Projecting Events From Survival Probabilities

From Survival to an Expected Event Count

Event-driven trial plans often specify a target number of events, but planning also requires asking how many events may have occurred by a given follow-up time. If a model supplies the probability that a participant remains event-free through that time, the complementary probability is the chance that the participant has had the event. Multiplying those probabilities across a group gives an expected event count.

This calculation is a projection under assumptions, not a count observed in advance. It is useful for understanding whether a proposed follow-up duration might produce enough events to support a planned analysis. Its usefulness depends on whether the survival probabilities represent the trial population, treatment arms, endpoint, and time origin being planned.

Definition: Let \(T\) be the time from a defined origin to the first occurrence of a specified event. The event-free survival probability at follow-up time \(t\) is \(S(t)=P(T>t)\). If every participant's event status is ascertained through \(t\), the probability of having experienced the event by \(t\) is \(1-S(t)\). For \(n\) participants with this same event probability, the expected number who experience the event by \(t\) is \(n[1-S(t)]\).

The time origin must be explicit. It might be randomization or another defined start of follow-up. A projection at 12 months after randomization is not automatically a projection at 12 months after enrollment if those dates differ. Likewise, specify what counts as an event: for example, first disease progression, death from any cause, or a defined composite endpoint.

Why the Formula Works

For participant \(i\), define an indicator \(I_i(t)\) that equals 1 if that participant experiences the event by time \(t\), and 0 otherwise. The total event count is the sum of these indicators. The expectation of an indicator is its probability of being 1. Consequently, the expected count follows directly from linearity of expectation.

D(t)=\sum_{i=1}^{n} I_i(t), \qquad E[D(t)]=\sum_{i=1}^{n}E[I_i(t)] =\sum_{i=1}^{n}P(T_i\leq t) =n[1-S(t)].

The final equality assumes that all \(n\) participants have the same event probability through \(t\). Independence among participants is not required for this expectation calculation: linearity of expectation holds whether or not the indicators are independent. Independence can matter for the distribution and variability of the count, which are separate questions from its expected value.

For multiple arms, use the survival probability appropriate to each arm. If arm \(j\) has \(n_j\) participants and event-free survival \(S_j(t)\), its expected count is \(n_j[1-S_j(t)]\); the expected total is the sum over arms. This does not require the arms to have the same survival probabilities.

E[D_{\mathrm{total}}(t)] =\sum_{j=1}^{J}n_j[1-S_j(t)].

Worked Projections

Worked Example: One Group With a Specified Survival Probability

Suppose a hypothetical study follows 120 participants from randomization for 18 months. A planning assumption is that event-free survival at 18 months is \(S(18)=0.82\). Assume the event is defined consistently for every participant and that event status is ascertained through 18 months.

The assumed probability of an event by 18 months is the complement of survival:

P(T\leq 18)=1-S(18)=1-0.82=0.18.

Therefore, the expected event count is

E[D(18)]=120(0.18)=21.6 \text{ events}.

The projection is 21.6 expected events, not a claim that a realized study can observe a fraction of an event. An actual count is an integer and may be above or below this expectation.

Worked Example: Combining Two Trial Arms

Consider a hypothetical two-arm trial with 90 participants per arm and a common follow-up duration of 12 months from randomization. Assume event-free survival at 12 months is 0.76 in the control arm and 0.84 in the treatment arm. Suppose the same first-event endpoint and ascertainment rules apply to both arms.

For control, the event probability is \(1-0.76=0.24\), so the expected count is

E[D_C(12)]=90(0.24)=21.6 \text{ events}.

For treatment, the event probability is \(1-0.84=0.16\), giving

E[D_T(12)]=90(0.16)=14.4 \text{ events}.

Adding the arm-specific expectations gives \(21.6+14.4=36.0\) expected events in total by 12 months. The expected counts need not be whole numbers. The arithmetic also does not establish that treatment causes better survival; it is a projection conditional on the stated survival assumptions.

Obtaining Survival Probabilities From a Hazard Model

A survival probability may be supplied directly, or calculated from an assumed hazard model. The hazard describes the instantaneous event rate among participants who have not yet experienced the event. For a constant hazard \(\lambda\) over time, the exponential survival model gives \(S(t)=\exp(-\lambda t)\). The units must match: if \(\lambda\) is measured per year, \(t\) must be in years.

A constant-hazard assumption is convenient, but it is not required for projecting events. Under a time-varying hazard \(\lambda(u)\), survival through \(t\) is the exponential of the negative cumulative hazard through \(t\). This lets different hazard patterns, including a treatment lag or other non-proportional hazards, produce arm-specific survival probabilities.

S(t)=\exp\left(-\int_0^t \lambda(u)\,du\right).

Worked Example: Piecewise Hazards With a Treatment Lag

Suppose a hypothetical trial plans to project events by 2 years after randomization, with 80 participants in each arm. Assume a constant control-arm hazard of 0.20 events per person-year. In the treatment arm, assume the hazard is 0 during the first 0.5 year and 0.12 events per person-year from 0.5 through 2 years. This is an intentionally simplified lag model; it assumes the specified hazards apply to participants still event-free, with the same first-event endpoint in both arms.

For control, the cumulative hazard is \(0.20(2)=0.40\). Thus

S_C(2)=\exp(-0.40)=0.6703, \qquad E[D_C(2)]=80[1-0.6703]=26.38 \text{ events}.

For treatment, the first 0.5 year contributes no cumulative hazard under the stated assumption. The remaining 1.5 years contribute \(0.12(1.5)=0.18\), so

S_T(2)=\exp(-0.18)=0.8353, \qquad E[D_T(2)]=80[1-0.8353]=13.18 \text{ events}.

The projected total is \(26.38+13.18=39.56\) expected events by 2 years. The treatment hazard is not a constant fraction of the control hazard throughout follow-up: it is zero at first and positive later. Therefore, a single constant hazard ratio would not describe the assumed pattern. The projection uses the survival probability implied by each arm's own hazard history.

Assumptions, Diagnostics, and Alternatives

The calculation is simple, but the probability supplied to it carries substantive assumptions. Before using a projection, check that the survival probability refers to the intended endpoint, population, treatment arm, time origin, and follow-up duration. If survival values come from a fitted or external model, the projection inherits that model's assumptions. For instance, an exponential model assumes a constant hazard over the period considered; a piecewise model assumes the specified hazard is constant within each piece.

Also check whether the survival quantity matches the event being counted. For a first event that includes any death, ordinary event-free survival can be defined using time to the first qualifying event. If the endpoint is a particular event type and other event types prevent it from occurring, the probability of that specific event by \(t\) is a cumulative incidence probability. It should not automatically be calculated as one minus a cause-specific or “net” survival estimate that treats competing events as censored.

This tutorial's basic formula assumes complete event-status ascertainment through the stated follow-up time for the participants being counted. If participants can be lost before that time, the observed information does not simply provide their event status through \(t\); handling that situation requires an explicit adjustment. Similarly, when participants have different amounts of follow-up because they entered at different calendar times, a single common \(S(t)\) may not represent everyone at a calendar cutoff.

Useful diagnostic questions include: Does the assumed survival curve agree with credible data over the relevant time range? Does it show a pattern inconsistent with the chosen hazard model, such as a delayed treatment effect when constant hazards were assumed? Are the event definition and time origin consistent between the model and the planned trial? A model should not be accepted just because it gives an easy calculation. If its assumptions are implausible, the projected count may be systematically misleading.

A direct survival-probability projection is transparent and quick when plausible probabilities at the intended follow-up time are available. A parametric hazard model can supply survival probabilities at times not directly tabulated, but makes its functional assumptions explicit and consequential. An empirical survival estimate can reflect observed time-to-event patterns, but its relevance depends on the data and population from which it was estimated. None of these approaches turns an expected count into a guaranteed count. Prediction intervals and more detailed event-count uncertainty are distinct from the expectation calculation.

Define the projection.
State the event, time origin, follow-up time, and number of participants in each arm.

Choose the survival input.
Use a survival probability or a clearly specified model that represents the relevant arm and population.

Calculate expected events.
For each group, multiply its size by one minus its event-free survival probability; add group expectations if a total is needed.

Qualify the result.
Report it as an expectation conditional on the assumptions, not as a promised observed count.

Common pitfall: Treating survival itself as the event probability reverses the calculation. If \(S(t)=0.80\), the event probability by \(t\) is \(1-0.80=0.20\), not 0.80. The expected event count among \(n\) participants is \(0.20n\), subject to the stated endpoint, follow-up, and ascertainment assumptions.

Check Your Understanding

Answer each question, then compare with the solutions below.

A hypothetical group has 75 participants and assumed event-free survival of 0.88 at 1 year. What is the expected number of events by 1 year?
Two hypothetical arms each have 50 participants. At a specified follow-up time, their assumed survival probabilities are 0.70 and 0.90. Find the expected total event count.
Does the formula for expected events require independent event indicators? Briefly explain.
A model assumes a constant hazard of 0.15 per year. Calculate the expected event count among 40 participants followed for 2 years, using the exponential survival model.
Why might one minus a cause-specific survival estimate fail to equal the probability of a particular event type when competing events are possible?

Solutions. (1) The event probability is \(1-0.88=0.12\), so the expectation is \(75(0.12)=9.00\) events. (2) The arm-specific expectations are \(50(1-0.70)=15\) and \(50(1-0.90)=5\), for 20 expected events in total. (3) No. Linearity of expectation gives the sum of event probabilities without requiring independence. (4) The survival probability is \(\exp[-0.15(2)]=\exp(-0.30)=0.7408\). The event probability is \(1-0.7408=0.2592\), so the expected count is \(40(0.2592)=10.37\) events. (5) A competing event can prevent the event of interest from occurring. A cause-specific survival calculation that treats competing events as censored does not, by itself, give the actual cumulative probability of that particular event.