 Hi, today I'll be talking about this package BNMA which is short and for Bayesian network material analysis using JAGS. My name is Michael Saw, I have been working on this project with Professor Chris Schmidt from Brown and I'm currently a PhD student in the University of Bern Switzerland. Recently, there has been many packages of Bayesian NMA in R, including GMTC, PCNet, MetaBoxNet, BNMA and Multianna. The goal of this presentation is not to compare different packages but to go over the package we have developed. BNMA implements models described in nice documents. This document provides detailed description of the Bayesian NMA model. BNMA can model normal binomial and multinomial outcomes. Required input includes outcome, study indicator, stream indicator and total number of observations or standard error based on the specified input BNMA creates JAGS code and initial values and automatically runs Bayesian NMA model. We will demonstrate this package using the smoking cessation data set. This consists of 24 studies including comparing four smoking cessation counseling programs. This is the first five studies of the smoking data set and we will briefly go over the Bayesian NMA model since our data set is binomial, we fit a binomial model and the i refers to the trial or k refers to the arm. The link function is eta i plus delta i1k, eta i are the trial specific baselines and delta i1k are the trial specific treatment effects. For a random effect we can model the delta i1k with a normal distribution with average treatment effect d and heterogeneity parameter sigma squared. NMA requires consistency equations to hold and the equal variances are commonly assumed. So this model can be fitted using BNMA package as follows. We first set up the model using a function called network.data and we specify the input response and type. Then we can run the model using a function network. Then we can summarize the result using the summary function and this is the summary we get. So the odds ratio for treatment for the group counseling is 3.01 so the model estimates a 200% increase in the odds of creating smoking for group counseling compared to no intervention. There are some pre-built summary visualization functions and this is the fourth spot. Additional details that we have included includes checking convergence. So even if the user didn't specify the initial values, we generate the dispersed initial values and then we use the Gell-Mann-Rubin diagnostics to test conversions or parameters such as the eta i, the baseline risk, the average treatment effect, d, and the heterogeneity. We check convergence every satisfy the iteration and once the samples are converged it keeps the last half of the converter sequence. If the user requires more sample size we can specify through the parameter called m.1. So that was a basic NMA model. We have extended this basic model to, for instance, have a different assumption on baseline risk. Instead of having baseline risk to be independent we can assume it to be exchangeable and so we model the baseline risk using a normal distribution with a common mean and between study variance e and sigma e squared. This extra assumption of random intercept should lead to a greater precision. However, this comes at the price of using the between study information meaning the treatment effects are no longer estimated only by the usual differences within studies but also by the differences between studies. We can fit this model by adding parameter baseline risk and setting it to equal exchangeable. This is the summary result for fitting this model. Now the common mean for the baseline risk is estimated to be minus 2.4912 which is equivalent to a baseline probability of quitting of 0.076. So another model that we have added is we can use baseline risk as a meta regression similar to how we include covariate information. So by adding this baseline risk as a meta regression we can explain possible source of heterogeneity. There are three different assumptions we can make. It's common exchangeable and independent and if the user wants to assume a common effect treatment and covariate interaction then the user can just simply add this parameter baseline and set it equal to common and they will fit this model. So this is the summary of the result and now we have a new parameter BBL which is the regressor for the baseline risk regression coefficient. So now the the interpretation for the odds ratio which is 3.17 is now the treatment effects of the group counseling for patients with baseline probability of quitting of 0.066. To summarize we quickly showed how to fit simple Bayesian NMA using smoking dataset with the BNMA package. We demonstrated how to incorporate baseline risk via a exchangeable assumption or a meta regression and the following slide is uploaded in my private website micjsco.github.io and here are the references. Thank you.