Why Check the Original Scale?
A transformed regression model can fit the transformed data well and still imply a pattern that is hard to defend in context. A straight line on a transformed scale may become a curve after back-transformation. That curve can level off, grow without bound, approach zero, or even become undefined. The practical question is not only whether the transformed-scale fit looks reasonable, but also what the fitted relationship says about the quantities actually measured.
This check matters both within the observed predictor range and when someone proposes using the model beyond it. Within-range predictions can still conflict with known physical or practical limits. Outside that range, the curve’s shape is an assumption being extended, not evidence supplied by the observed data.
For a model that transforms only the predictor, interpretation is different. If the response is modeled directly as \(E(Y\mid X=x)=m(x)\), then \(m(x)\) is the conditional mean on the original response scale. The transformed predictor determines the curve’s shape; it does not change the scale on which the response is modeled.
A Practical Check
Start by identifying exactly what is being predicted. State the predictor’s observed range, the outcome’s units and any meaningful bounds, and whether the response or predictor (or both) was transformed. Then calculate a few predictions at relevant values, including the range’s ends. If considering extrapolation, explicitly mark it as such.
Keep track of which quantity was transformed and the units in which the variables are expressed.
Apply the inverse transformation, and identify the prediction target. Do not label an inverse-transformed transformed-scale mean as the original-scale arithmetic mean without justification.
Evaluate predictions across the observed range. Look for impossible values, sharp changes, asymptotes, and regions where a formula is undefined.
Predictions inside the observed range are supported by data over that range, subject to model adequacy. Predictions beyond it depend on extending the model’s assumed shape.
Worked Examples
Worked Example: A Power Pattern and Extrapolation
Suppose a hypothetical analysis relates \(Y\), a daily operating cost measured in dollars, to \(X\), machine operating time measured in hours. The observed operating times range from 5 to 40 hours. A log-log model is fitted:
Assume the conditional median on the log scale is the fitted linear expression. Exponentiating that expression gives the fitted conditional median cost, in dollars:
At 20 operating hours, the calculation is
The fitted conditional median is about $9.474 at 20 hours. At 40 hours, doubling the operating time multiplies the fitted median by \(2^{0.35}\approx1.2746\), giving about \(9.474(1.2746)\approx12.08\) dollars. The model therefore implies increasing costs, with a proportional increase for a given proportional increase in operating time.
The observed data cover only 5 to 40 hours. At 400 hours, which is far outside that range, the same formula gives approximately
The arithmetic is straightforward, but the prediction is not validated merely because the formula produces a number. A machine operated for 400 hours may face different conditions, maintenance cycles, or cost mechanisms from one operated for at most 40 hours. The power curve keeps rising, and the observed range provides no direct check on whether that pattern continues. In this context, treat the 400-hour value as an extrapolation requiring substantive justification, not as a dependable consequence of the in-range fit.
Worked Example: A Reciprocal Model’s Asymptote
In a hypothetical treatment process, let \(Y\) be the concentration of a substance in milligrams per liter after \(X\) hours. Concentrations are positive. Suppose the reciprocal outcome is modeled as
Assume the fitted conditional mean of \(1/Y\) is \(0.10+0.02x\), and this fitted value is positive in the range being considered. Inverting the fitted reciprocal gives the back-transformed fitted value
This inverse of a conditional mean on the reciprocal scale is not generally the conditional arithmetic mean of \(Y\). Here it is used to inspect the pattern implied by the fitted reciprocal-scale location. At 0, 10, and 20 hours, it gives, respectively,
These values decrease as treatment time increases. The curve approaches zero as \(x\) increases, so zero is a horizontal asymptote: the formula gets closer to zero without reaching it at any finite \(x\). To see the pattern in the formula, as \(x\) becomes large, \(0.10+0.02x\) grows without bound, and its reciprocal tends to zero.
Suppose the measurements were taken only from 0 to 20 hours and the process is known to leave a background concentration near 1 mg/L. The fitted value at 50 hours is
This is below the stated background level, and 50 hours is outside the observed range. Both points prompt caution: the model’s implied approach toward zero may not reflect the process’s practical lower limit. Also, its denominator is zero at \(x=-5\), making the fitted curve undefined there. That point lies outside the nonnegative treatment times considered here, but it illustrates why the domain of a back-transformed formula should be checked rather than assumed.
Worked Example: A Log-Transformed Predictor and Changing Absolute Effects
Suppose a hypothetical educational program relates practice time \(X\), in hours, to a test score \(Y\), in points. The response is modeled on its original scale, while the predictor uses a logarithmic transformation:
Assume the conditional mean is correctly described by this expression over the range of interest. For nonnegative practice time, the logarithm’s argument is positive, so the formula is defined. It predicts 5 points at 0 hours, about \(5+3\log(2)=7.079\) points at 10 hours, and about \(5+3\log(4)=9.159\) points at 30 hours.
The model does not imply a constant gain in score for each additional hour. From 0 to 10 hours, the fitted mean increases by
From 10 to 20 hours, the increase is
The marginal rate of change follows by differentiating the fitted mean:
It is positive but gets smaller as practice time grows. This describes diminishing absolute gains per additional hour on the original score scale. It does not mean that the score is approaching a finite ceiling: the logarithmic expression continues to grow, however slowly. If the score has a fixed maximum, the model would eventually predict values above it if extended far enough. Therefore, predictions should be checked against that bound and the observed practice-time range.
Why Effects Depend on Scale
A transformation changes the shape of the fitted relationship, so a coefficient’s interpretation depends on the scale where it appears. For the log-log model in the first example, changing \(x\) to \(cx\) changes the fitted median by a factor of
Thus, the fitted median changes by the same factor for the same proportional change in \(x\). By contrast, in the third example, equal additions of one hour do not produce equal changes in the fitted mean: the derivative \(3/(10+x)\) decreases with \(x\). A statement such as “the effect increases” is incomplete unless it says what effect is meant and on which scale: a ratio, a difference, or a local rate of change.
For an outcome transformation \(g\), if the fitted transformed-scale mean is \(m(x)\), then the inverse-transformed fitted value is \(g^{-1}(m(x))\). That operation describes the transformed-scale location mapped back to the outcome’s units. It does not generally turn the transformed-scale mean into \(E(Y\mid X=x)\). In particular, averaging and nonlinear inverse transformation do not generally commute. The conditional residual distribution is needed to determine the original-scale mean; methods such as smearing estimates address that target.
Common Mistakes
- Calling every inverse-transformed fit a mean. The inverse of a transformed-scale conditional mean is not generally the original-scale conditional arithmetic mean. Name the target correctly, and use a suitable adjustment when an arithmetic-mean prediction is required.
- Assuming a plausible in-range curve remains plausible forever. Extrapolation extends the selected formula beyond the predictor values that informed its fit. Check context and physical limits before relying on it.
- Ignoring asymptotes and undefined regions. Reciprocal forms can approach a limit or become undefined when their denominator is zero. Check both the limiting behavior and the domain relevant to prediction.
- Treating a transformed-scale slope as a constant original-scale change. A slope on a log scale often describes a ratio or proportional change; a transformed predictor can make absolute changes depend on the current predictor value.
- Checking only one predicted value. A single prediction may conceal a rapidly changing curve or a bound violation. Compare values across the observed range and at any proposed extrapolation points.
Check Your Understanding
Use the models’ stated scales and contexts when answering.
- A log-outcome model has a fitted transformed-scale mean of 2.0 at a particular predictor value. What does exponentiating 2.0 directly give, and what additional information is needed to obtain the original-scale conditional arithmetic mean?
- A power model is fitted using predictor values from 3 to 25. Why is a prediction at 200 not supported just because the model’s formula returns a finite number?
- For the reciprocal fitted curve \(\widehat{Y}(x)=1/(0.2+0.04x)\), state its horizontal asymptote and determine whether the formula is defined for nonnegative \(x\).
- In \(E(Y\mid X=x)=4+2\log(1+x)\), explain why equal increases in \(x\) do not imply equal increases in the fitted mean.
- A fitted model predicts a response above a known physical maximum at an \(x\) value inside the observed range. What should be checked before using that prediction?