Both cohorts started at 15,843. Here is what was left at each window.

Under the alternative censoring scheme, the treated arm falls to 793 and the cost difference widens to $607. The effect size and the attrition rate move together, in the same direction, at every step. By the window that carries the policy conclusion, roughly one treated patient in twenty is still being observed.

The first window is worth noting on its own. At three to six months, with the largest surviving sample, there is no effect in either analysis. The primary gives −$37 (p=0.484); the secondary gives positive $7 (p=0.889). The signal only appears after the cohort has already halved.

The people still on tirzepatide at eighteen months are not a random sample of the people who started it. They tolerated the drug, could afford the refills, navigated prior authorization, and kept showing up. The healthy adherer effect is one of the better-replicated findings in pharmacoepidemiology. A 2006 BMJ meta-analysis of 46,847 patients found that good adherence to placebo was associated with an odds ratio of 0.56 for mortality. Adherence is a marker for the kind of person you are, not only for the drug you take.

The paper is honest that its estimate applies only to persisters. Then it hands the number to Medicare anyway.

Follow-up was censored at death, disenrollment, pregnancy, switching drugs, or discontinuation. Discontinuation is defined as a gap of more than 45 days after the supply runs out. That criterion can only fire in the treated arm. Controls are not taking anything, so they cannot stop.

This is the structural asymmetry underneath everything else. The treated arm is continuously filtered toward persistence. The control arm is not filtered at all on that dimension, which is why 19 percent of controls survive to the final window against 7.5 percent of the treated. The authors saw this coming and built two corrections for it. Inverse-probability weighting reweights the survivors to stand in for the departed. Pairwise censoring throws away a control whenever their partner leaves.

Both corrections assume censoring is explainable by measured variables. If the reason someone quits tirzepatide is nausea, cost, or a life falling apart, and the claims data cannot see it, no weighting scheme recovers it. The corrections disagree with each other by a factor of two at the final window. That spread is itself the finding: the estimate is not stable under reasonable analytic choices.

Age, sex, race, census region, payer type, index month, Charlson score, comorbidity counts, prior utilization, prior cost, and BMI. That is a serious list, and it addresses the crude version of the affordability objection. These are not rich people compared to poor people. Baseline spending was balanced.

Income, education, health literacy, food environment, and lifestyle behaviors. The authors say so directly in the limitations. Lab values including HbA1c, lipids, blood pressure, and weight were left out of the propensity model because too many were missing, which means the two arms were matched on diagnosis codes rather than on physiology.

Six covariates still exceeded the balance threshold after matching. The treated group had more Medicare coverage (38.5 versus 31.9 percent), more sleep apnea, more comorbidities, fewer people in the lowest BMI band, and was whiter. Those residual gaps were adjusted for in one analysis and folded into weights in the other.

The deeper problem is selection into treatment itself. To appear in the treated arm you had to obtain a paid pharmacy claim for branded Zepbound during a period when Medicare did not cover it and shortages were common. That is a filter for insurance generosity and administrative persistence, and no covariate in this model captures it. Anyone who bought vials directly from Lilly is invisible in the data and may be sitting in the control arm.

Internal age balance is excellent. Mean 64.5 years treated against 64.7 control, standard deviations of 6.8 and 7.0. Nothing to complain about in the numbers themselves.

How that balance was produced is more interesting, and it is visible only in the attrition figure. Matching was not performed on this cohort. It was performed on the full adult population, 171,925 treated against an eligible control pool of 5,009,184, a ratio of roughly twenty-nine to one. That produced 171,299 matched pairs. The over-55 cohort was then carved out by keeping only pairs in which both members were over 55, which left 15,843 pairs, about nine percent of the matched set.

Age was a covariate in the propensity model, so this is a defensible way to get a balanced subgroup, and the paper calls it pre-specified. But the calipers were tuned to the whole adult distribution, not to older adults, and the both-members rule discards precisely the pairs that straddle the age boundary. The balance is a byproduct rather than a design target.

The cohort is defined as over 55 with a mean of 64.5, so roughly half of it is younger than 65. The Bridge program serves beneficiaries 65 and older plus certain disabled enrollees. Only 38.5 percent of the treated arm had Medicare at all. A majority, 55.4 percent, were commercially insured.

The eligibility criteria diverge further. Bridge excludes people with type 2 diabetes, moderate-to-severe sleep apnea, and MASH, on the logic that those conditions already have covered indications. This study also excluded diabetes, which lines up. But 31.5 percent of the treated arm had sleep apnea, a group Bridge largely routes elsewhere. Bridge also covers only the tirzepatide KwikPen and applies tiered BMI and comorbidity requirements.

So the population that generated the $319 is younger than the Bridge population, mostly not on Medicare, and includes a substantial slice that Bridge would not enroll. The paper's own limitation section concedes the findings cannot be extrapolated beyond this age-selected subgroup, then extrapolates them to a program with different age criteria.

Total cost is medical cost plus pharmacy cost. Split the final window into its two parts and both fall apart.

In the primary analysis the two components sum to exactly $319. Medical does not reach significance. Pharmacy lands precisely on the threshold. Their sum is comfortably significant. In the secondary analysis neither component clears p=0.05 and the total does.

Nearly half the $607 is pharmacy spending, and pharmacy spending here excludes tirzepatide by design. So the claim embedded in that figure is that starting tirzepatide made the untreated comparison group's other drug costs explode.

In the secondary analysis the control arm's pharmacy cost runs $191, $204, $227, then $477. It more than doubles in the final window, at n=793. The treated arm sits flat at $221. Under the primary analysis the same control-arm series is $207, $215, $235, $297, a rise of 26 percent rather than 110 percent.

Two censoring schemes applied to the same underlying people produce control-arm pharmacy trajectories that differ by a factor of four in their final step. That is the signature of a handful of expensive specialty claims landing in a small denominator, not of a treatment effect.

One more thing the discussion never mentions. At three to six months the treated arm's pharmacy costs were higher by $34 per month, and that one was statistically significant at p=0.046.

Lilly's explanation for the savings is fewer hospitalizations. The primary analysis supports it: inpatient and emergency visit rates give incidence rate ratios of 0.86, 0.76, and 0.69 across the three windows, all significant.

Same outcome, same data, other censoring scheme. IRRs of 0.87 (p=0.101), 0.83 (p=0.065), and then 0.99 with a confidence interval of 0.60 to 1.64 and p=0.980.

At twelve to eighteen months the two arms are indistinguishable. The control arm's rate falls from 34 to 26 per thousand person-months and meets the treated arm exactly. The difference is two visits per thousand person-months.

Now hold that against the cost result from the same analysis and the same window. The saving is $607, the largest figure in the paper and the one Lilly's announcement leads with. So in the secondary analysis the biggest cost saving coincides with a complete absence of the effect that is supposed to be causing it.

Those two results cannot both be describing tirzepatide keeping older adults out of the hospital. One of them is noise, and the small denominator makes it hard to say which.