56 📚 Tracing paths and calculating robustness

56.1 Summary

56.1.1 How to trace paths and calculate robustness

Tracing paths allows you to view full causal pathways and to analyse the relationships between specific causal factors. You will find the trace robustness filter in the analysis filters section.

Irgj26EgMI

This will open the trace robustness panel which asks you to select how many steps to trace and which factors to trace from and to. Type one or more factors to begin your path (these are the origin factors) and which factor or factors you want the path to end at (the target factors). If necessary, move the slider to select how many steps down you want the path to go.

If you want you can also leave either (but not both) factor selectors blank, e.g.

  • to search from some specific factor or factors to “anywhere”, or
  • to search from “anywhere” to some specific factor or factors.

If you want to search from or to the main drivers and/or outcomes:

  • put “main_drivers” in the first box
  • and/or put “main_outcomes” in the second box.

“main_drivers” uses an algorithm which identifies 1-3 factors with the highest driver_score , i.e. and likewise for outcomes.

Once you click apply filter your map will appear, as below. Orange indicates the origin(s) of the path, purple the target factor(s) and green any factors in between.

image-20211016141204775

In this example, there are causal paths from the two external factors to the two outcomes. In real-life examples, these paths may be much harder to follow, so their are best seen and shown by using path tracing, as it simplifies the map and highlights the intervening factors between the two factors.

A path length of 1 will only show the one step in the causal chain from/to your chosen factor, i.e. A ➜ B. A path length of 2 will also show the next step in the causal chain (if there is one!), i.e. A ➜ B ➜ C.

Path tracing is the prerequisite for calculating Robustness.

It is also possible to count, for example, how many sources mentioned a complete path.

56.1.3 How to calculate robustness

Once you have applied the trace path filter, you can find the robustness calculations in the robustness tab. Here is the table for the example above.

UpSR4OnxP6

image-20211017130218716

In this tab you will find a table displaying the number of links between your origin and target factors. A larger number means more evidence; more pieces would have to be deleted until we gave up the hypothesis that the Origin factors influence the Target factors at all.

For example, in the top row of the table above you can see that there is more evidence that rainfall influenced both of the targets compared to the influence of loss of forests. In the first column, you can see that there is more evidence of the influence of external factors on people moving away than on people getting angry.

The following section will expand on how to understand these figures.

56.2 Understanding robustness

Robustness identifies how many pieces of evidence (individual mentions of individual links) would have to be deleted so that there is no longer any path from the Origin factors to the Target factors. It is important to understand a large number does not necessarily mean that this path is strong in the sense that the Origin factors have a large effect on the Target factors. It says there is plenty of evidence for the path, whatever its strength.

Subtotals are provided for each Origin factor, when there is more than one, and/or for each Target factor, when there is more than one; it can be useful to compare the amount of evidence for the impact of each Origin factor (and/or Target factor).

These numbers are not very useful on their own. But they are useful for comparisons.

56.3 Amount of overlap

If you look at the robustness of the path from interventions B and C to an outcome E, one thing you can do is compare the sum of the score for the two sets of paths and compare it with the overall score. If the overall score is more or less the same as the larger of the score for B and the score for C, you can assume that the (evidence for the) intervention paths overlap - so if the score for B is larger, you could say that C uses many of the same paths. But if the overall score is more like the sum of the two, you can deduce that the interventions work more independently. So for example if you are already funding intervention B, and you have to decide whether to fund C or D, and your causal map says that the score from C to E and from D to E are similar, you’d want to look at the pairs - you’d want to look at B plus C to E and compare it with B plus D to E. B + D might have a much larger score than C +D, because C shares many paths with B, but D doesn’t. So you’d want to fund D, as the combination of B and D would be more robust.

Of course, in principle you can see all of this just by looking at the maps and to what extent B, C and D share paths to E. But sometimes it is hard to see and this metric anchors a narrative description in actual numbers.