OBJECTIVES: Erlotinib, a tyrosine-kinase inhibitor, has been recommended for use in non-small cell lung cancer patients harbouring an EGFR mutation. The present study is retrospective in nature; using published clinical trial data for erlotinib, it demonstrates how an analysis of Phase II data could be used to identify what data Phase III studies should focus on collecting. METHODS: Phase II data were identified through a targeted literature search and used to determine the cost-effectiveness of erlotinib, utilising a simple Markov model framework. Data from the literature were mapped to model inputs using statistical data analysis methods. Value of Information (VOI) analysis tools, such as Expected Value of Perfect Information (EVPI) and Expected Value of Partial Perfect Information (EVPPI) were used to identify those uncertain parameters having research value in a prospective Phase III program. These findings helped to optimise the design of a hypothetical Phase III trial for erlotinib that could collect data that is most relevant for reimbursement decision-making.    RESULTS: At a cost-effectiveness threshold of £30,000 per quality-adjusted life year (QALY) gained, the population EVPI was £3,269,358, indicating that further research is valuable. The EVPPI identified the log hazards of erlotinib (intervention) and gefitinib (comparator) for progression-free survival and overall survival as the parameters for which uncertainty was the most valuable. The value of the uncertainty associated with other parameters, such as utilities and costs, was much lower. Hence, subsequent studies should focus on providing further information on efficacy parameters rather than on utilities and costs.  CONCLUSIONS: Undertaking VOI analysis on data collected at Phase II can help ensure that Phase III trials are designed efficiently, in turn ensuring that uncertainty in future decision-making is minimised. This model demonstrated the VOI from a public policy perspective. This could be extended to other perspectives to ensure greater relevance in different settings.