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)
- [[2026-07-30-virta-drcp-5yr-itt-remission-rate]] established Virta's three denominators from the primary paper: 20.0% of 120 completers-with-data, 14.2% of the 169 extension consenters (the paper's labelled ITT worst-observation-carried-forward figure), and 9.2% of the 262 originally enrolled (own arithmetic). It also recorded the comparator sentence in the Virta discussion verbatim: "Look AHEAD 7.3% at 4 years, DiRECT 13% at 5 years, Esposito Mediterranean 5% at 6 years," with the explicit caveat that "the comparators' denominators are not normalized here."
- [[2026-10-11-virta-drcp-68-unapproached-participants-denominator]] resolved where Virta's missing 93 went (68 never approached, 25 declined), found the 68 were directionally worse-prognosis but not significantly so, and recommended 9.2% as the payer-facing anchor with 14.2% carried only as a labelled consent-conditioned figure. It left this exact comparator question open and named it as queued.
- [[2026-05-25-virta-health-5yr-cohort-reversal-durability]] is the grandparent, which first flagged that Virta's press surface leads with completer numbers.
- Two sibling briefs close the thread's other open vectors the same night: [[2026-10-12-glp1-discontinuation-cohort-past-18-months]] (behavioral-substitution durability past 18 months) and [[2026-10-12-virta-drcp-medication-effect-score-itt-reconstruction]] (medication-reduction ITT reconstruction, which reuses this brief's denominator-tier framework).
- The vault had no prior coverage of Look AHEAD or DiRECT as primary sources. A QMD sweep returned only the Virta briefs themselves. Everything in the next section is new to 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.
- DiRECT's 13% is explicitly NOT intention-to-treat. It is a with-data figure on a twice-selected subgroup. Verbatim from the statistical analysis section: "No assumptions were made regarding missing data, so remission status was calculated for participants with available data for both HbA1c and medication use." Abstract, verbatim: "At 5 years, DiRECT extension participants (n=85) lost an average of 6.1 kg, with 11 (13%) of 85 in remission."
- The DiRECT denominator cascade is structurally identical to Virta's. Verbatim: "149 participants were randomly assigned to the intervention group... After 2 years, all intervention participants still in the trial (101 [68%] of 149) were approached to receive low-intensity support for a further 3 years. 95 (94%) of 101 were able to continue and consented and were allocated to the DiRECT extension group." So the 85 are the subset of 95 consenters with year-5 data, who are a subset of the 101 still in trial at 2 years, who are a subset of 149 randomized. The paper also states plainly that "The DiRECT extension group was self-selected by remaining engaged with the intervention at 2 years."
- DiRECT publishes its own randomized-denominator figure, and the Virta paper did not pick it up. Discussion, verbatim: "the 5-year results (known remissions in 11 (13%) of those with data, or 7% of the original 149 intervention group) remain remarkable." 11/149 = 7.4% (own arithmetic, matching the paper's rounded 7%). The honest randomized-population DiRECT number was printed one clause after the 13% and is roughly half of it.
- Look AHEAD's 7.3% is also not ITT, and the paper says so in its own limitations. Verbatim: "Our analyses were not conducted using a pure intention-to-treat approach, as participants who underwent bariatric surgery after randomization, had normal glucose levels at baseline, or were missing data were excluded from the analyses. This was a deliberate analytic decision based on the desire to estimate the absolute frequency of diabetes remission that could be attributed to lifestyle intervention."
- Look AHEAD's denominator detail. Of 5145 randomized, 355 were excluded (normal baseline glucose and related criteria) and a further 287 "who were missing relevant outcome data at all follow-up visits or who underwent bariatric surgery," leaving an analytic sample of 4503 (ILI n=2241, DSE n=2262). The 7.3% is a 100-dataset multiple-imputation estimate on that analytic sample. The paper also prints the raw figure: year 4, "7.3% (95% CI, 6.2%-9.4%) (150/2056)." So 2056 of 2241 ILI participants had year-4 data, and the imputed and raw estimates coincide at 7.3%.
- Look AHEAD's timepoint is year 4, not year 5, and there is no year-5 remission figure in this paper. Its sustained-remission numbers decay the same way everything else does: 9.2% held remission at least 2 years, 6.4% at least 3 years, 3.5% a full 4 years (ILI, imputed, n=4503 basis).
- Both comparators have control arms. Virta does not. DiRECT original control: "Original control participants (n=149) had mean weight loss 4.6 kg (n=82), and 5 (5%) of 93 were in remission" at 5 years. Look AHEAD DSE: 2.0% at both year 1 and year 4. Virta's DRCP is a single-arm non-randomized study with no counterfactual at any timepoint.
- Remission definitions are close enough not to drive the result, with one caveat. DiRECT: "HbA1c <48 mmol/mol (<6.5%) without glucose-lowering medications," annual snapshot. Virta: 2021 international consensus, HbA1c <6.5% persisting at least 3 months off glucose-lowering pharmacotherapy. Look AHEAD: ADA 2009 partial-or-complete remission, which requires both fasting plasma glucose below 126 mg/dL and HbA1c below 6.5% off all antihyperglycemic medication - marginally stricter on the glycemic side, with no duration requirement. All three require being off all glucose-lowering drugs including metformin. [inferred] Net definitional drift across the three is small relative to the denominator spread, and Virta's paper separately notes metformin was not deprescribed in its cohort, which if anything understates its own number.
- Population differences cut in Virta's favour, and this is the one dimension where the favourable comparison gets stronger. DiRECT enrolled age 20-65, under 6 years' diabetes duration, BMI over 27, and excluded insulin users, with roughly 3 years' average duration. Virta's CCI cohort had about 8.2 years since diagnosis and 26.2% on insulin at baseline (per [[2026-10-11-virta-drcp-68-unapproached-participants-denominator]] and [[2026-07-30-virta-drcp-5yr-itt-remission-rate]]). Short diabetes duration is among the strongest predictors of remission in both comparator papers, so Virta hit its rate in a harder population. Offsetting this: Virta's cohort is self-selected for motivation in a way a cluster-randomized primary-care trial is not.
Convergences and contradictions
The question's hypothesis is wrong, and wrong in Virta's favour. Neither 7.3% nor 13% is an ITT figure. Both are with-data or analytic-sample figures computed on exactly the kind of selected denominator that Virta's 20% is computed on. The specific artifact the follow-up suspected - ITT comparators versus a completer Virta number - does not exist. The comparison in Virta's discussion is closer to basis-matched than the parent brief feared.
But the parent brief's instinct was right about the denominator discipline, just aimed at the wrong paper. DiRECT is the more honest of the two reports: it prints its 13% and its 7%-of-149 in the same sentence. Virta's discussion quoted the 13% and dropped the 7%. That is the laundering step in this chain, and it happens in the citing paper rather than the cited one.
Virta's edge survives every normalization tier, and dissolves into overlapping confidence intervals at the honest one. Matched-tier comparison, all own arithmetic with Wilson 95% intervals:
Denominator tier Virta, 5y DiRECT, 5y Look AHEAD, 4y Completers / with data 20.0% (24/120), CI 13.8-28.0 12.9% (11/85), CI 7.4-21.7 7.3% (150/2056), CI 6.2-8.5 Extension consenters / analytic sample 14.2% (24/169), CI 9.7-20.3 11.6% (11/95), CI 6.6-19.6 6.7% (150/2241), CI 5.7-7.8 Original enrollment / randomized 9.2% (24/262), CI 6.2-13.3 7.4% (11/149), CI 4.2-12.7 ~5.8% (150/~2570), CI 5.0-6.8 [inferred denominator] Virta is numerically highest at all three tiers. The ratio against DiRECT compresses from 1.55x at the completer tier to 1.24x at the enrollment tier, and at that tier the intervals overlap almost completely (9.2%, CI 6.2-13.3 versus 7.4%, CI 4.2-12.7). The Look AHEAD randomized denominator of 2570 ILI participants is [inferred] from the standard Look AHEAD design rather than printed in Gregg 2012; the fully derivable analytic-sample figure is 6.7%.
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
- Is there a 5-year Look AHEAD remission figure anywhere? Gregg 2012 stops at year 4 by design. Gregg EW et al., Diabetologia 2024 ("Impact of remission from type 2 diabetes on long-term health outcomes: findings from the Look AHEAD study," cited by Virta as ref b0150) covers the same cohort over a longer horizon and may carry later-year remission prevalence. It was not opened in this run and would remove the one timepoint mismatch.
- What is the usual-care 5-year remission rate in an unselected type 2 population? This is the missing counterfactual for Virta's single-arm design, and it is the quantity that decides whether Virta's 9.2% represents a 4-point or a 9-point incremental. Virta's own reference list points at two candidate sources it did not use this way: Karter et al., Diabetes Care 2014 (incidence of remission in the Diabetes and Aging Study) and Dambha-Miller et al., PLoS One 2023 (population-based remission trajectories). Both are population cohorts, both are plausibly accessible, and neither has been read into the vault.
- Does the Esposito 2014 Mediterranean-diet comparator (5% at 6 years) have the same structure? It is the third comparator in Virta's discussion sentence and the only one not resolved. It is also the only one at a longer horizon than Virta, which makes it the comparator least flattered by timepoint choice and therefore the most informative of the three.
- Can DiRECT's control-arm contamination be bounded? The 5-year paper warns that "control group participants were strongly encouraged to control their weight after the end of the 2-year randomised trial, so their outcomes do not reflect usual care," and that "differences between groups beyond the 2-year point might not reflect intervention." If DiRECT's 5.4% control rate is inflated by that encouragement, DiRECT's incremental is understated and the normalized Virta-versus-DiRECT comparison tilts further against Virta.
- Should the vault carry standalone primary notes for Look AHEAD and DiRECT? Both are now read and neither has a note. They are the canonical comparators for any future behavioral-versus-pharmacological question, not just this one, and re-deriving their denominators from inside a Virta brief is how the next thread loses them.
Related
- [[2026-07-30-virta-drcp-5yr-itt-remission-rate]] - parent brief; this closes its open follow-up #4
- [[2026-10-11-virta-drcp-68-unapproached-participants-denominator]] - sibling brief; resolved Virta's missing 93 and queued this question
- [[2026-10-12-glp1-discontinuation-cohort-past-18-months]] - sibling brief closing the behavioral-substitution-durability vector of the same thread
- [[2026-10-12-virta-drcp-medication-effect-score-itt-reconstruction]] - sibling brief closing the medication-reduction ITT vector, reusing this brief's denominator-tier framework
- [[2026-05-25-virta-health-5yr-cohort-reversal-durability]] - grandparent; first flagged the completer-framing problem
- [[2026-05-21-lilly-glp1-longevity-thesis]] - the live position this calibrates
- [[2026-05-24-virta-health-payer-contracts-lly-thesis-input]] - payer-economics arm, still completer-only
- [[2026-07-18-capital-rx-reverse-lilly-relationship]] - substitution-versus-complement resolution on the payer side
- [[2026-05-22-tim-ferriss-sami-inkinen-virta-t2d-rowing]] - Inkinen anchor interview
Sources
Primary (fetched and read in full this run):
- Lean MEJ, Leslie WS, Barnes AC, Brosnahan N, Thom G, McCombie L, Peters C, Zhyzhneuskaya S, Al-Mrabeh A, Hollingsworth KG, Rodrigues AM, Rehackova L, Adamson AJ, Sniehotta FF, Mathers JC, Ross HM, McIlvenna Y, Welsh P, Kean S, Ford I, McConnachie A, Messow C-M, Sattar N, Taylor R. "5-year follow-up of the randomised Diabetes Remission Clinical Trial (DiRECT) of continued support for weight loss maintenance in the UK: an extension study." Lancet Diabetes Endocrinol 2024;12(4):233-246. DOI 10.1016/S2213-8587(23)00385-6. PMID 38423026. Full text obtained from the University of Glasgow Enlighten repository: https://eprints.gla.ac.uk/315287/7/315287.pdf - abstract, Statistical analysis section, Results, Table 3, Table 4 and Discussion read directly. Access note: Europe PMC holds no full-text copy (inEPMC N, isOpenAccess N); the repository deposit is the accessible route.
- Gregg EW, Chen H, Wagenknecht LE, Clark JM, Delahanty LM, Bantle J, Pownall HJ, Johnson KC, Safford MM, Kitabchi AE, Pi-Sunyer FX, Wing RR, Bertoni AG, for the Look AHEAD Research Group. "Association of an intensive lifestyle intervention with remission of type 2 diabetes." JAMA 2012;308(23):2489-2496. DOI 10.1001/jama.2012.67929. PMID 23288372, PMC4771522. Full text obtained via NCBI efetch (db=pmc, id=4771522) - Methods, Results, Figure 2 and 3 notes (raw cases and denominators), Limitations read directly.
- Crossref works API for DOI 10.1016/j.diabres.2024.111898 - Virta DRCP 2024 reference list retrieved to verify that Gregg 2012 (key b0025) and Lean 2024 (key b0045) are the cited sources for the 7.3% and 13% figures respectively.
Primary (quoted from a prior direct read, not re-fetched):
- McKenzie AL, Athinarayanan SJ, Van Tieghem MR, Volk BM, Roberts CGP, Adams RN, Volek JS, Phinney SD, Hallberg SJ. "5-Year effects of a novel continuous remote care model with carbohydrate-restricted nutrition therapy including nutritional ketosis in type 2 diabetes: An extension study." Diabetes Res Clin Pract 2024;217:111898. Figures and verbatim quotes taken from the direct full-text read recorded in [[2026-07-30-virta-drcp-5yr-itt-remission-rate]]. The publisher page remains behind an automation block.
Not opened (flagged, not retried):
- The DiRECT 5-year appendix (https://eprints.gla.ac.uk/315287/4/315287_supplement.pdf) - contains the per-year remission series and the sensitivity analysis detail; not needed for the denominator question.
- Gregg et al., Diabetologia 2024 (Look AHEAD long-term outcomes) and Esposito et al., Diabetes Care 2014 (Mediterranean diet) - queued as follow-ups above.
Own arithmetic (labelled, not from any paper):
- Virta: 24/120 = 20.0%, 24/169 = 14.2%, 24/262 = 9.2%.
- DiRECT: 11/85 = 12.9%, 11/95 = 11.6%, 11/101 = 10.9%, 11/149 = 7.4% (the paper prints this one rounded to 7%); control 5/93 = 5.4%, 5/149 = 3.4%.
- Look AHEAD: 150/2056 = 7.3% (the paper prints this), 150/2241 = 6.7%, 150/2570 = 5.8% with the 2570 ILI randomized count inferred rather than printed in Gregg 2012.
- All 95% intervals are Wilson score intervals computed in this run; none are from the source papers except Look AHEAD's, which are the papers' own multiple-imputation intervals where quoted as such.
- Ratio compression: 20.0/12.9 = 1.55x at the completer tier, 9.2/7.4 = 1.24x at the enrollment tier.
- Incremental-over-control: DiRECT 12.9 - 5.4 = 7.5 points with-data, 7.4 - 3.4 = 4.0 points randomized; Look AHEAD 7.3 - 2.0 = 5.3 points.