From Log Differences to Exposure Ratios

Why a Difference on the Log Scale Becomes a Ratio

Pharmacokinetic exposures such as AUC are positive, and formulation effects are often naturally described in relative terms: one formulation produces exposure that is some proportion of another. On the log scale, however, the basic contrast is a difference. The link between these descriptions is exact: subtracting mean log exposures and exponentiating the result gives the ratio of the corresponding geometric mean exposures.

The identity is useful because it gives a direct interpretation to a treatment contrast estimated on the log scale. It does not, by itself, establish that a formulation difference is statistically significant or clinically important. Those judgments require additional context and, for statistical inference, methods beyond this algebraic conversion.

Definition: Let \(X\) denote a positive exposure measurement, such as AUC, recorded in the same units and for the same endpoint across formulations. Its population mean log exposure is \(\mu_{\log X}=E[\ln(X)]\), and its population geometric mean exposure is \(G=\exp(E[\ln(X)])\). For positive observations \(x_1,\ldots,x_n\), the sample mean log exposure is \(\overline{\ln x}=\frac{1}{n}\sum_{i=1}^n\ln(x_i)\), and the sample geometric mean is \(GM=\exp(\overline{\ln x})\).

Here \(\ln\) is the natural logarithm, and \(\exp\) is its inverse. These definitions require positive exposure values because the natural logarithm is defined only for positive arguments. The population definition describes a target distribution; the sample definition summarizes the observed values. The algebra connecting differences and ratios holds for either population quantities or sample summaries.

The Difference-to-Ratio Identity

Let \(T\) and \(R\) label two formulations, and let \(\mu_T=E[\ln(X_T)]\) and \(\mu_R=E[\ln(X_R)]\) be their mean log exposures. Their geometric means are \(G_T=\exp(\mu_T)\) and \(G_R=\exp(\mu_R)\). Define the contrast with \(T\) in the numerator and \(R\) in the denominator:

\Delta=\mu_T-\mu_R

The contrast is on the log scale. Exponentiating it yields:

\exp(\Delta) =\exp(\mu_T-\mu_R) =\frac{\exp(\mu_T)}{\exp(\mu_R)} =\frac{G_T}{G_R}

Thus, \(\Delta\) is the natural logarithm of the ratio of geometric means, and \(\exp(\Delta)\) is the geometric mean ratio \(G_T/G_R\). For observed data, the same result holds when \(\mu_T\) and \(\mu_R\) are replaced by the respective sample mean log exposures. No normality assumption is needed for this identity: it follows from the definitions and the logarithm rules. Statistical assumptions may be needed to estimate or test a population contrast, but they are separate from this algebra.

Key result: For the contrast \(T-R\) on the natural-log scale, exponentiation gives the ratio \(T/R\) of geometric mean exposures. The order must stay consistent: reversing the contrast gives the reciprocal ratio, \(G_R/G_T\).

The ratio can be expressed as a relative percentage change. If \(r=G_T/G_R\), then the percent change from \(R\) to \(T\) is \(100(r-1)\%\). A ratio above 1 indicates higher geometric mean exposure for \(T\); a ratio below 1 indicates lower geometric mean exposure for \(T\). For example, \(r=1.20\) means \(T\)'s geometric mean is 20% higher than \(R\)'s. A ratio of \(0.84\) means it is 16% lower, because \(100(0.84-1)=-16\%\). These are relative comparisons, not differences in exposure units.

Worked Formulation Comparisons

Worked Example: A 20% Higher Geometric Mean Exposure

Suppose a hypothetical comparison records AUC values, in the same units, for three observations under each formulation. Formulation \(R\) has values \(80, 100, 125\), and formulation \(T\) has values \(96, 120, 150\). Calculate the difference between the sample mean log exposures and convert it to a ratio.

Check the conditions. All six AUC values are positive, and the same endpoint and units are used for both formulations. It is therefore valid to take natural logarithms and compare the resulting summaries.

For \(R\), the sample mean log exposure is:

\overline{\ln x}_R =\frac{\ln(80)+\ln(100)+\ln(125)}{3} =\frac{4.3820+4.6052+4.8283}{3} =4.6052

For \(T\), it is:

\overline{\ln x}_T =\frac{\ln(96)+\ln(120)+\ln(150)}{3} =\frac{4.5643+4.7875+5.0106}{3} =4.7875

With the contrast defined as \(T-R\), the difference is \(4.7875-4.6052=0.1823\). Exponentiating gives:

\exp(0.1823)=1.200

The sample geometric means are \(GM_R=\exp(4.6052)=100\) and \(GM_T=\exp(4.7875)=120\), so their ratio is \(120/100=1.200\), as the identity predicts. In this hypothetical data set, the geometric mean AUC for \(T\) is 20% higher than for \(R\). The result is a relative comparison; the difference between the geometric means is 20 AUC units.

Worked Example: A Ratio Below One and the Importance of Direction

In another hypothetical comparison, suppose the sample geometric mean AUC is \(100\) units for \(R\) and \(84\) units for \(T\). Find the log-scale contrast in the stated order \(T-R\), then interpret it.

By the difference-to-ratio identity, the contrast is the natural logarithm of the geometric mean ratio:

\Delta =\ln\left(\frac{GM_T}{GM_R}\right) =\ln\left(\frac{84}{100}\right) =\ln(0.84) =-0.1744

Exponentiating returns \(\exp(-0.1744)=0.8400\), the ratio \(T/R\). Since this ratio is below 1, the geometric mean exposure for \(T\) is lower. The relative change from \(R\) to \(T\) is \(100(0.8400-1)=-16.00\%\), so \(T\)'s geometric mean is 16% lower than \(R\)'s.

If the contrast had instead been defined as \(R-T\), its value would be \(+0.1744\), and exponentiating would give \(100/84=1.190\). That is the ratio \(R/T\), not \(T/R\). A positive contrast does not inherently mean that a particular formulation has higher exposure; it means the formulation named first in the contrast has the higher geometric mean.

Worked Example: Geometric Mean Ratio Versus Arithmetic Mean Ratio

Consider hypothetical positive AUC values in a common unit. Formulation \(R\) has \(1, 4, 9\), and formulation \(T\) has \(2, 4, 18\). Calculate the geometric mean ratio and compare it with the ratio of arithmetic means.

The sample mean log exposures are:

\overline{\ln x}_R =\frac{\ln(1)+\ln(4)+\ln(9)}{3} =\frac{0+1.3863+2.1972}{3} =1.1945

\overline{\ln x}_T =\frac{\ln(2)+\ln(4)+\ln(18)}{3} =\frac{0.6931+1.3863+2.8904}{3} =1.6566

The log-scale difference is \(1.6566-1.1945=0.4621\), so the geometric mean ratio is:

\exp(0.4621)=1.587

The same result follows directly from the sample geometric means: \(GM_R=\exp(1.1945)\approx3.302\), \(GM_T=\exp(1.6566)\approx5.242\), and \(5.242/3.302\approx1.587\). In contrast, the arithmetic means are \((1+4+9)/3=4.667\) and \((2+4+18)/3=8.000\), whose ratio is \(8.000/4.667\approx1.714\). These ratios differ because an arithmetic mean and a geometric mean summarize the data differently. The log contrast specifically gives the ratio of geometric means; it should not be described as the ratio of arithmetic means.

Interpretation, Assumptions, and Checks

A careful interpretation begins by stating the contrast order and endpoint. For instance, “The \(T-R\) mean-log AUC contrast is \(0.1823\), corresponding to a \(T/R\) geometric mean AUC ratio of \(1.200\)” is unambiguous. The ratio has no units when both exposures use the same units. The geometric means themselves retain the AUC units.

The essential requirements for this conversion are straightforward:

Positive values: every exposure included in a mean log exposure must be greater than zero. A zero or negative value cannot be logged directly; it must not be silently replaced by an arbitrary small positive value.
A common endpoint and scale: the two summaries must refer to the same exposure measure, such as AUC, using comparable definitions.
Consistent units: both formulations must use the same units. Changing the common units does not change the ratio, but mixing units between formulations can create a spurious contrast.
Explicit contrast order: a \(T-R\) difference corresponds to \(T/R\); changing the order changes the sign and takes the reciprocal ratio.

The ratio is invariant to a common change of exposure units. If all values are multiplied by a positive constant \(c\), then \(\ln(cX)=\ln(c)+\ln(X)\). The mean log exposure for each formulation increases by \(\ln(c)\), which cancels in their difference:

\{\mu_T+\ln(c)\}-\{\mu_R+\ln(c)\} =\mu_T-\mu_R

Equivalently, both geometric means are multiplied by \(c\), leaving their ratio unchanged. For example, expressing both formulations’ AUC values in a unit 1,000 times smaller adds \(\ln(1000)\) to each mean log exposure, but the \(T-R\) difference and \(T/R\) ratio remain the same.

Common Errors and Limitations

Confusing a log difference with an exposure-unit difference. A contrast of \(0.1823\) is not an increase of \(0.1823\) AUC units. Exponentiating gives a dimensionless ratio of \(1.200\).
Reversing the ratio. If the contrast is \(T-R\), report \(T/R\). The inverse ratio \(R/T\) corresponds to the opposite contrast.
Calling a ratio a percent increase without subtracting one. A ratio of \(1.20\) corresponds to a 20% increase, not a 120% increase. Use \(100(r-1)\%\).
Assuming the identity requires lognormal data. It does not. The identity follows from the definition of the geometric mean for positive data. A lognormal distribution may be relevant to other modeling questions, but it is not required for this algebra.
Interpreting the ratio as a causal or inferential conclusion. The calculation describes the summaries compared. By itself, it does not show that formulation caused a change, quantify uncertainty, or establish equivalence.

The main limitation is that a geometric mean ratio answers a particular question: how the geometric means compare. It does not generally equal the ratio of arithmetic means, and it does not convey all features of the exposure distributions. Nor does a point ratio indicate how precisely the population ratio has been estimated. For inferential conclusions, uncertainty and study design must also be considered.

Key takeaway: Define the log contrast in a stated order, then exponentiate it. A \(T-R\) difference between mean natural-log exposures gives the \(T/R\) geometric mean exposure ratio, with percent change calculated as \(100\{\exp(\Delta)-1\}\%\).

Check Your Understanding

For each question, use the stated contrast order and keep the interpretation tied to the geometric means.

A \(T-R\) mean-log exposure contrast is \(0.09531\). What is the \(T/R\) geometric mean ratio, and what relative change does it represent?
The geometric mean exposure is 72 units for \(T\) and 90 units for \(R\). Calculate the \(T-R\) contrast and interpret its direction and relative magnitude.
Why does changing both formulations’ exposure values from one common unit to another leave their geometric mean ratio unchanged?
Can the exponential of a mean-log contrast be interpreted automatically as the ratio of arithmetic means? Explain.

Solutions. (1) \(\exp(0.09531)\approx1.100\), so \(T\)'s geometric mean exposure is about 10.0% higher than \(R\)'s. (2) The ratio is \(72/90=0.8000\); the contrast is \(\ln(0.8000)=-0.2231\). The geometric mean for \(T\) is 20.0% lower than for \(R\). (3) Multiplying both sets of values by the same positive conversion constant adds the same logarithm to both mean log exposures; that addition cancels when they are subtracted, so exponentiation gives the unchanged ratio. (4) No. It gives the ratio of geometric means by definition. The ratio of arithmetic means is a different quantity and need not have the same value.