Difference-in-Differences for Platform Control Drift

Why Compare Changes Rather Than Levels?

In a platform trial, the outcome among control participants can drift over calendar time. A new experimental arm may open after that drift has begun, so a post-opening comparison with earlier controls can mix the treatment effect with changes that would have occurred anyway. Difference-in-differences (DiD) removes the observed change in a comparison group from the observed change in the target group. Its usefulness depends on a specific assumption about how the groups would have changed without the intervention.

The contrast is an adjustment for a particular form of control drift, not a substitute for concurrent randomized controls. It is most interpretable when the target group and comparison group are defined in advance, outcomes are available in the same pre- and post-opening windows, and the comparison group remains untreated by the added intervention.

The notation describes group-level potential outcomes; it does not require observing the same people in both periods. DiD can be calculated from repeated cross-sections or from longitudinal data, but the interpretation depends on how the groups and populations are defined. In particular, the pre-period target-group outcome is untreated, and the comparison group's outcomes are untreated in both periods. The same outcome definition and measurement scale must be used across all four group-period cells.

The Contrast and Its Assumption

Write \(\mu_{gt}\) for the observed mean outcome in group \(g\) and period \(t\). If the target group is untreated before the arm opens and treated afterward, while the comparison group remains untreated, the contrast is

In population notation, the corresponding contrast uses conditional expectations instead of sample means. The key identifying assumption is parallel trends in untreated potential outcomes: over the selected periods, the target and comparison groups would have had the same mean change if neither had received the added intervention.

Parallel trends does not require the groups to have the same baseline mean. For example, groups could begin at different average symptom scores and still satisfy the assumption if their untreated mean scores would change by the same amount. The assumption concerns changes, not levels. It is also scale-specific: this version concerns additive changes in the outcome mean.

The conclusion follows by separating the target group's observed change into its untreated change and its treatment effect. The comparison group's observed change supplies the untreated trend under parallel trends.

The bracketed difference in untreated trends is zero under parallel trends. The remaining term is the target group's post-period treatment effect. This argument also shows the consequence of a violation: the DiD contrast equals the treatment effect plus the difference between the two groups' untreated trends. A trend mismatch therefore becomes bias, not merely a loss of precision.

Assumptions to Check in a Platform Comparison

The algebra is simple; the design conditions are substantive. Before interpreting the contrast as an effect, ask whether each of the following is credible for the particular arm and time window.

• Comparable untreated trends: absent the added intervention, the groups' mean outcomes would have changed in parallel over these periods. This is not guaranteed by similar baseline means or by a good fit to measured covariates.
• No anticipation: the target group is not affected by knowledge of, preparation for, or access to the intervention before the designated opening period. Otherwise the pre-period outcome may already reflect treatment-related changes.
• Unexposed comparison group: the comparison group does not receive the added intervention and is not indirectly affected by it in a way that changes its outcomes.
• Stable interpretation of the cells: outcome measurement, eligibility, follow-up, and group composition are sufficiently comparable over time. Large changes can make a mean change reflect who was observed rather than a temporal trend in a comparable population.
• Appropriate uncertainty calculation: the four cell means are estimates. Their dependence structure and the trial's sampling or randomization features matter when calculating a standard error or interval. A DiD point estimate alone does not establish statistical significance.

These conditions are not proven by the contrast itself. Earlier material on calendar-time confounding motivates checking the outcome's trajectory and the composition of the comparison. Examining pre-period data can reveal obvious differences in trends, but agreement before arm opening cannot establish that untreated trends would have remained parallel afterward.

Arms Opening at Different Times

When platform arms open at different dates, there is generally no single pre/post contrast that automatically applies to every arm. For each arm, specify its own opening time, pre- and post-periods, target population, and comparison group. The parallel-trends condition is then a claim about that arm's target and comparison groups over those particular periods. A group that is an appropriate comparison for one opening may be unsuitable for another because the time windows or populations differ.

A simple pooled before-and-after calculation can also use the wrong calendar periods. If one arm opens earlier, its post period may overlap another arm's pre period. Treating the latter as an untreated comparison is valid only if it truly remains unexposed and supplies a credible untreated trend for the earlier arm. If comparison units later receive the intervention, their data after exposure cannot be treated as untreated outcomes for that contrast. Arm-specific contrasts keep the time comparison explicit; they do not, by themselves, resolve differences in populations or calendar conditions.

Diagnostics, Alternatives, and Limits

Useful checks focus on whether the design makes the assumptions plausible and whether the data contradict them. Plot group means over several pre-opening periods when available; look for visibly different slopes or abrupt changes. Check that the chosen comparison group is observed over the same calendar windows, and examine whether group sizes or measured characteristics shift substantially across periods. Confirm that the outcome is defined and measured consistently. These checks can identify warning signs, but none can reveal the target group's unobserved post-period outcome without treatment.

A tempting diagnostic is to see whether a pre-period DiD estimate is near zero. A nonzero estimate warns against parallel trends over that earlier interval. A near-zero estimate is not proof of parallel trends after opening: the post-opening trend could diverge, and pre-period estimates may be imprecise. A plot or test should therefore be treated as evidence about plausibility, not verification of the identifying assumption.

The main alternatives answer different questions. A post-period comparison alone does not adjust for baseline group differences or drift. A target-group before-and-after change adjusts for neither shared secular change nor other time changes. A concurrent randomized comparison is generally preferable for estimating a randomized treatment effect because randomization supports comparability within its trial population; DiD is relevant when a specified comparison must account for a pre/post drift and its assumptions are defensible. Regression adjustment can encode time and group terms, but a regression specification does not make parallel trends true. The earlier material on regression adjustment addresses that modeling approach; here the central issue is the identifying contrast.

DiD has important limitations. It cannot separate treatment effects from a group-specific shock that begins at the same time as arm opening. It may also be biased by changing group composition, anticipation, spillover, or different untreated trends. With arms opening at different times, a contrast based on an inappropriate time window or an already-treated comparison group can have no valid untreated-trend interpretation. The method is thus an explicit assumption-based adjustment, not a general correction for every kind of calendar-time bias.