06-reference/research

virta drcp 68 unapproached participants denominator

2026-10-11·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
investinglly-longevity-v1virta-healthdenominator-attritionprimary-source

The 68 never approached are almost certainly the 68 who had already left the trial by year 2 (inferred), and the published flow arithmetic is consistent with it

Numbers-honesty brief for an investing calibration. Nothing here is medical advice or a clinical recommendation.

The question

Verbatim: "In the Virta DRCP extension, why were 68 of 262 CCI participants never approached for the extension? If they were disproportionately non-responders, the 262-anchored 9.2% is closer to truth than the paper's 169-anchored 14.2%."

CCI = Continuous Care Intervention, the Virta / Indiana-and-Franciscan-Health trial arm. DRCP = the three-year diabetes-reversal continuation (extension) phase reported in the 2024 five-year paper. ITT = intention-to-treat. This is follow-up #2 from [[2026-07-30-virta-drcp-5yr-itt-remission-rate]], which named it "the single biggest remaining uncertainty in the calibration."

What we already know (from the vault)

What the web says

Three new primary sources were pulled in this run. The 2024 extension paper itself was not re-fetched (its prose is already on the record from the parent brief's direct read) and its publisher page is now behind a hard automation block; see Sources.

The reconciliation: 262 to 169 to 24, line by line

Every line is marked [sourced] with its source, or [inferred], or [unaccounted]. No line is estimated.

  1. 262 enrolled in the CCI arm (349 total T2D enrollees: 262 CCI, 87 usual care). [sourced: Hallberg 2018, Results]
  2. Minus 44 lost in year one, leaving 218 enrolled at 1 year (83% of 262). [sourced: Hallberg 2018, Results + Table 1 columns 262 / 218 / 44]
  3. Minus 24 more lost in year two, leaving 194 enrolled at 2 years (74% of 262); cumulative "dropout or missing data" = 68. [sourced: Athinarayanan 2019, Results + Table 1 columns 262 / 194 / 68]
  4. 194 approached for the three-year extension. [sourced: McKenzie 2024, Methods]
  5. The 68 never approached are the same 68 who were no longer retained at 2 years. [inferred, high confidence] Basis: exact numeric identity of the 194/68 split across the two papers, plus the explicitly stated invitation rule in the prediabetes arm of the same registry and extension ("We invited the 72 of 96 (75%) participants who completed the initial two-year study," Zoller 2026). Not printed as a statement in McKenzie 2024.
  6. Minus 25 who declined the extension, leaving 169 consenters = the published ITT worst-observation-carried-forward denominator. [sourced: McKenzie 2024, Methods; the 25 is 194 - 169, own subtraction]
  7. Minus 47 who dropped out before 5 years, minus 2 with missing 5-year HbA1c, leaving 120 with 5-year data; 122 "remained in the study." [sourced: McKenzie 2024, Methods + Table 1 footnote]
  8. 24 met the 2021 international consensus remission definition at 5 years. [sourced: McKenzie 2024, Results + Table 1]

Consistency check (own arithmetic): 68 + 25 + 47 + 2 + 120 = 262, with no residual. Note what this does and does not prove. It is an accounting identity almost everywhere: 25 is defined as 194 - 169, and 120 is 169 - 47 - 2 from the same footnote. The one genuinely independent cross-paper check is 262 - 194 = 68, matching the 2-year paper's own dropout column. Matching counts are consistent with the groups being the same people; they do not demonstrate it.

What remains unaccounted, precisely:

Convergences and contradictions

Synthesis for RDCO

The answer is reconstructed from primary sources rather than speculated, and it points the calibration the way the question suspected, but the load-bearing step is an inference and should be carried as one. [inferred, high confidence] The 68 were never approached because the extension's invitation rule required having completed the original two-year trial, and 68 of 262 had not: 44 left inside year one, 24 more during year two. That is sourced from the cohort's own 1-year and 2-year papers, whose Table 1 structures carry the identical 262 / 218 / 44 and 262 / 194 / 68 splits, and it is confirmed as a rule in words by the prediabetes arm of the same registry and the same extension, published June 2026: "We invited the 72 of 96 (75%) participants who completed the initial two-year study." Two things here are inference rather than quotation: the identity statement (that the 68 unapproached are the 68 two-year losses), and the transfer of the prediabetes arm's stated invitation rule to the type 2 diabetes arm. Both are tight but neither is printed: the counts match exactly, and the rule is stated only for the sibling arm.

So the 68 are not an administrative or random subsample. [inferred, following from the identity above] They are people for whom the program did not run to completion. That is not the same thing as biological non-response, and the brief should not pretend otherwise: the authors list trial-visit burden and intervening life events alongside insufficient motivation as dropout reasons, and never count them, so an unknown fraction of the 68 left for reasons unrelated to the treatment working. Of those three, only trial-visit burden is specific to running a trial. [inferred, not sourced] For the question the thesis actually asks - what does a payer observe across an unselected book of business - the distinction largely collapses, because a member who stops engaging delivers no remission regardless of why, and a commercial deployment substitutes its own frictions for the trial's rather than removing them. That step is reasoning, not evidence, and it is where a reader who disagrees should push. The 25 who declined the extension are a separate and milder case: they did complete two years, their baseline and two-year characteristics matched the consenters except for lower beta-hydroxybutyrate and higher fasting insulin at two years, and they are the group for which "mildly enriched against" is the honest description.

The verdict on the two anchors: prefer 9.2% for payer-facing, unselected-population comparisons, and use 14.2% only as a consent-conditioned durability figure, explicitly labelled as such. 14.2% is the remission rate conditional on having completed two years of the program and then consenting to three more. That is a legitimate figure for exactly one question - how durable is remission among people who opt into continuation - and it is the right figure to put beside other trials' randomized-population ITT rates only if those comparators are similarly consent-conditioned. Whether Look AHEAD and DiRECT (the comparators the 2024 paper's discussion cites) are ITT or completer-anchored is not established here and is already a separately queued question. One point in 14.2%'s favour that should not be lost: it is itself worst-observation-carried-forward, so it already gives zero credit to the 47 extension dropouts. It is not a completer rate. Its defect is the consent gate on its denominator, not its handling of missing data. 9.2% (24/262, own arithmetic) is the floor and the better payer-facing anchor, and it has a direct precedent in this cohort's own literature: the same authors published exactly that construction at two years ("the percentage of all CCI enrollees (N = 262) with verified ... remission ... was 14.9%"). For a defensible band rather than a point: at two years the gap between the authors' 262-anchored zero-credit figure (14.9%) and their 262-anchored model-based figure (17.6%) was +18% relative. Granting the same model generosity to the five-year number gives roughly 10.9%. So carry 9.2% as the anchor, ~10.9% as one illustrative upward scenario rather than a demonstrated ceiling, and 14.2% as a labelled consent-conditioned figure. Three caveats on the 10.9%, all of which cut against treating it as a bound: it is my own arithmetic; the 2-year side uses the looser 2009 partial-or-complete remission definition rather than the 2021 consensus definition used at five years; and it cannot be decomposed per participant, because the 2-year model estimates are covariate-adjusted marginals, not imputed counts. Most importantly, the unobserved share is roughly twice as large at five years (142 of 262, 54%) as at two years (68 of 262, 26%), so if an imputation gap scales with missingness the true adjustment would be larger than +18% relative, not smaller.

Two things this does not change, and one new datum it adds. It does not change the direction of the parent brief's verdict: behavioral remission at five years is a real minority outcome that decays inside a single cohort (24.3% to 14.2% ITT worst-observation-carried-forward, both on the 169 denominator, year two to year five, per [[2026-07-30-virta-drcp-5yr-itt-remission-rate]]), so it thins and delays prescriptions rather than replacing them, and the bear vector stays trimmed. It does not rescue the payer-savings arm, whose medication-reduction figures remain completer-only with no published ITT. What it adds is a reusable research gate: when a paper reports an extension-phase ITT, the extension's denominator is almost never the trial's enrollment, and the missing participants are usually recoverable from the predecessor paper's Table 1 rather than from the registry. In this case the registry was a dead end - no posted results, no flow table, no protocol document, no posting obligation - and the predecessor papers closed the arithmetic to the participant. Check the earlier paper's completer column before checking the registry.

Why this is in the vault

This resolves open follow-up #2 of [[2026-07-30-virta-drcp-5yr-itt-remission-rate]], pending only the supplementary-appendix check named below, and fixes which remission figure the behavioral-substitution bear vector in [[2026-05-21-lilly-glp1-longevity-thesis]] is weighted against: 9.2% on 262, not 14.2% on 169, with 14.2% retained only as a labeled consent-conditioned figure. It also supplies the auditable reconciliation (262 to 218 to 194 to 169 to 122 to 120 to 24) that any future challenge to the thesis's Virta input can be checked against without re-reading four papers.

Open follow-ups

Related

Sources

Primary, fetched and read in full this run:

Primary, read in a prior run and relied on here without re-fetch:

Unreachable this run, with exact failures:

Own arithmetic, labelled, not printed in any source: