Progression-Free Survival (PFS) in Oncology Trials: SAS Code, Data, Output and Interpretation

A step-by-step worked example of a time-to-event analysis for progression-free survival: censoring rules, Kaplan-Meier estimates, the log-rank test and the Cox hazard ratio, with a dummy ADTTE dataset, annotated SAS code, output and interpretation.

Progression-free survival (PFS) is one of the most widely used endpoints in oncology trials. It measures how long patients live without their cancer getting worse, and it is usually available earlier than overall survival. Because not every patient has progressed or died when the analysis is performed, PFS is a time-to-event endpoint and needs methods that handle censoring correctly.

This article walks through a complete PFS analysis on a dummy dataset: the definition and censoring rules, the statistical concepts, the SAS code, the output, and the interpretation of each result.

What is progression-free survival?

PFS is the time from randomisation to the first documented disease progression (typically assessed by RECIST 1.1) or death from any cause, whichever occurs first. Patients with neither event are censored: their true PFS time is known only to be longer than the time observed.

Typical event and censoring rules

The rules below follow the approach commonly used for FDA submissions. The exact rules must be pre-specified in the protocol and statistical analysis plan.

SituationOutcomeDate used
Documented progressionEventDate of the assessment showing progression
Death without prior progressionEventDate of death
No progression and no death at data cut-offCensoredDate of last adequate tumour assessment
New anticancer therapy started before progressionCensoredLast adequate assessment before the new therapy
Progression or death after two or more missed assessmentsCensoredLast adequate assessment before the missed visits
No baseline or no post-baseline assessmentCensoredDate of randomisation

Statistical concepts behind the analysis

1. Censoring

A censored patient still contributes information: they are known to be event-free up to their censoring time. Dropping censored patients or treating their censoring time as an event time would bias the result. The standard methods assume censoring is non-informative, meaning that being censored says nothing about the patient's future risk of progression.

2. The survival function and the Kaplan-Meier estimate

The survival function S(t) is the probability of remaining event-free beyond time t. The Kaplan-Meier (product-limit) method estimates it as a step function. At each event time, the estimate is multiplied by (1 − d/n), where d is the number of events at that time and n is the number of patients still at risk. Censored patients leave the risk set without causing a step.

3. Median PFS and its confidence interval

Median PFS is the earliest time at which the Kaplan-Meier estimate falls to 50% or below. Its confidence interval is obtained with the Brookmeyer-Crowley method, usually with a log-log transformation of the survival function, which keeps the limits between 0 and 1 and behaves better in small samples. If the curve never reaches 50%, the median is "not estimable".

4. The log-rank test

The log-rank test compares the whole survival curves of two arms. At each event time it compares the number of events observed in an arm with the number expected if both arms had the same risk, then combines these differences over time into a chi-square statistic. It is most powerful when the hazards are proportional.

5. Hazard and hazard ratio

The hazard is the instantaneous rate of the event among patients still at risk. The hazard ratio (HR) compares the hazard in the experimental arm with the hazard in the control arm. An HR below 1 favours the experimental arm; an HR of 0.70 means a 30% reduction in the rate of progression or death at any time.

6. The Cox proportional hazards model

The Cox model estimates the hazard ratio without assuming any shape for the underlying hazard, which is why it is called semi-parametric. It uses a partial likelihood built only from the order of the event times. The model coefficient is the log hazard ratio, so HR = exp(coefficient). When several events share the same time, a tie-handling method is needed; the Efron method is generally preferred to the SAS default (Breslow).

7. The proportional hazards assumption

The Cox model assumes the hazard ratio is constant over time. If the curves cross, or separate only late (as often seen with immunotherapy), a single hazard ratio can be misleading. The assumption is checked with log(−log(survival)) plots, Schoenfeld residuals or the ASSESS statement, and supportive measures such as restricted mean survival time may be reported.

8. Stratified analysis

When randomisation is stratified (for example by ECOG status or region), the primary analysis usually uses a stratified log-rank test and a stratified Cox model. Each stratum has its own baseline hazard, and the treatment effect is combined across strata.

9. Landmark rates

The PFS rate at a fixed time, such as 6 or 12 months, is read directly from the Kaplan-Meier curve with its confidence interval. Landmark rates are easy to communicate and complement the median, which describes only one point on the curve.

Study design and dummy data

The example is a randomised, two-arm oncology trial with 40 patients, 20 per arm. The dataset follows the ADaM time-to-event (ADTTE) structure. The values are simulated and are for illustration only.

VariableMeaning
USUBJIDUnique subject identifier
TRT01PNPlanned treatment: 1 = Control, 2 = Experimental
AVALPFS time in months
CNSRCensoring flag: 0 = event (progression or death), 1 = censored
proc format;
   value trtf  1 = 'Control'
               2 = 'Experimental';
run;

data adtte;
   length usubjid $8 paramcd $8 param $40;
   input usubjid $ trt01pn aval cnsr;
   paramcd = 'PFS';
   param   = 'Progression-Free Survival (months)';
   format trt01pn trtf.;
   datalines;
PFS-001 1   8.49 0
PFS-002 2  12.05 1
PFS-003 1  15.74 1
PFS-004 2  12.26 0
PFS-005 1   8.98 0
PFS-006 2  11.33 1
PFS-007 1   1.34 0
PFS-008 2   7.83 0
PFS-009 1   2.35 0
PFS-010 2  18.16 0
PFS-011 1   1.72 0
PFS-012 2  10.08 0
PFS-013 1   7.68 0
PFS-014 2   3.08 0
PFS-015 1   5.20 1
PFS-016 2  18.36 1
PFS-017 1   5.66 0
PFS-018 2   9.20 0
PFS-019 1  11.29 0
PFS-020 2   0.49 0
PFS-021 1   0.47 0
PFS-022 2   5.32 0
PFS-023 1   6.44 0
PFS-024 2   1.70 0
PFS-025 1   7.51 0
PFS-026 2   0.67 0
PFS-027 1   5.74 0
PFS-028 2  11.05 0
PFS-029 1   0.26 0
PFS-030 2  17.49 1
PFS-031 1   1.66 0
PFS-032 2   4.07 0
PFS-033 1   1.67 0
PFS-034 2   0.59 1
PFS-035 1   4.00 0
PFS-036 2   2.98 0
PFS-037 1   3.04 1
PFS-038 2  14.97 0
PFS-039 1   6.98 0
PFS-040 2  17.85 0
;
run;

SAS code, step by step

Step 1: Kaplan-Meier estimates, median PFS and log-rank test

In the TIME statement, cnsr(1) tells SAS that a value of 1 means censored. CONFTYPE=LOGLOG requests the log-log transformation for confidence limits, and the PLOTS option draws the Kaplan-Meier curve with the number at risk below the axis.

ods graphics on;
ods output Quartiles=quart CensoredSummary=censum HomTests=logrank;

proc lifetest data=adtte method=km conftype=loglog alpha=0.05
              plots=survival(atrisk(outside)=0 to 18 by 3 test);
   time aval*cnsr(1);
   strata trt01pn / test=logrank;
run;

Step 2: Hazard ratio from the Cox proportional hazards model

The control arm is set as the reference, so the hazard ratio is experimental versus control. TIES=EFRON selects the Efron method for tied event times.

ods output ParameterEstimates=hr;

proc phreg data=adtte;
   class trt01pn (ref='Control') / param=ref;
   model aval*cnsr(1) = trt01pn / ties=efron risklimits alpha=0.05;
   hazardratio 'Experimental vs Control' trt01pn / diff=ref cl=wald;
run;

Step 3: Landmark PFS rates at 6 and 12 months

proc lifetest data=adtte method=km conftype=loglog
              timelist=6 12 reduceout outsurv=landmark noprint;
   time aval*cnsr(1);
   strata trt01pn;
run;

data landmark2;
   set landmark;
   rate  = 100 * survival;      /* PFS rate (%)        */
   lower = 100 * sdf_lcl;       /* 95% CI lower limit  */
   upper = 100 * sdf_ucl;       /* 95% CI upper limit  */
   keep trt01pn timelist rate lower upper;
run;

proc print data=landmark2 noobs;
   format rate lower upper 6.1;
run;

Step 4: Stratified analysis (when randomisation is stratified)

The dummy data have no stratification factors. In a real trial with factors such as ECOG status and region, the stratified versions are:

/* Stratified log-rank test */
proc lifetest data=adtte method=km conftype=loglog;
   time aval*cnsr(1);
   strata ecog region / group=trt01pn test=logrank;
run;

/* Stratified Cox model */
proc phreg data=adtte;
   class trt01pn (ref='Control') / param=ref;
   model aval*cnsr(1) = trt01pn / ties=efron risklimits;
   strata ecog region;
run;

Step 5: Check the proportional hazards assumption

/* Log(-log(survival)) plot: roughly parallel curves support PH */
proc lifetest data=adtte method=km plots=(lls);
   time aval*cnsr(1);
   strata trt01pn;
run;

/* Supremum test based on cumulative martingale residuals */
proc phreg data=adtte;
   class trt01pn (ref='Control') / param=ref;
   model aval*cnsr(1) = trt01pn / ties=efron;
   assess ph / resample seed=20261118;
run;

Output

Summary of events and censoring

TreatmentTotalEventsCensoredPercent censored
Control2017315.00
Experimental2015525.00
Total4032820.00

Kaplan-Meier quartiles of PFS (months)

Treatment25th percentileMedian95% CI of median75th percentile
Control1.675.741.67 – 7.687.68
Experimental3.0810.083.08 – 17.8517.85

Log-rank test of equality over strata

TestChi-squareDFPr > Chi-square
Log-rank5.286810.0215

Events in the experimental arm: 15 observed against 20.87 expected under the null hypothesis. Events in the control arm: 17 observed against 11.13 expected.

Cox proportional hazards model

ParameterEstimateStandard errorChi-squarePr > Chi-squareHazard ratio95% CI
Experimental vs Control−0.865010.385525.03430.02490.4210.198 – 0.896

Landmark PFS rates

TreatmentTime (months)PFS rate (%)95% CI (%)
Control647.724.4 – 67.8
Control126.00.4 – 23.7
Experimental663.338.1 – 80.6
Experimental1242.220.4 – 62.6

Summary table for the clinical study report

Control (N=20)Experimental (N=20)
Patients with event, n (%)17 (85.0)15 (75.0)
Patients censored, n (%)3 (15.0)5 (25.0)
Median PFS, months (95% CI)5.74 (1.67, 7.68)10.08 (3.08, 17.85)
Hazard ratio (95% CI)0.421 (0.198, 0.896)
Log-rank p-value (two-sided)0.0215
6-month PFS rate, % (95% CI)47.7 (24.4, 67.8)63.3 (38.1, 80.6)
12-month PFS rate, % (95% CI)6.0 (0.4, 23.7)42.2 (20.4, 62.6)

Interpretation

  1. Events and maturity. 32 of 40 patients (80%) had progression or death. The data are mature, so the medians and the hazard ratio are estimated from most of the patients rather than extrapolated.
  2. Median PFS. Median PFS was 10.08 months in the experimental arm and 5.74 months in the control arm, an improvement of about 4.3 months. The confidence intervals are wide and overlap, which is expected with only 20 patients per arm. Overlapping intervals for the medians do not mean the curves are the same; the formal comparison is the log-rank test.
  3. Log-rank test. The two-sided p-value is 0.0215, below 0.05. The PFS curves differ significantly. The experimental arm had fewer events than expected under no difference (15 against 20.87) and the control arm had more (17 against 11.13).
  4. Hazard ratio. The hazard ratio is 0.421 with a 95% confidence interval of 0.198 to 0.896. The experimental treatment reduced the rate of progression or death by about 58%. The upper limit is below 1, which agrees with the significant log-rank test, but the interval is wide: the true reduction could plausibly be anywhere from about 10% to 80%.
  5. Landmark rates. At 12 months, 42.2% of patients in the experimental arm were alive and progression-free compared with 6.0% in the control arm. The curves separate more clearly later in follow-up.
  6. Overall conclusion. In this dummy example the experimental treatment shows a statistically significant and clinically meaningful improvement in PFS. In a real trial this would be read alongside the proportional hazards check, sensitivity analyses of the censoring rules, and overall survival.

Why the log-rank and Cox p-values differ slightly. The log-rank test gives 0.0215 and the Wald test from the Cox model gives 0.0249. Both test the same hypothesis but are calculated differently: the log-rank test is the score test of the Cox model, while PROC PHREG reports the Wald test for the parameter. By convention the p-value is reported from the log-rank test and the hazard ratio from the Cox model.

Practical points

  • Investigator versus independent review. PFS can be assessed by the investigator or by blinded independent central review (BICR). State which is primary and report the other as a sensitivity analysis.
  • Assessment schedule. Progression is detected only at scheduled scans, so PFS times are interval-censored in reality. Keep the scan schedule identical in both arms to avoid bias.
  • Sensitivity analyses. Repeat the analysis with alternative censoring rules, for example counting new anticancer therapy or events after missed visits as events, to show the result does not depend on one set of rules.
  • One-sided or two-sided. Many oncology protocols specify a one-sided alpha of 0.025. The one-sided p-value is half the two-sided value when the effect is in the expected direction.
  • Event-driven timing. The analysis is usually triggered by a pre-specified number of events, not a calendar date, because power depends on the number of events.
  • Units. ADTTE often stores AVAL in days. Convert to months with AVAL / 30.4375 for reporting.

References

  • US FDA. Clinical Trial Endpoints for the Approval of Cancer Drugs and Biologics, Guidance for Industry. 2018.
  • Eisenhauer EA, Therasse P, Bogaerts J, et al. New response evaluation criteria in solid tumours: revised RECIST guideline (version 1.1). European Journal of Cancer. 2009;45(2):228–247.
  • Kaplan EL, Meier P. Nonparametric estimation from incomplete observations. Journal of the American Statistical Association. 1958;53(282):457–481.
  • Cox DR. Regression models and life-tables. Journal of the Royal Statistical Society, Series B. 1972;34(2):187–220.
  • Brookmeyer R, Crowley J. A confidence interval for the median survival time. Biometrics. 1982;38(1):29–41.

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.