 Hello everybody, my name is Wolfgang Fichtbauer and I will now be giving my talk on automated report generation using the metaphor package. So to start, let me say a few words about the metaphor package itself. This is an R package for conducting meta-analyses. It grew out of code that I wrote from my dissertation research. Eventually this turned into a full R package in 2009 and I've been updating the package ever since. Hello here are some relevant resources related to this package. Now what is the usual workflow when you're conducting a meta-analyses? Well first you need to search for the relevant studies, collect them, extract information from these studies and then eventually you can compute effect size estimates which you can use as input to some meta-analytic model like a random effects model and then you can do some further analysis steps like checking for outliers, drawing some pretty pictures, maybe checking for publication bias and eventually you need to write this all up and then the very last step is to send your paper to science or nature and get it rejected. Okay, so let me give you an example. BCG is a vaccine against tuberculosis and if you want to find out if this vaccine is really effective you can do a study where you vaccinate one group of individuals and not another group and then you check for the infection risk in the two groups. So here are the results from one of these studies. So we have four TB cases in the vaccinated group out of 123 people, 11 cases in the not vaccinated group out of 139 cases. We have the infection risk in the two groups. The risk ratio which indicates that the risk of infection was quite a bit lower in the vaccinated group, the log risk ratio which we really need for the meta-analysis and the corresponding sampling variance. Now in total 13 studies have been conducted on the effectiveness of the BCG vaccine. So if you want to do a meta-analysis of these 13 studies, we first need to compute the log risk ratio and the corresponding variance of these 13 studies which we can do using the ESCalc function. And then we can use these log risk ratios and variances as input to the RMA function which is one of the main model fitting functions. And by default it will fit a random effects model and then we can look at the output. So we see here our estimate of tau square. So the amount of heterogeneity I squared, we get the Q test for heterogeneity. We get the estimated average log risk ratio and the corresponding confidence interval. And then we can do things like forest plots, funnel plots. We can check for outliers and we can run the rank or the regression test for funnel plot asymmetry which may be indicative of publication bias. So that is the usual workflow but the metaphor package has a function called report term that can automate quite a bit of this process. So what it will do is it will write a report for you describing the statistical methods used and it will provide the results in terms of natural language including a forest plot, a funnel plot, it will give references for all of the methods used and then you either get an HTML file, a PDF or a Word document. So let me demonstrate this, how this works. So we first load the metaphor package, then I'm gonna again compute the log risk ratios and variances and then I could fit a random effects model and then look at the results. And then instead what I could do is I could run the reporter function. So if I do this by default, you will get HTML output. So I get this analysis report. So what does it tell me? Well, it will tell me what kind of effect size estimate was used for this analysis, log risk ratios, a random effects model was fitted. How was tau square estimated using restricted maximum likelihood estimation? Then it provides references for the Q test for I squared. It describes how the data were checked for potential outliers or overly influential studies. It provides information, how the data were checked for publication bias. And then you get the results. So 13 studies were included in this analysis. We get the range of the observed log risk ratios. We get the estimated average confidence interval. We get the funnel, the forest plot. We get the Q test, I squared. We have no indication in these data that there may be outliers. And we get a funnel plot and we get the results from the rank and regression test for funnel plot asymmetry. And then some notes down here and references for all of the methods used. So this report is dynamically generated. So if the results, of course, were different, then it may indicate that they are outliers and whether there is heterogeneity or not, while you will get different output depending on the results. So that is what the function can do. Now it's not complete yet. So right now the results are given in terms of the effect size measure that you use as input. So if you meta analyze log risk ratios, then you will also get an average log risk ratio. But usually we want to back transform this to risk ratios. So in the future, it may be nice to add the option to sort of transform the results. Also, at the moment, the reporter function does not work with meta regression models. So this is something that I would like to add. And also to extend this function to work with other model objects, like the arm a dot h function, which uses the mental Hensel method. So right now it only works with the arm a function and maybe some more customization, maybe an option to include explanatory footnotes. So to make this even more useful. But already right now this function is quite useful for teaching purposes, learning purposes, and just really for automating report generation. So that's all. Thanks for your attention and I'll be happy to answer any questions.