1  Introduction

1.1 Background

In a social landscape where the public is growing progressively disillusioned with its representation in politics (Landemore 2026), it is important to consider ways to foster more inclusive decision-making practices. One way of achieving this is using sortition to create panels that are more representative of the population for political discussions and deliberations. Adopted recently by an increasing number of institutions, including the Parisian (Sortition Foundation 2021), Irish (Walsh and Elkink 2021), and British Columbian (Ward 2008) governments, sortition as a part of democratic process is gaining traction. As a result, the need for fair and representative methods of building panels is only growing, and progress is being made in improving the structure of selective algorithms. However, there are still many challenges to overcome in order to ensure that these panels consistently achieve the level of representation that they envision. And there remains little procedure on measuring representativeness beyond the “surface” marginal level of demographics.

Sortition is a method of selection that uses random sampling to create panels that should be descriptively representative of a population. It has been used in various contexts, reaching as far back as ancient Athens (where it was how jurors were selected), and now into modern times for citizens’ assemblies. The idea is that randomly selecting individuals from the population can create panels that are more representative and inclusive than those selected through traditional methods, like elections or open-ended surveys. Due to its random nature, sortition also reduces dominance by elites or highly engaged groups, enabling broader participation across social classes (Giuliani 2025), making it more likely that decisions are in the relevant population’s best interest.

The representation problem

One of the main roadblocks in sortition is ensuring that the selection process is transparent and fair, and that the panels are truly representative of the population in terms of demographics and perspectives. But this raises the questions of what it means to be fair, what it means to be representative, and how these look from a systematic perspective. A fair lottery is not, by itself, proof of a representative outcome – it answers whether everyone had a reasonable chance of being chosen, not whether the people who actually get chosen reflect the community they are meant to stand in for. This thesis is concerned less with whether the process was fair, and more with a plainer question underneath it: who actually ends up on a panel?

As a process, sortition moves through several distinct stages between defining a population and convening a panel (Chapter 2 breaks these down in full), and each one has the potential to introduce bias that compounds by the end – yet there is currently no rigorous evaluation of representation achieved before a deliberative process is run. It is therefore crucial to measure how the level of representativeness (or potential representativeness) flows through each stage, and to understand how different factors may affect it.

Progression of sortition methods

Sortition methods have evolved substantially over time, moving from simple self-selection and proportional sampling toward increasingly systematic algorithmic approaches – LEGACY, MaxiMin, MiniMax, and most recently LexiMin and Goldilocks – each refining how fairly and robustly individual selection probabilities are equalised across a respondent pool (Chapter 2 gives the full history and formal definitions). Of these, LexiMin is the algorithm in practical use today, running inside Panelot (Flanigan et al. 2020), the tool currently used to assemble real sortition panels, and is the algorithm this thesis works with directly.

1.2 Thesis scope

All of the aforementioned algorithms only focus on optimising for maximal selection likelihoods as their sole objective functions. This means that in its discovery and construction of feasible panels, the existing LexiMin algorithm may be prone to distortion of intersections and intersectional association trends present within population distributions. This presents a significant risk for the validity of representation achieved by panels output by the algorithm – the extent and reasoning of which will be explored further in the paper.

Additionally, there is limited literature on the visualisations of the outputs of these algorithms. This presents a great opportunity to demystify LexiMin and sortition as a process, improving transparency for practitioners, facilitators, participants, and broader community stakeholders alike. The comparative visualisation methods and tools also allow a general comparison to be made between any selected group and any broader sample it is pulled from, with any algorithm or selection process. This could be useful as a mechanism for quick sanity checks of panel configurations, and as a way for practitioners to evaluate the validity of panels before engaging in deliberation.

To investigate this empirically, this thesis builds a simulated sortition pipeline: synthetic populations with a known, deliberately introduced two-variable association (and a matched population with none), survey samples drawn under both equal-chance and differential response, and panels selected from those samples via LexiMin. Statistical tests then compare each panel’s joint two-variable structure against its own sample, isolating whether a panel’s intersectional composition diverges from the sample it was drawn from beyond what sampling variability alone would produce. The resulting findings motivate a diagnostic Shiny app, Intersection Inspector, that lets practitioners check this directly for their own panels.

Thesis structure

The remainder of this thesis is structured as follows: Chapter 2 provides a literature review of the existing research on sortition and representation, and identifies gaps in the current knowledge. Chapter 3 outlines the simulation-based methodology used to evaluate panel representativeness, including how differential survey non-response is modelled alongside an equal-chance baseline, and the statistical tests used to assess whether a panel’s joint category distribution preserves the association present in the sample it was drawn from. Chapter 4 presents the results of the analysis, showing near-total intersectional association distortion in selected panels, well beyond what a random-sampling benchmark would produce. Chapter 5 introduces “Intersection Inspector”, a tool developed in response to those results to help practitioners assess panel representativeness directly. Chapter 6 concludes by proposing a framework and recommendations for practitioners, discussing the limitations of this thesis, suggesting directions for future research, and reflecting on what genuine representation demands beyond marginal fairness – namely, keeping sight of who the process is meant to represent, and whether it actually does.

Flanigan, B., Gölz, P., Gupta, A., Procaccia, A. D., and Rusak, G. (2020), “Panelot.”

Giuliani, A. (2025), “Sortition in politics: From history to contemporary democracy,” Parliaments, Estates and Representation, Routledge, 45, 317–335. https://doi.org/10.1080/02606755.2025.2530916.

Landemore, H. (2026), “Opinion No Shy Person Left Behind,” The New York Times.

Sortition Foundation (2021), “Paris creates a permanent Citizens’ Council,” Sortition Foundation.

Walsh, C. D., and Elkink, J. A. (2021), “The dissatisfied and the engaged: Citizen support for citizens’ assemblies and their willingness to participate,” Irish Political Studies, Routledge, 36, 647–666. https://doi.org/10.1080/07907184.2021.1974717.

Ward, I. (2008), “An Experiment in Political Communication: The British Columbia Citizens’ Assembly on Electoral Reform,” Australian Journal of Political Science, Routledge, 43, 301–315. https://doi.org/10.1080/10361140802035796.