Time-to-Event Endpoints: Writing Censoring Rules Readers Can Follow

Consider a hypothetical participant, not a real record, on the experimental arm of a randomized progression-free survival (PFS) trial. At week 8 the participant permanently discontinues the investigational agent for grade 3 transaminitis. Scheduled tumor evaluations at weeks 12 and 16 are missed. At week 18 a community oncologist starts a commercially available later-line regimen. At week 20 restaging scans show disease progression. The last scan that documented no progression was at week 6. At database lock, is this record an event or a censored observation, which date is used, and which analysis-dataset field records the decision?

In many protocols, this question cannot be answered because the protocol text contains only a generic summary statement: Participants who remain alive without disease progression will be censored at the date of their last adequate tumor assessment. That statement is not a censoring rule. It is a broad statement of statistical intent that provides zero guidance when an actual participant departs from the scheduled protocol pathway. It fails to define what constitutes an 'adequate' assessment, fails to state whether missed visits invalidate subsequent event observations, fails to address intervening non-protocol interventions, and fails to instruct the statistical programmer which derivation rule takes precedence.

The root of this operational gap is visible in the core statistical guideline. A full-text search of the ICH E9 guideline on Statistical Principles for Clinical Trials (Step 4, 5 February 1998) finds survival-analysis methods and no occurrence of the word censoring. ICH E9 therefore does not supply censoring-rule templates. Teams take row-level examples from therapeutic-area guidances, including the FDA oncology endpoint guidances and the EMA anticancer appendix discussed below, and store the resulting dates in CDISC analysis datasets. Those sources are nonbinding examples for defined settings. They are not a universal censoring code for every time-to-event endpoint.

To be executable, a censoring rule must be written as an instruction set, not an analytical remark. It must establish time zero, define the event criteria, specify what qualifies as an adequate observation, and enumerate every plausible early-exit pathway in an explicit, ordered matrix. Each situation must resolve to exactly one dated outcome—either a failure event on a verifiable date or a censored observation on a verifiable date.

Every time-to-event endpoint definition—whether tracking progression-free survival, overall survival, disease-free survival, time to treatment failure, or time to symptom resolution—requires four foundational components before a single participant is enrolled:

1. Time Origin (T₀): The calendar date from which elapsed time is counted. For a randomized comparison, FDA's December 2018 cancer-endpoint guidance defines progression-free survival from randomization. EMA Appendix 1 defines progression-free survival and disease-free survival from randomization, or from registration in a non-randomized trial. Using first dose as time zero in a randomized analysis can create immortal time if participants who die or withdraw between randomization and first dose are dropped from the analysis set. Some single-arm plans, including the publicly posted plan discussed later, measure PFS from first dose. That choice has to be written down. The protocol also has to say how people who are randomized or enrolled but never dosed are classified.

2. The Adjudicated Event Definition: The clinical, laboratory, or radiological criteria that count as failure. FDA's December 2018 cancer-endpoint guidance defines PFS as time from randomization to objective tumor progression or death, whichever occurs first, and time to progression as time to objective tumor progression with deaths censored. It states that censoring deaths in a time-to-progression analysis is informative censoring, and that PFS is the preferred regulatory endpoint of the two. Those oncology definitions do not travel unchanged. A cardiovascular major-adverse-cardiovascular-event endpoint still has to state whether non-cardiovascular death is part of the composite, a competing event, or a censored observation. The non-small cell lung cancer tables below are PFS examples, not a template for that choice.

3. The Definition of an Adequate Assessment: The criteria required to affirm that the event has not occurred. Appendix B of FDA's April 2015 non-small cell lung cancer endpoint guidance says that, when there is no evidence of progression, censoring often uses the last adequate tumor assessment, and that adequacy includes target-lesion measurements plus radiological tests sufficient to evaluate nontarget lesions and to look for new lesions. A physical examination can be a source of a progression observation — the 2018 guidance notes that progression data may come from exams as well as scans — but an exam alone does not establish the absence of progression for a censoring date. A partial scan, such as a chest CT without the abdominal CT the protocol required, is a case the statistical analysis plan has to classify. Appendix B says the analysis plan should describe how progression status is calculated when tumor data from an otherwise adequate visit are partially missing.

4. The Enumerated Early-Exit Matrix: The comprehensive catalog of early discontinuation and deviation pathways. Participants withdraw consent, experience dose-limiting toxicities, miss scheduled visits due to illness, seek alternative therapy, or remain in continuous follow-up past the data cutoff date. Every pathway must be assigned a deterministic date and a deterministic status.

Two FDA documents describe progression and censoring dates, and they do not collapse into one mandatory rule. The December 2018 cancer-endpoint guidance recommends assigning the progression date to the earliest time any progression is observed without prior missing assessments, and censoring at the date the last radiological assessment showed no progression. It points to Appendices C and D of the April 2015 non-small cell lung cancer guidance for example tables. Appendix B of that 2015 guidance offers two ways to date progression: the first time progression can be declared, or the protocol-scheduled clinic visit after the assessments that collectively document progression, which it recommends for a sensitivity analysis. For a new lesion, the first-time rule uses the date the lesion was first observed. For target lesions measured on different days, it uses the first observation that meets the predefined increase. On completely missing assessments, Appendix B says that taking a later progression or death after a gap can be informative censoring, that a later assessment can be reasonable when evaluations are frequent and only one visit was missed, and that censoring at the last adequate tumor assessment can be more appropriate when two or more visits were missed. It also says all efforts should be made to keep following patients for progression no matter how many visits were missed.

The example primary schemes are Tables C1 and C2 in Appendix C of FDA's April 2015 guidance, Clinical Trial Endpoints for the Approval of Non-Small Cell Lung Cancer Drugs and Biologics. The guidance calls them examples of a prespecified censoring scheme, and the document contains nonbinding recommendations. They are written for PFS. Both tables use three columns — situation, date of progression or censoring, and outcome — and they are not a single default. They agree on several rows and diverge on discontinuation for toxicity or another reason, new anticancer treatment, and death or progression after more than one missed visit. Table C2's divergent toxicity and new-treatment rows are conditioned on protocol-specified continued follow-up in all treatment arms. Appendix D, discussed later, is a set of sensitivity analyses, not a second pair of primary schemes.

The split is narrower than 'any early exit.' Discontinuation for undocumented progression is censored in both Table C1 and Table C2. The schemes diverge for discontinuation because of toxicity or another reason, for a new anticancer treatment, and for death or progression after more than one missed visit. In those rows Table C1 censors at the last assessment that still showed no progression. That choice avoids attributing a later progression to the randomized treatment, and it creates informative censoring if the arm with more toxicity or more early switching is the arm that stops being scanned. Table C2 counts a later documented progression as the event. For the toxicity and new-treatment rows, the date column states the condition in the guidance: protocol-specified continued follow-up in all treatment arms. Table C2 does not move the event to the discontinuation date or to the new-therapy start date.

Table C2 is an analysis rule and a follow-up requirement together. If sites stop scans when study drug stops, the documented progression date that Table C2 names is never collected, and those rows cannot be executed as written. Continued follow-up is also what Appendix B asks for: efforts should be made to keep following patients for progression irrespective of missed visits. EMA Appendix 1, adopted 13 December 2012 and effective 1 July 2013, states its own primary-analysis default in similar terms. The progression or recurrence date should be assigned from the first documented evidence that the criteria were met, including when progression is observed after one or more missed visits, after treatment discontinuation, or after new anticancer treatment. Outcome data should be collected on the intended schedule for all randomized patients. A different primary approach needs a justification, and the appendix recommends CHMP scientific advice at the planning stage. These documents are guidance. They do not make the non-small cell lung cancer example tables binding for overall survival, symptom endpoints, or non-oncology time-to-event analyses.

Apply the tables to the hypothetical participant. Toxicity discontinuation is the 'toxicity or other reason' row, not undocumented progression. The missed week-12 and week-16 evaluations are more than one missed visit. New therapy starts at week 18, and progression is documented at week 20. The last assessment with no progression is week 6. Under Table C1, the toxicity row, the new-treatment row, and the missed-visit row each censor at that last non-progression assessment. They agree: censored at week 6. Under Table C2, the toxicity row and the new-treatment row are events at the documented progression date if continued follow-up was specified for all arms: an event at week 20, not at week 8 and not at week 18. The missed-visit row in Table C2 also records an event at the documented progression date. If scans stopped at discontinuation, that date was not collected and Table C2's condition is unmet. A composite sensitivity rule is a different table. In Appendix D, Table D2 treats discontinuation for toxicity as progression on the discontinuation date, week 8, and new anticancer treatment as progression on the start date, week 18. Both situations apply here, so the statistical analysis plan has to rank them. The earliest composite failure would be week 8. A rule that counts only new therapy, and not discontinuation, would use week 18. Neither date is what Table C2 says. If the protocol never chooses among these rows, the hazard ratio is not yet defined. It moves with whichever rule is selected once the number of records like this one is known.

The December 2018 FDA cancer-endpoint guidance is labeled nonbinding. It does not say that a later change of censoring scheme is forbidden. It says FDA and the applicant should agree prospectively on the items below, and that the applicant should finalize them before the study starts, to the extent possible:

• The study design. Visit windows make that design executable, but they are not a separate bullet in the 2018 list.

• The definition of progression. The guidance says there are no standard regulatory criteria. RECIST is one option applicants use, and the protocol has to add the detail the published criteria leave open.

• The statistical analysis plan. Aligning the estimator with an ICH E9(R1) estimand matters, but that addendum was adopted in November 2019 and is not one of the 2018 bullets.

• The methodology for handling missing data and censoring methods. The same section says the protocol should define an adequate assessment visit and that the analysis plan should specify the primary analysis and one or more sensitivity analyses.

• The data to be recorded on the case report form. To execute a censoring table, those fields need to show whether each assessment was adequate and the dates of non-protocol therapy. That field list is what the table requires. It is not a quotation of the 2018 bullet, which names the CRF data without listing those fields.

• The operating procedures of an independent endpoint review committee, if applicable. The guidance appendix says the independent review charter should be in place before initiation when an independent review is used.

Appendix 1 to the EMA guideline on the evaluation of anticancer medicinal products in man (EMA/CHMP/27994/2008 Rev.1, adopted 13 December 2012, effective 1 July 2013) is methodological guidance for confirmatory PFS and disease-free survival analyses. It is not an inspection manual. The text below is the appendix adopted in 2012 and effective in 2013; confirm that this is still the current EMA wording before a protocol copies a sentence from it. On withdrawals the appendix states: "Events of withdrawal from study therapy prior to adjudicated progression are likely to be informative and the adequacy of censoring these events in the statistical analysis should always be questioned." It also says commonly used methods for comparing survival times are valid only if censoring is uninformative, and that informative censoring may lead to incorrect conclusions about the extent of the treatment difference.

For patients with incomplete follow-up at the cutoff — premature censoring, as distinct from administrative censoring of patients with complete follow-up — the appendix says the uninformative-censoring assumption should be examined systematically, and the extent of informative censoring should be discussed in the clinical trial report. The examples it gives are below. A separation of time-to-censoring curves is a reason for that discussion. It is not, by itself, a finding that the log-rank test is invalid.

1. Time-to-censoring curves: The appendix's example is to investigate time-to-censoring with Kaplan-Meier curves, including differences by treatment arm. One practical display codes censoring as the event and treats progression and death as censored. Earlier or heavier censoring on one arm means the independent-censoring assumption has to be discussed. The appendix does not require a hypothesis test on those curves, and a separation does not automatically discard the primary comparison.

2. Patterns across covariates: The appendix also says to examine patterns of censoring across covariates and the association between censoring and covariates that may be related to the PFS event, such as tumour burden, including differences by arm. Age, stage, performance status, prior therapy, and biomarkers fit that examination when they were prespecified as prognostic factors. The appendix does not prescribe a multivariate Cox model. An association between baseline severity and being censored is evidence that the assumption may fail. It is not a correction.

The appendix states the limit in one sentence: "There is no satisfactory way to correct for informative censoring, which should be minimised by adequate design and conduct of the study." It says similar considerations apply to overall survival. Sensitivity analyses can explore the assumption. They are not described as a repair for a trial that stopped follow-up. The appendix also says post hoc analyses are of limited value in compensating for detection bias from unscheduled assessments. The control it names is design and conduct: collect outcome data after treatment discontinuation, and prespecify censoring rules that minimise bias and loss of information. ICH E8(R1) critical-to-quality factors and the trial monitoring plan are where that follow-up obligation becomes operational. They do not add a statistical correction the appendix says is unavailable.

The ICH E9(R1) addendum on estimands and sensitivity analysis, adopted 20 November 2019, separates two problems that censoring tables often mix. It does not contain situation-by-situation censoring templates. Administrative censoring and intercurrent events are classified differently, and the strategy chosen for an intercurrent event decides whether later observations are used.

Administrative censoring is a missing-data problem, not an intercurrent event. The addendum's example is a participant for whom no outcome can be observed because the trial has ended. That observation is right-censored at the cutoff. It is not right-truncated. Truncation would mean the participant was included only because the event fell before a boundary, which is a different sampling mechanism. The addendum also says that neither study withdrawal nor other reasons for missing data, such as administrative censoring in survival trials, are in themselves intercurrent events. Someone who withdraws may already have had an intercurrent event, such as treatment discontinuation, before the withdrawal. Loss to follow-up is missing information about the event time. It does not name the strategy for a treatment change.

Treatment discontinuation, switching, and additional anticancer therapy are intercurrent events when they change how the endpoint is interpreted. E9(R1) describes five strategies and does not require the same strategy for every intercurrent event. For a treatment change, three strategies usually rewrite the censoring table: treatment policy, composite variable, and hypothetical. A fourth, while on treatment, is easy to confuse with hypothetical censoring and belongs in the same draft. Principal stratum strategies define a target population, such as participants who would not discontinue. They do not by themselves assign an event date or a censoring date. The censoring rule has to name which strategy each intercurrent event is serving.

• Treatment-policy strategy: The intercurrent event is irrelevant to the definition. The value of the variable is used whether or not the event occurs. For PFS, that means continued assessment and the date of documented progression or death. This is the question Table C2 and the EMA primary default are built to answer, and only if the later scans exist. E9(R1) notes that a treatment-policy strategy cannot be implemented for a terminal event such as death when the variable cannot be measured afterward. PFS already handles death by counting it as an event, which is a composite of progression and death, not a treatment-policy strategy for death.

• Composite-variable strategy: The intercurrent event is part of the outcome. E9(R1)'s example is discontinuation for toxicity treated as another mode of failure. In a time-to-event composite, the date is the date of that failure. That is Table D2's dating — discontinuation date, or new-therapy start date — not Table C2's later progression date. PFS is already a composite of progression and death. Adding discontinuation answers a further, different question.

• Hypothetical strategy: A scenario is envisaged in which the intercurrent event would not occur, and the value of interest is the value the variable would have taken in that scenario. E9(R1) says the scenario has to be stated. Censoring at the last adequate assessment before the intercurrent event is one implementation teams use. It is not the definition of the strategy, and it assumes the remaining risk set stands in for the hypothetical world. The addendum also says some hypotheticals — imagining, for example, that a participant who stopped for an adverse event would not have had the event — are of limited clinical and regulatory interest.

• While-on-treatment strategy: The question is the response before the intercurrent event, not the outcome had the event never occurred. For a time-to-event endpoint, observation stops at the named intercurrent event, which can look like censoring at discontinuation. The estimand is 'while exposed,' not 'what would have happened if treatment had continued.' E9(R1) says particular care is required when the intercurrent event occurs at different rates in the arms, which is the informative-censoring problem above.

flowchart TD
  A["Time origin stated in the protocol"] --> B{"Assessment shows the endpoint event?"}
  B -->|"Yes, and the rule counts it"| C["Event date, CNSR = 0"]
  B -->|"No"| D{"Stop, switch, or new therapy?"}
  D -->|"No: event-free at cutoff"| E["Administrative right-censoring"]
  D -->|"Yes"| F{"Strategy named for that event"}
  F -->|"Treatment policy"| G["Keep follow-up; use documented progression or death"]
  F -->|"Composite"| H["Event on the intercurrent-event date"]
  F -->|"While on treatment"| I["Question ends at the intercurrent event"]
  F -->|"Hypothetical"| J["Target is the outcome had the event not occurred"]

Censoring at a treatment change is often treated as the rigorous default. A simulation in BMC Medical Research Methodology (Jamoul and colleagues, 2022) is narrower than that slogan. The authors modeled a six-month period with no enrollment and no tumor scans, and they varied when that shutdown started, in an advanced-cancer PFS setting. The treatment-policy analysis ascribed events to the time they were observed. The hypothetical analysis censored patients who had events during the shutdown at the last assessment before that period. The paper is about missing scans, not a direct simulation of subsequent anticancer therapy.

Compared with no shutdown, the treatment-policy analysis slightly overestimated median PFS in proportion to when the shutdown fell. Power was not affected and, except in one scenario, the hazard ratio was not affected. The hypothetical analysis had a larger effect: power decreased, and median PFS was overestimated by more. The authors suggest that the treatment-policy approach, which follows the intention-to-treat principle, should be the primary analysis in that setting so the trial does not give up power unnecessarily and so bias in the median is smaller. The paper does not show that treatment policy removes median bias, and it does not measure effectiveness outside the simulation.

A censoring rule that a programmer cannot derive is not yet a rule. In a CDISC analysis submission, a time-to-event dataset is expected to use the ADaM basic data structure for time-to-event analyses: one record per subject per parameter such as PFS, overall survival, or event-free survival. The standard defines the columns. It does not decide whether Table C1, Table C2, or a therapeutic-area variant is the right clinical rule. That choice stays in the statistical analysis plan. Variable names below follow that time-to-event structure as implemented in CDISC-aligned tooling, including CNSR equal to 0 for an event and positive integers for censoring.

Each subject contributes one record per analysis parameter (PARAMCD). The protocol table has to say which source date fills ADT and which integer fills CNSR. The mapping below is the data contract, not a second clinical rule.

Positive values of CNSR keep censoring reasons distinct. They are not a CDISC-mandated code list. One workable map, which the plan has to publish, is 1 for administrative censoring at cutoff, 2 for loss to follow-up, 3 for subsequent non-protocol therapy, 4 for two or more missed visits, and 5 for no baseline or no evaluable post-baseline assessment. Another map is acceptable if every code is defined once and used the same way in the table, the dataset, and the reviewer's guide. The test is whether a programmer can derive the row without inventing a reason at unblinding. See what the statistical analysis plan has to lock before that derivation is treated as final.

The publicly posted statistical analysis plan for NCT04119453 (Elevar Therapeutics, RM-202), section 6.2.1.1, is one illustration, not a regulatory template. It measures PFS from the first dose of rivoceranib, not from randomization, and it calculates days as the event or death date minus the first-dose date plus 1. Subjects with neither progression nor death at the analysis are censored at the last response assessment. If progression or death occurs after two or more missed visits, censoring is at the latest assessment before those missed visits. Two missed visits are defined from the schedule, every 8 weeks in the first year and every 12 weeks thereafter, plus the protocol window. Subjects without documented progression who start subsequent anticancer therapy are censored at the last response assessment before that therapy. Subjects with no evaluable visits or no baseline assessment are censored on Day 1, unless death occurs within two visits of baseline, in which case the death date is used. Leaving out that death exception would not be an implementation of this plan.

The standard supplies the columns. It does not choose Table C1, Table C2, or a therapeutic-area variant as the primary analysis. If the plan omits the date, the outcome, and the order of overlapping rows, a programming dry run will not produce one defensible analysis dataset.

Independent censoring is an assumption, so both FDA and EMA ask for prespecified sensitivity analyses. They do not publish one shared list of four methods. Appendix D of the 2015 non-small cell lung cancer guidance gives three example tables. It says these sensitivity analyses are exploratory and supportive of the primary analysis, and that efficacy may not be claimed based on sensitivity analysis alone. EMA Appendix 1 lists a separate set of supportive examples. The four items below keep those sources apart.

1. FDA Table D1, scheduled dates only: Censoring and event dates are taken at scheduled visits. Progression documented between visits is dated at the next scheduled visit. The guidance notes that this can introduce bias if progression occurred closer to the last visit. The table is a check on ascertainment timing, not a more accurate clock. Discontinuation, new anticancer treatment, and death or progression after more than one missed visit stay censored at the last visit with an adequate assessment.

2. FDA Table D2, any change counted as progression: Discontinuation for undocumented progression, and discontinuation for toxicity or another reason, are events on the discontinuation date. New anticancer treatment is an event on the start date. Death or progression after more than one missed visit is an event on the date of the first missed visit. If a primary PFS result depends on censoring those changes, this is the table that shows the dependence. It is still a sensitivity analysis.

3. FDA Table D3, investigator clinical progression: An investigator claim of clinical progression is an event at the scheduled visit, or at the next scheduled visit if the claim falls between visits. Discontinuation for toxicity, and new anticancer treatment without a claim of progression, are censored at the last visit with an adequate disease assessment. Death between adequate visits, or after a single missed visit, is an event on the death date. Death after two or more missed assessments is censored at the last adequate visit. Table D3 adds clinical progression. It does not turn every treatment change into an event, which is what Table D2 does.

4. EMA Appendix 1 supportive analyses: For the primary analysis, the appendix prefers the documented time of progression, including an unscheduled assessment, because that dating is closer to the intention-to-treat principle and may be less prone to informative censoring. Supportive analyses may assign the progression date to the scheduled clinic visit, use an interval-censored analysis, use a single prespecified time point after randomization, or assume censored subjects are at high risk or at low risk of an event. Interval censoring treats progression as having occurred somewhere between the last negative assessment and the first positive one. The appendix says interval censoring is ignored in the primary analysis it describes. It does not define a tipping-point procedure, a worst-case assignment for one arm and a best-case assignment for the other, or a threshold at which statistical significance is lost. Those displays can be programmed. They are not the analysis the appendix specifies. The same-estimand sensitivity logic is developed further in the site's missing-data sensitivity analysis.

The same assumption matters when the comparison uses external data. Rudra Gupta and colleagues, ESMO Open (2025;10:104094), simulated informative censoring of external-control data in cancer trials. The designs included model-based replicates, in silico glioblastoma trials resampled from completed phase III trials, and both single-arm and randomized designs that use external data. Informative censoring was dependence between censoring times and survival times, with scenarios for a positive association and a negative association.

The simulations showed that informative censoring of the external-control data can bias the trial result, and that the direction of the bias depends on the censoring mechanism. The same censoring can inflate or reduce type I error and power. The authors conclude that choosing external-control data, and deciding to use those data, should account for this risk, and that analyses to detect informative censoring are recommended when the design uses external data. The paper does not estimate dropout in electronic health records, and it does not claim that external controls are always censored more heavily than trial participants. The implication for rule writing is narrower: the external cohort still needs a situation table, because an unexamined censoring rule can move the treatment-effect estimate in either direction.

To ensure that your trial's time-to-event endpoints withstand scrutiny from clinical operations, data management, independent monitors, and regulatory inspectors, follow this 6-point implementation checklist:

1. Enumerate the situations before the case report form is frozen: For a PFS analysis modeled on the non-small cell lung cancer examples, separate incomplete or missing baseline assessments, progression between visits, no progression at cutoff, discontinuation for undocumented progression, discontinuation for toxicity or other reasons, new anticancer treatment, death before the first progression assessment, death between adequate visits, and death or progression after more than one missed visit. Other endpoints need their own rows. Pasting the PFS rows into a symptom endpoint or a cardiovascular composite answers a question the table was not written for.

2. Assign Exactly One Dated Outcome Per Situation: Ensure every row of the table explicitly defines whether the outcome is an Event or a Censor, and names the exact calendar date (e.g., date of randomization, date of last adequate assessment, date of objective progression, date of new therapy).

3. Establish Deterministic Rule Precedence: When a participant experiences multiple deviations—such as missing a visit and then starting subsequent therapy—the protocol must establish strict sequential priority to prevent conflicting programming derivations.

4. Quantify Missed-Visit Windows in Concrete Days: Define what constitutes 'two or more missed visits.' In protocols with 6-week imaging intervals, specify whether two missed visits equals 12 weeks plus the protocol window (e.g., 84 days + 14 days = 98 days) from the last adequate scan.

5. Keep the follow-up the chosen row assumes: EMA Appendix 1 says there is no satisfactory way to correct informative censoring and that it should be minimised by design and conduct. FDA Appendix B says efforts should be made to keep following patients for progression irrespective of missed visits. Monitoring instructions have to tell sites to continue the event assessment after investigational therapy stops, on the schedule Table C2 or the EMA primary default actually uses.

6. Map Table Rows to CDISC ADaM ADTTE Variables: Verify before database unblinding that every rule row maps directly to ADaM variables (STARTDT, ADT, AVAL, CNSR, EVNTDESC, CNSDTDSC), ensuring absolute traceability from electronic case report forms to final submission tables.

A time-to-event endpoint is executable when every early-exit situation has one prespecified date and one outcome, the row names the estimand strategy it serves, and a programmer can fill the analysis date and the censoring indicator without reinterpreting the protocol. The non-small cell lung cancer examples that make this concrete were issued in April 2015, and the EMA appendix took effect in July 2013. Confirm that both documents are still the current text before those rows are copied into a protocol, and write a new situation list when the endpoint is not PFS.