What the STAR*D Reanalysis Debate Teaches About Trial Fidelity

The number everyone quotes, and why it is contested

For nearly two decades, one figure anchored how clinicians talked about treating depression: after up to four sequential antidepressant steps, about 67 percent of patients in the STAR*D trial reached remission. A 2023 reanalysis of the same patient-level data, published in BMJ Open by Pigott and colleagues, reported a cumulative remission rate near 35 percent, roughly half. The original investigators replied in the American Journal of Psychiatry that their data remain strong. Both camps worked from the same trial. No new data separates their answers. Two methodological choices do: which depression scale counts as the outcome, and which patients count at all. That gap is the whole lesson.

STAR*D (Sequenced Treatment Alternatives to Relieve Depression) was a large NIH-funded effectiveness trial designed to mirror real-world care. Patients who did not remit on one treatment moved to the next step. The headline appeal was practical: keep trying, and most people eventually get better. When a single trial carries that much clinical weight, how its primary number was constructed deserves scrutiny, and that scrutiny is what the reanalysis debate provides.

Choice one: which scale is the outcome

The trial's own protocol named the Hamilton Rating Scale for Depression (HRSD), administered by trained assessors, as the primary outcome measure, and it specified that clinic-administered ratings such as the self-report QIDS-SR would not be used to declare research outcomes. The widely cited 67 percent figure, however, came from analyses built on the QIDS-SR, a self-report scale collected during clinic visits.

This is not a trivial swap. A blinded, clinician-administered instrument and an unblinded self-report instrument can diverge, and the direction of that divergence matters. The reanalysis argued that reverting to the protocol-stipulated HRSD, the measure the investigators themselves designated in advance, is the correct basis for the primary result. Using that measure and the protocol's rules, the cumulative remission estimate falls substantially. The original investigators, in their reply, defend the broader analytic approach and the inclusiveness of the reported cohort. Readers do not need to pick a winner to see the structural point: when a headline number rests on a measure the protocol excluded from outcome reporting, the choice of instrument is doing heavy lifting, and it should be stated plainly rather than assumed.

Choice two: which patients count

The second dispute is about the denominator. The reanalysis excluded patients who, per the protocol's own entry criteria, should not have contributed to the outcome. Two groups drew attention: patients who already scored as remitted on the HRSD at study entry, and patients who scored as remitted when starting their next treatment step. Counting people who were not depressed by the study's own threshold, the reanalysis argued, inflates apparent success.

The original investigators counter that the reanalysis stripped out the records of a large block of participants and that this pruning misreads the study's intent, which was to reflect an inclusive, real-world population. Here reasonable methodologists can genuinely disagree about intent. But the disagreement itself is the teaching case: applying a protocol's stated inclusion criteria versus preserving a broader as-enrolled sample are two defensible analytic postures that produce materially different headline rates from identical raw data.

Choice three: how to treat the people who left

A third factor compounds the first two. STARD had heavy attrition across steps. The 67 percent estimate is a theoretical cumulative figure that, in effect, assumes patients who dropped out would have remitted at rates similar to those who stayed. The reanalysis instead treated dropouts conservatively, as non-remitters, which is the more cautious convention for missing outcomes. Optimistic and conservative handling of missing data are both recognized approaches, and each has defenders. In a trial with high dropout, that single assumption can swing a cumulative number by tens of percentage points. The original 2006 STARD report itself presented the cumulative figure as a theoretical estimate, so the reanalysis is not inventing a new dispute so much as pressing on an assumption that was always load-bearing.

What trial fidelity actually means

Put the three choices together and a general principle emerges. Trial fidelity is not a slogan about honesty; it is a specific discipline of matching the analysis to the pre-registered protocol: the outcome measure you named, the population you defined, and the missing-data rules you committed to. Each departure from that plan is defensible in isolation and may even be reasonable. The problem is cumulative. Swap the scale, widen the sample, and impute optimistically, and a modest effect can be reported as a strong one, without any single step looking like misconduct.

This is why pre-registration, statistical analysis plans, and blinded outcome adjudication exist. They are not bureaucracy. They are the mechanism that keeps the headline number from drifting toward the most flattering of several defensible analyses. The STAR*D exchange is unusually clean as a teaching example precisely because both sides accept the underlying data and argue only about the rules applied to it.

How to read any trial after this

For anyone appraising evidence, the debate offers a short, portable checklist. Ask what the protocol named as the primary outcome, and whether the reported result uses that measure or a substitute. Ask how the analyzed population compares with the pre-specified inclusion and exclusion criteria. Ask how dropouts and missing data were handled, and whether a more conservative rule would move the number. When a widely cited figure is a theoretical or completer-based estimate rather than an intention-to-treat result, note that explicitly. None of these questions requires taking a side in the STAR*D dispute. They simply make visible the choices that turn raw data into a headline, which is the only reliable defense against being persuaded by a number before understanding how it was built.

This article is educational and is not medical advice.