Why Time at Risk Needs Its Own Definition
A calendar window tells us how much time has passed. It does not, by itself, tell us how long an entity was genuinely able to experience the event being studied. A policy may not yet have started, a membership may be suspended, or a person may leave the study. Even while a contract remains active, time after a first target event is no longer time at risk of that first event.
For predictive analysis, exposure must match the target and the population definition. If the question concerns the first covered loss, the relevant period is when the coverage is active and the policy has not yet had that loss. Counting all calendar days instead can make entities with different coverage histories appear to have had comparable opportunities for an event when they did not.
The phrase “at risk” is specific to the outcome. For a target defined as the first covered claim, a policy is at risk only while coverage applies and before that first claim. For a target defined as any claim occurrence, a later period might remain relevant after an earlier claim, depending on the target definition. The target must be stated before exposure is calculated.
This tutorial focuses on measuring the time interval. It does not assume a particular event distribution, or calculate an event rate. Those are separate modeling choices. Here, the essential task is to identify and add the eligible, event-free portions of the observation window.
Representing Eligible Time with Intervals
Use a consistent time scale, such as days from a stated starting date. We will represent periods with half-open intervals \([s,e)\): the start is included and the end is excluded. The interval has duration \(e-s\). This convention allows adjacent intervals such as \([0,10)\) and \([10,20)\) to meet without counting time at day 10 twice.
Let the study window be \([a,b)\). Suppose an entity is eligible and observed during one or more intervals \([s_j,e_j)\). If the target is the first event and that event occurs at time \(T\), the entity is at risk only before \(T\). If no target event occurs during the eligible period, the event time does not shorten that period. The exposure is the total duration in the intersection of the study window, the eligible intervals, and the time before the first target event.
Here, \(E\) is exposure on the chosen time scale. If there is no target event during the study window, treat the event cutoff as occurring after the window, so it does not reduce the overlap. Before using the sum, ensure that the eligible intervals do not overlap. If records describe overlapping periods for the same entity, merge them first; otherwise, the same time could be counted more than once.
Why the Interval Formula Works
The exposure contributed by one eligible interval is the length of its overlap with the study window and the pre-event period. The overlap begins when all required conditions have become true: the study has begun, the eligible interval has begun, and the entity has not yet reached the event cutoff. That is the latest start time, \(\max(a,s_j)\). The overlap ends as soon as any of those conditions stops being true: the study ends, eligibility ends, or the event occurs. That is the earliest end time, \(\min(b,e_j,T)\).
If the calculated end is later than the start, subtracting start from end gives the duration of the overlap. If the end is equal to or earlier than the start, there is no overlap and the contribution is zero. This gives the positive-part expression \(\max(0,\min(b,e_j,T)-\max(a,s_j))\). Adding the contributions is valid because the eligible intervals are disjoint.
This is a duration calculation, not a claim that every instant has the same probability of an event. It only establishes how much eligible time is present under the stated definitions. The data design must still ensure that the recorded intervals reflect actual coverage or membership and that the event time is defined consistently.
Worked Examples
Worked Example: A Policy with a Gap and a First Claim
A hypothetical policy is observed in the study window \([0,200)\), with time measured in days from the start of the window. Its active-coverage intervals are \([0,70)\) and \([80,160)\). The first target claim occurs on day 110. Find the exposure time for the first covered claim, and distinguish it from calendar time and active coverage after the claim.
Method. Intersect each coverage interval with the study window and the period before day 110. The first interval is entirely before the event, so its contribution is \(70-0=70\) days.
For the second interval, the relevant portion ends at the earlier of its coverage end, day 160, and the event time, day 110. Its contribution is \(110-80=30\) days. Total first-event exposure is therefore \(70+30=100\) days.
The calendar time from day 0 through the event on day 110 is 110 days. The 10-day gap \([70,80)\) is included in that calendar duration but is not covered time. The check is \(100+10=110\) days: the counted covered portions are exactly \([0,70)\) and \([80,110)\).
Coverage remains active from day 110 to day 160, a further \(160-110=50\) days. Those 50 days are after the first target event, so they are not exposure for the first-event target. The calculation is consistent: before the event, the 110 calendar days comprise 100 covered days and 10 uncovered days; after the event, the remaining active coverage is 50 days.
Check and common mistake avoided. The full active coverage totals \(70+(160-80)=150\) days, but using all 150 would wrongly include the 50 days after the first claim. Likewise, the post-event period is 50 days, not 40. The correct exposure is 100 days.
Worked Example: Membership, Temporary Suspension, and Event
A hypothetical member is followed from day 0 to day 100. Membership begins on day 15, is suspended during \([45,60)\), and ends on day 90. The target is the member’s first qualifying event, which occurs on day 75. Find the member’s exposure for this target. Use half-open intervals and measure all durations in days.
Method. Membership is active on \([15,45)\) and \([60,90)\). For a first-event target, stop counting at day 75. The first active interval contributes \(45-15=30\) days. The event-free part of the second active interval is \([60,75)\), contributing \(75-60=15\) days. Total exposure is \(30+15=45\) days.
As a check, calendar time from membership entry on day 15 to the event on day 75 is \(75-15=60\) days. Of these, 15 days are in the suspension \([45,60)\), leaving 45 active, event-free days. Membership after the event continues from day 75 to day 90, or \(90-75=15\) days, but those days are not exposure for the first-event target.
Check and common mistake avoided. Counting from day 0 would incorrectly include the 15 days before membership began. Counting continuously from day 15 to day 75 would incorrectly include the 15-day suspension. The correct exposure is 45 days.
Worked Example: Summing Exposure Across Three Entities
A hypothetical study observes three vehicles during \([0,100)\), with time measured in days. Vehicle A is covered on \([0,30)\) and \([40,100)\), with a first target event on day 70. Vehicle B is covered on \([10,90)\) and has no target event during the study. Vehicle C is covered on \([0,100)\) and has its first target event on day 25. Calculate each vehicle’s exposure and the total across the three vehicles.
Method. For vehicle A, the first interval contributes \(30-0=30\) days. In its second interval, only \([40,70)\) is before the event, contributing \(70-40=30\) days. A therefore contributes \(30+30=60\) days.
Vehicle B has no event during the study. Its complete covered interval contributes \(90-10=80\) days. Vehicle C is covered from day 0 until its first target event on day 25, so it contributes \(25-0=25\) days.
Adding the three contributions gives \(60+80+25=165\) days of exposure in total. As checks, A’s 100-day calendar window contains 10 days before coverage starts, 10 uncovered days from day 30 to day 40, and 20 covered days after its event; the remaining \(100-10-10-20=60\) days are its exposure. B has 10 days before entry and 10 days after coverage ends, leaving 80 days. C’s 75 days after the event are not exposure for a first-event target.
Check and common mistake avoided. The arithmetic adds exposure across entities, not their calendar windows. Counting all 100 days for each vehicle would give 300 days, including periods when vehicles were not covered or were no longer at risk of a first target event. The correct total is 165 days.
Common Traps in Defining Exposure
- Using the study window as the exposure. A study may run for 100 days, while an entity enters late or exits early. Count only the eligible portion.
- Ignoring an inactive gap. A policy or membership record can span a long calendar period while containing a gap in active status. Split the record into separate intervals and exclude the gap.
- Counting time after the target event. For a first-event target, event-free exposure stops at the event even if coverage or membership continues.
- Changing the target without changing the exposure rule. Time after a first event is excluded for a first-event target; whether it belongs in a different analysis depends on that analysis’s outcome definition.
- Mixing time units or interval boundaries. Convert start and end points to one time scale and apply the same endpoint convention throughout. A difference in units can produce a large error even when the interval logic is correct.
- Double-counting overlapping records. If the same entity has duplicate or overlapping eligible intervals, resolve or merge them before summing durations.
Check Your Understanding
For each problem, state the relevant intervals before calculating. Durations are in days, and intervals use the half-open convention.
- A policy is covered on \([0,50)\) and \([60,100)\), with a first target event on day 75. What is its exposure for the first-event target? (Answer: 65 days)
- A member’s study window is \([0,80)\). Membership is active on \([12,72)\), except for a suspension during \([30,42)\). The first target event occurs on day 55. What is the member’s exposure? (Answer: 31 days)
- A study window is \([0,80)\). An entity has continuous coverage on \([5,65)\) and no target event during the study. What is its exposure? (Answer: 60 days)
- An entity has active intervals \([0,25)\) and \([35,90)\), and its first target event occurs on day 50. Calculate exposure and identify the post-event active time that is excluded. (Answer: 40 days of exposure; 40 active days after the event are excluded)
- Entity A is eligible on \([0,40)\) with no event. Entity B is eligible on \([10,70)\), with a first target event on day 30. What is their combined exposure? (Answer: 60 days)