ISPOR-SMDM Modeling Good Research Practices

What it is

— The ISPOR-SMDM Modeling Good Research Practices series is a seven-part set of consensus reports issued jointly in 2012 by the International Society for Pharmacoeconomics and Outcomes Research (ISPOR) and the Society for Medical Decision Making (SMDM), co-published in Value in Health and Medical Decision Making. It is the field's reference standard for the conduct and reporting of decision-analytic (health-economic) models. The series is organized by task force: TF-1 Overview (Caro), TF-2 Conceptualizing the model (Roberts), TF-3 State-transition/Markov models (Siebert), TF-4 Discrete-event simulation (Karnon), TF-5 Dynamic transmission models (Pitman), TF-6 Parameter estimation and uncertainty (Briggs), and TF-7 Model transparency and validation (Eddy). It is a good-practices and modeling-reporting framework — not an observational-study reporting checklist (STROBE/RECORD) and not a risk-of-bias tool. Its modern companions are CHEERS 2022 (reporting the economic evaluation that wraps the model) and the AdViSHE validation-reporting tool; ISPOR maintains the series under its Good Practices program.

When to use

— Use this series whenever a decision-analytic model generates the comparative cost and health-outcome estimates in an HTA/payer dossier (NICE, ICER, CADTH, G-BA, HAS), a manufacturer global value dossier, or a peer-reviewed cost-effectiveness/cost-utility, cost-benefit, or budget-impact analysis. The choice of which report governs is driven by model architecture: cohort state-transition/Markov model with memoryless health states → TF-3; need for patient history, time-since-event, or many interacting attributes that explode a Markov state space → microsimulation/DES under TF-4; infectious-disease or herd-immunity questions where one person's state changes another's risk (non-linear force of infection) → dynamic transmission under TF-5. TF-2 (conceptualization) and TF-6 (parameter uncertainty) and TF-7 (validation/transparency) apply to every model regardless of type. Distinguish from siblings: use CHEERS 2022 to report the surrounding economic evaluation and ISPOR budget-impact good-practices for affordability/budget-impact specifics; ISPOR-Modeling governs the engine, not the wrapper.

What it requires

— The series enforces good practice across the model lifecycle. Problem conceptualization (TF-2): an explicit decision problem, PICO/scope, perspective, time horizon, and a model structure justified against the disease process — structure follows the problem, not software convenience. Structure and assumptions (TF-3/4/5): correct cycle length and half-cycle correction for Markov models; justification of the Markov (memoryless) assumption versus the need for individual histories in microsimulation/DES; valid handling of the force of infection and contact structure in transmission models. Parameter estimation and uncertainty (TF-6): every input traceable to a source with a defensible distribution; deterministic one-way and scenario analyses plus probabilistic sensitivity analysis (PSA) with appropriate distributional choices and correlation; uncertainty characterized as parameter, structural, and methodological. Transparency and validation (TF-7): non-technical and technical documentation sufficient for independent reproduction, plus the validation hierarchy — face validity, internal/verification (the model does what was intended; debugging, extreme-value/null tests), cross-validation against other models, external validation against data not used to build it, and predictive validation. Where real-world data supply inputs (incidence, transition probabilities, costs, utilities, treatment effects), those inputs must themselves be fit-for-purpose, with transparent estimands and confounding control — the model inherits the credibility of its parameters.

When NOT to use — limitations and common misapplications

— (1) This is a framework for models, not for primary observational analyses. Following ISPOR-Modeling perfectly does not validate the upstream effectiveness estimate fed into the model; if a hazard ratio from claims is confounded, a beautifully validated Markov model launders bias into a decision. (2) It is not a risk-of-bias instrument or a quality score — there is no numeric grade; a model can satisfy every reporting item and still rest on an indefensible structure or cherry-picked inputs. (3) Validation-as-theater: reporting "the model was validated" without specifying which validation (face/internal/external/predictive) and showing the comparison is a frequent failure that HTA reviewers reject. (4) Wrong architecture: forcing a cohort Markov model onto an infectious-disease question (ignoring herd effects) or onto a problem requiring patient memory violates TF-3/4/5 and produces structurally biased results. (5) PSA omitted or cosmetic: point estimates with no probabilistic uncertainty, or PSA with arbitrary distributions and no parameter correlation, fail TF-6. (6) Do not substitute this series for CHEERS when the deliverable is the economic-evaluation manuscript — they are complementary, not interchangeable.

How it maps to this catalog

— The model engine and its uncertainty/validation requirements map to: markov-transition-probabilities-rwe (TF-3 state-transition structure, cycle length, transition-probability estimation from RWD), discrete-event-simulation-rwe (TF-4 DES/individual simulation), partitioned-survival-models-rwe and survival-extrapolation-hta-rwe (oncology model structures and the long-horizon extrapolation TF-2/TF-7 demand be justified and validated), health-economic-modeling-methods-rwe (umbrella modeling concept), probabilistic-sensitivity-analysis-hea-rwe (TF-6 PSA), discounting-costs-effects-rwe, qaly-utility-mapping-rwe, and icer-net-monetary-benefit-rwe (outputs the model produces); affordability questions route to budget-impact. For the parameters the model consumes, the fitness and causal-credibility requirements map to fit-for-purpose-data-assessment-rwe, estimands-ate-att-intercurrent-events-rwe, high-dimensional-propensity-score-hdps-rwe and active-comparator-new-user (defensible treatment-effect inputs), target-trial-emulation (when the effectiveness input comes from an emulated trial), diagnosis-phenotype-algorithm-1ip-2op-time-window-rwe and claims-analysis (where incidence, event rates, and resource-use parameters are estimated from claims/EHR), and attrition-and-loss-to-follow-up-rwe (so survival and transition inputs are not distorted by informative censoring). The wrapping economic-evaluation report is governed by cost-effectiveness, cost-utility, and CHEERS.

Applied note (claims/EHR/registry RWE)

When transition probabilities, time-to-event curves, costs (PPPM/PPPY), or utility decrements are estimated from routinely collected data and piped into a Markov or partitioned-survival model, treat each input as a small RWE study in its own right: pre-specify the phenotype and time-zero, control confounding for any comparative-effectiveness input, and carry that input's sampling uncertainty (and, ideally, structural uncertainty over alternative algorithms) into the TF-6 PSA — not just a fixed mean. Document the data source's fitness and known payer-specific quirks (e.g., Medicare FFS vs MA capture) so a reviewer can trace every model number back to an observable, defensible source.