Why the Timing of a Safety Analysis Matters
An exposure–safety analysis can use accurate drug concentrations and correctly recorded adverse events yet still give a misleading result if it pairs the wrong times. An exposure measured after an event cannot explain that event. Conversely, a short exposure snapshot may miss a harmful effect that emerges only after repeated exposure or a delay. The central task is to define, before interpreting an association, which exposure history could plausibly contribute to which event times.
A window is an analysis choice, not merely a date range in a database. It determines what exposure is summarized, which events count as occurring under that exposure, and which periods contribute to the comparison. A window should therefore reflect the event’s plausible biology and the way exposure and outcome times are recorded. The earlier material on delayed effects and hysteresis provides useful context for distinguishing an immediate exposure–response relationship from a response that lags behind exposure.
These intervals answer different questions. For example, an analysis might summarize exposure during days 8–14 before an event-risk window beginning on day 1. The seven-day gap is the lag; it does not mean the exposure window itself lasted seven days. The observation period may include many such intervals, as well as periods that are not eligible for the particular comparison.
Define the Time Relationship Before Calculating an Association
Specify the relationship in a consistent time direction. For a given analysis time \(t\), let \(E(t)\) denote the exposure history and let \(Y(t)\) indicate an adverse event. A simple acute-effect question might relate exposure immediately before \(t\) to an event beginning at \(t\). A delayed-effect question might relate exposure in an earlier interval to an event in a later interval. A cumulative-effect question might relate accumulated exposure over a longer lookback period to a subsequent event.
In each case, the exposure summary must be available before the event-risk period it is intended to explain. Using exposure measured during or after an event can create reverse temporal ordering. This is especially important when symptoms prompt treatment interruption, dose reduction, additional monitoring, or discontinuation. An exposure value recorded after symptoms begin may reflect the response to the event process rather than the exposure that preceded it.
For a simple descriptive comparison, an event risk can be calculated as the number of people with an event divided by the number of people in the corresponding at-risk group. The risk ratio is the risk in one group divided by the risk in a reference group. These summaries describe the chosen windows; by themselves, they do not establish that exposure caused the events. A causal interpretation additionally depends on appropriate design and control of relevant differences between exposure groups.
Windows for Acute, Delayed, and Cumulative Events
| Expected pattern | Exposure summary to consider | Risk-window question |
|---|---|---|
| Acute | Recent exposure immediately before the event-risk period, such as a concentration or exposure summary from the preceding hours | When could a rapid effect plausibly begin, and how long after exposure could it persist? |
| Delayed | Exposure from an earlier period, separated from the risk window by a biologically plausible lag | How long after the relevant exposure could the event first become observable? |
| Cumulative | Exposure accumulated over a prespecified lookback period, or a history that gives appropriate weight to prior exposure | What duration of accumulated exposure could plausibly contribute to risk, and when does that risk period begin? |
These are organizing patterns rather than universal recipes. The right duration depends on the drug, event, dosing schedule, pharmacology, event-onset process, and available measurement frequency. A rapidly developing event may require hourly timing, whereas a delayed event may be assessed over days or longer. A cumulative measure may be an area under the concentration–time curve over a fixed interval, a sum of daily exposure summaries, or another prespecified quantity. The metric and its units must be stated: a concentration, a recent maximum concentration, and accumulated exposure are not interchangeable.
For a delayed event, define whether the lag is fixed or whether the analysis considers a range of plausible lags. For a cumulative event, define both the length of the lookback and the way exposure is combined. A long lookback can dilute a short, relevant period; a short lookback can omit exposure that contributes to a slowly developing process. If event onset is known only to lie between visits, the risk-window boundary is uncertain too. That uncertainty should be described rather than treated as an exact onset time.
Worked Example: An Acute Event and a One-Day Exposure Window
Suppose a hypothetical safety analysis follows 200 patient-days. For each day, exposure is classified using the drug exposure summary from the preceding 24 hours, before that day’s risk period begins. Among 100 high-exposure patient-days, 12 include an event; among 100 low-exposure patient-days, 5 include an event. Assume event onset is recorded accurately enough to assign each event to the correct day, and each patient-day contributes only once to this comparison.
The event risks are \(12/100=0.12\) in the high-exposure group and \(5/100=0.05\) in the low-exposure group. The descriptive risk ratio is
Thus, in these hypothetical data, the event risk in the high-exposure group is 2.40 times the risk in the low-exposure group over the defined daily risk periods. The interpretation depends on the alignment: exposure is taken from before the day’s risk window, and events are assigned to that window. This ratio is descriptive; it does not show that high exposure caused the difference.
The timing assumption is that the exposure summary from the preceding 24 hours is relevant to an event that starts during the following day. If the event can begin rapidly but its onset is recorded only at a weekly visit, this daily classification may be unjustified. The analysis should then acknowledge the coarser timing or use a window compatible with the available observations.
Worked Example: A Delayed Risk Window and Reversed Association
Consider 200 hypothetical at-risk intervals. The proposed biology suggests that exposure during days 8–14 before an interval could contribute to events during that interval. In the aligned comparison, 15 of 100 intervals with high exposure in that earlier window have an event, compared with 5 of 100 intervals with low exposure. Assume the exposure classification precedes the risk interval, the event is counted once, and both groups use the same risk-window length.
The aligned risk ratio is
Now suppose symptoms can prompt clinicians to hold or reduce treatment before the event is formally recorded. If intervals are instead grouped using exposure measured at the time of the event-risk period, the same hypothetical intervals could be summarized as 3 events among 100 currently high-exposure intervals and 17 among 100 currently low-exposure intervals. That mismatched comparison gives
The contemporaneous summary points in the opposite direction: the event risk appears lower in the currently high-exposure group. The reversal is plausible because exposure measured during the risk period may already have been reduced in response to early symptoms. It does not prove that this mechanism occurred; it shows why the temporal definition must be justified. The aligned ratio of 3.00 is still an association, not proof of causation.
Worked Example: Defining a Cumulative Exposure Window
In a hypothetical analysis, exposure is summarized as daily area under the concentration–time curve, in mg·h/L, for each of the 14 days before a risk period. For one patient, the daily values are 80 mg·h/L on each day. The 14-day cumulative exposure is
Suppose 50 intervals have cumulative exposure at or above a prespecified threshold of 1000 mg·h/L, with 12 events, and 50 intervals are below the threshold, with 4 events. The descriptive risk ratio is
For this hypothetical comparison, the event risk is three times as high in the group above the stated threshold. The result is interpretable only with the window and metric attached: it compares event risks after a 14-day cumulative AUC classification using a 1000 mg·h/L cut point. The cut point is an analysis choice here, not a generally valid safety threshold. The cumulative summary must also end before the risk period; including exposure accumulated after an event begins would mix cause-relevant history with post-event treatment.
Why Mismatched Windows Distort Associations
A mismatch can weaken an association even when the exposure is measured without error. If an effect depends on exposure in a specific earlier period, classifying people by a much longer or later period can place relevant and irrelevant histories in the same group. The contrast between groups then no longer represents the intended exposure contrast. This can make a real temporal pattern harder to see; the direction and size of the distortion are not guaranteed.
A mismatch can also reverse an observed association. The delayed example illustrates one route: an early symptom or event process leads to a dose hold, so exposure measured at the time of recorded onset is lower among some people who experience an event. Other routes include assigning events to an exposure period that begins after their true onset, or comparing periods with different risk durations. These problems concern temporal ordering and classification. They are distinct from the exposure-estimation uncertainty discussed in the earlier material on measurement error in individual exposure estimates, although both issues can occur together.
Choosing a window after inspecting which one gives the strongest association is also hazardous. The result can become dependent on selective window choice. A defensible primary window should be clinically motivated and specified before examining the association where possible. Alternative windows can be useful as sensitivity analyses, but should be labelled as such and interpreted as checks on the assumptions, not as an invitation to report only the most striking result.
Diagnostics, Alternatives, and Limitations
Review the timeline as carefully as the model output. Plot or tabulate exposure and event timing around event onset, including dose changes and treatment interruptions. Check whether events cluster near window boundaries, whether the chosen lookback covers the proposed biological period, and whether the amount of observation time is comparable across groups. Verify that exposure estimates are available before the risk period and that event dates are not being inferred more precisely than the data allow. The earlier material on propagating pharmacokinetic uncertainty and on individual exposure measurement error is relevant when exposure histories themselves are uncertain.
A fixed window is straightforward to communicate and can be well suited to a specific biological hypothesis. Its limitation is that it may depend heavily on a boundary choice. A small change in the lag or duration can move events or exposure summaries across that boundary. A set of prespecified alternative windows can show whether the qualitative conclusion is robust, but each comparison should be interpreted in light of its own definition.
When a single fixed interval is not a credible summary of the biological timing, one alternative is to describe exposure over several prior intervals rather than compressing all prior history into one number. This can help examine whether associations are concentrated at particular lags, but it requires adequate data and careful interpretation; it does not remove confounding or repair inaccurate event times. A cumulative summary is another alternative when total prior exposure is the scientific question, but it can hide whether recent and remote exposure have different relationships with risk. No windowing strategy can compensate for inadequate follow-up, incomplete exposure records, or unrecorded factors that influence both exposure and safety outcomes.
Define the event and the time its risk period begins, using the precision supported by the data.
Choose an exposure metric and lookback interval consistent with the event’s expected timing; state any lag explicitly.
Confirm that exposure is measured before the risk period and that intervals included in the comparison are genuinely at risk.
Use clinically defensible alternative windows to identify boundary sensitivity, and report the definitions alongside each result.
A useful practical rule is to write one sentence that states what past exposure is being related to what future event period. If that sentence is ambiguous, the analysis window is not yet sufficiently defined. The window should also be treated as part of the estimand: changing it changes the question being answered, not just the calculation.
Check Your Understanding
Answer each question using the time ordering and definitions in the tutorial. Solutions follow the questions.
- An event is expected to occur within hours of a concentration peak. Should exposure be summarized using a long-term cumulative measure or a recent pre-event measure? What ordering must be preserved?
- A delayed event is hypothesized to follow exposure by 10 days. State one possible exposure window and risk window that preserve this lag.
- In a hypothetical analysis, 8 events occur among 40 high-exposure intervals and 4 among 40 low-exposure intervals. Calculate the descriptive risk ratio for the specified windows.
- Why can exposure measured on the event date suggest an association in the opposite direction from exposure measured before the risk period?
Solutions. (1) A recent pre-event measure is more directly aligned with the proposed acute effect. It must be measured before the event-risk period; exposure after the event begins must not be used to explain that event. (2) For example, use exposure during days 11–17 before the risk window, leave days 1–10 as the lag, and begin the risk window after those 10 days. The exact dates should be adapted to the stated time origin. (3) The high-exposure risk is \(8/40=0.20\), and the low-exposure risk is \(4/40=0.10\), so \(RR=0.20/0.10=2.00\). The event risk is twice as high in the high-exposure group for these hypothetical windows; this is not by itself a causal conclusion. (4) Early symptoms may prompt dose reduction or interruption before the event is formally recorded. The event can then be grouped with lower exposure measured at onset even if relevant exposure was higher beforehand.