Highly variable drugs (HVDs) are products whose within-subject variability for the reference product is 30% or more. For these drugs, the conventional 80.00–125.00% bioequivalence limits often demand impractically large studies, even when the test and reference products are truly equivalent. The US FDA addresses this with the reference-scaled average bioequivalence (RSABE) approach, in which the acceptance limits widen in proportion to the variability of the reference product.
This article walks through the complete FDA method on a dummy dataset: the study design, the statistical concepts behind the model, the SAS code, the output, and the interpretation of every number that matters for the decision.
When does the FDA RSABE method apply?
- The study uses a replicate design in which the reference product is given at least twice to each subject: a fully replicated design (TRTR / RTRT) or a partially replicated design (TRR / RTR / RRT).
- The within-subject standard deviation of the reference, sWR, is 0.294 or greater (equivalent to a within-subject CV of about 30%).
- The decision is made separately for each pharmacokinetic parameter (Cmax, AUC0-t, AUC0-inf). A parameter with sWR below 0.294 is assessed with the conventional unscaled method.
The RSABE criterion
On the natural-log scale, test and reference are bioequivalent under RSABE when:
(μT − μR)2 − θ · σ2WR ≤ 0, where θ = (ln 1.25 / σW0)2 and σW0 = 0.25, so θ = 0.7967
Because the true values are unknown, the FDA requires both of the following:
- The 95% upper confidence bound of the linearised criterion, calculated with Howe's approximation, is ≤ 0.
- The point estimate of the test/reference geometric mean ratio lies within 0.8000–1.2500.
Statistical concepts behind the model
The SAS code is short, but each line rests on a specific statistical idea. Understanding these makes the code easier to defend in a protocol, a statistical analysis plan or a regulatory query.
1. Log-normal data and the log transformation
Cmax and AUC are positive, right-skewed and behave multiplicatively: a formulation changes exposure by a percentage, not by a fixed amount. They are therefore 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 on the original scale. This is why every estimate is calculated on the log scale and then exponentiated to give the geometric mean ratio (GMR).
2. The crossover principle: each subject is their own control
Each observation can be written as a formulation mean plus a period effect, a subject effect and a within-subject error. Subjects differ widely from one another, but that between-subject variability cancels when test and reference are compared within the same subject. ILAT and DLAT are both within-subject contrasts, so the subject effect drops out of each and only within-subject sources of variation remain.
3. Variance components
A replicate design allows the total variability to be separated into components that a standard 2×2 crossover cannot distinguish:
| Component | Symbol | Meaning |
|---|---|---|
| Between-subject variance | σ2B | How much subjects differ from each other in average exposure |
| Within-subject variance, reference | σ2WR | How much one subject varies between two doses of the reference |
| Within-subject variance, test | σ2WT | The same for the test product |
| Subject-by-formulation interaction | σ2D | Whether the test-reference difference varies from subject to subject |
RSABE scales the limits using σ2WR only. This is why the reference product must be given at least twice: within-subject variability cannot be estimated from a single administration.
4. Why the DLAT mean square is divided by two
DLAT is the difference between two reference observations in the same subject. The subject effect cancels, leaving the difference of two independent within-subject errors, so its variance is 2σ2WR. The error mean square from the DLAT model estimates that quantity, and dividing by two gives s2WR. This is a method-of-moments estimate: it is obtained directly from an ANOVA mean square rather than by iterative likelihood fitting.
5. The role of sequence in the model
Sequence is the only fixed effect in both models, and it does two jobs. In the ILAT model, the period effects enter the two sequences with opposite signs, so averaging the sequence means with equal weights (the 0.5 0.5 in the ESTIMATE statement) cancels them and leaves an estimate of μT − μR that is free of period effects. In the DLAT model, fitting sequence removes the period difference between the two reference administrations, so it does not inflate the variance estimate. The equal weighting also keeps the estimate unbiased when the sequences are unbalanced.
6. Sampling distributions: t and chi-square
Under normality, the ILAT estimate divided by its standard error follows a t distribution with the error degrees of freedom (N minus the number of sequences; 22 here). This gives the 90% confidence interval. Independently, the variance estimate follows a scaled chi-square distribution: dfd · s2WR / σ2WR is chi-square with dfd degrees of freedom. This is the source of the cinv(0.95, dfd) term used to bound the variance part of the criterion.
7. Hypothesis testing and the 90% confidence interval
Bioequivalence reverses the usual logic of hypothesis testing. The null hypothesis is that the products are not equivalent, and the study must provide evidence to reject it. Conventional average bioequivalence uses two one-sided tests (TOST), each at the 5% level, which is mathematically the same as requiring the 90% confidence interval to lie within the limits. RSABE keeps the same 5% consumer risk but applies it as a single one-sided 95% upper bound on the scaled criterion.
8. Linearising the scaled criterion
The scaled criterion is naturally a ratio, (μT − μR)2 / σ2WR ≤ θ. Ratios of estimates have awkward distributions, so the inequality is rearranged into a linear form, (μT − μR)2 − θσ2WR ≤ 0. The two terms are estimated independently (x from ILAT, y from DLAT), which makes a confidence bound for their sum tractable.
9. Bias correction for the squared difference
Squaring an estimate overstates the squared true difference, because the expected value of the squared estimate equals the squared true difference plus the squared standard error. Subtracting stderr**2 in the calculation of x removes this bias.
10. Howe's approximation
There is no exact confidence bound for the sum of a squared-normal term and a chi-square term. Howe's method (a modified large-sample approach) builds one from the individual bounds: start from the point estimate of the sum, x + y, and add the square root of the summed squared distances between each component and its own 95% upper bound. This is exactly the critbound formula in Step 4.
11. Mixed scaling and the point estimate constraint
The FDA approach is a mixed procedure: unscaled limits apply below sWR = 0.294 and scaled limits apply above it. With scaling alone, a very variable reference could allow products with quite different average exposure to pass. The requirement that the GMR lies within 0.8000–1.2500 is a clinical safeguard against this, and it also helps to control the type I error around the switching point.
12. Method of moments versus REML
The scaled analysis uses method-of-moments estimates from PROC GLM on complete cases. The unscaled analysis uses restricted maximum likelihood (REML) in PROC MIXED. There, random trt / type=fa0(2) fits a between-subject covariance structure that allows a subject-by-formulation interaction, repeated / grp=trt allows different within-subject variances for test and reference, and Satterthwaite degrees of freedom account for the estimated variance components. REML also uses all available data when some periods are missing.
13. Converting a log-scale variance to a CV
For log-normal data, the within-subject coefficient of variation is CV = 100 × √(exp(s2WR) − 1). The regulatory cut-off sWR = 0.294 corresponds to a CV of 30%.
Study design and dummy data
The example is a single-dose, two-treatment, two-sequence, four-period fully replicated crossover study in 24 healthy subjects, 12 per sequence. The values below are simulated Cmax results (ng/mL) and are for illustration only. Columns C1 to C4 are the Cmax values in periods 1 to 4.
data pk;
input subj seq $ c1 c2 c3 c4;
datalines;
01 TRTR 484.9 153.5 344.3 424.9
02 RTRT 192.5 237.2 322.4 199.8
03 TRTR 135.3 233.2 154.5 240.7
04 RTRT 712.8 664.3 201.4 533.4
05 TRTR 132.1 218.7 283.7 505.6
06 RTRT 388.6 62.7 192.7 239.3
07 TRTR 310.6 175.1 309.6 306.9
08 RTRT 157.6 176.1 155.5 131.2
09 TRTR 232.8 122.3 158.0 260.7
10 RTRT 330.2 217.4 135.5 370.4
11 TRTR 382.5 368.8 218.7 224.0
12 RTRT 256.2 320.3 449.4 315.0
13 TRTR 271.9 204.5 228.8 212.3
14 RTRT 392.7 197.9 166.7 307.5
15 TRTR 432.4 389.0 598.7 538.9
16 RTRT 222.3 137.6 84.4 227.6
17 TRTR 232.2 189.1 225.3 141.6
18 RTRT 105.6 203.1 68.7 191.2
19 TRTR 395.5 115.5 249.2 292.3
20 RTRT 309.2 121.5 114.5 367.6
21 TRTR 570.6 429.6 271.8 143.4
22 RTRT 691.4 303.3 281.9 355.3
23 TRTR 391.4 413.2 361.4 165.4
24 RTRT 308.6 328.4 136.0 218.9
;
run;
SAS code, step by step
Step 1: Log-transform and build the subject-level contrasts
For each subject, two quantities are derived from the log-transformed data:
- ILAT: the mean of the two test values minus the mean of the two reference values. This carries the test-versus-reference difference.
- DLAT: the first reference value minus the second reference value. This carries the within-subject variability of the reference.
data wide;
set pk;
/* Map period values to T1, T2, R1, R2 on the log scale */
if seq = 'TRTR' then do;
t1 = log(c1); r1 = log(c2);
t2 = log(c3); r2 = log(c4);
end;
else if seq = 'RTRT' then do;
r1 = log(c1); t1 = log(c2);
r2 = log(c3); t2 = log(c4);
end;
/* Keep only subjects with all four observations */
if nmiss(t1, t2, r1, r2) = 0;
ilat = 0.5*(t1 + t2) - 0.5*(r1 + r2); /* T - R difference */
dlat = r1 - r2; /* R - R difference */
run;
Step 2: Estimate the test-reference difference (ILAT)
ILAT is analysed with sequence as the only fixed effect. The ESTIMATE statement averages over the two sequences and CLPARM with ALPHA=0.1 returns the 90% confidence limits.
proc glm data=wide;
class seq;
model ilat = seq / clparm alpha=0.1;
estimate 'average' intercept 1 seq 0.5 0.5;
ods output Estimates=iglm1;
run;
quit;
data iglm2;
set iglm1;
pointest = exp(estimate); /* GMR (T/R) */
x = estimate**2 - stderr**2; /* unbiased (T-R)**2 */
boundx = (max(abs(lowercl), abs(uppercl)))**2; /* its 95% upper bound*/
keep estimate stderr lowercl uppercl pointest x boundx;
run;
Step 3: Estimate the reference within-subject variance (DLAT)
The error mean square from the DLAT model, divided by two, is the estimate of s2WR. The error degrees of freedom are kept for the chi-square bound.
proc glm data=wide;
class seq;
model dlat = seq;
ods output OverallANOVA=dglm1;
run;
quit;
data dglm2;
set dglm1;
where source = 'Error';
dfd = df;
s2wr = ms / 2;
keep dfd s2wr;
run;
Step 4: Calculate the 95% upper confidence bound and the decision
data rsabe;
merge iglm2 dglm2;
theta = (log(1.25) / 0.25)**2; /* regulatory constant */
swr = sqrt(s2wr);
cvwr = 100 * sqrt(exp(s2wr) - 1); /* within-subject CV (%) */
y = -theta * s2wr;
boundy = y * dfd / cinv(0.95, dfd);
/* Howe's approximation: 95% upper bound of the linearised criterion */
critbound = (x + y) + sqrt((boundx - x)**2 + (boundy - y)**2);
length method $20 conclusion $24;
if swr ge 0.294 then do;
method = 'Reference-scaled';
if critbound le 0 and 0.80 le pointest le 1.25
then conclusion = 'Bioequivalent';
else conclusion = 'Not bioequivalent';
end;
else do;
method = 'Unscaled ABE';
conclusion = 'Use PROC MIXED result';
end;
run;
proc print data=rsabe noobs;
var swr cvwr pointest lowercl uppercl x boundx y boundy critbound
method conclusion;
format swr pointest lowercl uppercl x boundx y boundy critbound 8.4
cvwr 6.2;
run;
Step 5: Unscaled analysis when sWR is below 0.294
If sWR is below 0.294 for a parameter, RSABE does not apply and the conventional 90% confidence interval must fall within 80.00–125.00%. For replicate designs, the FDA recommends the following mixed model, run on a long-format dataset with one record per subject and period (variable LNCMAX holds the log-transformed value).
proc mixed data=pklong;
class trt seq subj per;
model lncmax = seq per trt / ddfm=satterth;
random trt / type=fa0(2) sub=subj g;
repeated / grp=trt sub=subj;
estimate 'T vs R' trt 1 -1 / cl alpha=0.1;
run;
Exponentiate the estimate and its confidence limits to obtain the geometric mean ratio and the 90% confidence interval. In this example the step is not needed because sWR exceeds 0.294.
Output
ILAT model (Step 2)
Dependent Variable: ilat
Source DF Sum of Squares Mean Square
Error 22 3.21235687 0.14601622
Standard
Parameter Estimate Error t Value Pr > |t| 90% Confidence Limits
average 0.11016279 0.07800006 1.41 0.1718 -0.02377455 0.24410012
DLAT model (Step 3)
Dependent Variable: dlat
Source DF Sum of Squares Mean Square
Error 22 9.71204169 0.44145644
s2wr = 0.44145644 / 2 = 0.22072822 dfd = 22
Final RSABE result (Step 4)
| Quantity | SAS variable | Value |
|---|---|---|
| Within-subject SD of reference | swr | 0.4698 |
| Within-subject CV of reference (%) | cvwr | 49.70 |
| Point estimate, T/R geometric mean ratio | pointest | 1.1165 |
| 90% lower limit of ln(T) − ln(R) | lowercl | −0.0238 |
| 90% upper limit of ln(T) − ln(R) | uppercl | 0.2441 |
| Estimate of (μT − μR)2 | x | 0.0061 |
| Upper bound of x | boundx | 0.0596 |
| −θ · s2WR | y | −0.1759 |
| Upper bound of y | boundy | −0.1140 |
| 95% upper confidence bound | critbound | −0.0880 |
| Method applied | method | Reference-scaled |
| Conclusion | conclusion | Bioequivalent |
Interpretation
- Is the drug highly variable? sWR = 0.4698, which is above the 0.294 cut-off, and the within-subject CV of the reference is 49.70%. Cmax therefore qualifies for the reference-scaled approach.
- Is the 95% upper confidence bound ≤ 0? The bound is −0.0880. It is negative, so the first RSABE condition is met.
- Is the point estimate within 0.8000–1.2500? The geometric mean ratio is 1.1165 (111.65%), so the second condition is also met.
- Conclusion. Both conditions are satisfied, so the test product is bioequivalent to the reference product for Cmax under the FDA RSABE criterion.
Why scaling mattered here. Back-transforming the 90% confidence limits of the ILAT estimate gives a conventional interval of 97.65–127.65%. The upper limit is above 125.00%, so the same study would have failed the unscaled criterion. With sWR = 0.4698, the scaled limits implied by the criterion are approximately 65.75–152.10%, which is why the reference-scaled test passes. The point estimate constraint prevents this widening from accepting products whose average exposure differs by more than 20–25%.
How the bound is built. The criterion is split into two parts. The first, x, estimates the squared test-reference difference and its upper bound comes from the 90% confidence interval of ILAT. The second, y, is the scaled reference variance and its upper bound comes from the chi-square distribution with the DLAT error degrees of freedom. Howe's approximation combines the two into a single 95% upper bound. A negative bound means the difference between products is small relative to the variability of the reference.
Practical points
- Partial replicate design (TRR / RTR / RRT). Define ILAT as T − 0.5·(R1 + R2), keep DLAT as R1 − R2, and change the ESTIMATE statement to
estimate 'average' intercept 1 seq 0.3333333333 0.3333333333 0.3333333333;The rest of the code is unchanged. - Each parameter separately. Repeat the analysis for AUC0-t and AUC0-inf. It is common for Cmax to be scaled while AUC is assessed with the unscaled method.
- Incomplete subjects. The RSABE calculation uses only subjects with all required observations. Subjects with missing periods are still included in the PROC MIXED analysis.
- Sample size. The FDA expects a minimum of 24 subjects in a replicate study intended for RSABE.
- Pre-specification. State in the protocol and statistical analysis plan that the reference-scaled approach will be used if sWR ≥ 0.294, and always check the product-specific guidance for the drug.
References
- US FDA. Draft Guidance on Progesterone (oral capsule), which contains the reference-scaled average bioequivalence method and SAS code.
- US FDA. Bioequivalence Studies With Pharmacokinetic Endpoints for Drugs Submitted Under an ANDA, Draft Guidance for Industry.
- Haidar SH, Davit B, Chen ML, et al. Bioequivalence approaches for highly variable drugs and drug products. Pharmaceutical Research. 2008;25(1):237–241.
- Howe WG. Approximate confidence limits on the mean of X + Y where X and Y are two tabled independent random variables. Journal of the American Statistical Association. 1974;69(347):789–794.
The dataset in this article is simulated and is provided for training purposes only.




