step is a frequent source of errors. In particular, exponentiating a fitted log response does not always give the arithmetic mean, a log-response coefficient is not itself a percentage, and transforming a predictor does not automatically stabilize response variance.
The distinctions matter in applications such as hypothetical insurance models relating claim costs to policy characteristics. A percentage interpretation can be materially different from a change in dollars, and a prediction intended to represent an average payment must target the arithmetic mean when that is the quantity required.
Log-Response Coefficients Are Not Percentages by Themselves
Suppose the log-response model is \(\ln Y=\beta_0+\beta_1x+\varepsilon\), with \(E[\varepsilon\mid x]=0\). For a change in \(x\) from \(x\) to \(x+\Delta\), the fitted log response changes by \(\beta_1\Delta\). Exponentiating this difference gives a multiplicative factor \(\exp(\beta_1\Delta)\). The corresponding proportional change is that factor minus one.
This is an exact percentage-change interpretation for the exponentiated fitted log response, or for a conditional median when the error has median zero. It is not generally an exact ratio of conditional arithmetic means unless the error distribution has the same form at both predictor values. Under a homoscedastic normal-error model, that ratio is also the ratio of conditional arithmetic means because the same retransformation factor applies at both values.
Worked Example: Converting a Log-Response Slope to a Percentage
A hypothetical model for positive claim severity, in dollars, is \(\ln Y=1.1+0.4x+\varepsilon\). Assume \(E[\varepsilon\mid x]=0\), and suppose the residual variance is the same at the two predictor values being compared. Find the modeled proportional change when \(x\) increases from 2 to 5.
Method. The predictor change is \(\Delta=5-2=3\). The fitted log response therefore changes by \(0.4(3)=1.2\).
The multiplicative factor on the exponentiated fitted log response is \(\exp(1.2)\approx3.3201\). Thus the proportional increase is \(3.3201-1=2.3201\), or \(232.01\%\), rounded to two decimal places.
As a check, the fitted log values are \(1.1+0.4(2)=1.9\) and \(1.1+0.4(5)=3.1\). Their difference is \(3.1-1.9=1.2\), giving the same factor \(\exp(3.1)/\exp(1.9)=\exp(1.2)\). This method avoids treating the slope \(0.4\) as a 40% increase for a three-unit change.
For a one-unit increase, the exact proportional change is \(\exp(\beta_1)-1\), not \(\beta_1\). The approximation \(\exp(\beta_1)-1\approx\beta_1\) can be useful only when \(\beta_1\) is small; it should not replace the exact calculation when precision matters. For a change of several units, use \(\exp(\beta_1\Delta)-1\), not \(\Delta\beta_1\) as an exact percentage.
What Does an Intercept Mean After Exponentiation?
In a log-response model, \(\beta_0\) is the fitted value of \(\ln Y\) at \(x=0\). Consequently, \(\exp(\beta_0)\) is the fitted value on the exponentiated-log scale at \(x=0\). Under mean-zero log errors it is the conditional geometric mean. It is also the conditional median if the log error has median zero, as it does under a normal distribution.
That does not make \(\exp(\beta_0)\) the conditional arithmetic mean of \(Y\). Also, the intercept refers to \(x=0\), which may not be a useful or observed predictor value. If the data cover only \(x\) values from 10 to 20, for example, the intercept is an extrapolation rather than a summary of a typical observation.
Worked Example: Interpreting an Intercept Under a Lognormal Model
A hypothetical model for a positive annual payment has \(\ln Y=\beta_0+0.2x+\varepsilon\), where \(\beta_0=\ln(100)\). At \(x=0\), assume \(\varepsilon\mid x\sim N(0,0.16)\). Find the conditional median and the conditional arithmetic mean at \(x=0\).
Method. At \(x=0\), the fitted log value is \(\ln(100)+0.2(0)=\ln(100)\). Since the log error is normal with mean zero, the conditional median on the original scale is \(\exp(\ln(100))=100\) dollars.
For a normal log error with variance \(0.16\), the arithmetic mean is \(\exp(\eta+0.16/2)\), where \(\eta=\ln(100)\). Therefore it is \(\exp(\ln(100)+0.08)=100\exp(0.08)\approx108.33\) dollars, rounded to the nearest cent.
The conditional arithmetic mean exceeds the median because the lognormal distribution is right-skewed. The calculation avoids labeling \(\exp(\beta_0)=100\) as the arithmetic mean. It also applies specifically at \(x=0\); if zero is not meaningful or lies outside the observed range, the intercept may have little practical interpretation.
Back-Transformed Predictions and the Arithmetic Mean
Let \(\eta=\beta_0+\beta_1x\) be the fitted log-scale value. Exponentiating gives \(\exp(\eta)\), but what that number estimates depends on the error assumptions and the target quantity. Under \(E[\varepsilon\mid x]=0\), it is the conditional geometric mean of \(Y\). Under the additional condition that the conditional median of \(\varepsilon\) is zero, it is also the conditional median. It is not automatically the conditional arithmetic mean.
If \(\varepsilon\mid x\sim N(0,\sigma^2)\), then \(Y\mid x\) is lognormal, and its arithmetic mean includes a variance adjustment. The reason is that averaging exponentials is not the same as exponentiating an average.
Write \(\ln Y=\eta+\varepsilon\), with \(\varepsilon\mid x\sim N(0,\sigma^2)\). Then \(Y=\exp(\eta)\exp(\varepsilon)\).
Because \(\eta\) is fixed at a specified \(x\), \(E[Y\mid x]=\exp(\eta)E[\exp(\varepsilon)\mid x]\).
For a mean-zero normal variable with variance \(\sigma^2\), \(E[\exp(\varepsilon)]=\exp(\sigma^2/2)\). Thus \(E[Y\mid x]=\exp(\eta+\sigma^2/2)\), as claimed.
If the log errors are not normal, the factor \(\exp(\sigma^2/2)\) need not be correct. One possible alternative is a residual-based smearing estimate: multiply \(\exp(\hat\eta)\) by the average of \(\exp(\hat\varepsilon_i)\) over the training residuals, subject to the assumptions and validation appropriate to the application. The essential point is to match the back-transformation method to the desired target and error distribution.
Worked Example: Predicting an Arithmetic Mean From a Log Model
A hypothetical model for a positive benefit amount gives a fitted log value \(\hat\eta=\ln(80)\) for a specified policy profile. Assume the conditional log error is normal with mean zero and variance \(0.09\). Calculate the back-transformed median and the estimated conditional arithmetic mean.
Method. The fitted log value exponentiates to \(\exp(\ln(80))=80\) dollars. Under the stated normal-error assumption, this is the conditional median.
The conditional arithmetic mean is \(\exp(\ln(80)+0.09/2)=80\exp(0.045)\). Since \(\exp(0.045)\approx1.0460\), the estimate is approximately \(80(1.0460)=83.68\) dollars, rounded to the nearest cent.
The answer is checked by noting that the mean exceeds the median by about \(83.68-80=3.68\) dollars, consistent with the positive skew of a lognormal distribution. This method avoids reporting the median as the arithmetic mean.
Logarithmic Predictors Change the Units of the Effect
A model with a logarithmic predictor, \(Y=\beta_0+\beta_1\ln x+\varepsilon\), is not a log-response model. If \(E[\varepsilon\mid x]=0\), the modeled conditional mean changes by \(\beta_1\ln(x_2/x_1)\) as \(x\) changes from \(x_1\) to \(x_2\). That difference is in the units of \(Y\), not a proportional change in \(Y\).
For a small relative increase \(r\) in \(x\), \(\ln(1+r)\approx r\), so the response change is approximately \(\beta_1r\) response units. This is sometimes called a semi-elasticity interpretation: a percentage change in \(x\) corresponds approximately to an additive change in \(Y\). A model with both \(\ln Y\) and \(\ln x\) would have a different, percentage-on-percentage interpretation.
Worked Example: A Logarithmic Predictor Is Not a Percentage Response Effect
A hypothetical model for annual service cost in dollars is \(Y=120+18\ln x+\varepsilon\), where \(x\) is a positive exposure measure and \(E[\varepsilon\mid x]=0\). Compare the fitted conditional means at \(x=100\) and \(x=200\), and interpret a 1% increase near \(x=100\).
Method. Doubling \(x\) changes \(\ln x\) by \(\ln(200)-\ln(100)=\ln 2\). The modeled response change is \(18\ln 2\approx18(0.693147)=12.4766\) dollars, or about \(12.48\) dollars rounded to the nearest cent.
For a check, the fitted means are \(120+18\ln(100)\approx202.89\) dollars and \(120+18\ln(200)\approx215.37\) dollars. Their difference is \(215.37-202.89=12.48\) dollars, allowing for rounding.
A 1% increase from \(x=100\) means moving to \(101\). The exact fitted response change is \(18\ln(101/100)=18\ln(1.01)\approx18(0.0099503)=0.1791\) dollars, approximately \(0.18\) dollars. This is an additive response change, not a 1% increase in \(Y\). The calculation avoids assigning a percentage interpretation to a model that does not log-transform its response.
Transforming a Predictor Does Not Guarantee Constant Variance
A predictor transformation changes the model’s representation of how the conditional mean of \(Y\) depends on that predictor. It may make a curved mean relationship more nearly linear. But the transformation does not, by itself, impose a constant conditional variance on the response errors. In a model \(Y=f(x)+\varepsilon\), constant variance is the separate assumption \(\operatorname{Var}(\varepsilon\mid x)=\sigma^2\) for all relevant \(x\).
A better mean specification can sometimes make residual patterns less pronounced, but this is not a guarantee. After fitting the chosen predictor transformation, residual plots and other appropriate diagnostics still need to be examined for changing spread. If the response variance remains nonconstant, a response transformation or a method that models the variance may be considered. A transformed predictor should not be presented as having solved heteroscedasticity without evidence.
Worked Example: Checking a Claim About Variance Stabilization
An analyst models hypothetical monthly repair costs using \(\ln(x)\), where \(x\) is equipment age. The fitted model is \(Y=\beta_0+\beta_1\ln(x)+\varepsilon\). In a residual plot, the residual standard deviation is about 3 dollars for younger equipment and about 11 dollars for older equipment. Does the predictor transformation establish constant response variance?
Method. The model uses the logarithm of age to describe the conditional mean pattern. The variance assumption would instead require the residual spread to be approximately the same across ages. The reported spreads, 3 and 11 dollars, are not similar: the older-equipment spread is \(11/3\approx3.67\) times as large.
Therefore, the transformation has not established constant variance; the residual pattern is evidence that the constant-variance assumption may be questionable. It does not by itself identify the best remedy. The analyst should examine the diagnostic pattern and consider an appropriate response transformation or variance model. This method avoids mistaking a change to the predictor scale for a variance assumption about the response.
Common Exam Traps
- Using the slope as the exact percentage. For a log response and a one-unit increase, the exact percentage is \(100\{\exp(\beta_1)-1\}\%\), not \(100\beta_1\%\). For a change of \(\Delta\) units, use \(100\{\exp(\beta_1\Delta)-1\}\%\).
- Calling the exponentiated intercept an average payment. It is the fitted geometric-scale value at \(x=0\). The arithmetic mean requires a retransformation adjustment under a stated error model.
- Back-transforming without naming the target. Decide whether the question asks for a conditional median, geometric mean, or arithmetic mean before exponentiating or adjusting.
- Treating \(\ln x\) as if the response were also logged. In a level-response/log-predictor model, predictor changes produce additive changes in the response.
- Assuming a transformed predictor fixes residual spread. A transformation may improve the mean shape, but constant variance remains a separate property to assess.
Check Your Understanding
For each problem, identify the scale of the response and the prediction target before calculating.
- A model for positive payment \(Y\) is \(\ln Y=0.5+0.25x+\varepsilon\), with a mean-zero error. Find the exact proportional change in the exponentiated fitted log response when \(x\) increases by 2. (Answer: \(\exp(0.5)-1\approx0.6487\), or a 64.87% increase.)
- A hypothetical log-response model is \(\ln Y=0.7+0.3x+\varepsilon\). Compare the exponentiated fitted log responses at \(x=1\) and \(x=4\). (Answer: factor \(\exp(0.9)\approx2.4596\), or an increase of about 145.96%.)
- A level-response model is \(Y=50+8\ln x+\varepsilon\), with mean-zero errors. What is the fitted conditional-mean change when \(x\) doubles? (Answer: \(8\ln 2\approx5.55\) response units.)
- At a given profile, a log-response model has fitted log value \(\ln(40)\), and the conditional log error is normal with mean zero and variance \(0.09\). Find the conditional arithmetic mean. (Answer: \(40\exp(0.045)\approx41.84\).)
- A model uses \(\ln(x)\) as predictor and leaves \(Y\) untransformed. The residual spread increases systematically with \(x\). Is constant variance established by the predictor transformation, and what should be checked? (Answer: No; examine residual diagnostics for nonconstant variance and consider an appropriate response transformation or variance model.)