Teaching
Courses, training and their materials
Five years of MSc statistics and econometrics at Justus Liebig University Giessen; week-long courses that hand fiscal incidence microsimulation and poverty mapping over to the finance ministries and national statistical offices that have to run them; and the World Bank’s own course on AI-assisted coding. Every course was taught in R, and each entry below says what was written for it — slides, worked exercises, downloadable projects, recordings — cohort by cohort, and where it lives.
Courses and training
Delivering it meant changing the toolchain as well as teaching it: I had Quarto and the Positron IDE made available as Bank-wide software so that participants could practise reproducible workflows rather than only hear about them.
Two days in the classroom with a self-study block between them. Before the first day, a setup guide for Stata, R and Python with Positron, Quarto, Git and GitHub Copilot, with troubleshooting for each step; day one, live demonstrations of cleaning data, reproducing an existing analysis and handling sensitive data under the Bank’s guidelines; four guided self-study exercises; day two, multi-step agentic tasks, external tools through MCP and prompts written for one’s own work. The April 2026 sessions at headquarters were recorded for Bank staff.
I wrote the software the course teaches before teaching it — R packages and structured data pipelines carrying the office’s survey and census inputs through direct estimation and Fay–Herriot modelling to a subnational poverty map — and then delivered it in person at the Poznań centre. The week ran against the office’s own data rather than a worked example, so the code was being improved and handed over at the same time, and it closed on the part that decides whether a toolbox outlives the project: how to maintain it and develop it further without me.
A Quarto project that takes the office’s statisticians through the pipeline one step at a time — the geometries, direct estimates from the survey, the auxiliary variables, Fay–Herriot models within a year and across years, their multivariate extension, an exploratory pass over the resulting poverty estimates and a comparison of the models — with the R functions and scripts behind each step and a pinned package environment, so that the office can rerun and extend it on its own.
I took the lead on the analysis these courses teach — country by country, the microsimulations that measure on household survey data how taxes and subsidies reach household welfare — and then adapted that methodology into web-based applications in R Shiny, deployed on Posit Connect and Shiny Server. What the client receives is therefore a tool rather than a report: ministries of finance and social policy simulate counterfactual tax and subsidy policies themselves and read off the distributional, poverty and inequality consequences for households. Each engagement hands over the simulation code in R and Stata alongside the dashboard.
In Indonesia and Georgia I joined the country missions to teach it: a week each, opening on the incidence theory and the mechanics behind the numbers, moving to the code behind the tools, and closing on the capacity to develop them further — in R or in whichever language the counterpart team already works in.
The GitHub organisation of the World Bank’s Equity Policy Lab, where the country microsimulation models and their Shiny applications are kept — most of them private to the client government — beside a public example fiscal microsimulation model written for training.
The R package the applications are built with. Its site documents how a country’s incidence analysis becomes an interactive tool that a ministry analyst can run, and the package source is beside it.
Many in the cohort had never programmed, so the course taught the statistics and the tooling together: descriptive statistics, sampling and estimation, inference and hypothesis testing, each carried out by the students in R on real data. They left able to clean, describe and test a dataset of their own.
Seven weeks of revealjs lecture decks, thirteen in all: population and sample; descriptive statistics; distributions, variance and quantiles; bivariate data, correlation and causation; cause and effect in randomised trials and the first hypothesis tests; the central limit theorem, the t-distribution and confidence intervals; simple and multiple regression, bias and the regression assumptions. With them, the PDF handouts as they were distributed to the 2023–24 and 2024–25 cohorts.
Eight weeks of revealjs decks on the same syllabus, with a longer regression block at the end: a recap of simple regression and its causal reading, linearity, bias and efficiency, and multiple regression worked through on land prices and on a production function.
The R Markdown (xaringan) decks of the early cohorts, in three blocks — introduction, ceteris paribus, sample versus population, descriptive statistics and hypothesis testing; simple and multiple linear regression and linearity; bias, collinearity, endogeneity and heteroscedasticity — and the homework: problem sets drawn from a bank of ten exercises, generated in five variants each, with their solutions.
Built on the Mostly Harmless Econometrics toolbox, it trained students to implement each method themselves rather than read about it, and to judge which one a question actually calls for.
A fifteen-week plan with each week’s materials beside it: thirteen lecture decks in HTML and PDF — introduction and selection bias, R and RStudio, multiple regression in two parts, interaction terms, difference-in-differences, omitted variable bias, panel regression and the production function, instrumental variables with worked examples, regression discontinuity — seven exercises as downloadable R projects with their data, from a randomised trial to instrumental variables, and recorded walkthroughs of the lectures and the exercises on YouTube.
Rebuilt in September 2026 as a course site organised by topic rather than by week. Seven topics — introduction, data wrangling, describing data, simple regression, multiple regression, panel regression, instrumental variables — each with its decks, its exercises and its download on one page: fifteen revealjs lecture decks; twenty R exercises, most of them published as rendered walkthroughs that show the code with its output; the exercises of each topic packaged as an RStudio project with its data in place, ready to open and run; a ten-part video series on getting started with R; and, for every topic, the reading, cheat sheets and tutorials that go with it.
Reusing the materials
The course sites are published so that students can return to them after the course and other instructors can build on them. Every site has its source repository beside it, holding the slide decks, the exercises and the data they use. Reuse or adapt any of it with credit and a link back to this site. Where a repository is still private, the entry above says so, and its addresses will start working when it is opened.