62 Continuity and tracing threads

62.1 Summary

The transitivity trap can make it a challenge to interpret your maps. Some people (or somebody) said that improved hygiene led to reduction in mosquito environments, and some people said reduction in mosquito environments led to improved health, but was this the same or different people?

image-20211215081735743

You set the filter trace threads after a trace paths or trace robustness filter:

trace paths length=3 from=WASH information to=Improved health
trace threads

You can also trace threads downstream if you don’t specify target factors (to = “anything”):

trace paths length=3 from=WASH information to=
trace threads

62.2 Tracing threads (aka tracing continuity)

If your map makes it look like Cash Transfer causally influenced Increased Wellbeing further downstream of it, you surely want to know if there are at least some sources who each mentioned the whole of one of the possible paths.

We call this “tracing threads” as opposed to “tracing paths”. So a thread is a chain of individual “original links” inside the bundles, whereas a path is just a chain of bundles, regardless of which sources mentioned each link.

Tracing threads is built on top of tracing paths. After you set a trace paths filter, you can then set a trace threads filter to look at the individual threads within the paths.

“Ah so there are plenty of paths from X to Y, but are there any threads? yes look, here is MXFG-1 who has a thread all the way down”.

We can ask “how many threads make it all the way down to Improved Wellbeing”? The threads are named by the source ids, so we can say “Let’s look at downstream thread MXFG-1”.

When we are tracing threads, we can trace source threads, or village threads, or whatever. But usually we will trace in terms of sources, so if we do not specify which field defines the context, we assume we mean sources.

So a thread is a continuous chain of links with the same ID along a path. Within a bundle of links, each thread is one individual link associated with one ID (or sometimes, more than one link, if the ID mentioned that bundle more than once).

You set the filter trace threads after a trace paths or trace robustness filter. This creates the following calculated fields:

  • Links fields
    • has_threads threads
  • Factor field
    • threads_surviving
    • threads

You can for example label the links like this label links value=unique: threads.

These fields are also available as has_downstream_threads, has_downstream_threads_surviving etc, but as this filter now works upstream too, we have removed the downstream part.

You can for example colour the factors like this color factors field=threads_surviving. So if you set a filter to trace threads downstream from Cash Transfer C, imagine factor C is coloured completely red, and imagine a factor F immediately downstream of it being coloured with an intensity of red according to the number of sources which mentioned the link from C to F. In fact every factor downstream of C will be coloured with a shade of red in the same way. The central question of whether any source mentioned a whole path from C to, say, “Increased wellbeing” can be answered simply by seeing whether there is any red in the colour of “Increased wellbeing” or whether it is completely white.

… So you can just look at C and watch its influence (in terms of source continuity) seeping downstream and spreading out along all source-complete paths.

It’s like ink starting from C and flowing downstream as long as there is source continuity.

The direct counterpart, tracing upstream threads is not implemented yet. It asks about causes of effects rather than effects of causes: “Increased wellbeing” is completely red, and the redness of “Cash Transfer C” depends on the number of sources which mention a complete path from C to Increased wellbeing.

Tracing paths anchors the filter to one or more factors, and if you use thread tracing to format an aspect of the map (e.g. to colour the links) then this formatting is only related to this specific set of origin and/or target factors and will change if you trace from and/or to different factors.

image-20220301095046531

In this example, the width of the links reflects the number of sources mentioning it, but the colour of the links (and factors) reflects how many sources have a continuous story starting from “Living in a refugee settlement” all the way up to that point. This difference is also shown in the pair of numbers on the links. So the top-right link was mentioned by one source, but the 0 shows that no single source mentioned a story all the way from “Living in a refugee settlement” which also includes this link.

All of this works for tracing influences upstream too:

trace paths length=3 from=WASH information to=Improved health direc
trace threads

62.4 Advanced diagnostic filter: Show continuity

Read on only if you are interested in advanced diagnostics!

62.4.1 Summary

Above, the links are labelled with the sources.

The ▭ open half-box at the end of the first link tells us that at least half but not all of these stories stop here: less than half the sources mentioned any link out of K.

The ◼ filled box at the start of the second link tells us that all of these stories are continuations: all these sources mentioned some link into K.

The ▂ filled half-box at the end of the second link tells us that at least half but not all of these stories continue: Bob mentioned some link out of L, but Carla did not.

The ▢ open box on the link from L to N tells us that this story is not a continuation: Donna did not mention any link into L.

There is no UI for this filter yet. You can just type

show continuity

in the advanced editor.


The four kinds of boxes are (possibly aggregated) indicators of continuity, with respect to sources, between stages in a path.

If you want to look at say statement continuity rather than source continuity (the default), type

show continuity field=statement_id

If you want to see numbers (see examples below) rather than symbols (see examples further below; symbols are the default) then type:

show continuity type=label

image-20211216162518175

Here, the 0.9 says that 90% of the sources mentioning the link to ~performed well also mention the link from ~performed well. The 1 says that 100% of the sources mentioning the link from ~performed well also mention the link to ~performed well. And the zeros below say that there is no source continuity at all.

What this doesn’t tell you is, when there are more than one incoming link, which of them have sources which continue to the outgoing link (that is what the bs and cs are for in mark_links). It’s just an aggregate.

But what happens with filters which actually transform the map: zoom, bundle factors and combine opposites? Zoom can create its own version of the transitivity trap, if we have:

eating lemons –> health; no scurvy

and

health; fitness –> fast runner

image-20211216190724734

we should be very careful when concluding (when zooming)

eating lemons –> health –> fast runner

… and indeed, showing continuity highlights this error:

image-20211216190609563

62.4.2 Showing continuity with arrowtypes

Printing actual numbers (from 0 to 1) on the arrows can be very confusing. So the default is to use symbols.

  • white box: 0
  • half white box: <= 0.5
  • half full box: > .5
  • full box: 1

image-20211220172600085

image-20211220171759409

image-20211216195218720

image-20211216200541572

62.4.3 Showing continuity with colours

https://causalmap.shinyapps.io/CM2test/?s=618

Using arrowheads gives you information about both upstream and downstream flows, but it can be a bit tricky to read. Instead you can use colours to display either downstream (effects of causes) or upstream (causes of effects) continuity.

image-20220117103703826

Here we see that not so many of the people who mentioned the link from business to income mentioned the link from purchasing power to business.

Same, but upstream continuity:

image-20220117104107156

https://causalmap.shinyapps.io/CM2test/?s=619

These values are set to 1 at the edges of the map where the metric has no meaning.

Note this is not the same as the non-causal question “how many of the people who mentioned factor C also mentioned factor E”.

.http://theorymaker.info/?permalink=transitivity

62.4.4 More about these metrics

Local Continuity factors (simple) Factors (ego network)
overlap between sources who mentioned links to this factor and sources who mentioned links from this factor overlap between sources who mentioned links to the causes of this factor and sources who mentioned links from the effects of this factor

And, with links:

Local Continuity Links
Upstream overlap between sources who mentioned this link and sources who mentioned links to the cause of this link
Downstream overlap between sources who mentioned this link and sources who mentioned links from the effects of this link

Each of these metrics can be expressed as a confusion matrix and can be cashed out as different ratios. We can therefore also interpret these metrics in terms of causal necessity and sufficiency. For example, above we can say that K is causally sufficient (with respect to sources) for M because all the sources who mention causes of M (along paths from K to M) also mention effects of K (along paths from K to M).

We need to say “with respect to sources” because all these ideas are generalisable to other fields such as, for example, village or question domain.

Because these metrics (confusion matrices) are defined in terms of source_id (or some other context-relevant link variable) they partly counter the problem with previous versions of these metrics in that they provide a denominator (number of sources) even if this has to be used with some care: as usual, the fact that source S does not mention link L does not mean they wouldn’t assent to it, it may just not have appeared in the stochastic interview process.

Many different metrics are possible. These (all?) also have corresponding non-causal counterparts as in QCA, for example:

Local continuity (non-causal) Factors (ego network)
overlap between sources who mentioned the causes of this factor and sources who mentioned the effects of this factor

These QCA-type metrics (confusion matrices) are inferior to their causal counterparts because they lose the information about what causes what and only use information about co-occurrence.