Back-Transforming Log-Scale Estimates

From a Log-Scale Contrast to an Exposure Ratio

In pharmacokinetic comparisons, a treatment contrast is often estimated on the log scale because exposure measurements are commonly analyzed after logarithmic transformation. The contrast itself is not an exposure ratio: it is a difference between log-scale treatment means. Exponentiating that contrast gives the corresponding ratio, but its uncertainty requires care. A point estimate, a standard error, and a confidence interval do not all transform in the same way.

Let \(\theta\) denote the population treatment contrast on the natural-log scale, with the contrast defined as test treatment minus reference treatment. Let \(\hat{\theta}\) be its estimate and \(SE(\hat{\theta})\) its standard error. For example, if the log-scale means are \(\mu_T\) and \(\mu_R\), then \(\theta=\mu_T-\mu_R\). Exponentiation converts the difference into a ratio:

Thus, \(\exp(\hat{\theta})\) estimates the ratio of the treatment geometric means, with the numerator and denominator determined by the contrast direction. It is dimensionless. A ratio of 1 indicates equal geometric means; a ratio above 1 indicates a larger geometric mean for the numerator treatment, and a ratio below 1 indicates a smaller one. The earlier material on geometric means and log-scale contrasts establishes this connection; here the focus is how to carry estimates and uncertainty back to the ratio scale.

Why the Point Estimate and Its Uncertainty Transform Differently

Exponentiation is nonlinear. The point estimate is transformed directly: calculate \(\exp(\hat{\theta})\). But a standard error measures spread around an estimate, not an endpoint or a quantity to exponentiate on its own. In particular, \(\exp(SE(\hat{\theta}))\) is not a standard error for the ratio. A local approximation to the spread can be obtained by examining how quickly the exponential function changes near \(\hat{\theta}\).

For a small change \(d\) in the log contrast, the corresponding change in the ratio is approximately \(\exp(\hat{\theta})d\), because the derivative of \(\exp(x)\) at \(x=\hat{\theta}\) is \(\exp(\hat{\theta})\). Therefore, a log-scale standard error of \(SE(\hat{\theta})\) corresponds approximately to a ratio-scale standard error of \(\exp(\hat{\theta})SE(\hat{\theta})\). This is the delta method, a first-order approximation. It is most useful when uncertainty is not so large that the exponential curve changes substantially across the likely range of \(\hat{\theta}\).

A confidence interval follows a different calculation. Suppose an appropriate log-scale interval is \([L,U]\), obtained as \(\hat{\theta}\) plus or minus a suitable critical value times its standard error. Because the exponential function is strictly increasing, the corresponding interval for the ratio is \([\exp(L),\exp(U)]\). This transformation preserves the interval’s coverage under the assumptions used to form the original interval. Its endpoints will generally not be equally distant from \(\exp(\hat{\theta})\), so the ratio-scale interval is asymmetric.

The delta-method standard error is useful for expressing a local measure of uncertainty in ratio units. The exponentiated confidence interval is usually preferable for communicating uncertainty because it transforms the interval limits themselves. One should not construct an interval by treating the delta-method standard error as if it were an exact symmetric description of ratio-scale uncertainty, especially when the log-scale uncertainty is appreciable.

Assumptions and a Short Justification

Back-transformation does not repair a poorly specified analysis. The log-scale contrast and its standard error must come from an analysis appropriate to the design and data. For interval inference, the log-scale estimator should have the distributional approximation used to construct its interval: for example, an approximately normal estimator with a valid standard error and an appropriate normal critical value, or a t-based procedure with its appropriate degrees of freedom and critical value. The log transformation must also be meaningful for the endpoint; an exposure measure being logged must be positive.

The monotone transformation of a confidence interval has a simple justification. If \(\theta\in[L,U]\), then monotonicity of the exponential function gives \(\exp(L)\leq\exp(\theta)\leq\exp(U)\). Therefore, whenever the original interval covers \(\theta\), the transformed interval covers \(\exp(\theta)\). Its coverage is inherited from the original interval; exponentiating does not create better distributional assumptions.

The delta-method standard error has a related but approximate justification. A first-order Taylor expansion around the estimated log contrast gives:

If the sampling error \(d\) has standard deviation \(SE(\hat{\theta})\), multiplying that error by \(\exp(\hat{\theta})\) gives the first-order approximation \(\exp(\hat{\theta})SE(\hat{\theta})\). Terms beyond the linear part are ignored. When uncertainty is large, those terms can matter, which is one reason the transformed confidence interval is not generally equal to a symmetric interval built from the delta-method standard error.

Ratio-Scale Interpretation and Diagnostics

A ratio can also be expressed as a percent change relative to the denominator treatment: \(100(R-1)\%\). For example, \(R=1.200\) corresponds to a 20.0% higher geometric mean, while \(R=0.9000\) corresponds to a 10.0% lower geometric mean. The percent change is calculated from the ratio, not by multiplying the log contrast by 100 except as a small-contrast approximation.

Before reporting a back-transformed result, check the following:

• Contrast direction: confirm which treatment is in the numerator. Reversing test-minus-reference to reference-minus-test changes the ratio to its reciprocal.
• Scale: confirm that the model contrast and standard error are on the natural-log scale if using \(\exp\). A base-10 log contrast requires a different inverse transformation.
• Uncertainty calculation: use the standard error on the log scale in the delta-method formula, or transform the endpoints of a valid log-scale interval. Do not exponentiate the standard error by itself.
• Units and interpretation: the ratio is dimensionless, while the original exposure measurements may have units. A ratio of 1.10 means 10% higher for the numerator treatment’s geometric mean, not 1.10 exposure units higher.
• Interval shape: check that a transformed interval has lower and upper endpoints in the same order as the log-scale limits and is not forced to be symmetric around the ratio estimate.

A very wide log-scale interval produces a wide, often strongly asymmetric ratio interval. In that situation, the delta-method standard error remains a local approximation and may be an uninformative summary of the full uncertainty. If suitable, simulation from the fitted estimator distribution followed by exponentiation can show the transformed sampling distribution more directly. It does not eliminate the need for a valid fitted model or appropriate uncertainty assumptions.

Common Pitfalls and Alternatives

The most common error is to exponentiate both the contrast and its standard error, reporting \(\exp(\hat{\theta})\) and \(\exp(SE(\hat{\theta}))\). That mixes a ratio estimate with a transformed number that is not its standard error. Another error is to report \(\exp(\hat{\theta})\pm SE(\hat{\theta})\), which mixes ratio units and log-scale units. The delta-method approximation, if a ratio-scale standard error is needed, is \(\exp(\hat{\theta})SE(\hat{\theta})\); for interval reporting, transform the log-scale interval endpoints.

A related mistake is to describe an interval containing 1 as proof of equivalence or to interpret a ratio estimate as an absolute exposure difference. A ratio compares relative geometric means. Whether it supports a particular regulatory or clinical conclusion depends on the applicable criteria and analysis, not on back-transformation alone.

The main approaches serve different purposes: exponentiating the contrast gives the ratio point estimate; the delta method gives a convenient approximate ratio-scale standard error; and transforming the limits gives the ratio-scale confidence interval while retaining the original interval procedure’s coverage. They are related but not interchangeable. Back-transformation also does not change what population, treatment contrast, or estimand the log-scale model represents.