06-reference/research

virta drcp medication effect score itt reconstruction

2026-10-12·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
investinglly-longevity-v1virta-healthmissing-data-methodsprimary-source

The Medication Effect Score escape hatch does not exist, but the prescription counts are recoverable exactly, and a bounded ITT reconstruction cuts the payer headline from -41% to -19%

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

The question

Verbatim: "Can an ITT-equivalent be reconstructed for the Virta DRCP medication-reduction outcomes (which the paper reports for completers only), e.g. from the Medication Effect Score mixed-effects model on n=169 — or does that require unpublished data?"

ITT = intention-to-treat. MES = Medication Effect Score, a continuous summary of glucose-lowering drug burden. CCI = Continuous Care Intervention, the Virta trial arm. This is follow-up #1 from [[2026-07-30-virta-drcp-5yr-itt-remission-rate]], and it is the one remaining gap in the payer-savings arm of the behavioral-substitution bear vector.

The answer has three parts. The hoped-for route is not available. A different and better-grounded route is available. And the reconstruction, once done, moves the payer number by more than the remission correction did.

What we already know (from the vault)

What the web says

Three primary sources were read in full this run: the one-year paper, the two-year paper, and the five-year paper's PubMed record. The five-year paper's own full text was not reachable; see Sources.

My reconstruction

Two layers. Layer 1 recovers the raw numerators from the published percentages, which is close to certain. Layer 2 builds the bounded ITT-equivalent, which rests on labelled assumptions. Everything in this section is RDCO-derived and appears in no paper.

Layer 1: the completer counts are recoverable exactly, and the denominator is 122

All four published medication percentage pairs resolve to integers over n=122, and over no other candidate denominator. Residuals are the gap between the implied count and the nearest integer, in percentage points:

Class Baseline % implies 5-year % implies max residual
Any diabetes medication 85.2 104/122 71.3 87/122 0.056
Non-metformin 55.7 68/122 32.8 40/122 0.046
Insulin 26.2 32/122 13.1 16/122 0.036
Sulfonylurea 27.0 33/122 4.9 6/122 0.060

All eight land within 0.06 pp of an integer. Tested against n=120, seven of eight fail (85.2% would be 85.0% or 86.7%, neither of which rounds to 85.2). Tested against n=194, seven of eight fail. So the medication analysis runs on 122, the count who "remained in the study," not the 120 with five-year HbA1c that the remission analysis uses. The closure check holds with no residual: 122 completers + 47 extension dropouts + 25 decliners + 68 never approached = 262.

Two notes on this recovery. It is independent corroboration that the prior run transcribed the percentages correctly, since a transcription slip would almost certainly break the integer fit. And the baseline non-metformin figure of 55.7% happens to equal the two-year paper's 194-column baseline of 55.67% exactly; I checked whether the 2024 baselines might therefore be on the 194, and they are not, because 2024's insulin baseline of 26.2% and sulfonylurea baseline of 27.0% match neither the 194 (29.38%, 25.77%) nor the 262 (29.77%, 23.66%). The 55.7% match is coincidence.

Layer 2: the bounded ITT-equivalent

Two bounds, no point estimate.

The neutral bound requires no arithmetic, because under missing-at-random the ITT-equivalent equals the published completer figure. If non-completers' year-5 medication distribution matched completers', the completer rate is already the population rate. So the published numbers are the optimistic end of the range, not a neutral midpoint.

The pessimistic bound is baseline-carried-forward, matching the paper's own convention for its dropout-sensitive endpoints. Assumptions, labelled:

Tier B, the 169 extension consenters:

Class 5-year ITT-equivalent vs baseline change relative published completer relative
Any diabetes medication 127/169 = 75.1% 85.2% -10.1 pp -11.8% -16.3%
Non-metformin 66/169 = 39.1% 55.7% -16.7 pp -29.9% -41.1%
Insulin 28/169 = 16.6% 26.2% -9.6 pp -36.8% -50.0%
Sulfonylurea 19/169 = 11.2% 27.0% -15.8 pp -58.4% -81.9%

Tier C, the 262 enrolled, against the published 262 baseline:

Class numerator build 5-year ITT-equivalent vs published 262 baseline change relative
Non-metformin 40 + 26 + 41 + 14 = 121 121/262 = 46.2% 56.87% -10.7 pp -18.8%
Insulin 16 + 12 + 21 + 7 = 56 56/262 = 21.4% 29.77% -8.4 pp -28.2%
Sulfonylurea 6 + 13 + 12 + 6 = 37 37/262 = 14.1% 23.66% -9.5 pp -40.3%

Numerator terms in order: five-year completers on the drug, then 47 extension dropouts, then 68 never approached, then 25 decliners.

And the MES, conditional on the unverified figure being real and being a completer mean. If 1.5 and 0.9 are the observed completer means on 122, then baseline-carried-forward gives (122 x 0.9 + 47 x 1.5) / 169 = 1.07 on the 169, a -28.9% reduction, and (122 x 0.9 + 140 x 1.5) / 262 = 1.22 on the 262, a -18.6% reduction, against the -40% the parent brief quoted. If instead 0.9 is already the n=169 model-implied marginal, the 169-tier figure cannot be computed, because that requires the completer-only mean, which is unpublished. The two readings differ by exactly the quantity being estimated, which is why this line is not usable in the thesis file.

What blocks a point estimate with an interval, precisely

Six quantities. The first three are hard blockers and all three are unpublished:

  1. The SD or SE of MES at baseline and at five years. Without dispersion there is no interval, only a mean. Not published in any reachable source.
  2. The per-timepoint N for the MES model. Table 1's footnote gives the series for HbA1c, 169 / 165 / ... / 120, but medication ran on a separate completers-only basis, so the MES N series is unknown.
  3. Whether 0.9 is an observed completer mean or a model-implied marginal on 169. Unresolvable from the figure alone, and it changes the answer by the whole effect.
  4. Baseline medication status of the 47 extension dropouts as a subgroup. Forces assumption A2.
  5. Baseline "any diabetes medication including metformin" on the 262. Forces A5 and blocks that row at tier C.
  6. Any post-dropout medication observation for the 140 non-completers. Non-existent by construction, for anyone including the authors.

Items 1 through 5 are answerable by the authors or by Table 2 of the five-year paper. Item 6 is irreducible, which is why this is permanently a bounding exercise and never an estimate.

Convergences and contradictions

Synthesis for RDCO

The headline for the thesis file is a replacement number, not a caveat. The payer-facing medication reduction is "non-metformin glucose-lowering prescriptions fell from 56.9% to roughly 46% across everyone who started, about a 19% relative reduction over five years," and not "roughly halved." The 41% relative figure that the parent brief flagged as "a fair read of the non-metformin line" is the completer number, and the completer number is identical to the ITT-equivalent only under missing-at-random, which this paper's own authors decline to assume for their dropout-sensitive endpoints. Carry the full range rather than a point: -41% relative under MAR on 122 completers, -30% under baseline-carried-forward on the 169 extension consenters, -19% under baseline-carried-forward on the 262 enrolled. Use the 262 tier for payer economics, consistent with the anchor choice [[2026-10-11-virta-drcp-68-unapproached-participants-denominator]] made for remission, because a payer does not get to re-consent its non-engagers out of the denominator. The insulin line deflates less violently and is the most robust of the four, -50% completer to -28% on 262, which matters because insulin is the expensive line item.

That has a specific effect on the bear case, and it is the same direction as the remission correction but larger in proportion. The remission figure moved from 20% completer to 9.2% on 262, a factor of 2.2. The non-metformin medication reduction moves by a factor of 2.2 as well, from -41% to -19% relative. The payer-savings arm of the behavioral-substitution vector was the one arm still resting on undeflated numbers after the last two briefs, and it now deflates by the same multiple as everything else in this cohort. A program that reduces glucose-lowering prescription prevalence by roughly a fifth across an unselected enrolled population, with no control arm to attribute even that to, is not a mechanism that erodes chronic pharmaceutical revenue. It thins and delays prescriptions, which is the conclusion [[2026-07-30-virta-drcp-5yr-itt-remission-rate]] reached from the remission decay curve, now independently reached from the medication side. The bear vector stays trimmed and this is the third consecutive brief that declines to re-inflate it. One honest counterweight: a 19% relative reduction in non-metformin prescribing sustained for five years is not nothing, and the sulfonylurea result survives deflation better than the others at -40% relative on 262, so the right posture is a small persistent margin drag rather than dismissal.

Do not put the MES figure in the thesis file. It fails verification on five independent checks, it is the one number in the parent brief without a verbatim quote, and the specific claim attached to it, that it is an ITT mixed-effects estimate on n=169, is contradicted by what I could verify about how this cohort analyses medication. It may well be a real table row that the prior run read correctly. But the useful thing to notice is that its "is ITT" label was an inference generated inside our own research chain and then carried forward as if sourced, which is the same failure mode the parent brief caught the paper committing when it wrote "20% of those enrolled." If the 1.5-to-0.9 pair is wanted, it needs Table 2 of the five-year paper in hand, and the specific thing to read off it is not the means but the N and the standard errors.

The reusable gate from this run is about escape hatches. The last two briefs produced gates about denominators: tables are load-bearing and discussion prose is not; re-derive a cited comparator's denominator from its own paper. This one is different. When a completer-only outcome has a tempting model-based sibling that looks like it would rescue it, verify that the outcome is actually inside the model before building on it, by reading the model table's row list rather than its footnote. The footnote here says ITT on 262 and means it; the row list says medication is not in scope. Both are true, they sit a centimetre apart, and only the row list answers the question. That check cost one Europe PMC full-text fetch and a regex. The related discipline: when a figure enters our own chain without a verbatim quote while its siblings all have one, that asymmetry is the signal, and the cheapest moment to catch it is before the next brief treats it as a premise.

Why this is in the vault

This resolves open follow-up #1 of [[2026-07-30-virta-drcp-5yr-itt-remission-rate]] and closes the last undeflated input to the payer-savings arm of the behavioral-substitution bear vector in [[2026-05-21-lilly-glp1-longevity-thesis]]: the medication-reduction figure a payer model should use is roughly -19% relative on the 262 denominator, not the -41% completer figure, and the Medication Effect Score route that was supposed to supply a published ITT number does not exist and should not be cited.

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:

Secondary, one search, no figure taken from it:

Unreachable this run, flagged and not retried:

Own arithmetic (labelled, in no paper):