Dropout Driven by Future Event Risk

Why an Event That Has Not Happened Can Matter

A participant can leave clinical follow-up before a recorded event, yet still be at elevated risk of that event. For example, a person may stop attending scheduled visits shortly before a hospitalization. The hospitalization cannot literally cause an earlier departure, but pre-event symptoms, care-seeking elsewhere, or practical changes associated with worsening risk may already be affecting attendance. Thus, future event risk can be related to present follow-up loss even when event status is not yet observed.

This is distinct from the earlier material on dropout driven by unobserved current health. The key emphasis here is temporal: the clinical event occurs later, but the participant’s chance of experiencing it over a future period may help describe who is more likely to drop out now. A statistical association with a later event is not, by itself, evidence that the event caused dropout or that a particular mechanism is responsible.

The definition concerns a relationship between processes, not a claim that future outcomes are known to participants or investigators. A participant can leave while event-free; an event that occurs later may still be known through a registry or other source. Alternatively, it may remain unobserved. Those two data situations have different implications for whether the relationship can be directly assessed.

A Fixed-Horizon Description

To state one concrete comparison, consider people who are event-free and still followed at time \(t\). Choose a future horizon \(h>0\). Define \(A=1\) if a participant is lost to follow-up before experiencing the event and by time \(t+h\); otherwise \(A=0\). Define \(B=1\) if the participant experiences the event by \(t+h\); otherwise \(B=0\). In this definition, event time is the participant’s event time whether or not the event is recorded by the study after dropout.

Within a specified observed-history stratum \(H(t)\), one fixed-horizon comparison is \(P(A=1\mid B=1,H(t),\text{event-free and followed at }t)\) against \(P(A=1\mid B=0,H(t),\text{event-free and followed at }t)\). The history \(H(t)\) represents information recorded by the landmark, such as prior measurements and visit history. The comparison asks whether dropout before the event by the horizon is more common among those who experience the event by that horizon than among those who do not.

The time ordering matters. Someone may drop out first and then experience the event within the horizon; someone else may experience the event first, in which case they are not counted as having dropped out before the event. If the study can no longer learn about events after dropout, future event status may be missing precisely for the people whose dropout is of interest. A fixed-horizon comparison then cannot be calculated from ordinary follow-up records alone without additional information or assumptions.

Worked Examples

Why the Fixed-Horizon Comparison Is Limited

The relationship between a particular pair of binary indicators is narrower than the relationship between the underlying times. To see this, work within one fixed stratum of observed history at \(t\), and write \(q_1=P(A=1\mid B=1)\), \(q_0=P(A=1\mid B=0)\), and \(p=P(B=1)\). If \(0<p<1\), then

Indeed, \(E(AB)=p q_1\), while \(E(A)=p q_1+(1-p)q_0\) and \(E(B)=p\). Substitution gives \(p q_1-p[p q_1+(1-p)q_0]=p(1-p)(q_1-q_0)\). Therefore, if \(q_1\ne q_0\), the two binary indicators are dependent in that history stratum. If \(q_1=q_0\), these two indicators are independent there, because for binary variables equality of the conditional probabilities implies the joint probabilities factorize.

That last conclusion applies only to \(A\) and \(B\) at this horizon. It does not imply that the exact dropout time and exact event time are conditionally independent. Dropout could be associated with whether an event happens early rather than late within the horizon, or with event timing outside the chosen horizon, even when the binary comparison shows equal probabilities. A different horizon or a time-based analysis can reveal patterns that this single comparison misses.

Assumptions, Diagnostics, and Alternatives

The fixed-horizon description is meaningful when the landmark risk set and horizon are defined consistently, dropout is distinguished from the event, and event times are measured on a common time scale. Interpreting the comparison as one based on actual future event status also requires event ascertainment after clinic dropout, or another defensible source of that information. If event status is unavailable after dropout, classifying a participant as event-free through the horizon just because no event was recorded would be incorrect.

Useful checks include comparing event dates with last-contact dates, confirming whether registry linkage continues after clinic follow-up ends, and reviewing how “lost to follow-up” was distinguished from withdrawal, transfer of care, or administrative end of study. Repeat the descriptive comparison at clinically meaningful horizons when appropriate; the pattern can depend on the horizon. Also report how many people have unknown post-dropout event status. A large or risk-related unknown group can make the apparent pattern hard to interpret.

The main alternatives address different questions. Independent dropout is a benchmark in which loss does not depend on future event timing after the relevant information is accounted for. Dropout driven by unobserved current health emphasizes current health not captured in the recorded history. This tutorial’s focus is a relationship between dropout and a later event or future event risk. These descriptions can overlap; a cohort may exhibit more than one pattern. None should be inferred solely from a difference in observed dropout percentages without considering event ascertainment, timing, and the information used to define risk groups.

A common mistake is to say, “The later hospitalization caused the earlier dropout,” because patients who later had a hospitalization dropped out more often. That reverses temporal order and confuses prediction with causation. The corrected statement is that dropout before the event was associated with event occurrence within the specified horizon in this comparison. Establishing a causal mechanism requires more than this association.

A second mistake is to treat equal dropout proportions across future-event groups at one horizon as proof that dropout and event times are independent. Equality says only that the selected binary indicators have equal conditional dropout probabilities at that horizon. It does not rule out associations involving exact timing, other horizons, or other aspects of the event and dropout processes.