Strong Cross-Pillar

Surrogate Endpoints vs Real Outcomes: When the Marker Moves but You Don't Get Better

Name: Surrogate Endpoints vs Real Outcomes: When the Marker Moves but You Don't Get Better
Creator: Realised
Published: 2026-06-01

Summary

A surrogate endpoint is a lab marker (LDL, HbA1c, bone density, blood pressure, tumour size) used as a stand-in for the thing you actually care about (heart attack, fracture, living longer, living better) — and the recurring, sometimes lethal lesson of medicine is that the marker can move in the "right" direction while patients do worse. Surrogates are genuinely useful and a few are validated, but a moved marker is a hypothesis, not a proven benefit, until it's tested against hard outcomes.

Why Strong

Strong Evidence. The canonical cases (CAST, torcetrapib, rosiglitazone, IMPROVE-IT) are landmark trials with exact magnitudes from NEJM/Lancet primary sources; the validation standard (Prentice criteria) is the textbook definition; the oncology surrogacy gap is documented across convergent systematic reviews. The governing principle — a moved surrogate is a hypothesis until validated against hard outcomes — is uncontested in evidence-based medicine. Confidence dips only on precise magnitudes of the more recent oncology summary statistics, which vary by drug class and review.

Practical takeaway

When a study or headline reports that a marker improved, run three checks:
• Name the endpoint. Was this a hard outcome (heart attack, fracture, death, quality of life) or a surrogate (a lab value, a scan, a score)? If you can't tell, assume surrogate — that's usually what gets reported first.
• Ask if the surrogate is validated for THIS intervention. LDL and blood pressure are broadly validated. Most others are not, and even validated ones can fail when a new mechanism moves them. "The marker improved" plus "validated surrogate, familiar mechanism" earns provisional trust; "the marker improved" alone earns a hypothesis flag, not a behaviour change.
• Look for the confirmatory outcome trial — or its absence. In oncology especially, ask whether the overall-survival data exist yet. "Approved on progression-free survival, survival data pending" is a different claim from "shown to extend life."

For your own tracking: a personal marker moving (HbA1c, resting heart rate, HRV, weight, blood pressure) is a real, encouraging signal worth following. The surrogate lesson doesn't tell you to dismiss it — it tells you to keep asking whether moving that marker, by that means, actually delivers the outcome you're after. Realised reads marker improvements as signals to confirm, never as the destination.

Evidence detail

Why This Entry Exists

Most studies you'll read, and most drug approvals, rest on a surrogate, because waiting decades for mortality data is slow and expensive. That's often reasonable. But the assumption baked into a surrogate — "if the marker improves, the patient improves" — is exactly the kind of plausible causal story that biology loves to break. The history here is not abstract: people died because a marker moved as designed and everyone trusted it.

This is a spoke under the evidence-literacy hub (rct_vs_observational_evidence). The hub covers study design (RCT vs observational); this entry owns one specific trap inside even good RCTs: what was actually measured, and does moving it reliably buy the outcome you want?

What bad advice this protects against, in all directions:
• "It lowers cholesterol / sugar / blood pressure, so it works." → treating any marker move as proven benefit.
• "This new drug shrinks tumours, so it extends life." → confusing a scan with survival.
• Over-cynicism: "surrogates are all lies, ignore every marker." → false; LDL and blood pressure are genuinely validated.
• Self-tracking nihilism: "my HbA1c dropping means nothing." → a moved personal marker is encouraging; the honest question is whether moving this marker by this means delivers the outcome.

EVIDENCE (the canonical cases as receipts)

1. CAST — the marker moved exactly as intended and the patients died. The Cardiac Arrhythmia Suppression Trial tested antiarrhythmic drugs (encainide, flecainide) that successfully suppressed PVCs — premature ventricular contractions, the surrogate, an arrhythmia marker assumed to predict sudden cardiac death. The drugs did their job on the marker and more than doubled death: arrhythmic death or cardiac arrest 5.7% vs 2.2% on placebo (RR 2.64, p=0.0004); all-cause death or cardiac arrest 8.3% vs 3.5% (RR 2.38, p=0.0001). The trial was stopped early in 1989. (Echt, Liebson et al., NEJM 1991; preliminary report NEJM 1989. Strong Evidence — landmark NHLBI-funded double-blind RCT, stopped early by the safety board, government-funded so no industry confound; the most-cited cautionary case in this field.)

2. Torcetrapib — a "good" lipid surrogate moving dramatically the right way while outcomes worsened. Torcetrapib raised HDL ("good cholesterol") by about 72% and lowered LDL by about 25% — a textbook-perfect lipid panel. The ILLUMINATE trial was nonetheless stopped early for excess death: 93 deaths (1.2%) vs 59 (0.8%), HR 1.58 (95% CI 1.14–2.19, p=0.006); cardiovascular events HR 1.25 (p=0.001). The harm traced to an off-target effect (raised aldosterone and blood pressure), not to HDL itself. (Barter et al., ILLUMINATE Investigators, NEJM 2007. Strong Evidence — large phase-3 RCT, ~15,000 patients, stopped early.) Both-ways note: later CETP inhibitors (anacetrapib, REVEAL) did not show this mortality signal, so the failure was torcetrapib-specific, not proof that raising HDL is universally harmful.

3. Rosiglitazone — better numbers, worse hearts. The diabetes drug improved HbA1c (the glycaemic surrogate), but a 2007 meta-analysis of 42 trials found increased myocardial infarction risk (OR 1.43, 95% CI 1.03–1.98, p=0.03) and a borderline cardiovascular-death signal. Better glucose control on paper did not translate to better cardiovascular outcomes, and the drug's use collapsed. (Nissen & Wolski, NEJM 2007. Strong-to-Moderate Evidence — meta-analysis of RCTs, but of trials not designed for cardiovascular endpoints; the wide CI and borderline p are part of the lesson.) This result directly catalysed the FDA's 2008 requirement that new diabetes drugs demonstrate cardiovascular safety, not just HbA1c lowering.

4. Oncology PFS vs OS — the largest contemporary version. Many cancer drugs are approved on surrogates — tumour response rate, progression-free survival (PFS) — that often correlate poorly with overall survival (OS) or quality of life. Reviews find only roughly a third of new cancer drugs have evidence of improving overall survival, and a large share of accelerated approvals show no confirmed survival benefit when finally tested; in some confirmatory trials the survival signal was actually unfavourable. A shrinking tumour or delayed radiographic progression is not the same as living longer or living better. (Kim & Prasad, JAMA Internal Medicine 2015; convergent accelerated-approval analyses; FDA 2025 overall-survival guidance shift. Strong Evidence — convergent systematic reviews, though magnitudes vary by tumour type.)

5. Validated surrogates DO exist — the anchor that keeps this honest. LDL cholesterol is the cleanest validated case: multiple independent mechanisms — statins, ezetimibe, PCSK9 inhibitors — all lower LDL and reduce hard cardiovascular events. IMPROVE-IT is the keystone: adding ezetimibe to a statin lowered LDL a further ~24% and produced a real outcome benefit (primary cardiovascular composite 32.7% vs 34.7%, p=0.016, NNT ~50), confirming "lower is better" for LDL via a non-statin route. (Cannon et al., IMPROVE-IT, NEJM 2015. Strong Evidence — ~18,000 patients, ~6-year follow-up.) Honest texture: the benefit was modest and there was no all-cause or cardiovascular mortality reduction — validation does not mean large benefit, only that the marker tracks the outcome.

6. The formal test — Prentice criteria. Prentice (1989) defined a valid surrogate as one where a treatment's effect on the surrogate is a valid test of its effect on the true endpoint — operationally requiring that the surrogate (1) correlate with the clinical outcome and (2) fully capture the intervention's net effect on that outcome. The second condition is the one that fails in the disasters: torcetrapib affected mortality through a pathway (off-target toxicity) the lipid panel never saw; CAST drugs killed through proarrhythmia the PVC count never saw. Correlation alone is never enough. (Prentice, Statistics in Medicine 1989; extended by Freedman et al. 1992. Strong Evidence — the foundational methodological definition.)

MECHANISM (the statistical and causal logic)

A surrogate is a bet on a causal chain: intervention → marker → outcome. The bet pays off only if two things hold. First, the marker must genuinely sit on the path to the outcome (not merely travel alongside it). Second — the harder condition — the intervention must affect the outcome only through that marker, with no side route.

The disasters are all failures of the second condition. The intervention reaches the surrogate by one mechanism and reaches the outcome by another, hidden one:
• CAST: suppressing PVCs did nothing about the drugs' own proarrhythmic toxicity, a separate lethal pathway.
• Torcetrapib: raising HDL was real, but the molecule also raised blood pressure and aldosterone — an off-target route to death the lipid panel couldn't register.
• Rosiglitazone: lowering HbA1c didn't touch (and may have worsened) cardiovascular risk through fluid retention and other effects.

This is why a new mechanism moving an old, trusted marker is the highest-danger pattern. "LDL down" earns trust when reached by statins because that route has been tested against outcomes thousands of times. The same LDL number reached by a brand-new molecule carries whatever unmeasured baggage that molecule brings. The marker looks identical; the causal chain behind it is not.

RISKS AND CONTRAINDICATIONS (the over-correction failure mode)

The vivid disasters here invite the wrong lesson, and it's worth naming the two ways this entry can be misused:
• Evidence nihilism: "surrogates are always lies, all marker-based medicine is fraud." False. Surrogates enabled statins, antihypertensives, and HIV therapy guided by viral load — enormous, real wins. LDL and blood pressure are genuinely validated. The point is don't assume validation, not assume fraud.
• Weaponising it against self-tracking: using "surrogate ≠ outcome" to dismiss every personal biomarker is a category error. A drug trial and a person watching their own HbA1c are in different epistemic situations; the principle sharpens the question ("does this marker move buy my outcome?") rather than dismissing the marker.
• Treating the corrections as flawless. Some surrogate-based approvals were overturned by contested evidence — the rosiglitazone meta-analysis had a wide confidence interval and borderline significance. The honest stance holds even here: the correction can be imperfect too. Skepticism cuts both ways.

Controversy

Nature of the dispute. Almost no one disputes that surrogates can mislead — the disasters are too well documented. The live disagreement is over how much default trust a surrogate earns and how readily drugs should be approved on one.

Position A (surrogates are useful and necessary): You cannot always wait 10–30 years for mortality data; many useful interventions would never reach patients. Some surrogates are genuinely validated across mechanisms (LDL via statins, ezetimibe, PCSK9; blood pressure across drug classes). For these, moving the marker is a reasonable proxy for benefit, and demanding mortality data for everything would freeze good medicine.

Position B (surrogates regularly and catastrophically mislead): The marker can move the right way while patients die — CAST more than doubled mortality (RR 2.38), torcetrapib raised death 58%, rosiglitazone raised MI risk, and a large fraction of oncology approvals show no survival benefit when finally tested. A surrogate is an unproven assumption about a causal chain until that chain is demonstrated against the outcome that actually matters.

Funding texture. The split is not random. Industry programs (torcetrapib, rosiglitazone) ran on surrogates; the corrective trials and meta-analyses were largely academic and regulatory (NHLBI-funded CAST, Cleveland Clinic's rosiglitazone analysis, Prasad's oncology reviews). A surrogate trial is shorter, smaller, and likelier to hit significance than a mortality trial — a structural incentive to ship on the marker.

Realised Position: A moved surrogate is a hypothesis, not a proven benefit. A validated surrogate (LDL, blood pressure) earns provisional trust; an unvalidated one earns a hypothesis flag. The danger is highest when a new mechanism moves an old marker — the same number reached by a different causal route can carry hidden harm. We read marker improvements as encouraging signals to be confirmed against hard outcomes, never as the destination. The honest stance is neither blanket trust nor blanket cynicism: surrogates are real tools that genuinely accelerate good medicine and a recurring source of confident error.

Cross-Pillar Connections

The surrogate-vs-outcome distinction governs claims in every pillar: a sleep supplement that shifts a sleep-stage number without improving how you feel or function; a training method that moves a performance marker without changing injury or longevity; a diet that moves a blood panel. The cleanest worked example lives in cholesterol_misinformation_correction (LDL as a validated-but-nuanced marker), and the self-tracking application connects to biomarker_tracking — a moved personal marker is encouraging but still subject to the same honest question.

What would change our mind

Falsifiability: explicit upgrade/downgrade criteria from source

The core claim — an unvalidated surrogate is a hypothesis until tested against hard outcomes — is essentially definitional logic, not an empirical bet, and would not be overturned. What would update the framing:
• A broad class of validated surrogates. If multi-omic or AI-derived markers were prospectively shown to fully capture treatment effects (the Prentice condition) across many drug classes, default skepticism toward those validated families could soften.
• A canonical case re-analysed. If the rosiglitazone cardiovascular signal were shown to be a pure statistical artefact (it was contested), that specific receipt would need softening — though CAST and torcetrapib are not seriously disputed.
• Regulatory reform closing the confirmatory gap. The FDA's 2025 overall-survival guidance direction, if it sticks, would shift the oncology framing from "systemic failure" toward "historical problem being corrected."

We would not change our mind on an anecdote of a surrogate that happened to track an outcome — the entry already concedes validated surrogates exist. The claim is about not assuming validation, which single successes don't refute.

Industry bias note

Structural incentives the evidence base may reflect

Strong and structural. Surrogate endpoints are the primary vehicle by which industry brings drugs to market faster and cheaper — a surrogate trial is shorter, smaller, and likelier to hit significance than a mortality trial. This creates a systematic incentive to (a) power trials on surrogates rather than hard outcomes, (b) emphasise the moved marker in marketing while confirmatory outcome data lag or never arrive (accelerated oncology approval is the clearest case), and (c) frame an unvalidated surrogate as if validation were settled. When a trial reports only a surrogate, the next question is cui_bono: who benefits from not measuring the hard outcome?

The counter-balance keeps this honest: industry also ran the trials that validated LDL against outcomes (IMPROVE-IT, the statin trials), and IMPROVE-IT's modest, mortality-null result was unflattering to its sponsor yet strengthens the validation argument. The relationship is incentive-shaped, not uniformly bad-faith — which is exactly why a structural lens (who pays, what gets measured) beats assuming villainy.

Sources (7)

Open in the Library: search, filter, every entry →