Michael Ganslmeier

Assistant Professor in Data Science · University College London

I search for correlation and causation.

About

Michael Ganslmeier

I'm an Assistant Professor in Data Science at University College London. I use causal methods to study how decisions affect people's live and how reliable such scientific findings are. I work with universities, governments, international organisations, and companies on applied data and evaluation problems.

Positions

  • Assistant Professor in Data Science, UCL (2026–)
  • Data Scientist & Consultant, World Bank (2023–)
  • Assistant Professor, University of Exeter (2024–25)
  • Postdoctoral Fellow, LSE (2022–24)

Education

  • DPhil, University of Oxford
  • MSc, London School of Economics
  • Visiting, Columbia University
  • BA, Zeppelin University
Download CV

Affiliated with

UCL
Oxford
LSE
World Bank
OneThousand

Research

Below, you can find a selection of my published papers and ongoing research projects.

Publications

Work in Progress

  • R&R, Nature Climate Change
    Empirical Identification of Feasible and Strategic Climate Policies

    Which climate policies are politically and institutionally viable? A data-driven map.

    Authors: Penny Mealy, Michael Ganslmeier, Stephane Hallegatte  ·  Funded by World Bank
  • Under Review
    Rising Heat in the Labor Market: Evidence from European Regions and Firms

    How rising temperatures affect productivity, employment, and output across European regions and firms.

    Authors: Luca Bettarelli, Thomas Farole, Michael Ganslmeier, Indhira Santos, Marc Schiffbauer
  • Under Review
    Harmonised Daily Climate Indicators at Country, State, and District Level, 1950–2025

    A research-ready database of daily climate indicators for 39,096 districts across 194 countries, spanning 1950 to 2025.

    Authors: Michael Ganslmeier, Luca Bettarelli, Davide Furceri, Marc Schiffbauer  ·  Funded by World Bank
  • Under Review
    Positional Inertia in Sequential Rank-Order Contests

    Why early advantages get locked in even under fair rules.

    Authors: Michael Ganslmeier
  • Under Review
    The Economic Costs of Temperature Volatility

    Volatile temperatures, not just rising averages, are damaging economic output. Estimates from US firms.

    Authors: Luca Bettarelli, Davide Furceri, Michael Ganslmeier, Marc Schiffbauer  ·  Funded by World Bank / IMF
  • Under Review
    Firm-Level Climate Change Adaptation: Micro Evidence from 134 Nations

    How are firms actually adapting to climate change on the ground? Survey evidence from 134 countries.

    Authors: Claudia Berg, Luca Bettarelli, Davide Furceri, Michael Ganslmeier, Arti Grover, Megan Lang, Marc Schiffbauer
  • In Progress
    Death by a Thousand Cuts: Pension Benefits Reduce Mortality

    Do higher pension benefits help retirees live longer? Causal evidence from a US pension reform.

    Authors: Michael Ganslmeier, Margaryta Klymak, Tim Vlandas
  • In Progress
    Political Integration and Trust in Foreigners

    Does deeper political integration between nations reduce distrust of outsiders? Evidence from European integration.

    Authors: Barry Eichengreen, Michael Ganslmeier, Orkun Saka
Full CV

Writing

I write about research and evidence for non-academic audiences. If you're a journalist or editor, feel free to get in touch.

Software

Some of my research has produced open-source tools that make advanced statistical methods easier to use. The packages/software below grew out of specific projects and are freely available.

sensiverse

sensiverse

An R package for sensitivity and model uncertainty analysis in the social sciences. Explore large model spaces, quantify robustness, and visualize uncertainty.

PNAS Paper  ·  GitHub  ·  Tutorial

Global Weather Map

Global Weather Map

An interactive climate data platform built for the World Bank. Maps temperature, precipitation, and extreme event trends using ERA5 weather data.

Open Tool

DRCC

DRCC

The Database of Regional Climate Change: a research-ready panel of daily climate indicators from 1950 to 2025 at country, state, and district level, built from ERA5 and CEMS data.

Website  ·  GitHub

Teaching

Data Visualization

Data Visualization
University College London

Research Design

Research Design
University College London

Causal Methods for Policy Evaluation

Causal Methods for Policy Evaluation
University of Exeter

Data Analysis in Social Science II

Data Analysis in Social Science II
University of Exeter

Introduction to Programming

Introduction to Programming
University of Exeter

Causal Inference for Observational Studies

Causal Inference for Observational Studies
London School of Economics

Research Design for Policy Evaluation

Research Design for Policy Evaluation
London School of Economics

Applied Regression Analysis

Applied Regression Analysis
London School of Economics

Consulting

I help organisations to use data to evaluate policies and make better decisions.

Policy & Decision Evaluation

Measuring whether individual and organisations' decisions work in practice, and what difference they make in the real world.

Statistical Analysis & Machine Learning

Turning complex real-world data into clear analysis for decision-making, using statistical models and machine learning.

AI & Data Training

Tailored training on statistical thinking, AI, and how to evaluate data-driven work.

Public Opinion

Understanding what drives public support and political behaviour.

Climate & Environment

Analysis of environmental policies and adaptation strategies.

Health

Analysis of public health interventions on health outcomes.

Past and current engagements

World Bank
IMF
EBRD
OneThousand
LSE
Oxford
UCL
KCL

Get in touch

I'd be happy to hear from you.

Get in touch

Contact

If you would like to discuss research collaboration, consulting, or public speaking, feel free to get in touch. I work with universities, international organisations, governments, NGOs, and companies.

Email

michael.ganslmeier@ucl.ac.uk

Twitter / X

@GanslmeierM

Google Scholar

Publication record & citations

LinkedIn

michael-ganslmeier

GitHub

MGanslmeier

Office

UCL, Foster Court
London WC1E 6BT

Latest on LinkedIn

All activity
A paper that finds nothing isn't what you picture when you start a PhD.You imagine the big effect and the clean story.For my last paper in my dissertation, I had a very simple question: if you give old people 10K/year more when they retire, are they more likely to vote?There are several good reasons to expect a yes.You have more to protect, so you turn out to defend it. You have more money and time, which makes taking part easier (yes, voting is costly). And getting something from the state sends a signal: you count, you matter … at least electorally.And we also know it works in reverse: if parties cut people's benefits, voters punish them in the next election, see the Austerity papers.So, to answer the question causally, we followed 93,606 voters in Illinois across twelve elections, from 1992 to 2020. Last week, I was working on a new paper idea where I used the (famous) Diff-in-Diff method (a core causal inference method). The results came out "clean"; there was just one problem: the test for the key assumption (the parallel trends assumption) failed. If this assumption cannot be "proven", then the results cannot be trusted. So far, so clear.So I though let’s ask Claude Code what’s going on. Claude came back with a reasonable list of reasons why the test fails, and then it suggested some easy fixes that I can try "to make it work".One of these reasons and fixes were reasonable: a couple of control units had their own (alternative) shock just before treatment, so it suggested to drop them, which I did. Things started to look better. But the test still fails.So Claude Code suggested another fix: a sub-group of my control units are just not good because they are “too different” (in terms of certain demographics) to the treatment group. Again, it suggested to remove these observations. When I applied both fixes, the parallel trends assumption looked valid now. In 2017 I was a predoc at the LSE who could barely code. A project needed synthetic control, and the only package that did it was badly broken. So I spent two weeks reading the source line by line until I could fix it.Somewhere in there, it clicked. I finally understood the method, the real thing.Nobody set that as an exercise but the bug set it for me.My students will never get that. They vibe code, vibe write, vibe design: they say what they want, a model writes it, done. And I really don’t mind if they use these tools.But here is what I do mind though. An exercise was never about what you hand in. You write an essay to force a fuzzy idea into a clear sentence, and find out it was wrong. The useful bit is the friction, and that friction was the teacher (a way better teacher than I will probably ever be). Let me ask you a question: when a kicker misses in a penalty shootout, is the next penalty more or less likely to go in?Got an answer? Hold that thought.The question hit me a few weeks ago, under the worst possible circumstances: Germany losing to Paraguay on penalties (it took me a while to digest this). Shortly before Manuel Neuer (thanks for everything in the last decades!) held the fifth shot, this random question popped up in my head.What struck me was that my brain instantly produced confident theories in both directions. More likely: the pressure is released, your own miss suddenly costs much less. Less likely: your kick can now decide things / put your team ahead, and that's exactly when people get nervous.Both seem plausible and both have actual psych literatures behind them (yes, I checked). That's the thing about how I believe humans think: our minds run on theoretical priors, and theories almost by construction link things to effects. Up or down, but rarely nothing. I've stopped betting against AI, with one exception.Every year someone says AI will never write code, never make art, never do anything truly complex. Every year a new model proves them wrong. I've learned not to bet against the machine.But one question keeps nagging at me: is there one thing AI will never do? Not "hasn't yet" but actually never.I think there is. AI can describe and predict the world better than anything we've built. What it can't do is tell you what causes what.Does a vaccine work? Not "do vaccinated people fall ill less", but does taking it actually change your odds of getting ill? To really know, you'd need to see the same person, at the same moment, with the vaccine and without. One of those worlds never exists. It's missing from every dataset, because it never happened. Nine years ago I stood outside the UCL Main Building in a suit that fitted nothing but the moment.I was waiting for an interview for a research assistant role at UCL.By then, I’d already collected over 200 rejections for similar pre-doc positions during my master’s. Yes, I am that kind of a guy.I dreamed of becoming a professor someday, but I knew I wasn’t ready even for a PhD.That first job changed everything.It showed me what research really feels like: constant doubt, dead ends, and those rare, electric moments when the data suddenly makes sense.That’s when I fell in love with science... Patricularly, because science treats mistakes (and there are many) as progress.Popper would approve.I don’t want to romanticize it though: science today is often polarized.We tend to favor evidence that confirms our existing beliefs over what might contradict them, and we prioritize visibility and speed over methodological rigor and depth.