06-reference/research

lookahead direct remission itt vs completer

2026-10-12·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
investinglly-longevity-v1virta-healthtype-2-diabetesprimary-source

Neither comparator is ITT - both are with-data figures, so Virta's edge survives normalization but shrinks to statistical noise

The question

"Are the Look AHEAD (7.3%) and DiRECT (13%) remission rates that the Virta DRCP paper compares itself to reported as ITT or completer figures? If ITT, the favourable comparison against Virta's cited 20% completer rate is partly an artifact."

Open follow-up #4 from [[2026-07-30-virta-drcp-5yr-itt-remission-rate]]. The premise under test is the comparator side of "Virta beats prior lifestyle interventions," which is the premise the behavioral-substitution bear vector in the LLY position rests on. Both comparator primary papers were obtained and read directly in this run. The DiRECT 5-year paper came from the University of Glasgow Enlighten repository PDF (publisher version, 11 pages, full text extracted); Look AHEAD's remission paper came from the PMC author manuscript via NCBI efetch.

What we already know (from the vault)

What the web says

Citation mapping (verified via Crossref reference list for DOI 10.1016/j.diabres.2024.111898). Virta's reference list contains both comparator papers, so the figures map unambiguously: the 7.3%-at-4-years is Gregg EW et al., JAMA 2012 (ref b0025), and the 13%-at-5-years is Lean MEJ et al., Lancet Diabetes Endocrinol 2024 (ref b0045). Both were resolved and read.

Convergences and contradictions

Synthesis for RDCO

The claim "Virta beats prior lifestyle interventions on absolute 5-year remission" survives denominator normalization. It does not survive being called a demonstrated advantage. Both comparators turn out to be completer-flavoured, so normalizing the denominators moves all three numbers down together and preserves the ordering. What normalization destroys is the size of the gap: at the only tier a payer would care about - everyone who ever started - Virta's 9.2% against DiRECT's 7.4% is a 1.8-percentage-point difference between a non-randomized single-arm program and a cluster-randomized trial, with confidence intervals that sit almost on top of each other. For the thesis file, the right sentence is "Virta's 5-year remission rate is in the same band as the best randomized lifestyle trial, not above it," and the wrong sentence is any version of "Virta roughly doubles DiRECT," which is what the completer-versus-completer framing invites.

Two artifacts are present and they are both independent of the ITT question. The second one is larger than the first. The first is the timepoint mismatch: Look AHEAD's 7.3% is at year 4 and is being set beside 5-year numbers, in a literature where remission decays monotonically in every cohort examined (Virta 24.3% to 14.2% ITT from year 2 to year 5; DiRECT 36% at year 2 to 13% with-data at year 5; Look AHEAD 11.5% at year 1 to 7.3% at year 4). A 4-year comparator is flattered by roughly one year of decay. To Virta's credit this one is disclosed - the paper writes "at 4 years" - so it is a reader trap rather than a misstatement. The second and much bigger artifact is that Virta has no control arm and both comparators do. DiRECT's control reached 5.4% with-data (3.4% randomized) remission at 5 years and Look AHEAD's reached 2.0% at year 4, so the incremental effect attributable to intervention is roughly 7.5 points for DiRECT at the with-data tier, 4.0 points at the randomized tier, and 5.3 points for Look AHEAD. Virta's incremental effect over usual care is unknown and is not recoverable from the DRCP cohort at any denominator. If usual-care 5-year remission is anywhere near DiRECT's control rate, Virta's incremental becomes roughly 4 points on the 262 denominator - indistinguishable from DiRECT's randomized incremental, from a design that cannot support a causal claim at all.

For the LLY position, this closes the comparator question without moving the weighting. The behavioral-substitution bear vector was already trimmed in [[2026-07-30-virta-drcp-5yr-itt-remission-rate]] on the strength of the decay curve, and nothing here argues for re-inflating it. What this adds is a defensive fact: the entire behavioral-remission literature, including its best randomized trial, lands single-digit-to-low-teens at 5 years once you anchor to the people who started. That is the ceiling on the substitution threat across the whole modality, not just across Virta's program. A modality whose best randomized result is 7.4% durable drug-free remission on the randomized denominator does not displace chronic pharmacotherapy; it thins and delays it. One genuine point in Virta's favour that should go in the file alongside the deflation: it reached its rate in a longer-duration, insulin-including population that DiRECT's protocol would have excluded outright, which makes the comparison harder than the headline numbers suggest even as the numbers themselves converge.

The reusable research gate from this run. The parent brief's gate was "tables are load-bearing, discussion prose is not," derived from Virta mislabelling its own completer rate as "those enrolled." This run extends it one hop outward: when a paper cites a comparator percentage, re-derive the comparator's denominator from the comparator's own paper, because the citing paper will quote whichever of the comparator's several published figures is most convenient. DiRECT printed 13% and 7% in the same sentence and Virta quoted only the 13%. That is not fabrication and it is not even unusual, but it is the step at which a number becomes unauditable, and it is cheap to catch - the fix here cost one repository PDF and one PMC author manuscript.

Why this is in the vault

This closes open follow-up #4 of [[2026-07-30-virta-drcp-5yr-itt-remission-rate]] and supplies the basis-matched comparator table that the "behavioral substitution" bear vector in [[2026-05-21-lilly-glp1-longevity-thesis]] needs in order to state how far Virta's 5-year remission actually sits above the best randomized lifestyle trial - which turns out to be within overlapping confidence intervals, not the 1.5x the completer framing implies.

Open follow-ups

Related

Sources

Primary (fetched and read in full this run):

Primary (quoted from a prior direct read, not re-fetched):

Not opened (flagged, not retried):

Own arithmetic (labelled, not from any paper):