OBJECTIVES: Digitisation/recreation of patient-level data (PLD) are frequently undertaken in HTA when ‘true’ PLD are unavailable. However, there is no published best practice guidance for digitisation, which can lead to variable methods and consequently inconsistent results. This study investigates the impact of using different digitisation techniques on survival outcomes.

METHODS: Two hypothetical case studies with ‘true’ PLD were generated to produce Kaplan-Meier (KM) estimates. Each KM estimate was digitised using the same software with four different digitisation methods: manual with minimal plotting (M-), manual with extensive plotting (M+), automated with minimal plotting (A-), and automated with extensive plotting (A+). After digitisation, pseudo-PLD were generated using the approach defined by Guyot et al., (2012), and parametric curves fitted. Extrapolated survival estimates were compared to the estimates using the ‘true’ PLD, with the choice of model informed by statistical goodness-of-fit scores, and differences in mean extrapolated survival expressed as a percentage.

RESULTS: Goodness-of-fit scores generally agreed and recommended the same model as the ‘true’ PLD for 5 out of the 8 scenarios. In the first case study with A- plotting, the discordance in survival estimates increased from 0.81% using the same parametric model to 18.14% with the parametric model chosen based on statistical goodness-of-fit scores. With the same parametric model as the ‘true’ PLD selected, the greatest difference was observed for the second case study with M- plotting (4.80%).

CONCLUSIONS: A large degree of variation in survival estimates was found dependent on the digitisation method. Lack of guidance and detailed explanation of digitisation approaches are expected to have contributed to uncertainty in the accuracy of digitised data and any associated statistical analyses. This has important implications for HTA and future research possibilities, as extrapolations of digitised data are often used to inform comparative efficacy estimates, and by extension calculations of life-years.