The Conditional Meaning Of A Cause-Specific Hazard

What Is Being Conditioned On?

A cause-specific hazard answers a time-specific question: among individuals who have not yet experienced any first event by time \(t\), how rapidly is a particular cause occurring at that time? The condition matters. The group being described is not everyone enrolled at the start, and it is not just people who will eventually experience the cause of interest. It is the group still in the event-free state at time \(t\).

Let \(T\) denote the time to the first event and let \(J\) denote its cause, with mutually exclusive first-event causes indexed by \(k\). Write \(E(t)=\{T\geq t\}\) for the event that an individual is event-free through time \(t\). Using the small-interval definition, the cause-specific hazard for cause \(k\) is

\lambda_k(t) = \lim_{\Delta t\to 0^+} \frac{P(t\leq T<t+\Delta t,\ J=k\mid E(t))}{\Delta t}.

The numerator inside the limit is a conditional probability: it concerns a first event of cause \(k\) in a short interval, given that the person was event-free at the interval's start. Dividing by the interval length expresses the local rate per unit time. Thus \(\lambda_k(t)\) has units of inverse time, such as per year. It is a population-level description of a changing risk set, not a claim about an individual's fixed personal rate.

Definition: A cause-specific hazard at time \(t\) is the instantaneous rate of a first event from cause \(k\), conditional on having experienced no first event by time \(t\). The conditioning group is the event-free risk set at that time.

Why the Risk Set Defines the Meaning

At time \(t\), someone who has already experienced any first event is no longer event-free. This is true whether that earlier event was the cause currently being studied or a competing cause. A person whose first event was cause \(k\) cannot later have a different cause as their first event; likewise, a person with a competing first event has left the event-free state. Neither belongs in the denominator for the cause-specific hazard after that event.

This conditioning is different from keeping the original enrollment count as the denominator at every time. The initial cohort includes people who may no longer be eligible to have a first event at time \(t\). The cause-specific hazard instead describes the rate in the event-free group that remains at that time. As time passes, both the size and the composition of that group can change.

In observed data, people known to be event-free and still under observation at \(t\) contribute to the risk set. A person censored before \(t\) is not known to be event-free at \(t\), so cannot simply be counted as if follow-up continued. Estimating hazards from censored data therefore requires appropriate handling of censoring; in standard analyses, the censoring process must not distort which individuals are represented in the observed risk set. The definition itself is about the event-free population, while an estimator must use the observed data in a way that represents that population.

Worked Example: A Local Rate Among Those Still Event-Free

Suppose, in a hypothetical cohort, 100 people begin follow-up. At 6 months, 64 are known to be event-free and under observation. During the next 0.1 years, 2 have a first event from cause A and 1 has a first event from cause B. Assume all 64 have complete follow-up through that interval, and the event times and causes are recorded correctly. Estimate the local cause-specific rate for cause A at 6 months.

The relevant risk set has \(Y=64\) people, not 100: the other 36 are no longer in the observed event-free group at the start of this interval. The observed conditional proportion with a cause A event during the interval is \(2/64=0.03125\). Dividing by the interval length gives the local rate approximation:

\widehat{\lambda}_A(0.5) \approx \frac{D_A}{Y\,\Delta t} = \frac{2}{64(0.1)} = 0.3125\ \text{per year}.

The estimate describes the rate of cause A events around 6 months among people event-free at that time. The cause B event is not counted in \(D_A\), but its occurrence does not make the person eligible to remain event-free after that event. The calculation is a local approximation based on a finite interval; interpreting it as an instantaneous rate assumes the interval is sufficiently short, or that the rate does not change materially over it.

From Conditional Probability to a Rate

The link between conditioning and the usual risk-set calculation follows directly from the definition. For a short interval beginning at \(t\), let \(p_k(t,\Delta t)\) be the probability of a first event from cause \(k\) during that interval conditional on being event-free at \(t\). The definition says that the limit of this conditional probability divided by the interval length is the cause-specific hazard.

p_k(t,\Delta t) = P(t\leq T<t+\Delta t,\ J=k\mid E(t)), \qquad \lambda_k(t) = \lim_{\Delta t\to 0^+} \frac{p_k(t,\Delta t)}{\Delta t}.

In a complete, uncensored interval, suppose \(Y(t)\) people are event-free at its start and \(D_k(t,t+\Delta t)\) have a first event from cause \(k\) during it. The observed conditional proportion for the interval is \(D_k/Y(t)\). Dividing that proportion by \(\Delta t\) gives the local rate approximation:

\widehat{\lambda}_k(t) \approx \frac{D_k(t,t+\Delta t)}{Y(t)\,\Delta t}.

This is an approximation because the hazard is defined as an interval shrinks toward zero, whereas real data are grouped over intervals of positive length. If the rate varies substantially within the interval, the calculation summarizes the interval rather than giving the exact hazard at its starting time. Using the risk set at the start is also essential: the denominator represents the people eligible to experience a first event at that time.

Worked Example: Why the Initial Cohort Is Not the Denominator

Consider a hypothetical study with 100 initially enrolled people. At time 0.5 years, 80 are event-free and under observation. Four have a first event from cause A during the next 0.1 years. At time 1.5 years, 20 are event-free and under observation; two have a first event from cause A during the next 0.1 years. Assume complete observation over both short intervals and accurate event classification.

For the first interval, the local rate estimate is

\widehat{\lambda}_A(0.5) \approx \frac{4}{80(0.1)} = 0.5\ \text{per year}.

For the second interval, it is

\widehat{\lambda}_A(1.5) \approx \frac{2}{20(0.1)} = 1.0\ \text{per year}.

Although fewer cause A events occurred in the later interval, the estimated local rate is higher because it is calculated among the much smaller event-free risk set at that time. If the initial enrollment count were used instead, the calculations would be \(4/[100(0.1)]=0.4\) and \(2/[100(0.1)]=0.2\) per year. Those numbers use a different denominator and do not estimate the cause-specific hazard at either time.

The comparison does not by itself explain why the later estimate is higher. The people remaining event-free at 1.5 years may differ in their characteristics from those at 0.5 years. The calculation supports a statement about the rate in each time-specific risk set, not a claim about what caused the rate to change.

Comparing Conditional Rates Across Groups

The same interpretation applies when comparing groups. Each group has its own event-free risk set at time \(t\), and a cause-specific rate in that group is conditional on membership in that group's risk set. Comparing rates therefore requires attention to both the time point and the group-specific conditioning sets. A contrast between two such rates is a contrast between rates among the people who remain event-free in each group at that time.

Worked Example: Two Group-Specific Risk Sets

Suppose a hypothetical study compares groups A and B over the same 0.1-year interval. At the interval's start, 30 people in group A and 50 in group B are event-free and fully observed. Three people in group A and two in group B experience a first event from cause K during the interval. Assume the event times and causes are known, and there is no loss to follow-up in either risk set over this interval.

For group A, the rate estimate is

\widehat{\lambda}_{K,A} \approx \frac{3}{30(0.1)} = 1.0\ \text{per year}.

For group B, it is

\widehat{\lambda}_{K,B} \approx \frac{2}{50(0.1)} = 0.4\ \text{per year}.

Over this interval, the estimated cause K rate among those event-free in group A is 0.6 per year higher than the corresponding rate among those event-free in group B. The conclusion is conditional and time-specific. It is not a comparison using all people initially assigned to each group, and it does not establish that group membership caused the difference.

Assumptions, Checks, and Common Pitfalls

Before interpreting a cause-specific hazard, check that its ingredients match the scientific question. The first event and its causes must be defined so that the causes are mutually exclusive; the time origin and time units must be clear; and event-free status at \(t\) must mean no first event of any cause by that time. For a calculation from observed records, the risk set must contain people known to be event-free and under observation at the interval start. The event count must include only first events of the cause being analyzed, and the interval length must be expressed in the time unit used for the reported rate.

A useful diagnostic is to trace a few records through time. A person with a competing first event should leave the cause-specific risk set at that event time. A person who is still event-free should remain eligible while observed. A person censored before the time of interest should not be treated as known event-free beyond their last observation. If the risk-set count cannot be reconciled with those statuses, the denominator may be wrong.

A common mistake is to reason, “There were 4 cause A events among the 100 initially enrolled, so the cause-specific rate is based on 100 people.” That uses the wrong conditioning group if only 80 were event-free and observed at the time in question. The correction is to identify the risk set at the start of the interval and use that count in the denominator, as well as the interval length. Another mistake is to report an interval estimate such as \(D_k/[Y(t)\Delta t]\) as an exact instantaneous rate without considering whether the interval is short enough for a local approximation.

A cause-specific hazard is also not interchangeable with every other summary of event occurrence. For example, a crude count divided by the original cohort size uses a baseline denominator and answers a different descriptive question. An absolute event-risk summary likewise has a different interpretation and must account for how event occurrence accumulates over time. Those quantities can be important, but they should not be substituted for the conditional rate defined here.

Key takeaway: To interpret a cause-specific hazard, name the cause, the time, and the condition: it is the instantaneous rate of that first-event cause among people still event-free at that time. The risk-set denominator is part of the meaning, not a bookkeeping choice.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

At time \(t\), 45 people remain event-free and under observation. One has a first event from cause R during the next 0.05 years. What is the local rate estimate for cause R?
Why is a person with a first event from a competing cause excluded from the cause-specific risk set after that event?
A study reports 6 cause S events in an interval, with 120 people initially enrolled but 75 event-free at its start. Which count belongs in the hazard denominator, and why?
What does a cause-specific hazard comparison between two groups condition on in each group?
Name one reason why an interval-based calculation might not equal the instantaneous hazard at the interval's start.

Solutions. (1) The estimate is \(1/[45(0.05)]=0.4444\) per year. (2) The person has experienced a first event and is no longer in the event-free state, so cannot experience a different cause as their first event. (3) Use 75, the event-free risk set at the interval's start; the initial 120 include people no longer eligible for a first event at that time. (4) Each rate is conditional on being event-free at the specified time within that group, so the two conditioning sets need not have the same size or composition. (5) The hazard may change materially during the interval; an interval-based rate is a local approximation rather than the limit over shrinking intervals.