D4.4 – Report describing the tool and some retrospective applications for validation

Before a health system decides to fund a new medicine, it needs to know not just how much it will cost, but when those costs will fall. A budget that looks affordable in year one can become a strain by year three, as more patients start the treatment and those already on it live longer. Predicting this changing pattern over time is difficult, and most standard forecasting methods simplify it away. This report sets out a new and more flexible way of making these predictions, and tests whether it actually works by applying it to a real case.

The method builds a budget impact analysis around a Markov model, a technique common in the economic evaluation of health technologies but rarely used for budget forecasting. Its central idea is to track not only how far ahead we are looking, but also how long each patient has been on a given treatment. This lets the model capture things that ordinary analyses miss, such as costs that change with treatment duration, populations that grow as survival improves, and payment by results agreements where the manufacturer is not paid if a patient stops treatment early. The report first presents a general version of the model, then a data parsimonious version that needs only the information a payer typically has when a product reaches the market, such as prevalence, incidence, median survival and treatment cost. It was validated retrospectively against real regional expenditure in the Veneto Region of Italy, using the cancer drug Lenvatinib, and the model’s forecast tracked actual spending well over a 36 month horizon. A worked hypothetical example then shows how the tool is used in practice and what its outputs reveal.

What the deliverable contains:

• A general Markov based framework for budget impact analysis that accounts for both forecasting time and treatment time
• A data parsimonious version requiring only inputs usually available to decision makers at launch
• A retrospective validation using Lenvatinib for liver cancer, comparing forecast against actual regional expenditure in Veneto
• A worked example with multiple treatment sequences, illustrating inputs, probabilistic sensitivity analysis and model outputs
• A discussion of the approach’s strengths and its limits, including where a disease specific model remains necessary