Progression-Free Survival vs. Overall Survival

Introduction

In oncology clinical trials, progression-free survival (PFS) and overall survival (OS) are two of the most important time-to-event endpoints. They answer related but fundamentally different questions.

PFS asks whether a patient remains alive without disease progression, whereas OS asks whether the patient remains alive regardless of disease status.

This distinction is clinically important because a treatment can substantially delay progression without producing the same magnitude of improvement in overall survival.

What Is Overall Survival?

Overall survival is one of the simplest survival endpoints conceptually. It is the time from a prespecified origin, usually randomization or treatment assignment, until death from any cause.

For patient \(i\), let \(T_i\) denote the time from randomization to death. Then the OS event is:

Patients who are alive at the analysis cutoff are generally censored at their last known date alive, provided the censoring assumptions and protocol-defined rules are satisfied.

Why OS Is Clinically Attractive

Death is an objective and clinically definitive endpoint. It does not require an investigator to determine whether a tumor has progressed according to radiographic criteria.

For this reason, OS is often regarded as the most direct measure of whether a treatment ultimately helps patients live longer.

What Is Progression-Free Survival?

Progression-free survival measures the time from a defined origin until the first occurrence of either disease progression or death.

For patient \(i\), define \(T_{PFS,i}\) as the time to the first qualifying progression or death.

Thus, a patient does not need to die for a PFS event to occur. A documented disease progression is sufficient.

Example

Suppose a patient enters an oncology trial on January 1. The patient's disease progresses on September 1, but the patient remains alive for several additional years.

The patient's PFS event occurs on September 1. The OS event occurs later, when the patient dies.

This illustrates why PFS and OS are not interchangeable.

PFS and OS Side by Side

Why PFS Often Provides Earlier Information

In many advanced cancer trials, progression occurs substantially more frequently than death during the period when the primary analysis is performed.

Consequently, investigators may accumulate PFS events more rapidly than OS events.

This can allow a randomized trial to estimate a treatment effect on disease control before enough deaths have occurred to provide a mature OS analysis.

Why PFS and OS Can Disagree

A common misconception is that an improvement in PFS must eventually produce an equal improvement in OS.

That is not necessarily true.

The main reason is that patients may receive additional treatments after progression. These subsequent therapies can affect survival after the initial trial treatment.

Consider a trial in which the experimental treatment delays progression but patients in both treatment groups receive effective therapies after progression. The experimental treatment may therefore produce a clear PFS benefit while the difference in OS becomes smaller.

The Time-to-Event Framework

Both PFS and OS are time-to-event endpoints and can therefore be represented using a survival function.

For a generic event time \(T\), the survival function is:

For OS:

For PFS:

The PFS survival function therefore represents the probability of being alive and progression-free beyond time \(t\).

Kaplan-Meier Curves for PFS and OS

The Kaplan-Meier estimator is commonly used to estimate the survival function for both endpoints.

For a treatment arm, the estimated survival probability is:

where \(d_j\) is the number of events at event time \(t_j\), and \(n_j\) is the number of patients at risk immediately before that time.

The same Kaplan-Meier machinery can be used for PFS and OS. What changes is the definition of the event.

Reading a PFS Kaplan-Meier Curve

Suppose the PFS curves for two randomized treatment groups separate early and remain separated throughout follow-up.

This suggests that the experimental treatment is associated with a longer time to progression or death.

However, the visual separation of Kaplan-Meier curves should not be treated as a substitute for the prespecified statistical analysis.

Investigators typically consider:

• The estimated hazard ratio
• The confidence interval
• The number of events
• The median PFS, when estimable
• The prespecified hypothesis test
• The maturity of the data

Reading an OS Kaplan-Meier Curve

An OS Kaplan-Meier curve estimates the probability of remaining alive over time.

A separation between treatment groups indicates a difference in the estimated survival distributions, but the magnitude and reliability of the treatment effect still depend on the number and timing of deaths.

An OS analysis can be immature even when the PFS analysis is already mature.

Hazard Ratios for PFS and OS

The hazard ratio is commonly used to summarize the relative event rate between two randomized treatment groups.

A Cox proportional hazards model may be written as:

For a binary treatment indicator \(X\), the hazard ratio is:

For PFS, the hazard corresponds to the instantaneous risk of progression or death. For OS, the hazard corresponds to the instantaneous risk of death.

Example: Interpreting a PFS Hazard Ratio

Suppose a randomized trial reports:

A conventional interpretation is that the estimated hazard of progression or death is approximately 30% lower in the experimental group than in the control group, under the assumptions of the fitted model and proportional hazards interpretation.

The calculation is:

Thus the commonly reported relative reduction is approximately 30%.

Example: PFS Benefit Without an Immediate OS Benefit

Imagine the following simplified results from a randomized oncology trial.

These results could indicate a clinically meaningful delay in disease progression without a comparably large observed difference in overall survival.

That pattern does not automatically mean the PFS result is irrelevant. It may mean that subsequent treatment, crossover, effective salvage therapy, or other post-progression factors reduced the eventual difference in OS.

Why Subsequent Therapy Matters

OS after randomization is influenced by what happens throughout the remainder of a patient's disease course.

After progression, patients may receive:

• Second-line chemotherapy
• Targeted therapy
• Immunotherapy
• Radiation
• Surgery
• Supportive care
• Another investigational treatment

These treatments can differ between patients and treatment groups.

Consequently, the OS comparison may reflect not only the randomized treatment but also the subsequent therapeutic pathway.

Crossover Can Dilute an OS Difference

Suppose patients randomized to the control group are allowed to receive the experimental treatment after disease progression.

If the experimental treatment is effective, crossover can improve outcomes in the control group after progression.

The original randomized treatment comparison can therefore show:

Median PFS and Median OS

The median survival time is the time at which the estimated survival probability falls to 0.50.

If the Kaplan-Meier curve crosses 0.50, the median can be estimated from the curve.

However, a median may be not estimable when the survival probability has not fallen below 0.50 by the analysis cutoff.

Why Median Survival Alone Can Be Misleading

Two treatment groups can have similar median survival times but meaningfully different survival curves.

Conversely, a large difference in medians does not necessarily summarize the entire treatment effect.

This is particularly important when hazards are not proportional.

Therefore, clinical interpretation should consider the entire Kaplan-Meier curve, event counts, follow-up duration, hazard ratio, confidence interval, and clinically relevant landmark survival estimates where appropriate.

PFS Is More Dependent on Assessment Schedule

Unlike death, tumor progression is generally observed only when disease is assessed.

For example, if imaging is performed every eight weeks, progression may be identified at one of those scheduled assessments even if biological progression occurred earlier.

This means PFS can be affected by:

• Imaging frequency
• Timing of assessments
• Assessment windows
• Radiographic criteria
• Investigator interpretation
• Independent central review procedures

What Counts as Progression?

The precise definition depends on the disease, protocol, and prespecified assessment criteria.

In oncology trials, progression may be defined using standardized tumor assessment frameworks, such as RECIST-based criteria where applicable.

The statistical analysis plan should specify how progression is determined, including rules for:

• Target lesions
• Non-target lesions
• New lesions
• Clinical progression
• Unscheduled assessments
• Missing assessments
• Assessment dates

Censoring in PFS

Censoring is especially important for PFS because patients may leave the study before progression or death is observed.

Examples include:

• Withdrawal of consent
• Loss to follow-up
• Study discontinuation
• Administrative database cutoff
• Missing tumor assessments

The exact censoring rules should be defined prospectively.

For example, a patient who has no documented progression before the analysis cutoff may be censored at an appropriate last disease assessment rather than simply assigned a PFS time equal to the database cutoff.

Progression-Free Survival and Treatment Discontinuation

Treatment discontinuation does not necessarily equal progression.

A patient may stop treatment because of:

• Adverse events
• Patient preference
• Investigator decision
• Clinical improvement
• Protocol-defined treatment duration
• Other medical reasons

If the patient has not progressed and remains alive, simply stopping study treatment does not automatically create a PFS event.

The subsequent follow-up and prespecified endpoint rules determine how the patient contributes to the PFS analysis.

PFS Versus OS: A Clinical Interpretation Framework

When reviewing an oncology trial, it is useful to ask the following questions.

Example: Three Possible Trial Patterns

Pattern 1: PFS and OS Both Improve

This is often the clearest scenario.

The treatment appears to delay progression and reduce the subsequent risk of death.

Pattern 2: PFS Improves but OS Does Not

Possible explanations include effective post-progression therapy, crossover, long post-progression survival, limited OS maturity, or other factors.

The correct interpretation depends on the complete trial context.

Pattern 3: OS Appears to Improve Before PFS

This pattern may occur in unusual circumstances and should prompt careful investigation.

Potential considerations include:

• Immature PFS data
• Nonproportional hazards
• Assessment-related issues
• Chance variation
• Differences in post-progression treatment
• Endpoint-specific data quality

The result should not automatically be interpreted as evidence that PFS is invalid. Rather, the two endpoints should be examined using the complete prespecified statistical analysis.

Landmark Survival Rates

A useful complement to medians and hazard ratios is the estimated probability of being progression-free or alive at a clinically meaningful time point.

For example:

could represent the probability of remaining alive and progression-free at 12 months.

Similarly:

represents the estimated probability of being alive at 24 months.

These measures can be easier to communicate clinically than a hazard ratio alone.

PFS2 and Other Extended Endpoints

Some oncology trials use additional time-to-event endpoints to examine what happens beyond the first progression.

One such concept is PFS2, in which the event is typically defined using progression on a subsequent line of therapy or death, according to the protocol's specific definition.

The purpose is to investigate whether an initial treatment affects the broader disease course rather than only the first progression.

PFS as a Surrogate for OS

Because PFS occurs earlier than OS, investigators have long considered whether PFS can serve as a surrogate endpoint for overall survival.

A treatment effect on PFS does not automatically establish a treatment effect on OS.

Surrogacy is a stronger statistical concept than simple correlation between two endpoints.

A convincing surrogate endpoint requires evidence that treatment effects on the surrogate reliably predict clinically meaningful effects on the true clinical outcome across an appropriate body of evidence.

Statistical Testing

For randomized time-to-event analyses, treatment groups are commonly compared using a stratified or unstratified log-rank test, depending on the design.

The Cox model may then be used to estimate a hazard ratio and confidence interval.

A generic null hypothesis for the hazard ratio is:

against an alternative such as:

The exact hypothesis, stratification factors, significance level, and analysis method should be prespecified in the protocol and statistical analysis plan.

Stratification Matters

Many oncology trials stratify randomization according to important baseline factors, such as disease stage, biomarker status, geographic region, or prior therapy.

If the primary analysis is stratified, the statistical methods should reflect the prespecified stratification factors.

A commonly used stratified Cox model estimates a treatment effect while accounting for these strata.

What Does a 95% Confidence Interval Tell You?

Suppose a trial reports:

The point estimate suggests a lower hazard of progression or death in the experimental group.

The confidence interval describes the statistical uncertainty around the estimated hazard ratio.

Because the interval lies below 1, the result is compatible with a lower progression/death hazard for the experimental treatment under the model and analysis assumptions.

The corresponding clinical importance still requires consideration of the absolute difference in survival, toxicity, quality of life, duration of benefit, and other clinical factors.

Relative Versus Absolute Benefit

Hazard ratios describe a relative treatment effect. They do not directly tell you how many additional months a patient will remain progression-free or alive.

For example, the same hazard ratio can arise in settings with very different absolute survival times.

When PFS May Be Particularly Informative

PFS can be especially informative when progression itself represents an important clinical deterioration and when subsequent treatment makes OS difficult to interpret.

It may also provide a more timely assessment when OS requires many years of follow-up.

However, the importance of PFS depends on the disease, treatment mechanism, available subsequent therapies, progression definition, and clinical context.

When OS May Be Particularly Informative

OS is particularly attractive when mortality is sufficiently frequent and the treatment's effect on survival is expected to be clinically meaningful.

It is also less dependent on radiographic assessment and progression definitions.

The tradeoff is that OS generally requires longer follow-up and may be strongly influenced by treatments received after progression.

Common Mistakes

1. Equating PFS with OS. PFS includes progression as an event; OS includes death only.
2. Interpreting a PFS hazard ratio as a percentage increase in survival time. A hazard ratio is not a direct ratio of survival times.
3. Assuming PFS improvement guarantees OS improvement. Subsequent therapies and other post-progression events can alter the OS comparison.
4. Ignoring the progression definition. PFS depends on how progression is defined and assessed.
5. Ignoring censoring rules. The handling of missing assessments and other censoring situations can affect the PFS analysis.
6. Reporting only the median. The median does not necessarily summarize the entire survival distribution.
7. Ignoring immature OS data. A nonsignificant OS result can be difficult to interpret when relatively few deaths have occurred.
8. Assuming a surrogate relationship without evidence. A PFS benefit is not automatically proof of an OS benefit.
9. Ignoring toxicity and quality of life. A longer PFS interval is not the only consideration when evaluating a treatment.

A Practical Workflow for Interpreting PFS and OS

Reporting PFS in a Clinical Trial

A clear statistical report should identify:

• The PFS definition
• The time origin
• The progression criteria
• The censoring rules
• The analysis population
• The number of PFS events
• The Kaplan-Meier estimates
• The median PFS, if estimable
• The hazard ratio and confidence interval
• The prespecified hypothesis test
• The follow-up duration

Reporting OS in a Clinical Trial

For OS, reporting should generally include:

• The time origin
• The number of deaths
• The median OS, if estimable
• Kaplan-Meier estimates
• The hazard ratio
• The confidence interval
• The prespecified hypothesis test
• The follow-up duration
• Subsequent anticancer therapy information when relevant
• Crossover information when applicable

A Compact Interpretation Example

Suppose an oncology trial reports:

A reasonable interpretation is that the experimental treatment is associated with a substantial reduction in the hazard of progression or death and a more modest reduction in the hazard of death.

The PFS result should not be translated into "patients live 32% longer." Instead, the HR of 0.68 describes the estimated relative hazard of progression or death under the fitted survival model.

The OS result should be interpreted separately, taking account of its maturity, confidence interval, subsequent treatment, and the prespecified analysis.

The Most Important Concept

The most important distinction is simple:

whereas:

Because progression usually occurs before death, PFS can provide an earlier assessment of treatment activity. But because progression is influenced by disease assessment and endpoint definitions, PFS is not equivalent to OS.

Likewise, because OS incorporates everything that happens after randomization, including subsequent anticancer treatments, an improvement in PFS does not necessarily translate into an equally large improvement in OS.

Summary Table

References

Food and Drug Administration. Clinical Trial Endpoints for the Approval of Cancer Drugs and Biologics. U.S. FDA.
Eisenhauer, E.A., et al. (2009). New response evaluation criteria in solid tumours: Revised RECIST guideline (version 1.1). European Journal of Cancer, 45, 228–247.
Broglio, K.R. & Berry, D.A. (2009). Detecting an overall survival benefit that is derived from progression-free survival. Journal of the National Cancer Institute, 101, 1642–1649.
Fleming, T.R., Rothmann, M.D. & Lu, H.L. (2005). Issues in using progression-free survival endpoint in randomized oncology clinical trials. Journal of Clinical Oncology, 23, 2356–2362.
Pazdur, R. (2008). Endpoints for assessing drug activity in clinical trials. The Oncologist, 13(Suppl 2), 19–21.