Why Analyze Exposure Over an Early Interval?
For a rapid-onset product, the timing of exposure can matter as well as the total amount of exposure. Two formulations can have similar total area under the concentration–time curve (AUC) but differ in how much exposure occurs soon after dosing. A partial AUC makes that early portion a defined endpoint, so the comparison directly addresses a prespecified interval of interest.
The interval must be selected before examining treatment results and justified by the clinical or pharmacological question. It is not a universal time window: an appropriate interval depends on the product and its intended onset. The analysis should ordinarily report both the partial AUC and total AUC, because the early comparison does not, by itself, describe exposure over the full profile.
With observed concentrations, the area is estimated numerically. Under the linear trapezoidal rule, two consecutive concentrations \(C_i\) and \(C_{i+1}\), measured at \(t_i\) and \(t_{i+1}\), contribute \((C_i+C_{i+1})(t_{i+1}-t_i)/2\). The interval boundaries should align with sampling times when feasible. If a boundary falls between samples, the interpolation method must be specified in advance and applied consistently to both treatments.
Calculating Partial AUC From Concentration Data
Worked Example: Early and Total AUC From One Profile
Suppose a hypothetical subject receives two formulations, T and R, in a crossover study. Concentrations are measured in mg/L at the same times after each dose. The study prespecifies \(AUC_{0-1}\) as its early endpoint and follows concentrations through 4 hours. The table shows one subject’s illustrative profiles.
| Time (h) | T concentration (mg/L) | R concentration (mg/L) |
|---|---|---|
| 0 | 0 | 0 |
| 0.25 | 20 | 10 |
| 0.5 | 35 | 25 |
| 1 | 40 | 38 |
| 2 | 30 | 32 |
| 4 | 15 | 16 |
For T, the first three trapezoids give the area from 0 to 1 hour:
For R, the corresponding calculation is \(1.25+4.375+15.75=21.375\ \text{mg}\cdot\text{h/L}\). In this subject, the early partial-AUC ratio is \(28.125/21.375=1.316\). This individual ratio is descriptive; it is not a study-level bioequivalence conclusion.
For comparison, the remaining trapezoids from 1 to 4 hours contribute \(35+45=80\ \text{mg}\cdot\text{h/L}\) for T and \(35+48=83\ \text{mg}\cdot\text{h/L}\) for R. Thus, the observed \(AUC_{0-4}\) values are \(108.125\) and \(104.375\ \text{mg}\cdot\text{h/L}\), respectively. This profile illustrates how a larger early area can coexist with a smaller difference in total area. It does not establish a population-level treatment effect.
Comparing Treatments in a Crossover Study
The earlier material on complete 2×2 crossover analysis provides the general framework for analyzing a two-period crossover. For a simple complete crossover, calculate the partial AUC for each subject under each treatment and compare the natural logarithms of the two values. The analysis assumes that the interval and AUC calculation are prespecified, measurements are collected comparably under both treatments, and the study design supports a within-subject comparison. It also requires independent subjects and an adequately behaved distribution of within-subject log differences for the small-sample \(t\) interval to be appropriate. Design problems such as carryover or serious period effects can undermine the comparison; a statistical calculation cannot repair a flawed study.
Define each subject’s difference as \(d_i=\ln(AUC_{T,i})-\ln(AUC_{R,i})=\ln(AUC_{T,i}/AUC_{R,i})\). The mean difference \(\bar d\) estimates the log of the geometric mean treatment ratio. A two-sided 90% confidence interval on the log scale is \(\bar d \pm t_{0.95,df}SE\), where \(SE=s_d/\sqrt{n}\), \(s_d\) is the sample standard deviation of the subject-level differences, and \(df=n-1\) for the simple paired analysis. Exponentiating the estimate and interval limits gives the geometric mean ratio and its 90% confidence interval.
In a formal crossover model, treatment, period, and sequence are represented as appropriate to the design, with subject accounted for as a source of within-subject variation. The treatment contrast and its standard error then come from that model. The paired calculation below is a transparent special case, not a substitute for a suitable model when the design or data require one.
Worked Example: Early and Total Exposure Meet Prespecified Limits
In a hypothetical 12-subject crossover study, the protocol prespecifies an early partial AUC and total AUC, and specifies ratio limits of 0.80 to 1.25 for each endpoint. These limits are assumptions for this example, not universal criteria. For the early endpoint, suppose the mean of the 12 subject-level log ratios is \(\bar d=\ln(1.12)=0.1133\), and their standard deviation is \(s_d=0.120\). Use a paired \(t\) interval with \(df=11\), \(t_{0.95,11}=1.796\), and \(SE=0.120/\sqrt{12}=0.03464\).
The 90% interval on the log scale is \(0.1133 \pm 1.796(0.03464)=0.1133\pm0.0622\), or \([0.0511,0.1755]\). Exponentiating gives a geometric mean ratio of \(1.12\) and a 90% confidence interval of \([\exp(0.0511),\exp(0.1755)]=[1.052,1.192]\). Because the entire interval lies within \([0.80,1.25]\), the early partial-AUC endpoint meets the example’s prespecified equivalence criterion.
For total AUC, suppose the geometric mean ratio is 1.02 and \(s_d=0.100\). Then \(\bar d=\ln(1.02)=0.01980\), \(SE=0.100/\sqrt{12}=0.02887\), and the log interval is \(0.01980\pm1.796(0.02887)=[-0.0320,0.0716]\). After exponentiation, the ratio is 1.02 with a 90% interval of \([0.9685,1.074]\), also inside \([0.80,1.25]\). Under the stated assumptions and criteria, both endpoints meet the example’s equivalence criterion. This conclusion concerns those measured AUC endpoints; it does not establish that every aspect of the formulations is equivalent.
Justifying the Log-Scale Interval and Equivalence Rule
The log transformation is useful because treatment comparisons are naturally expressed as ratios, and positive AUC values become differences on the log scale. For subject \(i\), let \(d_i=\ln(AUC_{T,i}/AUC_{R,i})\). If the subject differences are independent and approximately normally distributed with common population mean \(\mu_d\), then the paired \(t\) statistic has a \(t\) distribution with \(n-1\) degrees of freedom when the differences are normally distributed. The standard error of \(\bar d\) is \(s_d/\sqrt{n}\), giving the interval used above. Exponentiation preserves the order of the limits and transforms a difference in logs back into a ratio.
The connection between a 90% interval and two one-sided equivalence tests follows from the rejection conditions. Let \(L\) and \(U\) be the prespecified lower and upper limits on the log-ratio scale. Equivalence requires rejecting both \(H_0:\mu_d\leq L\) and \(H_0:\mu_d\geq U\), each at \(\alpha=0.05\). For a \(t\)-based analysis, those two rejections require the lower confidence limit to exceed \(L\) and the upper confidence limit to be less than \(U\). Those are exactly the conditions for the two-sided 90% confidence interval to lie entirely within the log-scale bounds. Exponentiating returns the equivalent criterion on the ratio scale.
This justification depends on the specified model and its assumptions. If subject-level log differences are markedly skewed, dominated by unusual observations, or poorly estimated because the sample is small, the nominal interval may not have its intended coverage. A prespecified model suited to the design, adequate sampling, and transparent diagnostics are therefore part of the analysis, not optional finishing steps.
When Early Exposure Does Not Meet the Criterion
Worked Example: Total AUC Meets the Criterion but Early AUC Does Not
Consider a hypothetical eight-subject crossover study with the same illustrative limits of 0.80 to 1.25. For early partial AUC, suppose the geometric mean ratio is 1.15 and the standard deviation of paired log ratios is \(s_d=0.200\). With \(df=7\), \(t_{0.95,7}=1.895\). The mean log ratio is \(\ln(1.15)=0.1398\), and \(SE=0.200/\sqrt{8}=0.07071\).
The 90% log interval is \(0.1398\pm1.895(0.07071)=0.1398\pm0.1340=[0.0058,0.2738]\). Exponentiating gives a ratio of 1.15 and an interval of \([1.006,1.315]\). The upper limit exceeds 1.25, so equivalence is not demonstrated for the early endpoint. The point estimate alone being inside the limits is insufficient.
Suppose the total-AUC ratio is 1.01 and \(s_d=0.080\). Then \(\bar d=\ln(1.01)=0.00995\), \(SE=0.080/\sqrt{8}=0.02828\), and the 90% log interval is \(0.00995\pm1.895(0.02828)=[-0.0437,0.0636]\). Exponentiating gives \([0.957,1.066]\), which is inside the example’s limits. This hypothetical result supports equivalence for total AUC under those criteria, but not for early partial AUC. It does not show that the formulations are conclusively different in early exposure; it shows that the early endpoint did not meet the specified equivalence test.
Diagnostics, Alternatives, and Limitations
A partial-AUC analysis can go wrong before any model is fitted. Verify that the same prespecified boundaries, units, integration rule, and interpolation choices were used for both treatments. Inspect the concentration–time profiles and individual partial-AUC values for missing boundary samples, implausible concentrations, data-entry errors, or unusually influential subjects. Check the distribution of paired log differences and the consistency of the treatment contrast across subjects. These checks can reveal fragile results, although they do not justify changing the interval or exclusions after seeing which choice produces a preferred conclusion.
The most important design safeguard is prespecification. Selecting the early interval after comparing observed profiles can make the nominal confidence interval and equivalence decision misleading because the endpoint was chosen using the same data. Sampling must also be dense enough in the interval to estimate the early area meaningfully. Sparse samples can miss a rapid concentration rise or peak, and interpolation cannot recover information that was never collected. Partial AUC is sensitive to its boundaries, so a result for one interval should not be presented as a result for all possible definitions of early exposure.
Partial AUC, total AUC, and \(C_{\max}\) answer related but different questions. A prespecified early partial AUC summarizes integrated exposure within a time window; total AUC summarizes exposure across the measured full profile; \(C_{\max}\) is the observed or estimated maximum concentration. A \(C_{\max}\) comparison may be informative about peak exposure but does not replace an interval-specific measure when the question concerns early integrated exposure. Conversely, a favorable early partial-AUC result does not establish equivalence in total exposure or peak concentration. The endpoints should be interpreted together without treating one as a proxy for the others.
A nonsignificant test of a difference is not an equivalence result: a study may be too imprecise to detect a difference while also being too imprecise to place the full interval inside the equivalence limits. Similarly, a ratio near 1 does not establish equivalence if its confidence interval extends beyond a limit. Use the protocol’s endpoint, model, limits, and decision rule; report estimates and intervals even when the criterion is not met.
Check Your Understanding
Answer each question using the stated assumptions and criteria. Brief solutions follow each problem.
- A hypothetical study reports \(AUC_{0-1,T}=36\) and \(AUC_{0-1,R}=30\), both in mg·h/L. What is the individual ratio, and does it alone establish bioequivalence?
- A paired analysis has \(n=16\), mean log ratio \(\bar d=0\), and standard deviation \(s_d=0.16\). Using \(t_{0.95,15}=1.753\), calculate the 90% ratio interval. Does it meet illustrative limits of 0.80 to 1.25?
- Why is it not valid to choose the early interval after inspecting which interval gives the most favorable treatment comparison?
- In a hypothetical study, the early partial-AUC interval lies within prespecified equivalence limits but the total-AUC interval does not. What can and cannot be concluded?
Solutions. 1. The individual ratio is \(36/30=1.20\). It describes one subject only; bioequivalence is a study-level conclusion requiring the prespecified analysis and uncertainty assessment. 2. \(SE=0.16/\sqrt{16}=0.0400\). The log interval is \(0\pm1.753(0.0400)=[-0.07012,0.07012]\), so the ratio interval is \([\exp(-0.07012),\exp(0.07012)]=[0.9323,1.073]\). It lies within the illustrative limits, so the endpoint meets that criterion under the model assumptions. 3. Selecting the interval using the observed treatment results introduces data-dependent endpoint choice; the ordinary interval and equivalence rule do not account for that selection. 4. The early endpoint meets its stated criterion, but the total-AUC endpoint does not. The early result cannot be substituted for total exposure, and failure to meet the total-AUC criterion does not by itself prove that treatments differ.