Independent Dropout as a Benchmark

Why Use Independent Dropout as a Benchmark?

When participants stop attending visits, later outcomes are missing for them. A useful starting point is to ask what would happen if loss to follow-up were unrelated to participants’ longitudinal outcomes. Under that benchmark, those still observed at a visit provide an unbiased representation of the outcomes of the cohort at that visit, although fewer observations mean less precision. The benchmark is simple, but it is an assumption about why outcomes are missing, not a fact guaranteed by the data.

The earlier material on informative follow-up loss motivates the concern: if people with different health trajectories have different chances of remaining under observation, the observed trajectories can differ systematically from the cohort trajectories. Independent dropout provides a clear comparison case. It lets us separate the effect of having fewer observations from the additional distortion that can occur when observation depends on outcomes.

Definition: Let \(Y_i(t)\) be participant \(i\)'s longitudinal outcome at time \(t\), and let \(D_i\) be the time of loss to follow-up. Define \(R_i(t)=1\) if participant \(i\) is still under observation at \(t\), and \(R_i(t)=0\) otherwise. Marginal independent dropout means \(R_i(t)\) is independent of the participant’s outcome trajectory \(\{Y_i(s):s\text{ in follow-up}\}\) at each relevant time. A conditional version means this independence holds after conditioning on stated baseline covariates \(X_i\).

For dropout defined by a single loss time, observation is typically monotone: a participant who has left follow-up does not return for later measurements, so \(R_i(t)\) can change from 1 to 0 but not back to 1. The assumption concerns loss to follow-up, not a terminal event that makes later outcomes undefined. It also does not say that all participants have the same outcomes or that everyone remains observed.

A key implication at a fixed visit is

E\{Y_i(t)\mid R_i(t)=1\}=E\{Y_i(t)\},

provided the marginal independence assumption holds and the relevant means exist. In words, the mean outcome among participants observed at that visit equals the cohort mean at that visit in the population model. For the conditional version, the corresponding statement is

E\{Y_i(t)\mid R_i(t)=1,X_i=x\}=E\{Y_i(t)\mid X_i=x\}.

This conditional statement does not by itself say that the unadjusted observed mean equals the unadjusted cohort mean. If observation rates differ across baseline-covariate groups, the mix of groups among those observed may change.

What the Mean-Preservation Result Says

The central implication follows directly from independence. If \(R_i(t)\) is independent of \(Y_i(t)\), then conditioning on \(R_i(t)=1\) does not change the distribution of \(Y_i(t)\). In particular,

E\{Y_i(t)\mid R_i(t)=1\}=E\{Y_i(t)\}.

This is a population statement about the expected outcome among those observed. In a particular sample, the retained mean will generally not equal the cohort mean exactly: random retention selects a particular subset, and that subset can be atypical by chance. The assumption rules out systematic selection based on outcomes; it does not eliminate sampling variation.

There is also an exact finite-cohort version when every participant has the same independent probability \(q\) of being retained. Suppose the cohort has fixed outcomes \(y_1,\ldots,y_n\), and each participant’s retention indicator is independently Bernoulli with probability \(q\). Conditional on exactly \(m>0\) participants being retained, every subset of size \(m\) is equally likely. Each participant therefore has probability \(m/n\) of appearing in the retained subset. The expected retained mean, conditional on \(m\), is

E\left\{\frac{\sum_{i=1}^n R_i y_i}{m}\,\middle|\,\sum_{i=1}^n R_i=m\right\} =\frac{1}{m}\sum_{i=1}^n y_i\frac{m}{n} =\frac{1}{n}\sum_{i=1}^n y_i.

Averaging over all possible positive values of \(m\) gives the same result, conditional on at least one person being retained. This exact result relies on equal retention probabilities and outcome-independent selection. When retention probabilities differ across groups, a ratio formed from expected group counts is generally a large-sample retained mean, not automatically the exact expectation of a random retained sample mean.

Worked Example: Equal-Probability Retention

Consider a hypothetical cohort of four participants with fixed biomarker values at a visit of 2, 4, 8, and 10 units. Each participant independently has probability \(q=0.75\) of still being observed. Check what the independent-dropout benchmark implies for the retained mean.

The cohort mean is

\bar{y}_{\text{cohort}}=\frac{2+4+8+10}{4}=\frac{24}{4}=6\text{ units}.

The expected number retained is \(4(0.75)=3\), and the expected retained outcome total is \(0.75(2+4+8+10)=18\) units. Their ratio is \(18/3=6\) units. More importantly, the finite-cohort argument above shows that the exact expected retained mean, conditional on at least one participant being retained, is also 6 units: conditional on any positive retained count, the retained subset is a uniformly selected subset. A realized retained mean need not be 6; for example, retaining only the participant with value 10 gives a retained mean of 10 units. Such differences are chance variation, not evidence by themselves that dropout depends on outcomes.

When Independence Holds Only Within Baseline Groups

Sometimes the benchmark is plausible only after accounting for baseline characteristics. For example, age may be related both to the chance of remaining in follow-up and to the outcome level. In that case, outcomes can be independent of retention within age groups while the overall retained group has a different age composition from the original cohort. The conditional independence statement protects comparisons within groups; an unadjusted comparison can still reflect the changed composition.

Worked Example: Different Retention Rates Across Baseline Groups

Suppose a hypothetical cohort has 100 participants in each of two baseline groups. The mean outcome is 4 units in the first group and 8 units in the second. Assume retention is independent of each participant’s outcome within their group, but the retention probabilities are 0.90 and 0.50, respectively. What does the large-sample retained mean imply?

The cohort mean, with equal group sizes, is

\frac{100(4)+100(8)}{200}=\frac{1200}{200}=6\text{ units}.

Expected retained counts are \(100(0.90)=90\) and \(100(0.50)=50\). The corresponding expected outcome totals are \(90(4)=360\) and \(50(8)=400\) units. The mean formed from these expected counts and totals is

\frac{360+400}{90+50}=\frac{760}{140}=5.429\text{ units}.

This is the large-sample limiting mean among retained participants under the stated group-specific retention rates; it is not an exact finite-sample expectation of the random retained mean. Within each group, retention does not select on outcome, so the retained group means remain 4 and 8 units in expectation. The overall difference arises because the retained group contains a larger share of the first group: \(90/140=0.643\), compared with \(100/200=0.500\) in the cohort. Thus “independent within baseline groups” should not be casually replaced by “the unadjusted retained sample represents the cohort.”

Independent Dropout Reduces Precision

Even when dropout does not distort the mean, it removes information. Consider a hypothetical longitudinal study with 100 participants and a biomarker whose cohort mean follows \(10+2t\) units, where \(t\) is measured in years. Suppose the outcome standard deviation at year 2 is 4 units, and dropout is independent of the outcome trajectory. If 60 participants are observed at year 2, the observed mean estimates the year-2 cohort mean of 14 units without systematic selection under the assumption. Its standard error, using the observed sample size, is \(4/\sqrt{60}\), rather than \(4/\sqrt{100}\) if all 100 were observed.

Worked Example: Attrition and the Standard Error

In a hypothetical study, 100 participants have a year-2 outcome with standard deviation 4 units. Assume the 60 participants actually observed at year 2 are an outcome-independent subset, and use the usual independent-participant standard-error calculation. Compare the standard error with complete follow-up.

With 60 observed outcomes,

SE_{\text{observed}}=\frac{4}{\sqrt{60}}=0.5164\text{ units}.

With all 100 outcomes,

SE_{\text{complete}}=\frac{4}{\sqrt{100}}=0.4000\text{ units}.

The standard-error ratio is \(0.5164/0.4000=1.291\), so this amount of attrition produces about a 29.1% larger standard error under these assumptions. This calculation treats the observed count as 60 and uses the stated standard deviation; an expected retained count is not necessarily the count in a particular study. The independent-dropout assumption supports the mean interpretation, while the reduced sample size explains the loss of precision.

Diagnostics, Alternatives, and Limits

Independent dropout is a benchmark, not something that can usually be confirmed from observed outcomes alone. Several checks can reveal reasons for concern:

Compare follow-up rates across baseline characteristics and across groups defined by observed outcomes at earlier visits.
Summarize when participants leave and, when recorded, their stated reasons for leaving. Patterns tied to worsening health deserve particular scrutiny.
Compare observed longitudinal histories of those who later leave with histories of those who remain, while remembering that such comparisons cannot reveal the outcomes that would have been measured after departure.
Check whether a proposed independent-dropout analysis changes substantially under reasonable alternative assumptions about the unobserved outcomes.

These checks are diagnostic, not proof. If those who leave have similar recorded histories to those who remain, their unobserved later outcomes could still differ. Conversely, different retention rates across observed baseline groups do not automatically rule out independence after conditioning on those groups. Independence is an assumption about the joint relationship between retention and outcomes, including outcomes not observed after dropout.

The main alternative is to use an analysis that allows dropout to be related to the outcome process, or to model the observation process using measured predictors. Such approaches require additional assumptions and information; they do not make missing outcomes directly observable. Independent dropout remains valuable as a reference point because it makes a simple prediction: less data should reduce precision, but should not systematically change the outcome distribution among those observed. If that prediction is implausible, treating the benchmark as the sole analysis can be misleading.

Common pitfall: “The expected retained count in group A is 90 and in group B is 50, so the exact expected raw retained mean is the ratio of the expected outcome total to the expected count.” This reasoning is generally incorrect because the denominator is random, and the expectation of a ratio need not equal the ratio of expectations. That ratio describes a large-sample limiting mean in the group-specific example. To claim an exact finite-sample expectation, account for the distribution of the random retained count and specify what happens when no one is retained.

Check Your Understanding

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

If retention is independent of a participant’s outcome at a visit, what equality holds between the retained-population mean and the cohort mean?
A cohort has fixed outcomes 3, 5, and 10 units, and each person is independently retained with the same probability. What is the expected retained mean conditional on at least one person being retained?
Two equally sized baseline groups have means 2 and 6 units. Their retention probabilities are 0.80 and 0.40. What is the large-sample limiting retained mean based on expected group counts? Is it the cohort mean?
Why can an analysis have an unbiased observed mean under independent dropout but still have wider uncertainty intervals?

Solutions. (1) \(E\{Y(t)\mid R(t)=1\}=E\{Y(t)\}\). (2) The cohort mean is \((3+5+10)/3=6\) units. Equal, independent retention gives a uniformly selected subset conditional on its positive size, so the exact expected retained mean is 6 units. (3) Expected retained counts are proportional to 0.80 and 0.40. The limiting mean is \(\{0.80(2)+0.40(6)\}/(0.80+0.40)=4/1.20=3.333\) units, whereas the cohort mean is \((2+6)/2=4\) units. The group-specific retention rates change the group composition. (4) Independent dropout can preserve the mean while reducing the observed sample size; a smaller sample size generally increases the standard error, as \(s/\sqrt{n}\) illustrates.