Rare Disease Clinical Trials: Primary Endpoint and Statistical Analysis Using SAS

How to choose a primary endpoint in a small rare disease trial and analyse it in SAS: ANCOVA, exact Wilcoxon and Fisher tests, and MMRM, with a dummy dataset, annotated code, output and interpretation.

Rare disease trials face a problem that common-disease trials do not: there are very few patients to study. A pivotal trial may enrol 20 to 60 patients in total, the disease often varies widely between patients, and there may be no validated endpoint from earlier trials. In this setting, the choice of primary endpoint and the statistical analysis largely decide whether a truly effective treatment can be shown to work.

This article explains how a primary endpoint is chosen in a rare disease trial and then works through a complete analysis on a dummy dataset: the statistical concepts, the SAS code, the output, and the interpretation.

Why rare disease trials are different

  • Small populations. In the US a rare disease affects fewer than 200,000 people; many affect only a few hundred. Large trials are not feasible.
  • Heterogeneity. Age of onset, severity and rate of progression often differ widely between patients, which increases variability.
  • Limited natural history data. How the disease progresses without treatment may be poorly documented, which makes it hard to choose an endpoint and a trial duration.
  • Few established endpoints. There may be no endpoint with regulatory precedent, so it has to be justified from first principles.

The regulatory standard does not change: approval still requires substantial evidence of effectiveness. What regulators offer is flexibility in how that evidence is generated, which makes efficient design and analysis essential.

Choosing the primary endpoint

A good primary endpoint measures something that matters to patients (how they feel, function or survive), changes within the duration of the trial, and can be measured reliably.

Endpoint typeExampleStrengthLimitation
Functional clinical outcomeSix-minute walk distance (6MWD), motor function scaleDirectly meaningful to patientsEffort-dependent and variable
Biomarker or surrogateEnzyme activity, substrate level, protein expressionObjective, responds quickly, low variabilityMust be reasonably likely to predict clinical benefit
Patient- or clinician-reported outcomeSymptom score, global impression of changeCaptures the patient's experienceNeeds validation in the specific disease
Time to eventTime to loss of ambulation, time to ventilationClear clinical milestoneNeeds long follow-up and enough events
Composite or multi-domainMulti-domain responder index, global rankHandles heterogeneous diseaseHarder to interpret; components may disagree

Three practical rules apply. Prefer a continuous endpoint to a yes/no responder endpoint, because it carries more information per patient. Use natural history data to confirm the endpoint changes enough during the trial. Agree the endpoint and its clinically meaningful difference with regulators before the trial starts.

Statistical concepts behind the analysis

1. The estimand

Before choosing a method, define what is being estimated: the population, the endpoint, how intercurrent events (such as stopping treatment or starting rescue therapy) are handled, and the summary measure. In this example the summary measure is the difference between treatments in mean change from baseline at Week 48.

2. Analysis of covariance (ANCOVA)

ANCOVA models the change from baseline with treatment as a factor and the baseline value as a covariate. Adjusting for baseline removes the part of the variability that baseline explains and corrects for chance imbalance between arms, which is common in small trials. The result is a more precise treatment estimate and more power from the same patients.

3. Least-squares means

The least-squares (LS) mean is the mean change each arm would have if both arms had the same average baseline. The treatment effect is the difference between the two LS means, reported with its confidence interval.

4. Assumptions matter more in small samples

ANCOVA assumes normally distributed residuals with equal variance. With hundreds of patients, moderate departures do little harm. With 15 per arm, one or two extreme values can change the conclusion, so the residuals should be examined and a distribution-free analysis should be pre-specified as a sensitivity analysis.

5. Exact and non-parametric methods

The Wilcoxon rank-sum test compares the arms using ranks rather than raw values, so it does not depend on normality and is resistant to outliers. In small samples the exact p-value should be used: it is calculated from all possible arrangements of the ranks instead of a large-sample approximation. For yes/no outcomes, Fisher's exact test plays the same role.

6. Responder analysis and loss of information

Classifying each patient as a responder or non-responder is easy to explain, but it discards information: a patient who improves by 31 metres and one who improves by 80 metres are treated the same. A responder analysis therefore has less power than the analysis of the continuous endpoint and is best used as a supportive analysis.

7. Missing data and repeated measures

Each missing value costs proportionally more in a small trial. When the endpoint is measured at several visits, a mixed model for repeated measures (MMRM) uses all available visits from every patient and is valid when data are missing at random. Sensitivity analyses such as multiple imputation or a tipping-point analysis show how robust the result is to that assumption.

8. Multiplicity

With few patients there is a temptation to test many endpoints. The overall type I error must still be controlled, usually with one primary endpoint and a pre-specified hierarchy for key secondary endpoints.

Study design and dummy data

The example is a randomised, double-blind, placebo-controlled trial in a rare neuromuscular disease with 30 patients, 15 per arm. The primary endpoint is the change from baseline in six-minute walk distance (6MWD, metres) at Week 48. A responder is defined as a patient who improves by at least 30 metres. The values are simulated and are for illustration only.

VariableMeaning
USUBJIDUnique subject identifier
TRT01PNPlanned treatment: 1 = Placebo, 2 = Active
BASE6MWD at baseline (metres)
AVAL6MWD at Week 48 (metres)
CHGChange from baseline (derived)
RESPFLResponder flag: Y if CHG is 30 metres or more (derived)
data adeff;
   length usubjid $8 trt01p $8 respfl $1;
   input usubjid $ trt01pn base aval;

   if trt01pn = 1 then trt01p = 'Placebo';
   else                trt01p = 'Active';

   chg = aval - base;                       /* change from baseline */

   if chg ge 30 then respfl = 'Y';          /* responder: +30 m     */
   else              respfl = 'N';

   datalines;
RD-001 1  305  282
RD-002 2  374  408
RD-003 1  301  260
RD-004 2  451  469
RD-005 1  392  357
RD-006 2  438  441
RD-007 1  417  400
RD-008 2  309  389
RD-009 1  314  302
RD-010 2  382  451
RD-011 1  419  439
RD-012 2  292  316
RD-013 1  338  366
RD-014 2  430  481
RD-015 1  270  352
RD-016 2  334  356
RD-017 1  315  365
RD-018 2  331  343
RD-019 1  480  434
RD-020 2  342  326
RD-021 1  347  366
RD-022 2  231  269
RD-023 1  332  283
RD-024 2  381  424
RD-025 1  357  310
RD-026 2  387  363
RD-027 1  354  368
RD-028 2  332  311
RD-029 1  291  239
RD-030 2  317  398
;
run;

SAS code, step by step

Step 1: Descriptive statistics

proc means data=adeff n mean std median min max maxdec=2;
   class trt01pn;
   var base aval chg;
run;

Step 2: Primary analysis with ANCOVA

Treatment is coded 1 = Placebo and 2 = Active, so the coefficients -1 1 estimate active minus placebo. The OUTPUT statement saves the residuals for the assumption check.

ods output LSMeans=lsm Estimates=est;

proc glm data=adeff;
   class trt01pn;
   model chg = trt01pn base / ss3 solution clparm alpha=0.05;
   lsmeans trt01pn / stderr cl;
   estimate 'Active vs Placebo' trt01pn -1 1;
   output out=resid r=resid p=pred;
run;
quit;

Step 3: Check the model assumptions

proc univariate data=resid normal;
   var resid;
   qqplot resid / normal(mu=est sigma=est);
run;

Step 4: Sensitivity analysis with the exact Wilcoxon rank-sum test

proc npar1way data=adeff wilcoxon;
   class trt01pn;
   var chg;
   exact wilcoxon;
run;

Step 5: Supportive responder analysis with Fisher's exact test

proc freq data=adeff;
   tables trt01pn*respfl / nocol nopercent fisher;
run;

Step 6: Mixed model for repeated measures (when there are several visits)

The dummy data contain only the Week 48 value. In a real trial with visits at, for example, Weeks 12, 24, 36 and 48 in a long-format dataset, the MMRM is:

proc mixed data=adeff_long method=reml;
   class usubjid trt01pn avisitn;
   model chg = trt01pn avisitn trt01pn*avisitn base base*avisitn
               / ddfm=kr solution;
   repeated avisitn / subject=usubjid type=un;
   lsmeans trt01pn*avisitn / cl pdiff;
run;

The treatment effect is the difference between the LS means at Week 48. An unstructured covariance matrix (TYPE=UN) is preferred; with very few patients it may not converge, so a simpler fallback structure should be pre-specified.

Output

Descriptive statistics for 6MWD (metres)

TreatmentNBaseline, mean (SD)Week 48, mean (SD)Change, mean (SD)Change, median
Placebo15348.80 (56.86)341.53 (60.33)−7.27 (40.95)−17
Active15355.40 (59.24)383.00 (63.36)27.60 (33.97)24

ANCOVA: Type III tests

SourceDFF valuePr > F
Treatment16.860.0143
Baseline 6MWD11.370.2522

Error degrees of freedom = 27, mean square error = 1396.83, Root MSE = 37.37, R-square = 0.2263. The baseline slope is −0.142 (standard error 0.122).

ANCOVA: Least-squares means of change from baseline

TreatmentLS meanStandard error95% confidence limits
Placebo−7.749.66−27.55 to 12.08
Active28.079.668.25 to 47.89

ANCOVA: Treatment difference

ParameterEstimateStandard errort valuePr > |t|95% confidence limits
Active vs Placebo35.8113.672.620.01437.76 to 63.86

Wilcoxon rank-sum test

TreatmentNSum of scoresExpected under H0Mean score
Placebo15177.0232.511.80
Active15288.0232.519.20

Statistic (S) = 177.0, Z = −2.2813, normal approximation two-sided p = 0.0225, exact two-sided p = 0.0209.

Responder analysis (improvement of 30 metres or more)

TreatmentResponders, n/NResponder rate (%)
Placebo2/1513.3
Active7/1546.7

Difference in responder rates = 33.3 percentage points. Fisher's exact test, two-sided p = 0.1086.

Interpretation

  1. Baseline comparability. Mean baseline 6MWD was similar in the two arms (348.80 and 355.40 metres). The small difference is adjusted for by the ANCOVA.
  2. Primary analysis. After adjusting for baseline, patients on active treatment improved by 28.07 metres on average, while patients on placebo declined by 7.74 metres. The treatment difference is 35.81 metres (95% confidence interval 7.76 to 63.86; p = 0.0143). The result is statistically significant at the two-sided 5% level.
  3. Clinical relevance. The estimated difference exceeds the 30-metre threshold defined as clinically meaningful. The confidence interval is wide, however: the lower limit of 7.76 metres is below that threshold. The trial shows that the treatment works, but the size of the benefit is estimated imprecisely, which is typical of small trials.
  4. Sensitivity analysis. The exact Wilcoxon rank-sum test gives p = 0.0209, in agreement with the ANCOVA. The conclusion does not depend on the normality assumption or on a few extreme values.
  5. Responder analysis. 46.7% of patients on active treatment were responders compared with 13.3% on placebo. The direction supports the primary analysis, but the difference is not statistically significant (p = 0.1086).
  6. Overall conclusion. In this dummy example the active treatment produces a statistically significant improvement in 6MWD at Week 48, supported by a consistent non-parametric analysis.

Why the responder analysis is not significant. The same 30 patients give p = 0.0143 for the continuous endpoint and p = 0.1086 for the responder endpoint. Nothing about the treatment has changed; splitting each patient into responder or non-responder simply discards information. This is the practical reason to choose a continuous primary endpoint in a small trial and to keep the responder analysis as supportive.

Effect of adjusting for baseline. The unadjusted difference in mean change is 27.60 − (−7.27) = 34.87 metres. The adjusted difference is 35.81 metres. Here the adjustment is small because baseline was well balanced and only weakly related to change. In trials where baseline is strongly related to the outcome, the adjustment gives a clearly smaller standard error.

Practical points

  • Pre-specify everything. The endpoint, model, covariates, handling of intercurrent events and sensitivity analyses must be fixed in the protocol and statistical analysis plan. Flexibility after unblinding is not credible in a small trial.
  • Limit covariates. Each covariate uses a degree of freedom. With 30 patients, adjust only for baseline and the stratification factors.
  • Standardise the assessment. For effort-dependent tests such as the 6MWD, train sites and use the same procedure at every visit to reduce variability.
  • Consider efficient designs. Crossover, randomised withdrawal, adaptive and externally controlled designs can reduce the number of patients needed when the disease and treatment allow.
  • Use natural history data. A well-run natural history study supports endpoint selection, sample size assumptions and, where justified, an external control.
  • Bayesian methods. Borrowing information from historical controls or adult data can increase precision, provided the approach and the prior are agreed with regulators in advance.

References

  • US FDA. Rare Diseases: Considerations for the Development of Drugs and Biological Products, Guidance for Industry. 2023.
  • US FDA. Adjusting for Covariates in Randomized Clinical Trials for Drugs and Biological Products, Guidance for Industry. 2023.
  • ICH E9(R1). Addendum on Estimands and Sensitivity Analysis in Clinical Trials. 2019.
  • European Medicines Agency. Guideline on Clinical Trials in Small Populations. 2006.

The dataset in this article is simulated and is provided for training purposes only.

Found this useful?Share it with your team or network.

Need help applying this to your study?

Our statisticians, programmers and quality specialists can review it with you.