Simulating a Health Technology Assessment submission for market access — from pivotal-trial evidence to reimbursement decision
Every HTA submission begins as a structured evidence package. Pivotal Phase III trial data must be transformed from raw case report forms into standardized, machine-readable datasets, then reorganized around the PICOS framework so reviewers at NICE, ICER, G-BA, or CADTH can evaluate exactly the population, intervention, comparator, and outcomes relevant to their jurisdiction. A single omitted subgroup or mismatched comparator can stall an entire submission cycle.
Regulatory-grade and HTA-grade evidence share a common data backbone even though the audiences differ:
CDISC SDTM (Study Data Tabulation Model): • Raw case report form data mapped to standardized domains — DM (demographics), AE (adverse events), LB (labs), EX (exposure), RS (response), TU/TR (tumor assessments in oncology) • Controlled terminology enforced via NCI/CDISC codelists; MedDRA used for AE coding, WHO Drug Dictionary for concomitant medications • ADaM (Analysis Data Model) datasets derived from SDTM to directly support the statistical analysis plan
ICH harmonization underpinning the whole pipeline: • ICH E6(R2)/(R3): Good Clinical Practice — data integrity and audit trail requirements • ICH E9: statistical principles for clinical trials, including estimands framework (what treatment effect is actually being estimated: treatment policy, hypothetical, principal stratum) • ICH E2B: individual case safety report (ICSR) format feeding pharmacovigilance databases (FDA FAERS, EMA EudraVigilance)
PICOS scoping for HTA (distinct from regulatory scoping): • Population: often narrower or broader than the regulatory label — payers frequently want subgroup evidence regulatory approval did not require • Comparator: must be the jurisdiction's standard of care, not necessarily the trial's comparator arm — this comparator mismatch is the single most common trigger for supplementary ITC work • Outcomes: HTA bodies weight quality-of-life instruments (EQ-5D-5L, SF-36) and overall survival more heavily than surrogate endpoints regulators may have accepted for approval
A dossier built without this translation step is rejected at intake — reviewers cannot map trial endpoints onto the economic model's required inputs.
Randomized trials answer efficacy under ideal conditions; HTA bodies increasingly demand effectiveness evidence from real-world settings:
• Claims data (e.g., US commercial/Medicare claims, German AOK sickness-fund data): captures actual utilization, adherence, and downstream costs at population scale • Disease registries: longitudinal, prospectively collected outcomes in unselected patients — critical for rare diseases where RCT power is inherently limited • Electronic health record (EHR) mining: enables construction of external/synthetic control arms when a randomized comparator arm is ethically or practically infeasible
Methodological rigor for RWE increasingly follows FDA's RWE Framework and EMA's DARWIN EU initiative, both of which require pre-specified analysis plans, active-comparator new-user designs, and negative-control outcome testing to rule out confounding before RWE is accepted alongside RCT data in a submission.
Reimbursement decisions require a treatment effect versus the payer's actual standard of care — which the pivotal trial frequently did not test directly. Network meta-analysis and indirect treatment comparison methods borrow statistical strength across a web of related trials, connected through shared comparator arms, to estimate that missing head-to-head effect.
Two complementary approaches dominate HTA-grade indirect comparison:
Bucher method (anchored, simple triangle): • Requires only three trials sharing one common comparator (e.g., Drug A vs. Placebo, Drug B vs. Placebo) • Indirect estimate: log(HR_AB) = log(HR_A vs C) − log(HR_B vs C), variance summed • Preserves within-trial randomization — the "anchored" indirect comparison is far more defensible than naive cross-trial comparison • Limitation: only works with a single shared comparator and cannot pool multiple trials per treatment efficiently
Bayesian network meta-analysis (multi-arm evidence network): • Models all trials simultaneously in a network diagram — nodes are treatments, edges are trials directly comparing them • Consistency assumption: direct and indirect evidence for any given comparison should agree; node-splitting and design-by-treatment interaction models test this • Random-effects model with vague priors (e.g., Normal(0, 10000) on treatment effects, half-Normal or Uniform priors on heterogeneity variance τ²) run via MCMC (WinBUGS, JAGS, or Stan) for 20,000+ iterations after burn-in • Outputs: relative effect estimates, 95% credible intervals, and surface under the cumulative ranking curve (SUCRA) scores ranking all treatments
NICE Decision Support Unit (DSU) Technical Support Documents (TSD 1–7) are the de facto international standard for how these networks must be specified, validated, and reported in a submission.
When no evidence network connects the intervention to the payer's comparator at all, population-adjustment methods are required:
• MAIC (Matched-Adjusted Indirect Comparison): individual patient data (IPD) from the sponsor's own trial are reweighted by propensity/entropy balancing so their aggregate baseline characteristics match the published aggregate data of the comparator trial, then outcomes are compared in the reweighted population • STC (Simulated Treatment Comparison): a regression model fit on the IPD trial predicts outcomes at the comparator trial's aggregate covariate values • ML-NMR (Multilevel Network Meta-Regression): extends population adjustment to a full evidence network rather than a single pairwise comparison, increasingly favored by NICE DSU TSD 18 (2020)
All unanchored comparisons (MAIC/STC without a common comparator) carry a strong caveat: they cannot control for unmeasured confounding, and HTA reviewers systematically discount their certainty relative to anchored comparisons — often dropping the evidence grade by one full GRADE level.
In its 2022 lorlatinib appraisal, NICE accepted a MAIC-derived indirect comparison against crizotinib in ALK-positive NSCLC only after the sponsor demonstrated covariate overlap (propensity score range) exceeding 80% and ran extensive scenario analyses varying which prognostic factors were included in the matching — illustrating how heavily committees scrutinize population-adjusted evidence before it can support a positive recommendation.
The economic model is the mathematical engine of the dossier: it converts clinical evidence into the single number most HTA bodies weigh most heavily — the incremental cost-effectiveness ratio (ICER), expressed as cost per quality-adjusted life-year (QALY) gained. Every methodological choice inside the model — time horizon, discount rate, utility source — can move the ICER by tens of thousands of dollars.
Two model architectures dominate oncology and chronic-disease HTA submissions:
Partitioned-survival model (PSM): • Health states (e.g., progression-free, progressed, dead) are defined directly by extrapolated Kaplan-Meier survival curves — progression-free survival (PFS) and overall survival (OS) — rather than by explicit transition probabilities • State membership at time t = area between the extrapolated PFS and OS curves • Parametric extrapolation candidates: exponential, Weibull, log-normal, log-logistic, generalized gamma — selected via Akaike/Bayesian Information Criterion (AIC/BIC) and clinical plausibility of the extrapolated tail • Dominant structure in oncology submissions because it maps directly onto trial-reported PFS/OS endpoints
Markov cohort model: • Explicit transition probability matrix between health states each cycle (commonly 1-, 3-, or 6-month cycles) • Half-cycle correction applied to avoid systematic bias from discrete-time approximation of continuous events • Preferred when disease has well-characterized recurrent or bidirectional states (relapsing-remitting conditions, cardiovascular event models)
Lifetime horizon (typically 20–40 years or "until 100% cohort death") is standard for both, since QALY and cost differences that emerge only after the trial's observed follow-up period frequently drive the majority of the modeled ICER — placing enormous weight on the chosen extrapolation method.
QALY = years lived × health-state utility (0 = dead, 1 = perfect health), summed and discounted across the model horizon:
Utility sourcing hierarchy (most to least preferred by NICE/ICER reference cases): 1. Trial-collected EQ-5D-5L mapped to a country-specific value set (e.g., UK, US, or Germany-specific EQ-5D tariffs — utilities are NOT transferable across countries without justification) 2. Vignette-based utilities from time-trade-off or standard-gamble studies when trial-collected EQ-5D is unavailable for a health state 3. Published literature utilities, used only with justification and scenario testing
Discounting: NICE reference case specifies 3.5%/year on both costs and QALYs; ICER (US) and most US payers use 3%/year; discounting compounds heavily over a 30-year horizon — a $50,000 cost in year 30 is worth roughly $17,800 in year-0 present value at 3.5%.
Probabilistic sensitivity analysis (PSA): • Every model input (transition probabilities, utilities, costs, hazard ratios) is assigned a distribution (Beta for probabilities/utilities, Gamma or log-normal for costs, log-normal for hazard ratios) reflecting its own estimated uncertainty • 10,000 Monte Carlo draws propagate joint parameter uncertainty through the full model, generating a cost-effectiveness plane (incremental cost vs. incremental QALY per iteration) and a cost-effectiveness acceptability curve (CEAC) — the proportion of iterations cost-effective at each possible willingness-to-pay threshold
CADTH's 2023 review of the same drug class found that switching only the survival-extrapolation function from generalized gamma to a piecewise (spline-based) model reduced the base-case ICER by over $38,000/QALY — with zero change to the underlying trial data. Extrapolation-method sensitivity is now a standard scrutiny point in nearly every appraisal committee meeting.
Cost-effectiveness answers whether a technology is worth its price at the margin; budget impact analysis (BIA) answers a separate, equally decisive question — can the payer actually afford it at population scale? A therapy can be judged highly cost-effective and still be delayed or restricted purely on affordability grounds, particularly for high-prevalence indications.
Unlike the cost-effectiveness model, BIA uses a static, short (typically 3–5 year), undiscounted, payer-perspective framework governed by the ISPOR Task Force good-practice recommendations (Mauskopf et al., 2007; Sullivan et al., 2014):
Core structure: • Eligible population estimation: total covered population × disease prevalence/incidence × diagnosis rate × treatment-eligible fraction — each multiplier independently sourced and justified • "World without" scenario: projected costs if the new technology is NOT introduced, using current standard-of-care market shares • "World with" scenario: projected costs after launch, using a market-uptake curve (typically S-shaped, peaking at a plateau share by year 3–5) • Net budget impact = World-with costs − World-without costs, reported both cumulatively and per-year
Cost components explicitly modeled on both sides: • Drug acquisition cost net of confidential rebates/discounts where permitted to model • Administration costs (infusion chair time, monitoring) • Offset costs — avoided hospitalizations, avoided disease progression events, avoided use of the displaced comparator therapy
Per-member-per-month (PMPM) impact is the metric US payers (commercial, Medicare Advantage, Medicaid MCOs) use operationally to compare a new technology's affordability against premium-setting assumptions — a PMPM impact above roughly $0.10–$0.50 for a single product is enough to trigger formulary-tier restriction or step-therapy requirements even for a cost-effective drug.
Market-uptake assumptions are frequently the single most contested input in a BIA review:
• Sponsors typically model faster uptake (aggressive S-curve, peak share reached by year 2–3) to demonstrate near-term clinical value • Payers typically stress-test slower uptake and higher eligible-population estimates to expose worst-case affordability exposure • Scenario analyses required: alternative price/rebate assumptions, alternative eligible-population definitions (broad label vs. narrow trial population), and alternative comparator mix
Many jurisdictions pair BIA output with explicit affordability mechanisms: • CMS (US) increasingly references budget-impact evidence in coverage decisions and in negotiated maximum fair prices under the Inflation Reduction Act drug price negotiation program • England's NICE applies a Budget Impact Test: if the technology's cost to the NHS exceeds £20 million in any of its first three years, commercial discussion with NHS England is triggered before routine funding is mandated, regardless of a favorable ICER • Germany's AMNOG (G-BA) process ties the annual price negotiation directly to realized sales volume once uptake data become available, effectively converting the BIA into a live price-adjustment mechanism
All prior workstreams converge at a single meeting: an independent, multidisciplinary appraisal committee reviews the full dossier, questions the sponsor's modeling choices directly, and issues a recommendation. This is simultaneously a scientific and a deliberative-democratic process — committees explicitly weigh evidence certainty, unmet need, and societal values alongside the arithmetic of the ICER.
HTA bodies increasingly formalize how much confidence to place in a body of evidence using the GRADE framework (Grading of Recommendations Assessment, Development and Evaluation):
Starting point: randomized evidence starts as "High" certainty; observational/RWE evidence starts as "Low"
Downgrading factors (each can drop certainty by one or two levels): • Risk of bias — open-label design, high dropout, unclear allocation concealment • Indirectness — population, comparator, or outcome in the trial differs from the decision question (the single most common downgrade trigger for ITC/MAIC-based comparisons) • Inconsistency — conflicting results across trials in a network, high I² heterogeneity • Imprecision — wide confidence/credible intervals crossing clinically meaningful thresholds • Publication bias — asymmetric funnel plots, selective outcome reporting
Upgrading factors (rare, mainly for observational evidence): large effect magnitude, dose-response gradient, or effect estimate that would be strengthened by accounting for plausible residual confounding.
Final certainty ratings — High, Moderate, Low, Very Low — are reported per outcome, not as a single dossier-wide score, and directly shape how much weight the committee gives the modeled ICER versus how much additional discount or evidence-development condition it attaches to a positive recommendation.
Appraisal committees weigh several dimensions simultaneously, not just whether the ICER sits below the willingness-to-pay threshold:
• Severity/end-of-life modifiers: NICE applies a QALY weighting (1.2×–1.7×) for end-of-life criteria (short life expectancy, meaningful life extension), effectively raising the acceptable threshold from £20,000–30,000 to as high as £50,000/QALY for qualifying terminal conditions • Unmet need and innovation: therapies addressing conditions with no existing effective treatment receive qualitative weight even when quantitative uncertainty is higher • Equity considerations: distributional cost-effectiveness analysis (DCEA) increasingly assesses whether a technology narrows or widens health inequalities across socioeconomic strata • Patient and clinical expert testimony: formal patient-organization submissions and clinical expert statements are entered into the record and can materially shift committee deliberation, particularly on quality-of-life dimensions the quantitative model may understate • Confidential commercial arrangements: many "rejected on cost-effectiveness grounds" appraisals are resolved through a subsequent confidential patient access scheme (PAS) or managed-entry price discount, re-presented to the committee at a materially lower net price
NICE's 2023 appraisal of a CAR-T therapy initially returned a deterministic ICER above £100,000/QALY, exceeding even the end-of-life-adjusted threshold — the committee issued a preliminary "not recommended" decision. A subsequent confidential commercial access agreement lowered the effective net price sufficiently that the resubmitted ICER fell under the modified threshold, and the technology was approved on appeal roughly five months later — illustrating how commercial negotiation, not new clinical evidence, frequently resolves a borderline appraisal.
A positive appraisal is rarely unconditional. Increasingly, HTA bodies issue conditional reimbursement bound to a managed entry agreement (MEA) — coverage with evidence development — that requires the sponsor to keep generating real-world data after launch, feeding directly into pharmacovigilance and re-appraisal cycles that can revise, confirm, or withdraw the original decision.
Managed entry agreements (MEAs) let payers approve a technology despite residual uncertainty, by shifting financial risk back onto the manufacturer:
Financial-based MEAs (majority of schemes, simpler to administer): • Confidential discounts / patient access schemes (PAS): net price reduced below list price, invisible to public ICER reporting • Price-volume agreements: unit price steps down once cumulative sales volume crosses defined thresholds — directly linking the budget impact model's uptake projection to the final negotiated price (core mechanism of Germany's AMNOG) • Dose/vial capping, free-goods schemes for early cycles
Outcomes-based MEAs (rarer, operationally demanding, used when uncertainty is clinical rather than purely financial): • Coverage with evidence development (CED): CMS's primary US mechanism — reimbursement continues only while the sponsor conducts a specified post-launch study (often a registry) reporting predefined outcomes • Pay-for-performance: rebate triggered if a patient does not achieve a predefined response threshold (used for several CAR-T and gene-therapy launches given single-administration, high-cost profiles) • Annuity/outcomes-based installment models for one-time curative gene therapies, spreading payment across years contingent on durability of response
These schemes are explicitly designed to convert an appraisal committee's "maybe" into a conditional "yes" — trading a lower initial evidence bar for binding downstream accountability.
Once reimbursed, the technology enters routine practice under continuous safety and effectiveness monitoring that closes the loop back into the HTA evidence base:
• Risk Management Plan (RMP): required under EMA GVP Module V, specifying additional pharmacovigilance activities (registries, targeted follow-up studies) and risk-minimization measures beyond routine spontaneous reporting • Signal detection: EMA EudraVigilance and FDA FAERS continuously screen individual case safety reports using disproportionality methods (e.g., Empirical Bayes Geometric Mean, proportional reporting ratio) to flag emerging safety signals against the established benefit-risk profile • PASS (Post-Authorisation Safety Study) / PAES (Post-Authorisation Efficacy Study): formally registered in the EU PAS register, with results feeding directly into periodic safety update reports (PSURs) reviewed by the EMA Pharmacovigilance Risk Assessment Committee (PRAC) • Re-appraisal trigger: if accumulated RWE shifts the effectiveness or safety profile enough to move the re-estimated ICER by a materially significant margin (commonly a working threshold near ±15%), the HTA body can reopen the appraisal — confirming, narrowing, or in rare cases withdrawing the original reimbursement decision
This closes the dossier lifecycle: the evidence package assembled in Stage 1 is never truly final — coverage with evidence development treats market launch as the start of a new, real-world evidence-generation study rather than the end of the assessment process.
England's Cancer Drugs Fund functions as a large-scale, standing managed-access scheme: technologies with promising but immature survival data are funded conditionally for a defined data-collection period (typically up to 24 months) while real-world outcomes accrue, after which NICE issues a final appraisal determination using the combined trial-plus-RWE evidence base — over 80 oncology indications have passed through this route since its 2016 relaunch.