In the last post, I had a closer look at the emergence of political epidemiology and the methodological challenges it presents. At the core of the issue is the question of how to address the wider societal context into our studies. In practice, this requires methods that can deal with interacting causes (e.g., policies), as well as layered networks between individuals, organisations, and institutions. We would also like to address the fact that social structures both shape individuals and are shaped by them in a process of feedback. These challenges have become increasingly well-recognised in the field of epidemiology over the past decade, which has led some epidemiologists to embrace the tools of complex systems science and systems modelling. This post will provide some insight into these discussions and tools, and reflect on the emerging best practices for their utilisation.
Are socio-political systems complex systems? Most epidemiologists would argue that yes, including Diez Roux:
[P]opulations are systems in which individuals interact with each other and with their environment, so understanding the drivers of health (and the most effective ways to improve health) requires consideration of the systems within which health is generated.
This realisation is important because the usual approach to science - one that also underlies randomised-controlled trials - is one of reductionism. At the core of reductionism is the belief that once we understand the constituent parts of a system, we also understand the system; the sum of the parts is the whole. Most commonly used statistical models in public health rely on reductionism. Practitioners need to assume noninterference (no interaction between individuals) and unidirectionality of effects to have the models work as intended. However, this systems-naïve approach has repeatedly failed to identify useful interventions for many contemporary health challenges, including obesity, addiction, gun violence, and others. As Marshall and Galea argue:
[I]t is increasingly clear that these problems are characterised by complex, multifactorial processes that are highly resistant to interventions that address only one or a few causal effects. Therefore, multiple, highly interdependent causal pathways, which are fundamental characteristics of complex systems, are integral to the challenges that public health now faces. To this end, the application of complexity theory to understand how disease is generated and reproduced within populations systems has begun to gain traction and currency.
In the same article, the authors call for the use of agent-based models (ABMs), a mathematical modelling approach popular in complex systems science, to begin to tackle these issues. Marshall and Galea find ABMs:
[methodologically congruent] with macro- and eco-social frameworks that position health as a production of intersecting and interacting biological, social, and environmental factors.
Indeed, AMBs have exciting properties. A key appeal is that they explicitly simulate the assumed mechanisms that generate the investigated phenomena. One example is the ability of ABMs to replicate emergent population-level behaviours like phase transitions, which are sudden changes in the systems structure and function. Other features include the ability of ABMs to:
- cope with the differences between agents and the types of agents (e.g., people and organisations),
- model agents adapting their behaviour to changes in the environment, and
- incorporate existing social networks between the agents.
Of course, not all is sunshine and rainbows. ABMs are not at all easy to design and build. The explicit simulation of mechanisms makes ABMs hungry for data and their flexibility makes them easy to misuse. This is why recent discussions about ABMs in epidemiology have been concerned with the accuracy of causal estimates they generate. A key publication in this space is the Murray et al. study that showed that ABMs have important disadvantages when applied to the standard causal inference tasks when compared to more standard approaches.
The invited commentaries by Edwards et al. and Keyes et al. responded to these results, arguing that causal estimation within a well-measured population contained in a single dataset is not really what ABMs are used for. Instead, one would use ABMs to extrapolate established causal effects to novel situations (albeit the results may not be unbiased), explore the possible mechanisms that may have generated the results (alethic possibilities), or identify gaps (missing data or theory) in our understanding of health phenomena. Keyes et al. used a particularly memorable metaphor:
The parametric g-formula is like a late model sedan—perhaps a long-range, electric luxury car—offering a high-technology yet practical way to get from point A to point B over established roads. By contrast, an ABM is an off-road vehicle, capable not only of following the same roads as the luxury car (albeit less comfortably) but also of taking riskier excursions into uncharted territories where the wheels might fall off. Epidemiologists whose scientific questions can be answered within the constraints of the g-formula can avoid hazardous and unsupported assumptions with a g-formula approach. By contrast, epidemiologists whose questions require stronger assumptions must accept the consequent inferential hazards.
Hernán framed this question in terms of relative reliance on data and theory. The standard epidemiological approaches are deeply (but not wholly!) reliant on data. When the questions are too complex, however, epidemiologists may need to rely more on theory and strong assumptions, and find good ways of dealing with the threat of model misspecification and biased estimation. There is a rich body of knowledge on this topic already developed in other disciplines (e.g., Windrum et al.). Diez Roux agrees, noting that establishing the relationship between the model and reality is at the core of the challenge. She also emphasises that the process of building the models, no matter how flawed they end up being, may be a great source of insight:
The beauty of systems modelling is that it can help us understand the plausible implications of the knowledge that we have and how pieces may act together in ways that we might not have predicted from our understanding of each component separately. It can help us integrate quantitative and qualitative information and explore basic dynamics. It can generate new questions or hypotheses that can be subsequently investigated through new observations or experiments. Perhaps most importantly, the process of building a system models forces us to think about dynamic relationships, feedback, adaptivity, interference, and the ways in which they might play a role in the process we are studying. In this sense, systems modelling is a healthy antidote to the obfuscation that can result from too much simplification, from thinking that the world really functions like a regression equation (rather than that regression equations are useful tools that can help us understand some aspects of how the world functions).
The key message is one of complementarity between traditional and systems approaches, especially as increasing amounts of data become available to parametrise ABMs. As Cerdá and Keyes reflect:
By combining publicly accessible central data repositories with area and person-level prediction and causal frameworks, applying careful and systematic validation methods, and integrating systems science modelling into the broader epidemiological armamentarium through an iterative data collection—modelling—data collection—modelling—intervention—evaluation—modeling process, we can develop valid and reliable systems models that can inform policy priorities.
Complex systems approaches, of course, go far beyond the use of specific tools like ABMs. For example, I have not touched upon the intriguing work that uses system dynamics models (see the wonderful review by Darabi and Hosseinichimeh). Instead of just adopting tools from other disciplines, Naimi argues that we should also seek a deeper engagement with complex systems science. He cites Shannon (the information theorist) to deliver a warning:
Shannon […] long ago cautioned that the increasingly widespread application of concepts and tools of his field to other scientific domains, although exciting, carried “an element of danger.” Establishing interdisciplinary applications of information theory, he argued, “is not a trivial matter of translating words to a new domain,” but requires “a thorough understanding of [its] mathematical foundation” and “the slow and tedious process of hypothesis and experimental verification.”
I, for one, am excited to embark on this journey of a deeper understanding of complex systems science. There are a lot of fantastic resources out there and I look forward to curating some of the content on this blog in the future.