Flexible generic framework for evidence synthesis in health technology assessment.

Hamza, Tasnim (2023). Flexible generic framework for evidence synthesis in health technology assessment. (Unpublished). (Dissertation, University of Bern, the Faculty of Medicine and the Faculty of Human Sciences)

Full text not available from this repository. (Request a copy)

To choose appropriate interventions for treating specific health conditions, all competing interventions need consideration. Network meta-analysis (NMA) methods compare multiple interventions simultaneously by synthesizing direct and indirect evidence from a network of studies. The included studies are more likely to vary in many important factors—some of which are related to study participants—and alter the relative effects. To account for these factors and avoid aggregation bias, individual participant data (IPD) from studies should be used rather than aggregate data (AD). However, it is common to only access IPD from a subset of studies and extract the rest from published reports as summary information—the AD. The three-level hierarchical model combines IPD and AD studies. In terms of study design, NMA data is usually derived from randomized clinical trials (RCT). However, despite their many advantages, RCT are prone to bias, include highly selected participants, and rarely compare active drugs. Non-randomized studies (NRS) overcome some of these limitations, yet their bias—in addition to RCT bias—needs acknowledged. To find optimal doses of pharmacological agents, dose-response analysis can estimate the curve that relates outcomes to doses. A single study rarely estimates a precise dose-response curve (DRC) to draw comprehensive conclusions. When aggregated-level data are available from multiple studies, dose-response models are combined across studies in pairwise metaanalysis (PMA) using frequentist one-stage or two-stage methods. However, the two-stage model discards studies that do not report as many dose levels—other than reference doses—as dose-response parameters. The one-stage model assigns a normal distribution to the summary measure of effect, such as log odds ratio. Such normal assumption requires approximating the variance-covariance structure to reflect correlations across dose levels, yet small studies can violate the assumption. In NMA contexts, considering the effect of doses reduces betweenstudy heterogeneity and between-comparison inconsistency. The dose-response relationship in NMA models has been expressed using the Emax function. However, Emax poorly reflects the dose-response relationship and requires five dose levels to estimate its four parameters.

Aim
In my thesis, I sought to develop new extensions to existing NMA models for combining different data types or including dose-response models. To achieve my goal, I set three key objectives: (1) building a single NMA model for combining IPD and AD studies from RCT and NRS and implementing these models in R; (2) developing a Bayesian dose-response model for PMA and investigating the performance of the different models in simulations; and (3) broadening the Bayesian model to include dose-response model in NMA.

Methods
I built a suite of Bayesian models by integrating four different approaches for synthesizing RCT and NRS evidence into a three-level hierarchical model for combining IPD and AD. These four approaches include: (1) the unadjusted model, which naïvely treats RCT and NRS equally; (2) the NMA model with RCT, which uses priors constructed from NRS evidence, while allowing for down-weighting or shifting of NRS contribution; (3) and (4) these models incorporate the overall assessment of risk of bias in a meta-regression model, where relative treatment effects and bias effects are combined across studies either separately using different univariate normal models (bias-adjusted model 1) or simultaneously through a mixture of two normal distributions (bias-adjusted model 2). I then implemented the different models in a new JAGS-based R package. To demonstrate the four modelled approaches, I used a network of three pharmacological therapies and placebo for treating individuals with relapsing-remitting multiple sclerosis (RRMS) along with analyzing a network of 21 antidepressants and placebo. For dose-response analysis, I first reviewed available meta-analysis models. I then formulated the one-stage frequentist model in a Bayesian setting with normal and binomial likelihoods. Following that, I expanded the model to dose-response NMA (DR-NMA) and built two further extensions to include dose-response network meta-regression (DR-NMR) and the cluster-effect model. I examined the performance of the one-stage frequentist model and the new Bayesian dose-response meta-analysis models in a simulation study with various scenarios. I also employed these Bayesian and frequentist models to a dataset of five serotoninspecific reuptake inhibitor (SSRI) antidepressants and placebo. I then analyzed the extended antidepressants dataset with 21 drugs and placebo using the DR-NMA model and the other two extensions of the model.

Results
When I included the bias risk in RRMS and antidepressants analysis, the estimated relative treatment effects did not materially change. In RRMS analysis, the results of NMR indicate that intervention efficacy declines as participant age increases. I found the Bayesian doseresponse meta-analysis model with binomial likelihood provides lower bias results than the frequentist one-stage model and Bayesian model with normal likelihood when studies have a small sample size. When the true shape is half-sigmoid or log-log, I observed the position of knots determines how close the true shapes to the estimated restricted cubic spline model. In all other scenarios, all models fit well and produced essentially the same results. The estimated DRCs for SSRI antidepressants indicates that the efficacy increases between zero and 30–40 mg of fluoxetine-equivalent dose and gradually decreases beyond that. Using the extended antidepressants dataset, I found doses beyond where each antidepressant efficacy no longer improves and the efficacy is greater than placebo effect. When covariates are included in the model, small sample size studies tend to overestimate antidepressant efficacy for many drugs.

Conclusions
I propose two major extensions to the standard Bayesian NMA model. The first extension enables the inclusion of studies of different formats, designs, and bias risks, while accounting for important study and participant factors. Besides accounting for study-level covariates and analyzing agents within classes, the second extension allows for simultaneous estimation of the dose-response model with multiple agents. The new extensions and tools enable the inclusion of more data, which has the great potential to address unresolved issues affecting patients, healthcare providers, health technology assessment agencies, and regulatory bodies.

Item Type:	Thesis (Dissertation)
Division/Institute:	04 Faculty of Medicine > Pre-clinic Human Medicine > Institute of Social and Preventive Medicine (ISPM)
Graduate School:	Graduate School for Health Sciences (GHS)
UniBE Contributor:	Hamza, Tasnim A. A., Salanti, Georgia
Subjects:	600 Technology > 610 Medicine & health 300 Social sciences, sociology & anthropology > 360 Social problems & social services
Language:	English
Submitter:	Doris Kopp Heim
Date Deposited:	22 Dec 2023 16:16
Last Modified:	22 Dec 2023 16:16
Additional Information:	PhD in Health Sciences (Epidemiology and Biostatistics)
URI:	https://boris.unibe.ch/id/eprint/190720