Worked Replicate-Study Analysis for an NTI Drug

From replicate data to a decision

A replicate crossover study provides repeated exposure measurements within subjects, including repeated measurements under the reference formulation. The repeated observations support an analysis that accounts for the study design and within-subject dependence. For a narrow therapeutic index (NTI) drug, the final comparison may use tighter acceptance limits than conventional limits. The analysis must therefore keep three things distinct: the model-based estimate of the test-to-reference contrast, the uncertainty around that estimate, and the acceptance limits specified for the drug and jurisdiction.

This worked analysis uses hypothetical data and illustrative limits. Actual requirements depend on the applicable regulatory guidance and the endpoints specified in the protocol. The limits are treated here as fixed: the analysis does not widen them because the study estimates high within-subject variability.

Definition: For an exposure endpoint such as AUC or Cmax, let \(Y=\ln(X)\), where \(X\) is the positive exposure measurement. The model-based treatment contrast is \(d=\mu_T-\mu_R\) on the natural-log scale, with \(\mu_T\) and \(\mu_R\) denoting the model-adjusted mean log exposures for test and reference. The geometric mean ratio (GMR) is \(\exp(d)\). A confidence interval for \(d\) is converted to a confidence interval for the GMR by exponentiating both endpoints.

Design, model, and assumptions

Suppose a hypothetical study uses a four-period, two-sequence replicate design: TRTR and RTRT, where T denotes test and R reference. Each subject receives both formulations, and reference is administered twice. Consider one endpoint at a time, such as AUC. Exposure is positive, the endpoint is log-transformed, and the treatment comparison is based on a linear mixed-effects model.

A suitable model includes fixed effects for sequence, period, and formulation, and a random subject effect nested within sequence. The model’s residual covariance structure must represent the dependence among repeated observations within a subject. The treatment estimate is the model-adjusted contrast between test and reference. Its standard error and degrees of freedom are obtained from the fitted model; in this example, the degrees of freedom are approximated using the Satterthwaite method. The precise covariance structure and degrees-of-freedom method should be specified and justified for the study.

The model relies on assumptions that should be checked, not merely asserted. The log-transformed endpoint should be reasonably described by the model’s conditional distribution; the fixed-effects structure should account for design features that affect the endpoint; and the covariance model should represent within-subject dependence adequately. The model also assumes that the observations used in the analysis are valid under the study’s data-handling rules. Randomization, adequate washout, and the absence of meaningful carryover are design considerations: a statistical model cannot repair a compromised crossover.

Model-based equivalence rule: Let \(\hat d\) be the estimated log-scale contrast, \(SE\) its standard error, and \(df\) its degrees of freedom. A two-sided 90% confidence interval is \( \hat d \pm t_{0.95,df}SE \). If the pre-specified ratio limits are \(L\) and \(U\), equivalence for that endpoint is concluded when the exponentiated interval lies entirely within \([L,U]\). This is equivalent to requiring the full log-scale interval to lie within \([\ln(L),\ln(U)]\).

Worked analysis: AUC under fixed tight limits

Worked Example: AUC in a Four-Period Replicate Study

A hypothetical study has 24 randomized subjects in the TRTR and RTRT sequences, with 12 subjects per sequence. Twenty-four subjects have usable AUC observations for all four periods. The descriptive geometric mean AUCs are 98.0 units for test and 100.0 units for reference. These summaries give a descriptive ratio of \(98.0/100.0=0.9800\); the inferential comparison, however, comes from the prespecified mixed-effects model rather than from these unadjusted summaries.

Assume the model has converged, its covariance structure is appropriate, and its diagnostics raise no material concern. The adjusted test-minus-reference contrast is \(\hat d=-0.02000\) on the natural-log scale, with \(SE=0.02500\) and \(df=11\). For illustration, the applicable fixed AUC limits are 90.00% and 111.11%, corresponding approximately to \(L=0.9000\) and \(U=1/0.9000=1.1111\).

For a two-sided 90% interval, \(t_{0.95,11}=1.796\). The log-scale interval is

-0.02000 \pm 1.796(0.02500) = -0.02000 \pm 0.04490 = [-0.06490,\ 0.02490].

The point estimate on the ratio scale is

\exp(-0.02000)=0.9802,

or a test-to-reference GMR of 98.02%. Exponentiating the confidence limits gives

[\exp(-0.06490),\ \exp(0.02490)] = [0.9372,\ 1.0252].

Thus the 90% confidence interval is 93.72% to 102.52%, which is wholly inside the illustrative 90.00% to 111.11% limits. The AUC endpoint meets this interval criterion. The result does not mean there is a 90% probability that the true ratio lies in this particular interval; it means that the specified confidence-interval procedure has 90% coverage under its assumptions.

Why the interval rule gives an equivalence decision

The interval rule is a confidence-interval form of two one-sided tests. On the ratio scale, the null hypothesis for equivalence is that the true ratio is at or beyond at least one acceptance boundary; the alternative is that it is between the boundaries. On the log scale, the boundaries are \(\ln(L)\) and \(\ln(U)\).

For a 90% two-sided confidence interval, each endpoint is formed using the 95th percentile of the t distribution. If the lower confidence limit exceeds \(\ln(L)\), the lower-bound one-sided test rejects the possibility that the contrast is at or below \(\ln(L)\) at level 0.05. If the upper confidence limit is below \(\ln(U)\), the upper-bound one-sided test rejects the possibility that the contrast is at or above \(\ln(U)\) at level 0.05. Both conditions must hold. Exponentiation preserves the order of endpoints because the exponential function is strictly increasing, so the equivalent ratio-scale check is direct.

\ln(L) < \hat d-t_{0.95,df}SE \quad\text{and}\quad \hat d+t_{0.95,df}SE < \ln(U).

This is why a point estimate inside the limits is not enough: the interval must also fit inside them. Likewise, an interval that overlaps a limit does not establish equivalence, even if most of the interval is within the acceptable range.

Endpoint decisions are separate

AUC and Cmax are analyzed as distinct endpoints when both are required. Each has its own model estimate, standard error, confidence interval, and pre-specified criterion. A satisfactory result for one endpoint cannot compensate for a failure on another endpoint required for the overall conclusion.

Worked Example: Cmax Does Not Meet the Tight Limits

For the same hypothetical study, suppose the adjusted Cmax contrast is \(\hat d=0.07500\), with \(SE=0.02500\) and \(df=11\). Use the same illustrative ratio limits, 0.9000 to 1.1111. The point estimate is

\exp(0.07500)=1.0779,

or 107.79%, which lies inside the limits. But the 90% log-scale interval is

0.07500 \pm 1.796(0.02500) = [0.03010,\ 0.11990].

On the ratio scale this is

[\exp(0.03010),\ \exp(0.11990)] = [1.0306,\ 1.1274],

or 103.06% to 112.74%. The upper confidence limit exceeds 111.11%, so Cmax does not meet the interval criterion. If both AUC and Cmax are required, the study does not meet the joint endpoint requirement, despite the AUC result and despite the Cmax point estimate being within the limits.

Diagnostics and common failure modes

Diagnostics help determine whether the model-based interval is interpretable. They do not provide a substitute decision rule, and they should be evaluated in the context of the design and data.

Check the fitted model. Review convergence messages, the estimated covariance parameters, and whether the standard error and degrees of freedom are plausible. A singular or unstable fit can make the interval unreliable.
Inspect residuals on the log scale. Look for strong skewness, heavy tails, or observations that dominate the treatment contrast. A flagged observation should be investigated against prespecified data rules, not removed simply because its removal changes the result.
Review period and sequence patterns. Large systematic patterns may indicate a model specification problem or a design issue. Consider washout and possible carryover using the study design and relevant evidence; a nonsignificant test alone does not prove carryover is absent.
Account for missing observations. Identify which subjects and periods are missing and why. A mixed model can use incomplete records under its assumptions, but it does not make data missing at random. A sensitivity analysis may be warranted if the missingness could depend on exposure or treatment response.
Confirm the endpoint and limits. Check units, transformations, analysis population, and that the correct authority-specific limits were specified before interpreting the interval.

A practical alternative for a balanced replicate study is to calculate within-subject test-minus-reference differences and analyze their mean. That approach can be transparent when its assumptions and weighting match the design. The mixed-effects analysis is generally more adaptable to period and sequence effects, repeated measurements, and incomplete data, but its conclusions depend on a correctly specified model. Neither approach justifies replacing fixed tight limits with limits widened in response to observed variability. The choice of design and estimation of within-subject variance are distinct from the fixed-limit decision considered here.

Common mistake: confusing precision with the acceptance rule

Common mistake: “The point estimate is close to 1, and the study has high variability, so the observed confidence interval should be judged against wider limits.” This changes the criterion after seeing the data. Under a fixed-limit approach, variability affects the interval width through \(SE\), while the limits remain those specified for the drug and jurisdiction. The correct decision is to compare the full interval with those pre-specified limits.

A wide interval may reflect genuine imprecision even when the estimated ratio is near 1. In that situation, the data may not establish equivalence. That is not the same as proving that the formulations differ; it means the required interval criterion was not met.

Worked Example: Reporting a Joint Conclusion

Worked Example: Combining AUC and Cmax Results

Suppose the hypothetical AUC interval is 93.72% to 102.52%, while the Cmax interval is 103.06% to 112.74%. Both use the same illustrative 90.00% to 111.11% limits. Compare each interval with its limits separately.

For AUC, \(93.72\%>90.00\%\) and \(102.52\%<111.11\%\), so the interval is contained within the limits. For Cmax, \(112.74\%>111.11\%\), so the interval is not contained within the limits. If both endpoints are required, the overall conclusion is that the study does not meet the joint bioequivalence criteria under these illustrative requirements. Report the endpoint-specific estimates and intervals, rather than reducing the result to the point estimates alone.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

A model gives \(\hat d=0.0100\), \(SE=0.0200\), and \(df=15\). Write the formula for its 90% confidence interval on the log scale.
Why must a confidence interval, rather than only the GMR point estimate, be compared with the tight limits?
An endpoint has a ratio-scale 90% interval of 92.0% to 110.0%, with limits of 90.0% to 111.11%. Does it meet the interval criterion?
If AUC meets its interval criterion but a required Cmax endpoint does not, what is the joint conclusion?

Solutions. (1) The interval is \(0.0100\pm t_{0.95,15}(0.0200)\), using the t critical value for 15 degrees of freedom. (2) The point estimate does not represent uncertainty; equivalence requires the entire interval to be inside both boundaries. (3) Yes. Both 92.0% exceeds 90.0% and 110.0% is below 111.11%. (4) The joint requirement is not met if both endpoints are required; success on one endpoint does not offset failure on the other.