Thread by @GerritsLasse

Can one combine agent-based modelling (ABM) with qualitative comparative analysis (QCA)? For the past few days, we’ve attempted such a mixing of methods. Three teams worked on different prototypes @lorentzcenter. A thread about some of our findings. 👇 #complexity

Why combine QCA & ABM? QCA shows configurations of conditions that co-occur with an outcome (i.e., is case-based) but not the mechanisms that drive the outcome. QCA is also (mostly) static. ABM is dynamic, highlights the mechanisms but misses the configurational, case-based bit.

The two main challenges: (1) QCA runs on a set-theoretical approach to causality; ABM runs on vector causality. (2) QCA reconstructs what has happened; ABM simulates what could happen given a set of assumptions. There are technical challenges, too, of course.

What did we do? Starting point: cases may shift from one configuration to another over time. It means that the solution pathways at t1 have changed at t2; as moving cases come with changes in set-membership scores, hence giving different weights to configurations of conditions.

It is a simple idea but it gets complex quickly. This was our workflow:

Of course, we only had a couple of days to hack something together so the model was full of assumptions, weird artefacts, and producing random outcomes. So let’s look at some of the issues we encountered when doing this.

Modelling like this needs a ton of theoretical and empirical underpinning. Not only does one need information to identify conditions and calibrate cases; one also needs a serious set of assumptions about how agents interact.

Casing becomes incredibly important. Are all agents separate cases? Or may they be grouped into archetypical cases? Can cases change constituent parts? Should we identify cases at different levels? (e.g., individuals in teams in organisations).

Sampling: A QCA default says that sample size is partially determined by one’s case-based knowledge. ABM doesn’t have an upper limit for sampling. This is interesting as agents follow the behavioural rules put there by the researcher (i.e., case-based knowledge for 10ks of cases)

We also saw how the different groups freak out over different things. Randomization? Curse you, ABM-people! No network rules? What do you QCA-people think you are doing? <-tongue firmly in cheek, of course. It was fun!😁😅

Other open issues include differences in jargon (e.g., calibration is a different operation in QCA compared to ABM), functional sequences of methods, and many more.

How far did our team get? We defined a topic, cases, conditions and calibration rules for cases (agents). We built some archetypical cases but resorted to some randomization for calibration so the results were strange.

It is a start, though! We’ll continue to work on this. Shout out to team 'GreenQCAgents' Timo, Diego, Barbara and Benoît @rihoux_methods - and big thumbs up to the organisers and hosts @lorentzcenter !