Standard error of the log difference in a balanced two-period crossover

In a two-treatment, two-period, two-sequence crossover with n subjects split equally between sequences, the variance of the estimated mean log difference is twice the within-subject variance divided by n, where the within-subject variance is the residual mean square from the analysis of the log data. The within-subject coefficient of variation on the original scale is the square root of exp(MSE) minus 1.

Signature

SE = sqrt(2 * MSE / n)
Inputs
InputsDefinitionUnit
MSEResidual mean square from the analysis of variance of the log-transformed measure, estimating the within-subject variancesquared natural log units
nSubjects completing both periods, divided equally between the two sequencescount of subjects
Output
SEStandard error of the test minus reference difference in log meansnatural log units

Function

Average bioequivalence and generic price function

For an ANDA, bioequivalence to the reference listed drug is shown when the 90% confidence interval for the ratio of geometric means of a pharmacokinetic measure, such as AUC or Cmax, lies within 80.00% to 125.00% after rounding. The interval is computed on the log scale and exponentiated. For economic models, the approval of generic competitors is then translated into a post-entry price.

Implementations

  • Excel

    Standard error from the residual mean square

    With the residual mean square in ResidualMS and the number of subjects in Subjects, Excel returns the standard error; the second formula gives the within-subject coefficient of variation.

    =SQRT(2*ResidualMS/Subjects); =SQRT(EXP(ResidualMS)-1)

Assumptions

  • Balanced sequences and no carryover

    The two sequences have equal numbers of subjects and there is no carryover from the first period. With unequal sequences the factor 2 divided by n becomes one half of the sum of 1 over n_1 and 1 over n_2.

  • Common within-subject variance

    Test and reference share the same within-subject variance. Replicate designs estimate them separately, as used for highly variable and narrow therapeutic index drugs.

Worked examples

  • Residual mean square of 0.0432 in 24 subjects

    A residual mean square of 0.0432 on the log scale, a within-subject coefficient of variation of about 21%, gives a standard error of 0.06 with 24 subjects, the value used in the passing example. The figures are illustrative.

    MSE = 0.0432; n = 24; SE = 0.06
  • Residual mean square of 0.0588 in 24 subjects

    A residual mean square of 0.0588 gives a standard error of 0.07, the value used in the failing example.

    MSE = 0.0588; n = 24; SE = 0.07

Common errors

  • Using the between-subject variance

    In a crossover each subject is their own control, so the between-subject variance cancels from the difference. Using the total variance gives a standard error that is too large and a sample size that is too high.

Sources

  • Textbook analysis of the standard two-period crossover

    Chow SC, Liu JP. Design and Analysis of Bioavailability and Bioequivalence Studies. 3rd ed. Boca Raton: Chapman and Hall/CRC; 2008. Chapters on the standard 2 x 2 crossover design: variance of the estimated formulation effect from the intra-subject mean square, and confidence intervals for average bioequivalence on the log scale.

    View source →

Canonical Identity

Standard error of the log difference in a balanced two-period crossover | HealthEconomics.wiki