Indirect Comparisons with Population Adjustment Methods using Single-arm Studies in Health Technology Assessment

In the development process of pharmaceuticals, a frequent step is that a therapy is administered to all patients within a study; which is known as a single-arm study. A particular feature of single-arm studies is that they provide no direct estimate of treatment effects owing to the lack of a comparator arm. Therefore, estimation of treatment effects from single-arm studies involves reference to an external comparator (unanchored indirect treatment comparison). Even though single-arm studies can be completed faster than randomised control trials (RCTs), they add complexity to indirect comparisons as both prognostic and effect-modifier variables need to be balanced to obtain a valid relative treatment effect estimate.

In health technology assessment (HTA), when companies analyse their intervention treatment from a single-arm study with comparator/comparators, access to individual patient data (IPD) in all studies of interest is a rare situation as sharing of clinical data is often limited. A middle-ground situation is more realistic where the company has access to IPD for its own study and aggregate data (AgD) for the comparator studies. Moreover, the companies often have to estimate relative treatment effects of a single-arm study treatment against multiple comparator treatments in a larger disconnected network of evidence. Therefore, the fundamental objective of this thesis was to assess whether the population-adjustment method matching adjusted indirect comparison (MAIC) is suitable to implement for a larger
disconnected network of evidence or not.

This thesis starts with a review on National Institute for Health and Care Excellence (NICE) single technology appraisal (STA)s to evaluate the methods with the single-arm study. Unanchored MAIC and simulated treatment comparison (STC) were found to be frequently used methods to estimate relative treatment effects with single-arm studies. It was found that unanchored MAIC was applied multiple times to estimate the relative treatment effect of a single-arm study intervention in a larger disconnected network of evidence. The relative effect estimates from this multiple MAICs were described as if the MAIC estimates made a set of coherent relative effect estimates ignoring the fact that these estimates were from different target populations. Additionally, using the IPD several times for conducting multiple MAICs
breaks the independence of the unit of analysis assumption. In order to assess the impact of this, a simulation study was designed with multiple MAIC estimates in a fixed and a random effects network meta-analysis (NMA).

The major impact of performing an MAIC-adjusted NMA was seen in the coverage of the NMA estimates where the coverage dropped below nominal level (95%). The violation of the independence assumption together with the sandwich estimator had a repercussion on the NMA estimate coverage. The deviation from the nominal level of coverage was more pronounced for the larger compared to a smaller disconnected network of evidence. Double-bootstrapping with MAIC was found to solve the problem of undercoverage both for fixed and random effects NMA. However, the biases were found to be comparatively high with low-overlap scenarios and a smaller sample size. The proposed double-bootstrapping method was also applied in a case study with asthma. The case study illustrates how to
make multiple comparisons simultaneously using double-bootstrapped MAIC-adjusted NMA
where a correct level of coverage for the NMA estimates can be preserved with the use of
double-bootstrapping. Therefore, this thesis recommends MAIC-adjusted NMA with doublebootstrapping approach when there exists a sufficient level of overlap between studies together
with a satisfactory sample size.