Chapter 2 Manifesto: piecing together fragments of causal information

Gone are the days when we could think of data or information as primarily about numbers (or perhaps also qualitative narratives). We need to be able to store, present, understand and manipulate causal structures. In particular, we need to be able to construct models of other people’s models of causal structures.

There are plenty of tools for storing and presenting causal structures, primarily of two types

  • graphical e.g. Kumo
  • theoretical e.g. dagitty.

What we lack are tools of either type which help us to construct

  • models which include other people’s models. These are really important if we want to model the behaviour of people, groups or systems which have their own implicit or explicit Theories of Change
  • different versions of the same or overlapping models, e.g.
    • different stakeholders’ (maybe partial) information on the same causal topic
    • different editions or versions of a Theory of Change.

But first we need to take a step back and look in general how we (should) code causal information, before we take the next step of coding causal information from multiple sources.

The fundamental, radical points are these:

  • Causal information is primary information. It isn’t something which exists only virtually as a potential conclusion on the basis of observations of non-causal variables. There is a fact of the matter about what causes what, just as there is a fact of the matter about the number of people on a training course.

  • Parallel to that, humans’ perception of causation is primary, as primary (and fallible) as our perception, say, of colour. All the things which we know, or think we know, about our world – from the colour of that dress1 to the way the wind shakes the trees – have already been through a lot of cognitive processing, and none of it is “secondary”. So when we ask stakeholders the “why question” (what causes what in this domain), we are not asking them, first and foremost, about what they deduce from their (non-causal) observations in the way we might as scientists or researchers. That would be a very shaky method; we might be better just to ask them to tell us about their non-causal observations and we could try deducing causality themselves. No, we are asking them about what causes what based on their underlying understanding of the causal structure of their world. That underlying understanding might be somewhat erroneous, but it’s certainly a lot better than ours. Copestake (Copestake, Morsink, and Remnant 2019): “..… attribution claims underpinning the QuIP do not require a control group, nor indeed variation in exposure to the intervention across the sample of respondents interviewed. Rather, causal claims rely on the integrity of ‘within-case’ statements made by respondents themselves”.

  1. reference↩