When we launched Seshat: Global History Databank more than a decade ago, we intended it to be a multi-purpose tool. But the central question that motivated this gargantuan undertaking was to test the multitude of theories proposed by thinkers of the past and social scientists of today to explain what I have called the Great Holocene Transformation—a many orders of magnitude increase in the scale and complexity of human societies during the past 10,000 years.

We’ve now made, what I consider, a big advance in answering this question. This advance is a result of multiple articles that we published over the past couple of years, and what I want to do today is explain how they are interconnected and how they support each other.

As an aside, readers of this blog may have noticed that recently I’ve been posting a lot of retrospectives. The reason is that over the past 12 months my life made a kind of a phase transition. Most obviously, last July I retired from UConn, where I’ve been working since 1994—nearly three decades! My main employer now is CSH Vienna, where I lead the Social Complexity and Collapse project. While this association lasts (currently, through 2026, but likely to be extended), I spend 6 months a year in Vienna, and the rest of time here in Connecticut, where I kept my connection to UConn (as an emeritus). In addition to these institutional changes, my empirical research is shifting from the emphasis on the rise of complex societies to why they experience recurrent periods of political disintegration. This is our work on CrisisDB, about which I’ll write more in the next weeks.

But back to the question of theories attempting to explain the Great Holocene Transformation. As I said above, we (the Seshat Project) have published a number of articles whose results interlock and mutually support each other. The main article, which we ended up referring to as the Big One was published in Science Advances last summer. But it required a lot of preparatory work to get there.

I should probably start the story by going back to my 2016 book, Ultrasociety, because it was where I explained (in non-technical language) the cultural multilevel selection theory, which (running a bit ahead) turned out to be best supported by Seshat data. But note that the theory and its predictions were published before we collected the bulk of data.

The first article fully utilizing Seshat data was published in 2018 in PNAS. It is foundational, because before we attempt to explain something, we need to unambiguously define what this “something” is. The 2018 PNAS article analyzed >50 Seshat variables, capturing such dimensions of social complexity as social scale (population and territory of polities; note that the unit of analysis in Seshat is not a vague thing like “society”, but a more clearly defined “polity,” which includes states and empires, but also chiefdoms, or even politically independent Neolithic villages), institutions (governance), information processing, and others. The main result was that, in a certain sense, all these measures tend to align together. In technical terms, we can define a quantitative measure of social complexity as the first principal component capturing more than three-quarters of variation in these data. So now we have the variable whose evolution we need to explain. Here’s what it looks like for several world regions:

Next, we needed to define variables that can serve as potential predictors. These are variables proposed by various theories. For example, functional theories for social evolution propose that large-scale complex societies evolved to fulfill certain function, such as provision of public goods and infrastructure. Accordingly, we collected data on many such functions in Seshat. Other theories point to class struggle as the engine of social complexity, and we coded class for all polities in Seshat.

Two kinds of variables required special attention, and we ended publishing separate articles devoted to each one. The first one was agriculture. Nearly all theories today acknowledge the key role of agriculture, as a necessary (but perhaps not sufficient) condition for the rise of complex human societies.

But how to measure it? First, we quantified the productivity of agriculture in tons produced per hectare per year. This required a lot of work, and was published in 2021 in Holocene. Next, some theories posit that it is not agriculture productivity, but its antiquity that’s important (because the longer time societies practice agriculture, the more time they have to evolve social complexity). Accordingly, we used data from the ArchaeoGlobe project to define this variable. As an aside, wherever possible we use data collected by other researchers; the problem is that the bulk of data we had to code ourselves. Finally, some theories suggest that it’s not how much is grown, but what is grown. Because grain in granaries is easier to take away from farmers by state agents, compared to root crops, this gives us another potential predictor variable to test in the analysis.

The other difficult variable, in addition to agriculture, is warfare. Many theories, including the one I explain in Ultrasociety, propose that it was severe competition between polities, most often taking the form of warfare, which weeded out smaller-scale and less complex ones. Quantifying warfare was the goal of the article in PLOS One, also published in 2021.

Another area where we did a huge amount of work was religion (talk about a hard variable to code!). This topic requires another blog post (but also see this one).

And here I am talking only about most important articles (for a complete list of Seshat publications go here). But once these foundations were laid (by 2021), we could finally get to the Big Article. In the end we were able to collect data on 17 predictor variables proxying mechanisms proposed by five major classes of theories, including agriculture, functional factors, religion, class, and warfare.

I won’t go into the details of our results (worth another blog), but just say that the results of our analysis were very clear-cut. Two classes of potential predictor variables were strongly supported, while the rest were not. There is a lot more work to be done (as we discuss at the end of the Big Article), but overall all this enormous work over more than a decade, was clearly worth the effort.