Sign Conventions for Missing-Outcome Shifts

Why the Sign Needs Context

An additive delta is a numerical shift applied to the reference distribution for missing outcomes in a pattern-mixture sensitivity analysis. The sign tells you whether the shifted values are numerically higher or lower than that reference. It does not, by itself, say whether the assumption is favorable, unfavorable, optimistic, or pessimistic. That clinical interpretation depends on what the outcome measures and how it is coded.

For example, a positive shift in systolic blood pressure increases the assumed blood pressure, which is usually less favorable when lower blood pressure is the desired outcome. A positive shift in symptom-free days increases a measure for which higher values are usually favorable. The same sign can therefore describe different clinical directions across outcomes.

The sign convention is defined on the numerical scale, not on a universal scale of clinical desirability. Before calling a delta assumption favorable or unfavorable, state whether higher or lower outcome values are desirable. Also state which treatment arm and missing-outcome group the assumed shift concerns; a delta describes an assumption about missing values, not a change to outcomes that were observed.

What the Shift Does to a Group Mean

Suppose a group has \(n\) participants, of whom \(m\) have a missing endpoint. Keep the observed outcomes fixed. If the MAR reference mean among the missing participants is \(\bar{Y}^{\mathrm{MAR}}_{\mathrm{mis}}\), applying a common additive delta gives a missing-group mean of \(\bar{Y}^{\mathrm{MAR}}_{\mathrm{mis}}+\delta\). The full-group mean changes by \(m\delta/n\), not by \(\delta\), because only the missing fraction receives the shift.

This identity follows by writing each full-group mean as the sum of the fixed observed-outcome total and the imputed missing-outcome total, divided by \(n\). The adjusted missing total is \(m(\bar{Y}^{\mathrm{MAR}}_{\mathrm{mis}}+\delta)\), which exceeds the MAR missing total by \(m\delta\). Dividing that difference by \(n\) gives the displayed result. It is an arithmetic consequence of the common shift; it does not establish that a particular delta is clinically plausible.

When Higher Values Are Favorable

For an endpoint such as symptom-free days, it may be more favorable to have a higher value. A negative delta then lowers the missing outcomes relative to MAR and is a less favorable assumption for that group. It remains a negative numerical shift; the clinical label follows from the endpoint's meaning.

Sign Changes When Outcome Coding Changes

Clinical records can describe the same underlying condition using differently oriented scales. Suppose one scale records symptom severity, so higher values mean worse symptoms, and another records symptom relief, so higher values mean greater relief. A shift that worsens symptoms has opposite numerical signs on these two scales. Sign interpretation must therefore start with the recorded endpoint, not with a verbal label such as “worse.”

Assumptions, Checks, and Common Errors

A delta analysis is a sensitivity analysis: observed outcomes and the reference model are held fixed while the analyst examines alternative assumptions for outcomes that are missing. The chosen delta is not generally identified from observed data alone. Interpreting it correctly therefore requires the following conditions and checks:

• State the direction of the endpoint. Specify what a larger recorded value means and which direction is clinically preferred. This is essential before labeling a delta assumption favorable or unfavorable.
• Apply the shift only where intended. The additive adjustment changes the reference values for the specified missing outcomes. It does not alter observed outcomes or automatically apply to every treatment group.
• Keep units consistent. A delta in mmHg must be added to a blood-pressure reference in mmHg; a delta in days must be added to a reference in days. Its magnitude should be interpreted on that endpoint's scale.
• Check the reference and adjusted values. Compare their ranges and summaries with plausible values for the endpoint. Even a correctly signed shift can produce implausible values if its size is unsuitable.
• Check the estimand's contrast direction. If the reported contrast is treatment minus control, determine which sign favors treatment for that endpoint. A positive delta increases the treatment-minus-control contrast if applied to treatment and decreases it if applied to control. Whether that numerical movement favors treatment depends on the endpoint direction.

For a treatment-minus-control contrast, if only the treatment group's missing outcomes receive a common delta and its missing fraction is \(m_T/n_T\), the contrast changes by \((m_T/n_T)\delta\), with the control group held fixed. If only the control group's missing outcomes receive the delta, the contrast changes by \(-(m_C/n_C)\delta\). These signs follow by subtracting the control mean from the treatment mean. They are not universal benefit signs: whether a larger contrast favors treatment still depends on whether higher values are desirable.

A frequent mistake is to say “positive delta means worse outcomes.” That is wrong as a general rule: positive delta means higher numerical values than the reference. The correction is to name the endpoint orientation. For blood pressure where lower is preferred, a positive delta shifts toward less favorable values; for symptom-free days where higher is preferred, it shifts toward more favorable values.

Another mistake is to describe a delta as though it were estimated from the observed outcomes or as though it changed the observed data. The sensitivity parameter specifies an assumption about missing outcomes relative to a reference distribution. Since multiple values may be compatible with the observed data, results across a justified range of deltas show how conclusions depend on those unobserved-outcome assumptions; they do not prove which delta is true.

The main alternative for this particular reference is the MAR analysis, represented by \(\delta=0\). It supplies a clear comparison point, but does not establish that missing outcomes actually follow the MAR reference. Reference-based imputation uses a different specified reference for missing outcomes; a delta adjustment instead expresses departure from its chosen MAR reference through a numerical shift. Neither sign nor direction can be compared meaningfully without identifying the outcome scale and the reference being shifted.

An additive shift also has limitations. Its clinical interpretation depends on the scale, and a common shift may be an oversimplification if the plausible departure differs across people or circumstances. For a bounded outcome, shifting values can create values outside the permitted range unless the model or implementation addresses that problem. Inspect the adjusted distribution, not only its mean, and do not infer plausibility from a favorable or unfavorable sign alone.