TALAPRO-3: reading a median that has not been reached

The essential results, before interpretation

Source: US Talzenna prescribing information, section 14.2, TALAPRO-3 and Table 14; revised October 2026. Both groups also received androgen deprivation therapy or had previously undergone bilateral orchiectomy. Event counts are not estimated risks at a common time.

The immediate interpretation is a favourable difference in the primary endpoint, with quantified uncertainty, even though the experimental group’s curve did not provide a median. The table does not permit calculating additional months by subtracting medians, and it does not replace the assessment of overall survival.

A median not reached, in a synthetic example

Five invented people. These are not TALAPRO-3 data.

Synthetic data (CSV) · Reproduce in R · Explore the worked example

1. What is new, and what was already known

On 26 October, the FDA approved the combination for adults with metastatic androgen pathway modulation-naïve or -sensitive prostate cancer, mAPMN/S, with homologous recombination repair —HRR— gene mutations detected by an FDA-authorised test. This is a US decision for a defined population. Regulatory notification.

The scientific results were published online on 30 May 2026 and appeared in the 30 July issue of The New England Journal of Medicine. Article by Agarwal and colleagues. The sponsor’s 30 May announcement also presented them at that time.

October’s development is regulatory. Distinguishing these dates prevents a recent approval from turning earlier results into data supposedly obtained this week. Approval, publication and continued trial follow-up are separate milestones.

2. The clinical comparison to keep in view

TALAPRO-3 is a phase III randomised, double-blind, placebo-controlled trial. It compares adding talazoparib to enzalutamide with adding placebo to the same background treatment. Participants were selected for HRR alterations. Registered design and endpoints.

The comparison concerns the full regimen in that population. It does not evaluate talazoparib alone, every hormonal combination or all patients with prostate cancer. Randomisation helps balance baseline prognostic factors; applicability to other populations still requires clinical and biological justification.

The primary endpoint is time from randomisation to radiographic progression or death, whichever occurs first. Progression is investigator-assessed using the specified criteria. Overall survival has a different event: death from any cause. FDA efficacy assessment.

Each participant contributes a follow-up trajectory. Visits and images supply information used to establish that trajectory; they are not independent observations that can be counted as additional patients.

3. What the median represents

A survival curve estimates S(t), the probability of remaining event-free beyond time t. Here the event can be progression or death, so “survival” means remaining free of that composite endpoint.

The median is located where the curve reaches 50%. A conventional definition based on the first crossing is:

Median = first time t for which S(t) ≤ 0.50.

If the level is not reached within the observable information, the estimate does not provide a median. This does not make the value zero, infinite or equal to maximum follow-up. Nor does it mean that no events occurred.

Survival curves are step functions. When a segment lies exactly at 50%, calculation conventions should be documented. For example, the R survival package documentation describes using the midpoint of a matching horizontal segment. This detail matters when reproducing results across software. Quantiles from an estimated curve.

“Not reached” therefore describes what this summary can estimate. It does not, on its own, determine whether follow-up was adequate for other objectives or whether a treatment provides clinically meaningful benefit.

4. Five people: events, censoring and a median that never appears

This example is entirely synthetic. It uses no TALAPRO-3 patients or results. Its sole purpose is to make the calculation visible.

Five people begin follow-up. One experiences the event at month 2; another ends observation without an event at month 4; a third experiences the event at month 6; the remaining two end observation without an event at months 8 and 10.

Kaplan–Meier multiplies the probability of passing each event time among those at risk at that time. Here:

S(6) = (1 − 1/5) × (1 − 1/3) = 0.80 × 2/3 = 0.5333…

Censoring reduces the number observed for subsequent steps. It does not itself produce a downward step. At the end, the estimate remains above 50%, so the median is not reached.

Two tempting shortcuts fail. Dividing the three people without an observed event by the original five gives 60%, which is not the Kaplan–Meier estimate. Taking the median of the two event times, 2 and 6 months, gives 4 months and excludes censored information: it is not the survival median either.

5. A flat curve still needs context

In this example, the curve remains at 53.3% at the end, but only one person is under observation immediately before month 10. The shape of the final segment does not generate additional information about the future.

A flat segment supported by many people followed for sufficient time differs from a tail supported by few late observations. Censoring marks, numbers at risk and confidence intervals help distinguish them.

Censoring means that a person is known to have remained event-free up to a time, while their subsequent event time is unknown. Administrative closure, loss of contact and observation ending after a clinical problem can have different implications.

The usual interpretation of Kaplan–Meier requires censoring not to select future event times in a way incompatible with the analysis. If people with worsening disease stop being observed earlier, a curve may become overly favourable. This needs investigation in the data; a median not reached alone cannot diagnose it. The Kaplan–Meier guide explores these concepts using synthetic information.

6. The median and its interval are related estimates

A median estimate and its confidence limits can have different availability. Conceptually, limits can be obtained from the 50% crossings of the curve’s confidence limits. One crossing may be observed while another is not. Methodological documentation for quantiles.

A table should preserve those differences. “Not reached”, “not estimable” and “not reported” require a legend explaining their meaning, rather than automatic conversion to the same missing value.

A median of 45.8 months with a 95% CI from 37.7 to not reached does not express symmetric precision. The upper limit is not estimated in that analysis. Displaying only 45.8 hides part of the uncertainty.

A median’s interval also has time units. An HR interval describes a ratio. Neither can replace the other or supply an invented confidence interval for additional months.

7. Why a comparison can be informative without two medians

The median summarises a particular region of a distribution. A survival comparison can use event times and risk sets accumulated before either curve reaches 50%.

For example, information arises when fewer events than expected under a no-difference comparison are repeatedly observed in one group. That information does not disappear because the median crossing is absent. The comparative question and the 50th-percentile question use different aspects of the same data.

The TALAPRO-3 prescribing information identifies a stratified Cox model for the HR and a stratified log-rank test. Strata include disease status, disease volume and BRCA versus non-BRCA status. Table 14 notes.

The log-rank test compares distributions using the event sequence. The Cox model estimates a hazard ratio under its formulation. Statistical significance does not turn an HR into an individual probability or guarantee that a single ratio describes the entire time course.

8. The HR and three-year benefit answer different questions

The HR of 0.48 is a relative measure of hazards: instantaneous event rates among people who remain event-free, within the model used. Interpreting it as a constant effect over time requires assessing proportional hazards.

The sponsor’s initial announcement reports three-year progression-free survival of 77% versus 56%. These rounded percentages allow a descriptive translation:

These calculations are not new estimates from individual participant data. The approximate three-year ratio is not the HR either. They are different functions of the distribution and can differ without contradiction.

The 21-point difference is neither 21 months nor a cure proportion or a guaranteed benefit for every patient. Estimating its uncertainty requires the appropriate estimates and method; the HR interval does not automatically provide that interval.

9. Why accumulated counts do not replace Kaplan–Meier

The first table’s event counts describe accumulated information, but dividing them by the number assigned does not directly reproduce a three-year risk. Participants may have different follow-up durations and censoring dates.

A crude percentage asks how many events were observed among those included by the cutoff. A fixed-horizon estimate uses timing information to address the estimated probability of an event by that horizon, under its assumptions.

Nor can the median be inferred by counting whether more or less than half of all participants experienced the event. The weighting implicit in risk sets matters: with censoring, a curve’s 50% crossing need not coincide with a crude 50% event proportion.

A report should therefore present participants, events, follow-up and time-specific estimates together, with clear denominators and dates.

10. “Not reached” and “immature data” are different descriptions

A median may remain unreached with a favourable curve and substantial follow-up. It may also be unavailable because observation is still short. The label alone does not identify the dominant situation.

Maturity depends on the question: accumulated events, duration of observation, the proportion retaining useful follow-up and the precision needed for the planned analysis.

The initial scientific announcement described median follow-up exceeding 37 months. That quantity is not the progression median. Event time and observation duration are different variables.

A more frequent endpoint may become informative before a later one. Reaching the main progression analysis does not automatically complete the mortality information. The FDA explicitly states that OS remained immature at the rPFS analysis.

11. The estimand: which effect do we want to know?

A treatment question should specify population, interventions, endpoint, summary and handling of intercurrent events. ICH E9(R1) provides this framework, separating the targeted effect, its estimation method and the numerical result. EMA-adopted document.

When reading this trial, it is useful to check what happens after discontinuing a drug, switching therapy or changing the regimen. Continuing outcome observation and stopping analysis at that point may address different questions.

Treatment discontinuation and missing data are also different events. A drug can be stopped while outcome assessments continue.

The exact rules must come from the protocol and statistical analysis plan; they cannot be inferred from an HR or the label “double-blind”. No unverified intercurrent-event strategy is attributed to TALAPRO-3 here.

12. Death changes its role with the question

In rPFS, death is part of the composite endpoint. Death before radiographic progression can constitute an event under the analysis rules; it should not conveniently become “absence of progression”.

A different question would be the incidence of radiographic progression before death. Death may then prevent progression from being observed and becomes a competing event. A cumulative incidence function addresses a different question from remaining free of the composite.

Censoring deaths and calculating one minus Kaplan–Meier does not automatically estimate progression incidence in the presence of that competition. Changing the handling of death changes the variable and its interpretation. The FDA oncology endpoints guidance helps establish the distinction between progression and survival. Austin, Lee and Fine’s methodological article explains cumulative incidence in competing risks.

13. What to examine when one HR is insufficient

If an effect changes over time, a single HR may conceal relevant differences. Review should consider curve shapes, the evolution of effects and appropriate diagnostics; an isolated proportionality test does not settle the whole interpretation.

Complementary measures include probabilities at clinically justified times and restricted mean survival time, RMST: the area under a curve up to a common horizon. For rPFS, it means average time free of progression or death within that period. Royston and Parmar, 2013.

The horizon needs follow-up support and justification. These measures should not be selected after exploring multiple options to retain the most favourable. No TALAPRO-3 RMST is calculated here from the reported medians.

14. Overall survival: keeping uncertainty visible

The published interim analysis reported an HR for death of 0.77, with a 95% CI of 0.56 to 1.04. The interval includes 1: the favourable point estimate retains relevant uncertainty. Article abstract.

Interpretation should avoid both declaring a conclusive improvement and treating the result as proof that no benefit exists. An endpoint that fails a statistical criterion can remain compatible with effects of interest and with insufficient information to distinguish them.

A full review also examines interim decision boundaries and multiplicity control. Endpoints, subgroups and analysis times do not all have the same confirmatory status. A conventional interval cannot replace examination of those rules.

Further follow-up may improve precision and change the estimate. It does not guarantee the direction of the final result.

15. Safety: magnitude, seriousness and exposure

The scientific abstract reports serious adverse events in 42% versus 32%, grade 3 or higher anaemia in 51% of the talazoparib group and two treatment-related deaths. Published results.

Seriousness and high severity describe different aspects: regulatory seriousness concerns an event’s consequences or clinical criteria, whereas a grade describes intensity according to the scale used. The percentages should not be mixed as interchangeable categories. ICH E2A safety definitions hosted by the FDA.

Safety assessment also needs exposure duration, interruptions, dose reductions, discontinuations and the analysis population. The same crude proportion may correspond to different observation periods. Exposure-adjusted measures can add context, but they do not by themselves remove all biases or replace clinical review.

Benefit–risk assessment requires considering efficacy and harms in a defined population alongside preferences and alternatives. A favourable progression HR alone does not settle that assessment.

16. Biomarkers and subgroups: identifying whom the result represents

HRR selection improves the biological definition of the population, but the group contains different alterations. The overall effect does not automatically establish an identical magnitude for every gene, clinical context or patient category.

Assessing heterogeneity requires attention to the interaction test, its precision and whether it was planned. Significance in one subgroup and its absence in another does not establish a difference between subgroup effects. With few events, subgroup intervals may be wide despite a precise overall result.

Different-looking point estimates are insufficient too. The prescribing information identifies exploratory analyses by BRCA status. Their exploratory category should be preserved rather than turned into a universal selection rule.

17. What a survival table reviewer should be able to reconstruct

Reproducibility begins with the decisions producing the variable. Two programs can return matching numbers while answering a poorly specified question.

A useful review connects four levels:

1. Clinical definition: population, time origin, event and endpoint components.
2. Derivation: date sources, valid assessments, censoring rules and discrepancy resolution.
3. Estimation: analysis population, model, strata, intervals and conventions for medians and quantiles.
4. Presentation: units, data cutoff, denominators, numbers at risk, notes and uncertainty.

Checks should cover situations capable of changing results: partial dates, events and assessments on the same day, missed visits, death without documented progression or initiation of another therapy. These need predefined rules, traceability and review rather than an improvised choice when producing the figure.

Beyond reproducing the main calculation, relevant assumptions should be challenged. ICH E9(R1) distinguishes sensitivity targeting the same estimand from supplementary analyses answering other questions. Sensitivity framework.

18. Practising the method without confusing simulation and trial evidence

The Kaplan–Meier, step by step guide explores steps, censoring and numbers at risk, with an exercise reproducible in R. Its data are synthetic and its clinical case is fictional, using mortality as the event. It does not reproduce TALAPRO-3 or its analytical decisions.

Three tasks connect the guide to this article:

• Locate the first 50% crossing in curves that reach it and describe what you would report if no crossing occurred.
• Compare a fixed-time estimate with the number of people still observed; explain why they are different quantities.
• Identify the information needed before interpreting a flat tail or calculating an absolute difference.

The article’s five-person example adds a check: two events and three censorings produce 53.3% at the end and a median not reached. Understanding that calculation makes it easier to recognise why an isolated label needs context.

When a table says “median not reached”, the next things to read are the curve, its uncertainty and the information supporting it. The comparative result, overall survival and safety then retain their own questions.

References

1. FDA. Approval on 26 October 2026 and efficacy assessment.
2. Talzenna US prescribing information, revised October 2026; section 14.2, TALAPRO-3.
3. Agarwal N et al. N Engl J Med. 2026;395:427–439. DOI: 10.1056/NEJMoa2604126; online 30 May 2026.
4. TALAPRO-3, NCT04821622.
5. Pfizer. Initial results announcement, 30 May 2026.
6. FDA. Oncology endpoints guidance, December 2018.
7. ICH E9(R1). Estimands and sensitivity analysis; EMA Step 5 version, 2020.
8. Therneau T. Quantile documentation for the survival package.
9. Royston P, Parmar MKB. Restricted mean survival time in trials with time-to-event outcomes. BMC Med Res Methodol. 2013;13:152.
10. ICH E2A. Safety definitions and expedited reporting.
11. Austin PC, Lee DS, Fine JP. Competing risks methodology. Circulation. 2016;133:601–609.