Be Careful Where You Tread

The year is 1915, passengers are gathering on the Chicago River wharf to board the Great Lakes excursion steamer Eastland. Within a matter of minutes, 844 passengers will have lost their lives. With 2,573 passengers and crew onboard and still tied to the wharf, the Eastland rolled over in 20 feet of water [1]. There was no time to launch the lifeboats, more passengers died than in either the Lusitania or Titanic disasters – no Hollywood movie has ever been made about the Eastland. What caused the Eastland to ‘turn turtle’? The 1912 sinking of the Titanic had given rise to the ‘lifeboats for all’ movement, and the US Congress passed a bill requiring lifeboats to accommodate 75 percent of the vessel’s passengers. The intervention intended to save lives had resulted in the Eastland becoming top heavy and a catastrophe waiting to happen.

In large-scale, sociotechnical systems where we have a desire to bring about improvements or a need to mitigate problems, rather than jumping to conclusions about what to do, it is best to start with the question, what is going on here? Stepping back and looking at the whole helps to situate ourselves and consider more clearly what we think we know, what we don’t know, and what we can know. We start by drawing a boundary around the problem space and considering what we can measure at that boundary. As we get a better sense of the situation we can move, extend or contract that boundary to understand the whole at different scales and levels – recognising always that our boundary is arbitrary and subject to change.

If we take the current pandemic as a case study, we could choose to draw the boundary at the level of a country; ideally, one with clear sovereignty and delineated geography as this reduces the number of variables to control for. We can treat what lies within the boundary as a ‘black box’ to reduce the level of complexity we have to reason about while not becoming simplistic (excessively simple or simplified: treating a problem or subject with false simplicity by omitting or ignoring complicating factors or details).

Taking our ‘black box’ country i.e., the system being examined, what can we know about it from a pandemic perspective? (For the sake of argument, we will ignore the flow of people into and out of the country – remember, we’re in a lockdown!) There is a finite set of people within this boundary, and there are people being born or dying on a fairly consistent basis (we are assuming a large, mixed population such that we can safely aggregate and average numbers), there is a total population of living people, an ‘input’ of new people (babies) and an ‘output’ of people that have died.

What can we measure and hence ‘know’ about this system (rather than speculate or have an opinion about)? A small set of measurements is sufficient for us to understand the system at this level, the input and output rates, time in the system, and net population:

• Arrival rate, births per day
• Departure rate, deaths per day
• Life expectancy, number of years between being born and dying
• Population, number of people alive

Note the binary nature of these measures, you are either alive or dead, people are counted in integers, time is a liner progression, etc.

We can now ask ourselves interesting questions about this system to better understand it, having first established our starting conditions i.e., the baseline measures. There are two parameters to consider, births and deaths, each can be in one of three states; increasing, decreasing or static. Mapping this out, we can expect our system to behave in this way:

– No change. ↓ Decreasing. ↑ Increasing. ? Either, dependent on relative rate of change

We can now reason about the possible consequences of disruptions to, and interventions in, the system, and be able to measure their consequences. Assuming a starting condition where the system is in balance i.e., the birth and death rates are not changing, how would our measurements change if a deadly virus were to be introduced into the system i.e., our country’s population becomes infected with a pathogen? We would expect the death rate to go up because more people are dying (obviously, that’s what deadly means!), our total population would decrease (more people are departing than arriving, hence decreasing the pool of living people), and life expectancy would decrease (people are dying sooner than before).

There are latencies in the system, for example it takes time for people to become infected, fall ill and then die, and there are delays in making measurements and getting results, so we should not expect measurements to change instantaneously. We are also dealing with large numbers and making aggregate calculations such as the average life expectancy (by the way, please state whether it is the mean, mode or median, and don’t quote an average without its associated standard deviation). This makes measurement tricky, although not insurmountable, we just need to be careful with the quality of our measurements and calculations, and most especially how and what conclusions we draw.

All else being equal, we can generally tell a lot from these simple measurements and have a better understanding of ‘what is going on here’. If life expectancy is reducing and the rate of death is going up (more people are leaving the system faster) then we have a sense of how deadly the virus is. Or perhaps these measures are not changing in the way we had expected even though we have evidence of people becoming ill and dying from the virus – in that case we do not know what is going on here and need to investigate further to find better explanations of what might be happening.

Opening the lid of the black box and arguing about what is happening inside is fraught with difficulties because it is like shining a flashlight inside a mechanical chronometer from very far away in a pitch-black room trying to discern what the mechanisms are, their interrelationships and how they operate together to track time. To make it even more challenging, if the chronometer is defective and you are trying to fix it with this level of knowledge your actions are likely to be ineffective or even counterproductive – you may do more harm than good. If you are able to shrink your boundary down to a single mechanism and a simple set of measurements, you would have a better chance of fixing at this local level, and might get the chronometer back in working order if you were lucky enough that the defect was localised to the area you chose; the odds would be against you though.

Things start to get exciting though when we look to intervene in the country suffering from a pandemic because we are opening the black box and trying to find ways to change what is known as a Complex Adaptive System (CAS) – ‘a system that is complex in that it is a dynamic network of interactions, but the behavior of the ensemble may not be predictable according to the behavior of the components. It is adaptive in that the individual and collective behavior mutate and self-organize corresponding to the change-initiating micro-event or collection of events’ [2]. Imagine trying to fix your defective chronometer but every time you touched it the cogs morphed shape, moved around and reconfigured themselves, like something out of a Terry Pratchett novel!

Successful intervention in a CAS to achieve sustainable desired outcomes is hard, really hard. You do not know what is really going on, every time you intervene the system adapts itself, changes you do make will not show an effect until some arbitrary time in the future [3], multiple variables and interactions are in play simultaneously, there are usually many other actors intervening at the same time and confounding factors screwing with your theories. We can too easily fool ourselves (cue Richard Feynman quote) and find that our interventions are more ritualistic than scientific – if we burn a Yule log on the night of the winter solstice more piglets will be born in spring.

Now imagine trying to build a model of a CAS that is sufficiently representative of your system being examined so that you can study what might be going on and test out interventions – good luck with that [4], the map is not the territory! Statistical models can help explore ‘what if’ scenarios, but they are not magical ‘answer generators’ [5] and are subject to the whims of the model builder and the chosen coefficients, so they can be wildly wrong and misleading [6].

So, what to do? Consider at what level you should draw your boundary, identify measures that are both dependable (repeatable, reproducible and robust) and meaningful (they help you explain what is going on here), develop explanations that are sufficiently rich to include the dynamic inter-relationships of entities within the system and how they may adapt their behaviour, instrument the system to sense more of what the system is doing and how it is responding, and use A/B testing to validate the hypotheses of proposed interventions (maybe burning Yule logs has no effect on pig fertility after all!).

Measurement can give us a sense of ‘what is going on here’, and carefully considered and executed interventions can help nudge a system into a preferable state, but we rarely operate on isolated systems, our boundary is drawn within a system of systems (Pratchett’s Discworld again!) and any intervention may have consequences at different levels over varying timescales. A balanced set of measurements is therefore critical to ensure that you do not over index on one measure at the expense of the others, and you have a better holistic sense of what is going on here. A simple, and all too common example of over indexing is measuring yourself by your wealth while ignoring your health.

If you do decide that you want to try intervening in a complex adaptive system then you are likely to find that the decision-making process itself falls within the system being examined and you have to confront challenges such as Arrow’s Theorem [7], which proves it is impossible to combine any group’s policy preferences into a single preference in a way that meets even the simplest of fairness criteria – I’m sure you can image the challenges that governments must face when figuring out how to respond to the pandemic, such as who should get the vaccine first. We are starting to stray into the area of morals and ethics that are derived from the collective values of a society [8], which introduces another level of complexity to be considered in the system and potential interventions, so I will close here.

How to intervene in complex adaptive systems is a fascinating challenge and one that the world has been forced to confront in the extreme over many months, and will continue to do so for what looks like to be many more. Hopefully, I have been able to give you a new lens to look at the world through, and at least a passing sense of how you might navigate this reality – but before you decide to intervene, be careful where you tread!

If you would like to explore this fascinating area of systems further, I recommend reading Baron Schwartz on queueing theory [9], Mervyn King & John Kay on radical uncertainty [10], Annie Duke on decision making in uncertainty [11], The Systems Thinker publication archive [12], and listen to Joshua Epstein on agent-based systems [13].

1. ‘The Eastland Disaster Killed More Passengers Than the Titanic and the Lusitania. Why Has It Been Forgotten?’, Susan Q. Stranahan. https://www.smithsonianmag.com/history/eastland-disaster-killed-more-passengers-titanic-and-lusitania-why-has-it-been-forgotten-180953146/
2. ‘Complex adaptive system’, Wikipedia. https://en.wikipedia.org/wiki/Complex_adaptive_system
3. ‘Fixes That Fail’, Gene Bellinger. https://www.systems-thinking.org/theWay/sff/ff.htm
4. ‘Of Bits, Bugs and Responsibility in The Public Square’, Chris Von Csefalvay. https://chrisvoncsefalvay.com/2020/05/09/imperial-covid-model/
5. ‘The “Thinking” In Systems Thinking: How Can We Make It Easier to Master?’, Barry Richmond. https://thesystemsthinker.com/the-thinking-in-systems-thinking-how-can-we-make-it-easier-to-master/
6. ‘The UK’s coronavirus policy may sound scientific. It isn't’, Nassim Nicholas Taleb & Yaneer Bar-Yam. https://www.theguardian.com/commentisfree/2020/mar/25/uk-coronavirus-policy-scientific-dominic-cummings
7. Arrow’s Theorem, Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/arrows-theorem/
8. ‘The moral roots of liberals and conservatives’, Jonathan Haidt. https://www.ted.com/talks/jonathan_haidt_the_moral_roots_of_liberals_and_conservatives
9. ‘Everything You Need to Know About Queueing Theory’, Baron Schwartz. https://cdn2.hubspot.net/hubfs/498921/eBooks/queueing-theory_1-1.pdf
10. ‘Radical Uncertainty: Decision-making for an unknowable future’, Mervyn King & John Kay. https://www.hachette.com.au/mervyn-king-john-kay/radical-uncertainty-decision-making-for-an-unknowable-future
11. ‘Thinking in Bets’, Annie Duke. https://www.penguin.com.au/books/thinking-in-bets-9780735216358
12. ‘The Systems Thinker’. https://thesystemsthinker.com/
13. ‘EP90 Joshua Epstein on Agent-Based Modeling’, The Jim Rutt Show. https://jimruttshow.blubrry.net/joshua-epstein/

? Edmund P. O’Shaughnessy, 2021