The LSE Data Science Institute (DSI) is proud to highlight the work of LSE students such as Georgia Meyer who has published this book review of Escape from Model Land. This book is authored by the DSI's Senior Policy Fellow Dr Erica Thompson.
Are you a student at LSE with work that you would like to share with the DSI? Please get in touch here.
Book review: Escape from Model Land
Georgia Meyer (LSE Department of Management)
Escape from Model Land provides a blueprint for the kind of critical thinking required for us to more responsibly continue to co-create realities through models. Long ubiquitously deployed in finance, economics and climate forecasting, their increasing adoption across all aspects of economies and societies by virtue of the increasing sophistication of machine learning and predictive analytics demand a much more intentional and reflective scrutiny about their nature and impacts than what is oft the case. Dr Thompson does just that in Escape from Model Land by: (i) teasing apart the complicated enmeshment of models as translators of and / or producers of realities; (ii) reviewing models’ necessarily incomplete and changing quantification and qualification of uncertainties; (iii) considering their value in terms of transparency of decision-making processes and (iv), asking the important question of who gets to make the models in the first place (and with what ends in mind)?
Dr Thompson says that she ‘started thinking about these questions as a PhD student in Physics ten years ago’ when she ‘conducted a literature review on mathematical models of North Atlantic storms’. She realised that given the fact that all these peer-reviewed and published studies had ‘conflicting rather than overlapping uncertainty ranges’ an additionally important question (alongside figuring out how to predict the behaviour of North Atlantic storms) was ‘how we make inferences from models’. She sets up the ambition of the book as her effort to try and find a balance between two ‘unacceptable alternatives’: (i) taking models literally - not accounting for their misrepresentations of reality; or (ii) abandoning models and losing ‘lots of clearly valuable information’. The question she poses is: ‘how do we navigate the space between these extremes more effectively?’.
Following this set up in Chapter 1 Locating Model Land, Chapter 2 Thinking Inside the Box puts forward a case for how models articulate, and constrain, the boundaries of our imaginations. In making this case she first discusses a necessary and careful interplay between simplification and complexity as prerequisites for model explanatory power, citing Von Neumann’s parable about fitting an elephant with four parameters whilst a model with five parameters would make it 'wiggle his trunk'. Moving beyond the Ockham’s Razor principle (go with the simplest explanation) Dr Thompson expands the discussion on what is in and out of scope of models by introducing The Phillips-Newlyn Machine (or MONIAC, Monetary National Income Analogue Calculator) which ‘conceptualises and physically represents money as liquid…and circulated around the economy at rates depending on key model parameters’. Her discussion of the opportunities and limitations provided by MONIAC - including the way it expanded the consideration of relations between things not just the things themselves - notes how ‘each choice of representation lends itself to certain kinds of imaginative extension’. It is one of my favourite passages in the book (there are many).
Another chapter of the book, Models as Metaphors, riffs off a statement by Nobel Prize winning economist Peter Diamond’s Nobel lecture: ‘taking a model literally is not taking a model seriously’. Dr Thompson uses this starting point to set out how models have multi-dimensional purposes and applications depending on the phenomena at hand and the tacit acceptance of the nature of the stereotyping (reducing to simplicity) that is taking place in any given context within which a model is being deployed. In other words, she argues that we see models less as representations of reality and more as a companion to the version of reality that is acceptable in any given social context at any particular time. This is where questions of values - whose values - are at play when these ‘implicit value judgements are being made’ comes to the fore.
This theme is expanded in various parts of the book to address the discrimination and dangerous stereotyping that arises when deeply flawed processes of constructing and applying models (and poor input data) are left unchecked, drawing on the polemical work of Cathy O’Neil, Emily Bender and Tinmit Gebru. What makes this chapter particularly interesting is the way Dr Thompson interrogates the value of model explainability across various contexts making the case that whilst in some (aforementioned e.g. discrimination) explainability is enmeshed with questions of accountability and fairness, in others (e.g. cyclone activity), the picture is less clear cut (if a model does more accurately predict cyclone activity but we don’t know how it was able to, is that a problem?). This is complex territory and the clarity and reflexivity that Dr Thompson deploys across this matrix of values, agency, context and decision-making reads like a ‘how-to’ for any researcher or practitioner working with models. Or, for that matter, for any human being with varied and changing levels of self-awareness about the internal working models we all use consciously and unconsciously to move through the world.
These questions of how values intersect with science are addressed throughout the book in later chapters including one called The Accountability Gap. Here Dr Thompson draws on Birhane’s work mapping the territory of evaluative tools used in ML papers - noting how often the extent to which societal needs are met are absent from the discussion. In revisiting this theme the notion of objectivity in science is gently teased apart revealing the many constructed components that are inescapably bound up with processes of description, measurement and prediction. Where this thread leads is an instructive set of examples of various articulations of some kind of ‘social objectivity’ - which rests on multiple contributory accounts that are ultimately shared and agreed upon as a basis from which to move forward. Of course this is a notably different kind of treatment of scientific knowledge than that which has long held popularised dominance in many key areas of policy and common parlance.
Dr Thompson ends the book with five principles for responsible modelling, making a point about the human cognitive counterpart to navigating the world with models: that, ‘we must ourselves navigate the real-world territory and live up to the challenge of making the best of our imperfect knowledge to create a future worth living in.’. Amidst many important takeaways from the book (who’s values?; complexity vs simplicity trade-offs; explainability as virtue of accountability by context; obfuscation and misrepresentation via compression of variables in models and measurement), this point about human capacities feels critical. That ultimately, we cannot cede too much explanatory power to models - divorcing those who create and use them from the kind of transparency and humility to their imperfections. That we cannot sidestep our responsibilities to interrogate the decisions taken on the basis of models as new information or additional viewpoints come to light. Finally, what I take from this book, is that we ought regularly escape our own (internal) model lands that shape how we develop priorities, apply principles and evaluate our interactions with the world around us.