Monotone Dropout Histories in Longitudinal Studies

Why the Shape of a Missingness History Matters

In a longitudinal study, the order in which outcomes are observed can carry useful structural information. A participant who provides outcomes at the first two visits and then leaves follow-up has a different observation history from one who misses the second visit but returns at the third. The first history is a dropout pattern; the second is intermittent nonresponse. Distinguishing them is necessary before describing the missing-data patterns or choosing a model for them.

This tutorial considers a fixed schedule of \(K\) planned visits. Its purpose is to describe and encode observation histories—not to explain why participants drop out or to specify a model for the dropout process. A monotone pattern is a property of the observed-versus-missing record, not evidence by itself about whether missingness is ignorable or depends on unobserved outcomes.

For a monotone history, the observed visits form an initial segment of the schedule. Define \(D_i\) as the number of consecutive observed visits from the start of follow-up. Thus \(D_i=0\) means no scheduled outcomes were observed, \(D_i=d\) for \(1\leq d<K\) means the first \(d\) outcomes were observed and all later ones were missing, and \(D_i=K\) means all scheduled outcomes were observed. Under this convention, \(D_i<K\) describes a history with incomplete follow-up; \(D_i=K\) is complete follow-up, not dropout before a scheduled visit.

This convention assumes a defined cohort of enrolled subjects and a common set of planned visits. If the first planned measurement is at baseline, \(D_i=0\) records that none of the scheduled outcomes was observed. It does not, by itself, tell us whether the subject withdrew before baseline, was not measured for another reason, or had an outcome that was not recorded. Those circumstances require study-specific information.

Encoding Histories and Checking Monotonicity

An observation mask lists the status of each planned visit in order. For example, with four planned visits, \((1,1,0,0)\) records two observed outcomes followed by two missing outcomes, so \(D_i=2\). In contrast, \((1,0,1,0)\) is not monotone: visit 3 is observed after visit 2 is missing. That history contains intermittent nonresponse and cannot be represented by a single last-observed-visit value without losing information.

Why a Last-Observed-Visit Label Is Enough

For a monotone history, \(D_i\) and the full mask contain exactly the same information about which visits were observed. To see this, take any monotone mask and let \(d\) be the number of initial ones. If \(d=0\), the first visit is missing, and monotonicity forces every later visit to be missing. If \(0<d<K\), the first \(d\) visits are observed; the next visit is missing, and monotonicity forces all subsequent visits to be missing. If \(d=K\), every visit is observed. In each case the mask must have the form \((1,\ldots,1,0,\ldots,0)\), with exactly \(d\) ones. Thus \(D_i\) uniquely determines the mask, and the mask uniquely determines \(D_i\).

The count includes the all-missing history and the complete history. If a study’s definition of the analysis cohort excludes subjects with no observed outcomes, the patterns represented in that cohort may instead have \(D_i\in\{1,\ldots,K\}\), giving \(K\) patterns. The cohort rule changes which patterns are present; it does not change the definition of monotonicity.

What the Pattern Does—and Does Not—Establish

A monotone pattern is useful because it reduces a potentially long visit-by-visit record to a single label, \(D_i\). It can help organize outcome histories, tabulate how many subjects are complete at each visit, and describe which later outcomes are unavailable after an observed prefix. In pattern-mixture terminology, the label can identify a group with a particular observation pattern. That descriptive grouping does not supply the missing outcome values or determine their distribution.

In particular, observing that a participant has \(D_i=2\) does not show whether the probability of leaving follow-up was related to the participant’s health, to an unobserved future outcome, or to other factors. The same observed mask could arise under different missingness mechanisms. Identification and the assumptions used to extrapolate beyond observed data require separate arguments, as emphasized in the earlier material on identification.

For each record, monotonicity can be checked by scanning adjacent visits: a zero followed later by a one is a violation. If the intended study process permits participants to miss visits and return, those histories should not be forced into a single-dropout representation. If the intended process is permanent withdrawal but the database shows later outcomes, the discrepancy may indicate a coding, timing, or data-linkage issue. Conversely, a string of missing outcomes is not enough to prove that a participant formally withdrew; a reason for missingness must come from additional records or a clearly stated operational definition.

Common Mistakes and Practical Limits

• Calling every missing value dropout. One absent visit followed by a recorded outcome is intermittent nonresponse, not a monotone dropout history.
• Assuming monotonicity reveals the missingness mechanism. A mask records which outcomes are observed, not why they are missing or how the unobserved outcomes are distributed.
• Using only \(D_i\) when histories are nonmonotone. A single last-observed visit cannot represent a participant who later returns. Preserve the full visit-level mask in that case.
• Confusing complete follow-up with dropout. Under this tutorial’s convention, \(D_i=K\) is complete follow-up; an incomplete history has \(D_i<K\).
• Ignoring visit definitions and data quality. The encoding assumes visits are ordered and their observation status is known. Unclear windows, delayed measurements, or inconsistent record linkage can create apparent violations that need investigation.

The main alternatives are to retain the full observation mask for arbitrary missingness patterns or to use the last-observed-visit label when the patterns are monotone. The first representation is more general and preserves intermittent returns; the second is more compact and organizes subjects by dropout history. Neither representation is a substitute for a model of the outcome distribution or missingness process. In particular, a standard monotone pattern description should not be applied if the visit schedule differs across subjects unless the schedule and the definition of the observation history are first made explicit.