Time-Specific Delta Parameters in Longitudinal Data

Why Let the Delta Change Over Time?

In a longitudinal study, a participant may have an outcome recorded at some visits and missing at others. The MAR reference model supplies a conditional distribution for a missing outcome given the information used in that model. A delta adjustment changes that reference prediction to represent a sensitivity assumption about outcomes that were not observed. If the same additive delta is used at every visit, the assumed departure from the reference distribution is constant over time. That may be too restrictive when the concern is that unobserved outcomes deteriorate, improve, or change direction as follow-up continues.

A time-specific delta sequence assigns a separate shift to each scheduled visit. Its entries describe departures at calendar visits, not necessarily the number of visits since dropout. This distinction matters: an outcome missing at week 12 receives the week-12 shift under a calendar-time specification, whether the participant first missed week 4 or week 8. The earlier material on delta adjustment establishes the reference distribution and the sign convention used here: for an outcome where higher values are better, a negative additive delta shifts the missing-outcome assumption downward.

Definition and Assumptions

Definition: Let $Y_{it}$ denote the outcome for participant $i$ at scheduled visit $t$, and let $R_{it}=1$ indicate that it is observed and $R_{it}=0$ that it is missing. For a missing outcome at visit $t$, let $m_{0,it}$ be its conditional mean under the MAR reference model, given the information and missingness pattern used by that model. An additive time-specific delta sequence $(\delta_1,\ldots,\delta_T)$ specifies the sensitivity-adjusted conditional mean $$ E_\delta(Y_{it}\mid \text{conditioning information},R_i)=m_{0,it}+\delta_t,\qquad R_{it}=0. $$ Each $\delta_t$ is expressed in the outcome’s units. Unless explicitly indexed by pattern, the same visit-$t$ shift is used for every missingness pattern in which visit $t$ is missing.

This is a sensitivity model, not a parameter identified from the observed outcomes alone. Its interpretation depends on a reference model that gives suitable conditional predictions for the missing visits, on using a consistent outcome scale and visit schedule, and on specifying the conditioning information clearly. The delta sequence does not repair a poorly specified reference model. It also does not, by itself, determine every feature of a full conditional distribution: the definition above states the conditional-mean shift. If predictive draws are generated by translating a reference distribution, that additional distributional choice should be stated.

The additive formulation is most directly interpretable for continuous outcomes measured on a scale where differences have meaning. A delta of $-2$ means two outcome units below the reference conditional mean, not a two-percent reduction. For an outcome where lower values are better, the clinical meaning of a positive or negative delta reverses; always interpret the sign in relation to the outcome’s direction.

What a Sequence Says

The entries of $(\delta_1,\ldots,\delta_T)$ encode the assumed time course of departure from the reference model. They need not be monotone. A constant sequence represents a constant offset. Increasingly negative entries for a higher-is-better outcome represent progressively poorer assumed outcomes relative to the reference. Negative entries that move toward zero represent a departure that diminishes over time. A change from negative to positive represents a reversal in the assumed direction of departure.

These interpretations concern the difference from the corresponding reference prediction at each visit. They do not, on their own, assert that the unobserved outcome itself declines or improves in absolute terms. For example, a sequence of progressively more negative deltas could accompany reference means that are rising; the adjusted means might still rise, just less than the reference means. Therefore interpret the delta sequence alongside the reference predictions, rather than treating the delta values as a predicted outcome trajectory.

Worked Example: A Progressively More Adverse Departure

Suppose higher scores indicate better functioning, measured in points. For a particular observed history, the reference model predicts conditional means of 20 points at week 8, 18 at week 12, and 16 at week 16. The participant has these three visits missing. Consider the time-specific sequence $\delta_8=0$, $\delta_{12}=-1$, and $\delta_{16}=-3$ points.

Apply the definition separately at each missing visit:

\begin{aligned} m_{\delta,8}&=20+0=20,\\ m_{\delta,12}&=18+(-1)=17,\\ m_{\delta,16}&=16+(-3)=13. \end{aligned}

The assumed conditional means are therefore 20, 17, and 13 points. Relative to the reference model, the departure is unchanged at week 8, one point lower at week 12, and three points lower at week 16. The increasingly negative sequence encodes a worsening departure from the reference over follow-up. The adjusted means decline by 7 points from week 8 to week 16, while the reference means decline by 4 points; those changes combine reference-model time trends and delta shifts.

A Short Derivation: Why the Adjusted Mean Adds the Delta

For a location-shift implementation, imagine a reference-model predictive variable $Y_{it}^{(0)}$ for a missing outcome with conditional mean $m_{0,it}$. Define the sensitivity-adjusted variable by adding the visit-specific constant: $Y_{it}^{(\delta)}=Y_{it}^{(0)}+\delta_t$. Conditional expectation is linear, so

\begin{aligned} E(Y_{it}^{(\delta)}\mid \text{conditioning information},R_i) &=E(Y_{it}^{(0)}+\delta_t\mid \text{conditioning information},R_i)\\ &=E(Y_{it}^{(0)}\mid \text{conditioning information},R_i)+\delta_t\\ &=m_{0,it}+\delta_t. \end{aligned}

This establishes the mean-shift formula under the stated location shift. It does not establish that a location shift is clinically appropriate, nor does it show that the observed data can estimate $\delta_t$. The sequence remains a sensitivity assumption to be varied and reported.

Worked Applications

Worked Example: Comparing Worsening and Diminishing Departures

Consider a hypothetical symptom-function score, where higher values are better. For one missing-outcome history, the reference conditional means at weeks 4, 8, and 12 are 50, 48, and 47 points. Compare two sensitivity sequences: a progressively worsening departure $(-2,-4,-6)$ and a departure that diminishes over time $(-6,-4,-2)$, with entries ordered by visit.

For the first sequence, the adjusted means are

(50-2,\;48-4,\;47-6)=(48,\;44,\;41)\text{ points}.

For the second sequence, they are

(50-6,\;48-4,\;47-2)=(44,\;44,\;45)\text{ points}.

Both sequences assume outcomes below the reference at every visit, but they make different assumptions about when that departure is greatest. The first has a larger adverse shift at later visits; the second has the largest adverse shift at week 4 and approaches the reference over time. Neither sequence can be described as “the outcome worsens” solely from its delta values: the adjusted means also depend on the reference means.

Worked Example: Applying Calendar-Time Deltas to Different Patterns

Suppose visits occur at weeks 4, 8, 12, and 16, and a higher score is better. Use the common calendar-time sequence $(0,-1,-2,-3)$ points. In one hypothetical pattern, a participant is observed through week 8 and then misses weeks 12 and 16; the reference conditional means at those missing visits are 30 and 28 points. In another pattern, a participant is observed only through week 4 and misses weeks 8, 12, and 16; the corresponding reference means are 32, 30, and 28 points.

For the first pattern, only the week-12 and week-16 shifts apply:

30+(-2)=28,\qquad 28+(-3)=25\text{ points}.

For the second pattern, the missing week-8, week-12, and week-16 outcomes receive their calendar-visit shifts:

32+(-1)=31,\qquad 30+(-2)=28,\qquad 28+(-3)=25\text{ points}.

The week-16 delta is $-3$ points in both patterns because the specification indexes the shift by scheduled visit. The reference means differ because the observed histories differ. This example also shows why the sequence is not a “delta since dropout” rule: the week-12 shift remains $-2$ points even when week 12 is the first missing visit.

Choosing and Checking a Sequence

A sequence should be written down explicitly, with visit labels and units. A verbal description such as “assume outcomes get worse” is not enough to reproduce the analysis. Before using a sequence, check the following:

Direction and sign: Confirm which outcome direction is favorable. For a higher-is-better score, negative shifts lower the assumed missing outcomes.
Visit indexing: State whether the index refers to scheduled calendar visits or time since a participant’s dropout. The definition in this tutorial uses scheduled visits.
Scale and units: Ensure each entry is on the outcome scale used by the reference model. A shift measured in points cannot be added to a reference mean recorded in a different scale without a stated conversion.
Pattern coverage: Verify which visits are missing in each pattern and apply a delta only to those visits. Do not alter observed outcomes through this missing-outcome adjustment.
Resulting values: Compare adjusted predictions with plausible outcome values and the clinical meaning of the scale. Large shifts can yield implausible predictions, especially on bounded scales.
Sensitivity to choices: Examine conclusions under more than one scientifically defensible sequence. A single selected sequence does not reveal how dependent a conclusion is on the unobserved-outcome assumptions.

A practical diagnostic is to tabulate, by visit and relevant missingness pattern, the reference conditional means, the delta entries, and the adjusted means. This makes sign mistakes and accidental use of the wrong visit’s delta visible. If adjusted values are implausible, first check the model inputs, scale, units, visit mapping, and sign convention. A plausible-looking table does not validate the unobserved outcomes: the delta itself is not empirically checked by observed outcomes at visits where the outcomes are missing.

How This Specification Differs From Alternatives

Specification	What varies	Interpretation
One common delta	No variation by visit	A constant departure from the reference at every missing visit
Pattern-specific delta	Missingness pattern	Different departures for different observed/missing visit configurations
Time-specific delta	Scheduled visit	Different departures at different points in follow-up, shared across patterns unless further indexed
Reference model without delta	No sensitivity shift	The MAR reference distribution is used for missing outcomes

Pattern-specific and time-specific indexing answer different questions. A pattern-specific parameter distinguishes participants by their observed/missing configuration; a time-specific parameter distinguishes visits. They can be combined as $\delta_{r,t}$, but doing so introduces more assumptions and parameters than the common-across-pattern sequence defined here. The key comparison is not that one indexing scheme is universally preferable. Rather, each encodes a different proposed reason for missing outcomes to depart from the reference.

The main limitation is fundamental: outcomes that are missing cannot reveal their own delta values. A time-specific sequence makes a temporal assumption explicit, but does not make it identifiable. It can also be overly simple if the same visit-specific shift is implausible across distinct missingness patterns. Conversely, a richly varying sequence may be difficult to justify and may obscure which assumption drives the result. Report the visit labels, values, units, sign interpretation, and whether the shifts are shared across patterns.

Common Mistakes

Using a single delta while describing a time-varying assumption. One number cannot encode progressively larger departures by visit. Specify the full ordered sequence.
Confusing calendar time with dropout duration. A delta for week 12 is not automatically the shift for “the second missed visit.” State the indexing rule.
Reading the delta as the adjusted outcome. The adjusted conditional mean is the reference mean plus delta; the delta alone is only the departure.
Assuming a negative delta always means worse. That is true only when higher outcome values are favorable. Check the outcome’s direction.
Treating the sequence as estimated from observed data. The observed data inform the reference model, but do not identify the missing-outcome departures represented by the deltas.

Key takeaway: A time-specific delta sequence assigns an explicit departure from the reference conditional mean to each scheduled visit. Its direction and progression encode the sensitivity assumption; they are not learned from outcomes that are missing.

Check Your Understanding

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

A higher-is-better outcome has a reference conditional mean of 42 points at week 10 and a specified $\delta_{10}=-3$ points. What is the adjusted conditional mean, and how should the shift be interpreted?
For visits 1, 2, and 3, explain the difference between the sequences $(-1,-2,-4)$ and $(-4,-2,-1)$, assuming higher values are better.
A participant misses weeks 8 and 12. Under a calendar-time sequence, should the week-12 delta depend on whether week 8 was the participant’s first missing visit? Explain.
Why does a time-specific delta sequence not establish that the unobserved outcomes actually follow the assumed trajectory?

Solutions. (1) The adjusted conditional mean is $42+(-3)=39$ points, three points below the reference prediction. (2) Both sequences are negative at all visits. The first encodes an increasingly adverse departure over visits; the second encodes an adverse departure that diminishes over visits. These describe departures from the reference, not necessarily the absolute outcome trend. (3) No. Under calendar-time indexing, week 12 receives $\delta_{12}$ in either case; the reference prediction may still differ with the conditioning history. (4) The delta concerns outcomes that are missing and therefore cannot be directly checked against their observed values. It is a sensitivity assumption, not an estimate identified by those missing outcomes.