Diabetes Trial HbA1c Missing After Adverse Events

Why Missing HbA1c After an Adverse Event Needs a Clinical Assumption

In a diabetes trial, an adverse event may lead a participant to interrupt or stop assigned treatment, miss a visit, or withdraw from follow-up. The HbA1c measurement that would have been recorded at the endpoint is then unavailable. A model can predict that value from observed information, but the prediction relies on an assumption about the unobserved outcome. In particular, an assumption that missing outcomes are like comparable observed outcomes after conditioning on recorded history may not capture the possibility that HbA1c would have been systematically higher after an adverse event.

A delta adjustment makes that uncertainty explicit. It starts from an imputation model under missing at random (MAR), then shifts the imputed HbA1c values for a specified missingness pattern by a clinically chosen amount. The adjustment does not establish what the missing HbA1c values really were. It answers a conditional question: how would the treatment contrast change if those outcomes averaged a specified amount above or below their MAR predictions?

Use the treatment-minus-control contrast throughout: \(\theta=\mu_T-\mu_C\), where \(\mu_T\) and \(\mu_C\) are the endpoint mean HbA1c values under the analysis being considered. Lower HbA1c is generally favorable, so a negative contrast favors treatment on this scale. The direction of benefit does not determine the direction of a plausible delta: an adverse event may or may not imply worse subsequent glycemic control.

Assumptions and the Contrast Shift

A practical analysis first specifies the endpoint, treatment contrast, and adverse-event pattern before choosing delta values. The MAR imputation model should use relevant observed information, such as baseline HbA1c, earlier HbA1c measurements, treatment assignment, and recorded adverse-event history, as appropriate to the trial and estimand. Within the model, the MAR assumption says that, conditional on the information included, the distribution of missing outcomes follows the distribution used to predict them from observed outcomes. This is a working identifying assumption, not something verified by the observed data.

The delta model adds further assumptions. It treats the chosen shift as appropriate for the particular missingness pattern and arm, and usually applies the same additive shift to all missing outcomes in that pattern. If the shift is added to each imputed value, it translates that pattern’s imputed distribution without changing its residual spread. A constant shift may be inadequate if the difference from MAR varies with baseline severity or time since the event; those possibilities call for a more tailored sensitivity model, not a claim that one delta is known to be correct.

For an intuitive calculation, let \(q_{g,A}\) be the fraction of all randomized participants in arm \(g\) whose endpoint is missing after an adverse event, and let \(\delta_{g,A}\) be their additive shift. If the MAR complete-data mean in arm \(g\) is \(\bar{Y}^{(\mathrm{MAR})}_g\), shifting those missing values changes the arm mean by \(q_{g,A}\delta_{g,A}\). Thus, for one such pattern in each arm, the adjusted treatment contrast is:

The subtraction for the control arm follows directly from defining the contrast as treatment minus control. With multiple patterns, add the shift contribution for each pattern in each arm, using the appropriate arm-specific missing fraction and delta. In a multiple-imputation analysis, the same adjustment is made to each imputed data set, and uncertainty is calculated by pooling estimates and their within- and between-imputation variability. The simple formula above explains the point-estimate movement; it does not replace that uncertainty calculation.

Choosing and Interpreting Delta Values

Delta is not estimated from the observed outcomes: the outcomes to which it applies are missing. Its values should therefore be justified as sensitivity assumptions, for example by clinical discussion of what might happen to HbA1c after treatment interruption, review of observed trajectories before the adverse event, and consideration of how large a difference from MAR would be clinically meaningful. Observed pre-event values can inform the MAR prediction, but they cannot reveal the unobserved post-event endpoint.

A transparent analysis reports the delta scale and direction, the affected pattern and arm, and the resulting contrast. It is often useful to evaluate a range of values, including zero as the MAR reference, rather than presenting one clinical judgment as uniquely correct. For instance, a positive delta may represent a scenario in which HbA1c after an adverse event would be worse than predicted under MAR. That scenario is not automatically conservative: its effect on the contrast depends on the fractions missing in each arm and on the control-arm assumptions as well.

A Tipping-Point Calculation

A tipping-point calculation asks which value of a sensitivity parameter changes a stated conclusion. The conclusion must be defined precisely. For example, the delta at which the point estimate reaches zero is not necessarily the delta at which a confidence interval first includes zero. The latter also depends on the uncertainty calculation at each delta.

Diagnostics, Alternatives, and Limitations

Before interpreting a delta analysis, check that the pattern is coded as intended and that the missing fractions use the correct denominators. In the examples, the denominator is the number randomized in the arm because the displayed means are complete-data means over those participants. If the estimand or analysis population differs, the denominator and weighting must match that analysis. Also verify that delta is applied only to missing outcomes in the designated adverse-event pattern—not to observed HbA1c values or to every missing value indiscriminately.

Inspect the imputed values after adjustment. Large positive shifts can produce values outside a clinically credible range or imply implausible changes from observed pre-event measurements. Such findings do not automatically invalidate a sensitivity scenario, but they signal that the constant additive shift, pattern definition, or chosen delta may not be a useful representation of clinical uncertainty. Check whether the chosen shifts are arm-specific as intended, whether all imputation runs completed, and whether the adjusted estimates move in the direction predicted by the weighted-shift formula. A mismatch can reveal a coding or denominator error. Sensitivity analyses should also report uncertainty, not just point estimates.

Delta adjustment is one pattern-mixture approach: it describes departures from MAR directly on the outcome scale for a specified missingness pattern. A selection model instead represents missingness through a model relating observation probability to the outcome, which is a different way to express sensitivity assumptions. Reference-based imputation uses assumptions tied to outcomes in a reference arm or treatment trajectory; it is not the same as choosing a constant adverse-event delta. These approaches answer related but distinct questions, and their assumptions should not be treated as interchangeable.

The main limitation is fundamental: observed data alone cannot identify the mean HbA1c among outcomes missing after adverse events without additional assumptions. A delta grid shows how conclusions depend on specified assumptions; it does not identify a correct delta. Clinical input is important, but disagreement among reasonable judgments should be shown rather than concealed. Finally, a constant shift in HbA1c may fail to represent heterogeneous consequences of an adverse event, and a pattern defined only by “after an adverse event” may combine clinically different events. Narrower patterns or varying shifts may be worth examining when the data support them, while recognizing that finer patterns can leave fewer observed data to estimate the MAR predictions.