Average bioequivalence (ABE) is the standard method the US FDA uses to decide whether a generic (test) product delivers the same rate and extent of absorption as the reference product. For most drugs it is assessed in a two-treatment, two-period, two-sequence (2×2) crossover study, and the conclusion rests on one result: the 90% confidence interval for the test/reference geometric mean ratio must lie within 80.00–125.00%.
This article walks through the complete analysis on a dummy dataset: the statistical concepts behind the model, the SAS code, the output, and the interpretation of each number that matters for the decision.
What the FDA requires
- Parameters. Cmax, AUC0-t and AUC0-inf, each analysed separately after natural-log transformation.
- Model. Analysis of variance with sequence, subject nested within sequence, period and treatment.
- Criterion. The 90% confidence interval of the geometric mean ratio (test/reference) must fall entirely within 80.00–125.00% for every parameter.
- No rounding in your favour. The limits are applied to the interval reported to two decimal places. A lower limit of 79.99% fails.
Statistical concepts behind the model
1. Log-normal data and the log transformation
Cmax and AUC are positive, right-skewed and behave multiplicatively, so they are assumed to follow a log-normal distribution. After taking natural logs the data are approximately normal with stable variance, and a difference of log means corresponds to a ratio of geometric means. Every estimate is therefore calculated on the log scale and exponentiated at the end.
2. The 2×2 crossover model
Each log-transformed observation is modelled as:
Y = overall mean + sequence + subject(sequence) + period + treatment + residual error
| Term | What it accounts for |
|---|---|
| Sequence | Differences between the group dosed TR and the group dosed RT (also where unequal carryover would show) |
| Subject(sequence) | Differences between subjects; this removes between-subject variability from the treatment comparison |
| Period | Systematic differences between period 1 and period 2 (for example assay drift or study conduct) |
| Treatment | The test-versus-reference difference: the quantity of interest |
| Residual error | Within-subject variability, used to build the confidence interval |
3. Each subject is their own control
Because every subject receives both products, the comparison is made within subjects. Between-subject variability, which is usually much larger than within-subject variability, drops out. This is why a crossover study needs far fewer subjects than a parallel-group study.
4. Why the limits are 80.00–125.00%
The limits look asymmetric on the original scale but are symmetric on the log scale: ln(0.80) = −0.2231 and ln(1.25) = +0.2231. A 20% decrease and a 25% increase are the same distance from a ratio of 1 once the data are log-transformed.
5. Two one-sided tests (TOST) and the 90% confidence interval
Bioequivalence reverses the usual logic of hypothesis testing. The null hypothesis is that the products are not equivalent (the ratio is below 80% or above 125%), and the study must reject it. Schuirmann's procedure does this with two one-sided t-tests, each at the 5% level. Rejecting both is mathematically identical to the 90% confidence interval lying inside the limits, which is why a 90% interval, not a 95% interval, is reported.
6. Standard error of the treatment difference
In a 2×2 crossover the standard error of the difference is √(MSE/2 × (1/n1 + 1/n2)), where MSE is the residual mean square and n1, n2 are the number of subjects in each sequence. The confidence interval uses the t distribution with the residual degrees of freedom (N − 2).
7. Type III sums of squares and the sequence test
Type III sums of squares test each effect adjusted for all the others, which gives valid results when the sequences are unbalanced. The sequence effect is a between-subject effect, so it must be tested against the subject(sequence) mean square, not the residual. This is the purpose of the TEST statement in the code. By convention it is assessed at the 10% level.
8. Least-squares means and geometric means
Least-squares means are treatment means adjusted for the other terms in the model. Exponentiating an LS mean on the log scale gives the geometric LS mean, and exponentiating the difference between the two LS means gives the geometric mean ratio.
9. Intra-subject CV
The residual mean square estimates the within-subject variance on the log scale. It converts to a coefficient of variation with CV = 100 × √(exp(MSE) − 1). This value drives the width of the confidence interval and is the main input for sample size calculation in future studies.
10. Statistical significance is not the same as inequivalence
The p-value for treatment tests whether the ratio differs from exactly 100%. Bioequivalence asks a different question: whether the ratio is close enough to 100%. A precise study can show a statistically significant treatment effect and still be bioequivalent. The example below shows exactly this for Cmax.
Study design and dummy data
The example is a single-dose, randomised, two-treatment, two-period, two-sequence crossover study in 24 healthy subjects, 12 per sequence (TR and RT). The values are simulated and are for illustration only. Cmax is in ng/mL and AUC0-t in ng·h/mL. The suffix 1 or 2 is the period.
data pk;
input subj seq $ cmax1 cmax2 auct1 auct2;
datalines;
01 TR 446.4 540.0 4319 4474
02 RT 380.0 303.7 2724 3759
03 TR 314.6 286.5 2668 2699
04 RT 397.3 439.9 3467 4000
05 TR 307.1 325.9 1729 2268
06 RT 487.0 333.0 2773 3182
07 TR 294.0 503.1 3460 3703
08 RT 423.1 496.2 3569 3304
09 TR 579.5 775.6 6291 5030
10 RT 529.5 380.4 4757 3972
11 TR 359.2 576.6 3232 3932
12 RT 585.4 533.0 4610 3835
13 TR 469.6 515.9 4854 6239
14 RT 362.4 392.9 4740 4700
15 TR 523.5 536.4 5048 3367
16 RT 265.4 182.5 2358 2404
17 TR 342.2 388.7 3846 3940
18 RT 356.4 391.1 3413 3187
19 TR 341.5 364.1 4288 4311
20 RT 558.6 479.8 4019 3950
21 TR 475.3 528.6 4986 4789
22 RT 288.9 267.3 3120 2360
23 TR 542.1 475.4 4473 4588
24 RT 370.6 494.0 5571 6000
;
run;
SAS code, step by step
Step 1: Restructure to one record per subject and period
The treatment received in each period is read from the sequence code, and both parameters are log-transformed.
data pklong;
set pk;
length trt $1;
array c{2} cmax1 cmax2;
array a{2} auct1 auct2;
do per = 1 to 2;
trt = substr(seq, per, 1); /* T or R */
cmax = c{per};
auct = a{per};
lncmax = log(cmax); /* natural log */
lnauct = log(auct);
output;
end;
keep subj seq per trt cmax auct lncmax lnauct;
run;
Step 2: Fit the ANOVA model with PROC GLM
Treatment levels sort alphabetically as R then T, so the coefficients -1 1 estimate test minus reference. CLPARM with ALPHA=0.10 returns the 90% confidence limits.
ods output OverallANOVA=anova Estimates=est LSMeans=lsm;
proc glm data=pklong;
class seq subj per trt;
model lncmax lnauct = seq subj(seq) per trt / ss3 clparm alpha=0.10;
/* Sequence is tested against between-subject variation */
test h=seq e=subj(seq) / htype=3 etype=3;
lsmeans trt / stderr;
estimate 'T vs R' trt -1 1;
run;
quit;
Step 3: Back-transform and apply the bioequivalence limits
data mse;
set anova;
where source = 'Error';
mse = ms;
keep dependent mse;
run;
proc sort data=est; by dependent; run;
proc sort data=mse; by dependent; run;
data abe;
merge est mse;
by dependent;
ratio = 100 * exp(estimate); /* geometric mean ratio (%) */
lower = 100 * exp(lowercl); /* 90% CI lower limit (%) */
upper = 100 * exp(uppercl); /* 90% CI upper limit (%) */
cv_intra = 100 * sqrt(exp(mse) - 1); /* intra-subject CV (%) */
length conclusion $20;
if round(lower, 0.01) ge 80.00 and round(upper, 0.01) le 125.00
then conclusion = 'Bioequivalent';
else conclusion = 'Not bioequivalent';
run;
proc print data=abe noobs;
var dependent ratio lower upper cv_intra conclusion;
format ratio lower upper cv_intra 8.2;
run;
Step 4: Geometric least-squares means
data geolsm;
set lsm;
geomean = exp(lsmean);
keep dependent trt lsmean geomean;
run;
proc print data=geolsm noobs;
format lsmean 10.6 geomean 10.2;
run;
Alternative: PROC MIXED when data are incomplete
With complete data, PROC GLM and PROC MIXED give the same estimate and confidence interval. When a subject is missing a period, PROC GLM effectively discards that subject for the treatment comparison, while PROC MIXED with subject as a random effect keeps the available observation. The method must be stated in the protocol in advance.
proc mixed data=pklong method=reml;
class seq subj per trt;
model lncmax = seq per trt / ddfm=kr;
random subj(seq);
lsmeans trt;
estimate 'T vs R' trt -1 1 / cl alpha=0.10;
run;
Output
ANOVA for ln(Cmax)
| Source | DF | Sum of squares | Mean square |
|---|---|---|---|
| Model | 25 | 3.03501179 | 0.12140047 |
| Error | 22 | 0.49895726 | 0.02267988 |
| Corrected total | 47 | 3.53396905 |
R-square = 0.8588, Root MSE = 0.1506.
Type III tests for ln(Cmax)
| Source | DF | Type III SS | F value | Pr > F | Error term |
|---|---|---|---|---|---|
| Sequence | 1 | 0.14411559 | 1.16 | 0.2928 | Subject(sequence) |
| Period | 1 | 0.01443432 | 0.64 | 0.4335 | Residual |
| Treatment | 1 | 0.14670652 | 6.47 | 0.0185 | Residual |
Type III tests for ln(AUC0-t)
| Source | DF | Type III SS | F value | Pr > F | Error term |
|---|---|---|---|---|---|
| Sequence | 1 | 0.08456141 | 0.57 | 0.4602 | Subject(sequence) |
| Period | 1 | 0.00020880 | 0.01 | 0.9096 | Residual |
| Treatment | 1 | 0.00237861 | 0.15 | 0.7021 | Residual |
Residual mean square for ln(AUC0-t) = 0.01584409 with 22 degrees of freedom.
Estimate of test minus reference (log scale)
| Parameter | Estimate | Standard error | t value | Pr > |t| | 90% lower | 90% upper |
|---|---|---|---|---|---|---|
| ln(Cmax) | −0.11056918 | 0.04347401 | −2.54 | 0.0185 | −0.18522032 | −0.03591804 |
| ln(AUC0-t) | −0.01407896 | 0.03633649 | −0.39 | 0.7021 | −0.07647394 | 0.04831602 |
Final bioequivalence summary
| Parameter | Geometric LS mean, test | Geometric LS mean, reference | Ratio T/R (%) | 90% CI (%) | Intra-subject CV (%) | Conclusion |
|---|---|---|---|---|---|---|
| Cmax (ng/mL) | 390.22 | 435.84 | 89.53 | 83.09 – 96.47 | 15.15 | Bioequivalent |
| AUC0-t (ng·h/mL) | 3753.76 | 3806.98 | 98.60 | 92.64 – 104.95 | 12.64 | Bioequivalent |
Interpretation
- Cmax. The geometric mean ratio is 89.53% with a 90% confidence interval of 83.09–96.47%. The whole interval lies within 80.00–125.00%, so the test product is bioequivalent to the reference for rate of absorption.
- AUC0-t. The ratio is 98.60% with a 90% confidence interval of 92.64–104.95%, well inside the limits. The products are bioequivalent for extent of absorption.
- Sequence effect. The p-values are 0.2928 for Cmax and 0.4602 for AUC0-t, both above 0.10. There is no evidence of a sequence or unequal carryover effect.
- Period effect. The p-values are 0.4335 and 0.9096. There is no systematic difference between periods. A significant period effect would not invalidate the study, because period is adjusted for in the model.
- Intra-subject variability. The CV is 15.15% for Cmax and 12.64% for AUC0-t. Both are well below 30%, so the product is not highly variable and the conventional unscaled limits are the appropriate criterion.
- Overall conclusion. Both parameters meet the criterion, so the test product is bioequivalent to the reference product. In a real submission AUC0-inf must also pass.
A significant treatment effect that is still bioequivalent. For Cmax the treatment p-value is 0.0185, so the ratio is statistically different from 100%. The confidence interval even excludes 100%. This does not affect the conclusion: the interval is entirely within 80.00–125.00%, and that is the only criterion. The study is simply precise enough to detect a difference of about 10% that is not considered clinically relevant.
Checking the standard error by hand. For Cmax, √(0.02267988 / 2 × (1/12 + 1/12)) = 0.0435, which matches the SAS standard error. With t(0.95, 22) = 1.7171, the interval is −0.1106 ± 1.7171 × 0.0435, giving −0.1852 to −0.0359, and exponentiating gives 83.09–96.47%.
Practical points
- Pre-specify the model. State the procedure (GLM or MIXED), the model terms and the handling of missing data in the protocol and statistical analysis plan.
- Subject exclusions. Decide exclusions such as vomiting or pre-dose concentrations above 5% of Cmax before unblinding the analysis, using criteria written in the protocol.
- Tmax. Tmax is not log-transformed and is not assessed with this model. It is summarised descriptively or compared with a non-parametric method when clinically relevant.
- Highly variable drugs. If the within-subject CV of the reference is 30% or more, a replicate design with the reference-scaled approach may be more suitable. See Reference-Scaled Average Bioequivalence (RSABE) for FDA.
- Narrow therapeutic index drugs. The FDA applies tighter, product-specific criteria. Always check the product-specific guidance.
References
- US FDA. Statistical Approaches to Establishing Bioequivalence, Guidance for Industry. 2001.
- US FDA. Bioequivalence Studies With Pharmacokinetic Endpoints for Drugs Submitted Under an ANDA, Draft Guidance for Industry.
- Schuirmann DJ. A comparison of the two one-sided tests procedure and the power approach for assessing the equivalence of average bioavailability. Journal of Pharmacokinetics and Biopharmaceutics. 1987;15(6):657–680.
- Chow SC, Liu JP. Design and Analysis of Bioavailability and Bioequivalence Studies. 3rd edition. Chapman and Hall/CRC; 2008.
The dataset in this article is simulated and is provided for training purposes only.




