Generative modelling has entered the field of geography and spatial analysis some 35 years ago, under three main forms: CA (cellular automata), LUTI (land-use transport interaction models), and ABM (agent-based models). Besides allowing spatial analysts and geographers to build operational models for transportation and planning, it has represented the opportunity to take causal inference from the traditional (statistical) analysis of empirical models to the design of causal mechanisms simulated in virtual environments. It is a major epistemological shift, whose scientific contribution still needs advocacy in the wider community of geographers and spatial analysts. In parallel, the expansion of both computing power and available data have made it easier for isolated teams and individuals to build ad-hoc models fit for their specific research questions, leading to a cacophonous development of generative models unrelated to one another.

Since the introduction of generative modelling in geography and spatial analysis, hundreds of models have been developed: all may have been wrong, but some were surely useful. However, most models built over the years have been abandoned, their program either inoperable with today’s technology or lost entirely. Apart from a few classics (such as MatSim or SLEUTH), dozens of models of urban growth, of transport networks, of negotiation between actors, etc. have been developed separately, been exploited for a few years, published and then forgotten or reinvented, leading to the YAAWN syndrome (O’Sullivan et al., 2016).

In 35 years, this should strike us as vastly wasteful of ideas, time and energy. By that time, my hope for geosimulation is that we have found a convincing way to build on each other’s work so as to demonstrate the benefits of generative modelling for geography and spatial sciences: 1/ by ensuring the sustainability of models and the traceability of contributions, 2/ by relying on modular combinations of theoretically-grounded mechanisms (or “building blocks”) and 3/ by making model building and model evaluation transparent and painless.

In this chapter, I will lay out possible paths towards this goal and the expected impacts in terms of geographical theory development and scientific reproducibility. Reproducibility and reusability of model building blocks are especially important for us geographers and spatial analysts since we are concerned with generic processes (segregation, polarisation, urbanisation, etc.) embedded in singular spatiotemporal contexts (e.g. region x during period p, street y at time t). Knowing the variation of model structures associated with varying spatiotemporal contexts can contribute to the identification of both universal properties of social phenomena and the peculiarities of places. In that way, generative modelling done right can both contribute to a new form of causal inference and to the larger programme of social sciences: the simultaneous search for generalised explanations of social phenomena and recognition of the uniqueness of historical events.

The most interesting and useful questions in spatial science are often truly about cause and effect: if we do X what will happen to Y. If spatial data scientists care about better understanding society and designing policies to improve it, we must concern ourselves with why spatial processes function as they do, not simply how. We propose doing so requires greater focus on research design, not ever more arcane spatial methods or datasets.

There is no doubt that ongoing developments in spatial data analysis and computational capacity have enabled a breathtakingly interdisciplinary scope of spatial enquiry. However, the causal questions they tend to answer are very limited, perhaps just shy of being deemed “mostly pointless”. Rather than continuing to lament this, we outline a manifesto for why spatial analysts should care about causal inference, and provide guidelines for the converted.

In essence, understanding causal relationships requires us to create, assume or discover random variation. It is not usually practical or ethical for spatial analysts to create random spatial assignment, so we commonly rely on assuming randomness (via exchangeability) using statistical adjustment. Better yet, we may search for so-called natural or quasi-experiments to remove the need to rely on statistical assumptions. We do so by identifying situations where someone (sometimes ‘God’; but more often a bureaucrat) has, without realising, cast geographically informed dice which allow us to investigate spatial processes.

We demonstrate the benefits of explicitly considering causal research questions, rather than avoiding the “C-word” (and furtively hoping readers infer it regardless). Throughout, we provide examples of studies which have explicitly exploited such designs, from considering the impact of neighbourhood homicides on exam results, to the impact of localised restrictions on COVID-related mortality. We use these studies as a roadmap to identify the most fruitful future research avenues for spatial analysts.