Treatment Contrasts Within Dropout Patterns

Why Separate Treatment Contrasts by Pattern?

In a longitudinal study, participants can follow different dropout patterns: some provide outcomes at every visit, while others stop contributing outcomes earlier. A treatment contrast calculated separately within each pattern describes how the treatment groups differ among participants assigned to that pattern. This can reveal whether the observed group difference is similar across patterns, but it is not automatically the treatment effect for the full study population.

The distinction matters because dropout pattern is typically determined after treatment assignment and may itself be related to treatment and outcomes. Comparing treatment groups after conditioning on a post-assignment pattern is therefore a conditional comparison, not generally a randomized comparison. The introductory material on pattern weights and the population mixture established how pattern-specific distributions contribute to population summaries; here the focus is the treatment contrast within each pattern.

Definition: Let $A=1$ denote assignment to treatment and $A=0$ assignment to control. Let $G=g$ denote membership in dropout pattern $g$, and let $Y_t$ be the outcome at visit $t$. When $Y_t$ is observed in pattern $g$, the pattern-specific treatment contrast is $$ \Delta_g(t)=E(Y_t\mid A=1,G=g)-E(Y_t\mid A=0,G=g). $$ Its direct sample estimate is the treated-group mean minus the control-group mean among participants in that pattern with observed $Y_t$. The contrast is expressed in the units of $Y_t$.

The definition specifies a conditional mean difference. It does not, by itself, specify a causal estimand. When a pattern does not include visit $t$, its participants do not have observed $Y_t$; a pattern-mixture model may provide a model-implied pattern-specific mean, but that value depends on the model and its assumptions. Keep observed pattern-specific summaries distinct from model-implied quantities for missing outcomes.

Estimating a Contrast Within One Pattern

Suppose $n_{1g}$ and $n_{0g}$ participants in pattern $g$ have observed outcomes at visit $t$, with sample means $\bar{Y}_{1g,t}$ and $\bar{Y}_{0g,t}$. The unadjusted estimate is

\widehat{\Delta}_g(t)=\bar{Y}_{1g,t}-\bar{Y}_{0g,t}.

A positive estimate means that the observed mean outcome is higher in the treatment group than in the control group within this pattern. The units are those of the outcome. This interpretation is descriptive of the participants in the specified pattern and observed at the specified visit.

For independent participants and finite within-group sample variances $s^2_{1g,t}$ and $s^2_{0g,t}$, the estimated standard error of this difference is

SE\{\widehat{\Delta}_g(t)\} = \sqrt{\frac{s^2_{1g,t}}{n_{1g}}+\frac{s^2_{0g,t}}{n_{0g}}}.

This calculation treats observations from different participants as independent. A confidence interval based on a $t$ reference distribution additionally relies on adequate normal approximation or an appropriate small-sample model. Repeated outcomes from the same person are not independent; a calculation for several visits must account for their within-person dependence rather than treating every visit as a separate participant.

Worked Example: An Unadjusted Contrast in a Dropout Pattern

Consider hypothetical outcome scores, measured in points, at a visit observed for everyone in pattern $g$. Three treated participants have scores 8, 10, and 9; three control participants have scores 5, 7, and 6. Assume different participants are independent.

The treated mean is $\bar{Y}_{1g,t}=(8+10+9)/3=9$ points. The control mean is $\bar{Y}_{0g,t}=(5+7+6)/3=6$ points. Thus

\widehat{\Delta}_g(t)=9-6=3\text{ points}.

Each group has sample variance 1 point squared: for example, the treated squared deviations from 9 are $1,1,0$, whose sum is 2 and whose variance is $2/(3-1)=1$. Therefore

SE\{\widehat{\Delta}_g(t)\} =\sqrt{\frac{1}{3}+\frac{1}{3}} =\sqrt{\frac{2}{3}} \approx 0.8165\text{ points}.

Using a two-sided $t$ interval with $df=4$, the 95% interval is $3\pm 2.776(0.8165)$, or approximately 0.733 to 5.267 points. Under the independent-sample and $t$-model conditions, this interval is a procedure that would cover the conditional mean difference in 95% of repeated samples. It does not turn the within-pattern comparison into a population-wide causal effect.

Adjustment for a Baseline Covariate

A treatment contrast may be adjusted for a baseline covariate $X$, such as baseline outcome score or age. The earlier material on baseline-covariate regression within patterns gives the regression framework; the key point here is how to interpret the treatment coefficient. Within pattern $g$, a linear model might be

E(Y_t\mid A,G=g,X)=\beta_{0g}+\beta_{Ag}A+\beta_{Xg}(X-c),

where $c$ is a chosen centering value. Under this model, the fitted treatment contrast at any fixed $X$ is $\beta_{Ag}$: subtracting the fitted control mean from the fitted treated mean cancels the intercept and covariate term. This is a common adjusted contrast only if the treatment contrast is constant over the modeled values of $X$. If that condition is not plausible, one coefficient does not describe all covariate-specific contrasts.

An alternative way to make the adjustment concrete is to standardize both treatment groups to the same baseline-covariate distribution. This avoids comparing a treatment mean concentrated among one covariate group with a control mean concentrated among another. The chosen standardization distribution defines the adjusted comparison and should be stated.

Worked Example: Standardizing to a Shared Baseline Distribution

Suppose a hypothetical pattern has two baseline-age groups. Among younger participants, the treatment and control means at an observed visit are 12 and 10 points; among older participants, they are 8 and 7 points. Define a target baseline distribution that is 60% younger and 40% older, applied equally to both treatment groups.

The standardized treatment mean is $0.60(12)+0.40(8)=7.2+3.2=10.4$ points. The standardized control mean is $0.60(10)+0.40(7)=6.0+2.8=8.8$ points. Hence the standardized contrast is

\widehat{\Delta}^{\,std}_g(t) =10.4-8.8 =1.6\text{ points}.

Equivalently, the age-specific contrasts are $12-10=2$ points for younger participants and $8-7=1$ point for older participants; their target-weighted average is $0.60(2)+0.40(1)=1.6$ points. This calculation assumes the age-specific means are adequately estimated and that both treatment groups have support in both age groups. It describes a contrast standardized to the stated 60%/40% distribution, not necessarily the unadjusted contrast in the pattern.

From Pattern-Specific Contrasts to a Population Contrast

A pattern-specific contrast and a population-level contrast use different conditioning and, in general, different weights. Write $\pi_{ag}=P(G=g\mid A=a)$ and $\mu_{ag}(t)=E(Y_t\mid A=a,G=g)$. If pattern-specific means for $Y_t$ are defined for all patterns under consideration, the law of total expectation gives

E(Y_t\mid A=a)=\sum_g \pi_{ag}\mu_{ag}(t).

Subtracting the control marginal mean from the treatment marginal mean gives the population-level arm contrast. Because the pattern distribution can differ by arm, the two arms generally use different weights. This identity applies to model-implied means for unobserved outcomes only when the pattern-mixture model supplies those means; the observed data alone do not reveal an outcome after a participant has dropped out.

Worked Example: Why Pattern Weights Change the Overall Contrast

Consider a hypothetical endpoint with two dropout patterns. Suppose a fitted pattern-mixture model supplies endpoint means of 10 points for treatment and 8 points for control in pattern 1, and 4 points for treatment and 3 points for control in pattern 2. Suppose the estimated pattern proportions are 0.80 and 0.20 in treatment, but 0.40 and 0.60 in control.

The within-pattern contrasts are $10-8=2$ points in pattern 1 and $4-3=1$ point in pattern 2. The treatment marginal mean is $0.80(10)+0.20(4)=8.0+0.8=8.8$ points. The control marginal mean is $0.40(8)+0.60(3)=3.2+1.8=5.0$ points. The population-level contrast is therefore $8.8-5.0=3.8$ points.

The 3.8-point marginal contrast is not the simple average of the two pattern-specific contrasts. For example, weighting both conditional contrasts by the treatment pattern proportions gives $0.80(2)+0.20(1)=1.8$ points, not 3.8. The discrepancy occurs because the marginal arm means use different pattern distributions and because the pattern means differ. The model-implied endpoint means for pattern 2 also rely on assumptions about outcomes that are missing there; the arithmetic does not validate those assumptions.

A different target can be formed by applying a common set of weights $w_g$ to the pattern-specific contrasts:

\Delta^{std}(t)=\sum_g w_g\Delta_g(t),\qquad \sum_g w_g=1.

This is a standardized contrast for the chosen pattern distribution. It is not generally equal to the raw marginal arm contrast, which uses $\pi_{1g}$ for the treatment mean and $\pi_{0g}$ for the control mean. State the target weights whenever reporting a standardized summary.

Assumptions, Diagnostics, and Interpretation

For an observed within-pattern mean difference to be calculated as stated, the outcome must be measured on a comparable scale in both arms, the pattern and visit must be defined consistently, and both arms must contribute observed outcomes to the pattern. For model-based adjustment, the regression form and covariate specification must be adequate for the intended contrast. Standard errors also require an appropriate account of independent participants and any model-based uncertainty.

The most consequential interpretive issue is conditioning on $G$. Even if treatment was randomized at baseline, $G$ can be affected by treatment and by factors related to the outcome. Within a particular observed pattern, the treatment and control participants need not remain comparable as they were at randomization. For that reason, $\Delta_g(t)$ is ordinarily a conditional association among pattern members, not a causal treatment effect for a subgroup that would have had that pattern under either assignment.

Before reporting a contrast, inspect pattern-by-treatment counts, missingness at the target visit, covariate overlap, and the uncertainty of each estimate. Very small or empty cells can make sample means unstable or undefined; regression adjustment cannot create information where a treatment arm has no support. For model-implied means in patterns where the outcome is unobserved, examine sensitivity to the pattern-specific assumptions rather than presenting the output as directly observed evidence.

The main alternatives answer different questions. A within-pattern contrast describes a conditional comparison. A marginal contrast combines pattern-specific means using arm-specific pattern weights and targets the population mixture represented by the model. A common-weight standardized contrast compares arms over a specified shared pattern distribution. These summaries should not be substituted for one another without stating the change in target. Assessing whether treatment contrasts differ across patterns is a separate question; the pattern-by-treatment interaction framework is not needed to calculate the contrasts themselves.

Common pitfall: “The within-pattern contrast is 2 points, so treatment improves the outcome by 2 points for everyone in the population.” This conclusion mistakes a conditional comparison for a population-wide causal effect. The correction is to name the pattern, visit, and conditioning explicitly, then calculate a population summary with its relevant pattern weights if that is the target. Because dropout pattern can occur after treatment assignment, even the conditional contrast need not have a causal interpretation.

Key takeaway: Calculate each pattern-specific contrast as a treatment-minus-control difference within that pattern, and report its outcome units and uncertainty. Then distinguish that conditional quantity from any population-level or standardized summary, stating how pattern weights and model-implied missing outcomes enter the latter.

Check Your Understanding

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

In one pattern, the treatment mean is 14 points and the control mean is 11 points at a visit observed in both groups. What is the estimated contrast, and what does it describe?
Two patterns have contrasts of 2 and 1 points. Why is their average not automatically the population-level treatment contrast?
A pattern contains 20 treated participants but no control participants with the target outcome observed. Can the observed within-pattern mean difference be calculated?
Why does random assignment at baseline not automatically make a contrast conditional on dropout pattern causal?

Solutions. (1) The estimate is $14-11=3$ points. It describes the treatment-minus-control observed mean difference among participants in that pattern with the outcome measured at that visit. (2) A population marginal contrast combines pattern-specific means using the pattern proportions in each arm, which can differ; a simple average uses unstated weights and need not match those arm-specific mixtures. (3) No. The control mean within that pattern is unavailable, so the observed difference is undefined; a model could supply a value only by adding assumptions. (4) Dropout pattern can depend on treatment and outcome-related factors. Conditioning on it can disrupt the baseline comparability created by randomization, so the resulting comparison is generally conditional and associational rather than automatically causal.