From Reference Variability to Ratio-Scale Limits
When EMA reference scaling applies, the acceptance limits are widened according to within-subject variability in the reference product. The calculation has two stages: multiply the reference standard deviation on the log scale by 0.760, then exponentiate the positive and negative results. The resulting limits are expressed as ratios, so they can be compared with a treatment-to-reference ratio.
The earlier material on deriving the log-scale expansion of equivalence limits explains the origin of the factor 0.760. Here the task is to calculate the limits accurately and interpret them. These are expanded acceptance limits before applying any separate cap or other requirements. They are not confidence limits for the treatment effect.
The quantity \(s_{WR}\) is a standard deviation of log values, not a standard deviation in the original measurement units. It is nonnegative and has no physical units on the log scale. The ratio-scale limits are dimensionless. If an analysis supplies \(s_{WR}\), use that value in the formula; do not substitute a treatment-group standard deviation or a raw-scale reference standard deviation.
The calculation presumes that EMA reference scaling is applicable to the endpoint and analysis in question, and that the supplied \(s_{WR}\) is the appropriate within-subject reference estimate from the prescribed design and model. It does not by itself establish eligibility for scaling, validate the variance estimate, apply a cap, or determine the complete bioequivalence decision. Those questions require the applicable regulatory rules and analysis.
Calculation in Two Scales
Write the calculation explicitly so the log-scale and ratio-scale quantities do not get mixed. First compute \(a=0.760s_{WR}\). The log-scale limits are \(-a\) and \(+a\). Then exponentiate each limit separately: \(L=\exp(-a)\) and \(U=\exp(a)\). Retain enough digits during the calculation and round only the reported bounds.
Use the estimated within-subject reference SD on the log scale, \(s_{WR}\), from the relevant analysis. Keep the input in its stated scale; if it is instead reported as a coefficient of variation, first use the appropriate conversion rather than inserting the CV directly.
Multiply \(s_{WR}\) by 0.760: \(a=0.760s_{WR}\). The log-scale bounds are \(-a\) and \(+a\).
Calculate \(L=\exp(-a)\) and \(U=\exp(a)\). Report these as the lower and upper ratio-scale limits, respectively.
For \(s_{WR}>0\), the lower bound must be below 1 and the upper bound above 1. Before rounding, \(L=1/U\). Check that the reported values retain this reciprocal relationship to the precision shown.
Worked Example: Calculating Limits From \(s_{WR}=0.400\)
Suppose a hypothetical analysis supplies \(s_{WR}=0.400\) on the log scale. Calculate the uncapped EMA expanded limits.
Log-scale half-width. Substitute the supplied standard deviation into the formula:
The log-scale limits are \(-0.3040\) and \(+0.3040\). Exponentiating gives:
Thus the calculated ratio-scale interval is approximately 0.7379 to 1.3553. The lower bound is below 1 and the upper bound is above 1. Their unrounded values are reciprocals; small differences when checking reciprocals from displayed values can arise from rounding.
Why the Exponential Transformation Works
The limits are first defined symmetrically around zero on the log-ratio scale. If \(R\) denotes a positive treatment-to-reference ratio, its log is \(\log(R)\). A log-scale limit of \(+a\) therefore corresponds to \(R=\exp(a)\), while a limit of \(-a\) corresponds to \(R=\exp(-a)\). Because the exponential function is strictly increasing, the negative limit maps to the lower ratio bound and the positive limit maps to the upper ratio bound.
The reciprocal relationship follows directly from the exponential identity \(\exp(-a)=1/\exp(a)\). It is a useful arithmetic check, not a second method for defining the unrounded lower bound. Calculate both exponentials from the unrounded \(a\), then round them for reporting. If \(s_{WR}=0\), both bounds equal 1; for positive \(s_{WR}\), the bounds straddle 1.
Worked Example: A Second Reference Standard Deviation
For a separate hypothetical calculation, suppose \(s_{WR}=0.385\). Find the log-scale half-width and the two ratio-scale bounds.
Multiply first.
Exponentiate each signed limit.
The resulting uncapped limits are approximately 0.7463 and 1.3399. They are wider than limits obtained with a smaller \(s_{WR}\), because the log-scale half-width increases in direct proportion to \(s_{WR}\). The bounds remain reciprocal before rounding.
Interpreting the Bounds and Checking the Arithmetic
The ratio-scale bounds describe the acceptance interval generated by this scaling calculation. For example, a lower bound of 0.7379 means that the lower edge of this interval is a treatment-to-reference ratio of about 0.7379; an upper bound of 1.3553 means the upper edge is about 1.3553. They are not a confidence interval for the estimated treatment-to-reference ratio, and calculating them alone does not complete a bioequivalence analysis.
The calculation is monotone in \(s_{WR}\). If \(s_{WR}\) increases, \(a\) increases, so \(\exp(-a)\) decreases and \(\exp(a)\) increases. This makes the interval wider in both directions on the ratio scale. It does not mean that an individual study’s treatment effect has changed: the formula changes the acceptance limits in response to the specified reference variability estimate.
Worked Example: Checking a Wider Pair of Limits
Suppose a third hypothetical analysis gives \(s_{WR}=0.500\). Calculate its limits and check their direction and reciprocal relationship.
The log-scale half-width is:
The ratio-scale limits are:
The bounds straddle 1, and \(1/1.4623\) is approximately 0.6839 at the displayed precision. Compared with the preceding example, the larger \(s_{WR}\) gives a lower lower bound and a higher upper bound, as expected.
A practical diagnostic is to record the input, product, signed limits, and exponentials in sequence. If the reported lower bound exceeds 1, the sign or label has likely been reversed. If both bounds are on the same side of 1, check whether the positive and negative signs were used. If the interval narrows when \(s_{WR}\) increases, check the multiplication and exponential steps. These checks can catch arithmetic errors, but they cannot establish that the input variance estimate or the regulatory analysis is appropriate.
Common Mistakes and Limits of the Calculation
- Using a raw-scale SD as \(s_{WR}\). The formula requires the within-subject standard deviation on the log scale. A standard deviation in concentration units cannot be inserted directly.
- Using a CV as though it were a log-scale SD. A coefficient of variation and a log-scale SD are different quantities. If the input is a CV, convert it using the appropriate relationship before applying the limit formula.
- Reversing the signs or the labels. The negative log-scale limit gives \(L=\exp(-a)\), the lower bound; the positive one gives \(U=\exp(a)\), the upper bound.
- Rounding too early. Keep the product \(0.760s_{WR}\) at full available precision, exponentiate that value, and round the bounds at the end. A reciprocal calculated from an already rounded upper bound need not match the direct exponential calculation at the last reported digit.
- Calling the expanded limits a confidence interval. They are acceptance limits calculated from reference variability. They do not express uncertainty around the estimated treatment effect.
- Treating these values as the complete decision rule. The calculation gives uncapped limits only. Other applicable analysis requirements, including the separate cap, must be considered where required.
The principal alternative for an endpoint that is not being analyzed with reference scaling is a fixed equivalence interval, such as the conventional ratio limits 0.80 to 1.25. These fixed bounds do not change with \(s_{WR}\); the expanded EMA bounds do. The calculation here does not determine which approach is appropriate for a particular endpoint or replace the requirements of the relevant regulatory procedure.
The limits also depend on the quality and definition of the supplied \(s_{WR}\). A numerical result can be calculated exactly from an input while still being unsuitable if that input was obtained from the wrong observations, scale, design, or model. Always report the value of \(s_{WR}\), the resulting log-scale half-width, and the ratio-scale limits so that the calculation can be checked and its basis is clear.
Check Your Understanding
Answer each question, then compare your work with the solutions below.
- For \(s_{WR}=0.300\), calculate the log-scale half-width and both ratio-scale limits.
- For \(s_{WR}=0.450\), should the lower bound be higher or lower than the lower bound for \(s_{WR}=0.300\)? Explain without calculating both exponentials.
- A calculation reports \(s_{WR}=0.400\), log-scale limits of \(-0.3040\) and \(+0.3040\), and ratio limits of 1.3553 and 0.7379 in that order. Identify the reporting error.
- Why is it incorrect to describe the calculated ratio-scale limits as a confidence interval for the treatment-to-reference ratio?
Solutions. (1) \(a=0.760(0.300)=0.228\). Thus \(L=\exp(-0.228)\approx0.7961\) and \(U=\exp(0.228)\approx1.2561\). (2) The lower bound is lower for \(s_{WR}=0.450\). A larger \(s_{WR}\) makes \(a\) larger, and \(\exp(-a)\) decreases as \(a\) increases. (3) The values are labelled in reverse order: 0.7379 is the lower bound and 1.3553 is the upper bound. The negative log limit exponentiates to the lower ratio bound. (4) The limits are calculated acceptance bounds determined by reference variability; they do not quantify uncertainty around an estimated treatment effect. A confidence interval is a different inferential quantity.