Portrait of Lorcan McLaren

Lorcan McLaren

PhD Researcher in Computational Social Science
University College Dublin

LinkedInGitHubScholarEmail

I am a PhD researcher in computational social science at University College Dublin, funded by a Taighde Éireann – Research Ireland Government of Ireland scholarship, and based in the Connected_Politics Lab under the supervision of James P. Cross. My research focuses on political communication, climate politics and public opinion, combining text-as-data methods, large language models and survey experiments. I also work on best practice for LLM-based text annotation and develop open-source tools for it.

Prior to my doctoral studies, I completed a BA in Computer Science, Linguistics and French at Trinity College Dublin and an MSc in Politics and Data Science at University College Dublin, complemented by two years of industry experience as a data analyst.

I will be on the academic job market in 2027.

Research

Working Papers

What to Compare and When to Stop: LLM Pipelines for Text Annotation in Political Science
with James P. Cross, Zuzanna Krakowska, Robin Rauner and Martijn Schoonvelde.
[Revise & Resubmit, Political Analysis]
Preprint

Political scientists increasingly use large language models (LLMs) to annotate text. Model, prompt, and inference choices can alter results and create researcher degrees of freedom. We ask how researchers can choose a defensible annotation pipeline without testing every configuration. We vary model choice and prompting condition across political science annotation tasks, using open-weight LLMs. Separate within-family comparisons examine model size. We hold hardware constant and freeze each task’s codebook. We then simulate selection strategies using human-coded validation data and evaluate selected pipelines on held-out texts. After the task itself, task–model interaction explains the largest share of performance variation. Overall model rankings are poor indicators of performance for a particular codebook. Prompting has little average effect but can substantially change performance for individual model–task combinations. Validation can screen out weak pipelines but often cannot identify a unique best one. Testing shortlisted models zero-shot yields nearly the same held-out performance as exhaustive search, with much less energy. We recommend screen-then-expand: stop when the study’s pre-set performance threshold is met, then select the most efficient configuration within a pre-set margin of the best validation score. We also propose annotation-specific reporting standards. Accompanying open-source tools support codebook development and pipeline benchmarking.


Here and Now or There and Then? The Psychological Distance of Climate Change in Parliamentary Speech
[Under Review]

When politicians describe climate change as a threat to future generations in faraway places, they render it abstract; when they speak of floods at home and livelihoods at stake today, they bring it close. Experimental research shows this distinction matters: proximate framing increases concern and support for action, but these literatures rarely speak to each other. What determines whether a legislator chooses to proximise? I address this question using over 35,000 European Parliament speeches from 2014 to 2024, classifying whether legislators mention environmental impacts, frame them as specific or universal, and situate them as proximate or distant. Gender and left-right ideology emerge as the most consistent predictors: female and left-leaning legislators discuss impacts more frequently and frame them as closer. Legislators respond to domestic climate salience, but EU integration position does not shape climate rhetoric. The paper shows that generative language models can efficiently label training data for fine-tuned classifiers.


When Climate Change Hits Home: Extreme Weather and Climate Rhetoric in the European Parliament
[Draft Available]

Climate change is often understood as a psychologically distant threat. This paper asks whether extreme weather near a politician’s geographic home changes how climate impacts are described in parliamentary speech. I combine sentence-level annotations of 28,448 European Parliament speeches with two exposure registers: a meteorological record of locally unusual warm spells and a disaster impact register of recorded floods, storms, and other climate-related events. Both are analysed at the MEP-speech-day level using three adjacent 30-day pre-speech windows, a 30-day future-exposure placebo, MEP and calendar-date fixed effects, and country-clustered inference. The primary meteorological estimate is small and statistically indistinguishable from zero (0.059 percentage points; p = 0.151); its 90% confidence interval excludes effects larger than 0.026 outcome standard deviations. The summer-heat specification is also null. In the disaster register, the primary occurrence-and-impact dose combines event frequency with the logged number of people affected and killed. Its immediate proximity estimate is positive but uncertain (0.060 percentage points; clustered p = 0.099; wild-cluster p = 0.112), with an equivalence bound of 0.024 outcome standard deviations. An unweighted event-count robustness is nominally significant (p = 0.030) but does not survive adjustment across temporal windows. The evidence rules out large changes in elite climate framing and provides limited evidence of a small immediate association following nearby recorded disasters.


Near or Distant? Elite Climate Rhetoric and Public Opinion in Ireland
with James P. Cross.
[Pre-registered]

Legislators can portray climate impacts as already affecting people close to home or as a threat to distant populations and future generations. Whether these portrayals change public opinion remains unclear. Previous experiments have yielded mixed findings, and differences in how they manipulate psychological distance, alongside limited power to detect small effects, make the evidence difficult to interpret.

Using a preregistered survey experiment with a nationally representative sample of adults in Ireland, we test whether near or distant portrayals shift climate policy support and behavioural intentions. A conjoint then examines what makes climate rhetoric persuasive: how messages describe who is affected and where, and how the speaker’s gender, nationality and political alignment shape those judgements.

Both stages pair text with AI-generated speech. In the conjoint, voices match the speaker’s stated gender and signal nationality through Irish or continental European accents. Respondents hear these cues as well as read them, allowing us to vary how the speaker sounds while holding the message constant.


No matching items

Publications

A Heuristic Method for Automatic Gaze Detection in Constrained Multi-Modal Dialogue Corpora
with Maria Koutsombogera and Carl Vogel.
Proceedings of the 11th IEEE International Conference on Cognitive Infocommunications (CogInfoCom), 2020.
Best Paper Award
PaperOpen AccessSource Code

Gaze, Dominance and Dialogue Role in the MULTISIMO Corpus
with Maria Koutsombogera and Carl Vogel.
Proceedings of the 11th IEEE International Conference on Cognitive Infocommunications (CogInfoCom), 2020.
PaperOpen Access

No matching items

Projects

CodeBook Lab
DocumentationPython Package

A Python package for computational social science that benchmarks local LLM annotation runs against human labels across model choice, prompt style, example use, chat mode, reasoning mode, and sampling settings.


CodeBook Studio
AppDocumentation

A browser-based annotation tool for computational social science. Define a codebook, annotate texts, adjudicate disagreements, and export labelled data for LLM benchmarking in CodeBook Lab.


No matching items

Teaching

Module Instructor

Workshop Instructor

  • LLMs for Text Annotation in Political Science, Political Studies Association of Ireland pre-conference workshop, Cork (October 2026)
    Workshop Materials

Teaching Assistant

  • Introduction to EU Politics, Undergraduate (Autumn 2023, 2024)
  • Research Methods in Political Science, Undergraduate (Spring 2024)