 Today, I'm going to talk to you about multi-NMA. It's a new R package for Bayesian network meta-analysis of individual and aggregate data. So standard network meta-analysis and indirect comparison method synthesize published aggregate data from each study, assuming constant C of relative effects. So for an indirect comparison, that means that the AB effect that we see in the AB study is exactly the same as we'd expect to see in the AC study if that study had included a B arm as well. And this assumption breaks down when we incur bias if there are differences in effect modifiers between studies. So population adjustment methods aim to relax this assumption using available individual patient data or IPD to adjust for differences in effect modifiers between studies. Now, ideally, we'd have IPD from every study, in which case the gold standard approach is an IPD network meta-aggression, but more typically, we only have IPD from a subset of studies and then aggregate data from the rest. Multi-level network meta-aggression, or MLNMR, is a recent extension of the NMA framework to synthesize mixtures of IPD and aggregate data in networks of any size, whilst performing population adjustment. And it works, first of all, by defining an individual-level regression model. And this is actually exactly the same model as we'd fit in an IPD network meta-aggression. And then for the aggregate studies, we're going to average or integrate this individual-level model over the aggregate study populations to form the aggregate level model. And we do this using efficient and general numerical integration techniques. These models are implemented in the multi-NMA package, which is a suite of tools for performing MLNMR and NMA with individual patient data, aggregate data, or mixtures of both for a range of different outcome types. The package includes functions that streamline the setup of NMA and MLNMR models, form-model fitting of psilocyte diagnostics, produce posterior summaries of quantities of interest, and create flexible graphical outputs that leverage the power of gg-blocks and gg-dist. Behind the scenes, models are estimated in a Bayesian framework using STAN, and these STAN models are pre-compiled on CRAN, which means that there's no user C++ toolchain or R tools required. So an outline of a network meta-analysis then in multi-NMA, first of all, we start with data in a long or tidy format. That's one row per arm or contrast, per aggregate study, or per individual in the IPD. We then define the network, so we use these set functions to set up the different data types. If we have more than one data type, we can combine them together. We then specify our prior distributions and run the analysis using the NMA function, using this week of fixed and random effects models, inconsistency models, and regression models, et cetera. And once the model's fitted, we then check convergence, check model fits, and produce posterior results. The only difference when we're doing multi-level network meta-aggression is that we need to add in the numerically integration points into the network, add integration function. The other steps remain the same. So let's apply this to a short example. First of all, we have some aggregate data at the top here from one study with four arms. And you can see that we have account outcome. And then we have means and standard deviations and proportions for covariates in each of these study arms. And then we have individual patient data below this, so one row here per each individual with their outcome and their covariates. Finding these together in a network, first of all, we set the individual patient data and the aggregate data and then combine them together. You can see here I'm setting the columns of the data for the studies and the treatments and for the outcomes and also a treatment class. And then I can print the details of the network below. I can also plot the network using the plot method. Here I'm waiting the nodes by the social sample size and the edges by the number of studies on each comparison. And I'm also showing the treatment classes by coloring the nodes as well. We can then add in the numerical integration points to perform this MLNMR model using the add integration function. So here, for example, duration of psoriasis, I'm giving a gamma distribution with mean and standard deviation taken from the data. And that's because duration is strictly positive and so is likely to be skewed. And I can go away and fit the model. So using the NMA function, here I'm choosing to fit a fixed effects model with a probit link. And then here I specify a regression model which includes main effects and interactions for each of the five covariates. And then I'm specifying normal priors on each of the model priors. To get to the model, I can then print a summary of the results. So here, you can see it's showing us the summary of the model that was fitted. So regression model is sent to the covariates automatically for us. And then we have some posterior summaries for each of the parameters and also convergence information. These plots are model parameters, just using the plot method again. And here, it's using the power of GG disks. So I can specify a half-i stat, for example, or other plot stats and it will change the look of the blocks that are generated. Here, it's giving me densities as well as points and meters. Can then produce population average relative effects using the relative effects function. And here I'm specifying a new data frame for new study population. And if I omit this, it will produce relative effects for each of the studies in the network. And then I can also plot this again using plot methods. Very similarly for the population average, predictive probabilities, using the predict function. This time, I'm also specifying a baseline distribution on the baseline and probate probability of response. And then again, I can plot these. As much more, we can produce ranks, rank probabilities and plot these as well. Model fit, so DIC residual deviants and we have plot methods to plot these as well. So I've really just scratched the surface here today. If you'd like to find out more, you can check out our package website. That contains illustrated documentation and detailed walkthroughs, for example, analysis. If you want to learn more about the methods, you can read our methods paper published in JRSS-A. And if you want to contribute ongoing developments, please check out our GitHub page. Thank you very much.