merlin for HTA
Statistical methods for health technology assessment — flexible parametric extrapolation, cure models, relative survival, multistate cost-effectiveness, multilevel structure for IPD meta-analysis — built on the merlin family with the people who designed it on contract.
simulated data, reproducible code
flexible parametric · cure
mixture · non-mixture
cure × RS × multilevel
Modern HTA submissions increasingly require methods that go beyond the standard partitioned survival template — flexible parametric extrapolation under multiple modelling assumptions, multistate cost-effectiveness models, cure models for long-term survival, relative survival for cancer-registry data, and rigorous handling of multi-trial IPD evidence.
The merlin family specifies and estimates these models in one coherent language. merlin fits the flexible parametric extrapolation, cure, and relative-survival models; pendragon assembles multistate cost-effectiveness models on top of them; morgana brings full Bayesian uncertainty for extrapolation under HTA-ready priors. The same diagnostic and prediction toolkit applies across all of them. (morgana is available in Stata, with R in development; pendragon is in active development in R and Stata.)
Four reasons HTA teams pick merlin
Each is a specific capability or relationship that's hard to assemble from the standard R / Stata HTA toolchain.
One framework for the whole submission
Survival extrapolation, multistate transition model, cure structure, joint biomarker model — all share the same specification language, fitting routine, and post-estimation toolkit. The dossier reads in one vocabulary.
Methods that hold up under scrutiny
merlin implements the flexible parametric methods NICE DSU TSD 14 and TSD 21 recommend, including Royston-Parmar splines. Reviewer-asked sensitivity analyses are straightforward to produce without rebuilding the model.
Reproducible, audit-ready outputs
Each case study ships with the full code used to produce it. The same code runs in R or Stata via the appropriate merlin implementation; differences are flagged. For qualified releases see merlin Validated.
Direct access to the team that built it
RDA designs and maintains merlin. For submissions where reviewer questions or non-standard model structures need a fast, authoritative response, the people who wrote the package are available to consult.
Cure × relative-survival × multilevel for cancer registry extrapolation
A combination we know of no counterpart for in any other R or Stata package — and the strongest single proof point for what merlin uniquely enables in HTA practice.
Lifetime survival extrapolation for an immuno-oncology cohort, with population-mortality offset and trial-as-cluster multilevel structure
A simulated immuno-oncology trial cohort with a 30% long-term-survivor fraction, fit as a non-mixture cure model on a Royston-Parmar baseline, with a national life-table population hazard supplied via the merlin_bhazard() helper. The multilevel extension adds trial-as-cluster random effects for IPD meta-analytic extrapolation across three trials. ICER consequences are estimated against a no-cure parametric counterfactual.
More HTA cases, same merlin engine
Each ships with reproducible code in R and (where applicable) Stata. PDF write-ups are available via the contact form.
Partitioned survival with flexible parametric extrapolation
Three-state partitioned survival model on a simulated oncology cohort, using Royston-Parmar splines with extrapolation calibrated to external registry data. Reconciled against the standard NICE DSU TSD 14 parametric set.
Multistate cost-effectiveness model
Four-state illness-death model with shared frailty on simulated cohort data, fitted as a single multistate likelihood. Transition-specific covariate effects, state occupancy probabilities for the economic model, and ICER uncertainty via integrated bootstrap.
Cure model for long-term survival in immuno-oncology
Both cure forms fitted on a simulated immuno-oncology trial with a long-term-survivor fraction. Covers covariate-dependent cure proportion estimation and the ICER impact of a cure vs no-cure parametric assumption.
Additional case studies in preparation: joint longitudinal–survival modelling of a continuous biomarker for surrogate-endpoint evaluation · anchored MAIC indirect comparisons · external validity for single-arm trials.
Short notes on recurring HTA questions
Methodological notes on topics that come up in nearly every NICE / CADTH / IQWiG / HAS submission. Each is short, code-illustrated, and aimed at analysts and reviewers.
Survival extrapolation under uncertainty
Model selection, model averaging, and how to present extrapolated estimates with credible uncertainty intervals.
Integrating external evidence
Registry data, expert elicitation, and informative priors for extrapolation calibration.
Multistate models for cost-effectiveness
When a multistate model is preferable to partitioned survival, and how to specify and validate one.
Cure models in immuno-oncology
What mixture vs non-mixture choice signals to reviewers, identification considerations, and the ICER consequences.
Relative survival for registry data
When net survival is the right quantity, life-table workflows via merlin_bhazard(), and observed-survival reporting.
Uncertainty quantification end-to-end
Propagating parameter uncertainty from estimation through to economic outputs without losing model structure.
A unified R framework for net-survival, life-expectancy, and excess-mortality analyses
Relative survival arrives with merlin v2 for R — combined with multiple-timescales support, cure models, and the merlin_bhazard() helper compatible with popEpi::popmort and relsurv::slopop — gives cancer registry and population-epidemiology teams a unified R framework for net-survival analysis, life-expectancy decomposition, excess mortality, and long-term-survivor extrapolation.
The methodological foundations span a substantial body of published work from the team: Bower / Andersson / Crowther / Lambert on life expectancy with excess mortality (American Journal of Epidemiology 2018); Bower et al. on potential life-years gained (The Breast 2019); Weibull / Lambert / Eloranta / Andersson / Dickman / Crowther on multistate models with excess hazards and multiple time-scales (Statistics in Medicine 2021); Skourlis / Crowther / Andersson / Lambert on timescale choice in competing-risks excess-hazard settings (Biometrical Journal 2022); Crowther / Royston / Clements on flexible parametric AFT with time-dependent acceleration factors (Biostatistics 2023).
For registry and population epidemiology teams
National cancer registries, IARC-affiliated groups, NORDCAN / EUROCARE collaborators, and population epidemiology teams using R for survival analysis — merlin's implementation is a direct fit for the methods this audience has been using for years, with the multilevel and cure extensions current R tools don't provide. Get in touch →
When the qualification matters too
For submissions where qualified, signed releases and validation documentation are part of the deliverable — typically pharma, HEOR, and CRO settings — a Validated engagement is in preparation — a qualified, signed release with its digest manifest, run record and requirements-to-test map, and named-expert support on contract. It is not open yet; the page takes registrations of interest.
The qualification is the wrapper. The methods support is the offering.
Validation isn't formally required by HTA agencies, but it's strong practice for serious analytical work and a near-universal requirement of internal QMS in pharma, CROs, and mature HEOR consultancies. merlin Validated supplies the documentation, the relationship, and the qualified release that make that practice straightforward. See merlin Validated →
The specification is the other half of the defence
A reviewer challenge to an extrapolation rarely turns on whether the model converged. It turns on whether the choice was made before the results existed, and whether that can be shown. We build software for the specification side of the same work: analysis plans written as a structured register rather than a prose document — the estimand, populations, endpoints and causal assumptions each a field — checked against the reporting guideline you write to, then hashed and time-stamped before any result exists.
Written for the plan, not the paper
Health economic analysis plans and SAPs come out of the same locked register, with a CHEERS 2022 conformance report behind them; a later change produces a dated amendment carrying the diff, rather than a quietly edited file. It is the tooling we use on our own submission work.