Context effects (aka Scope condtiions)

Analysing a data set, like the Krook dataset built into EvalC3, we can come up with a predictive model that performs well. In the Krook data set, the combination of “quotas for women in parliament” and “country in a post-conflict situation” is a good predictor of above-average levels of women’s representation in parliament.

But we then might want to identify how this predictive model is affected by other factors which might also present. These could be seen as contextual attributes of the cases (countries in the Krook dataset). They can also be described as “scope conditions“, in that they define the scope within which a particular model performs.

If we manually add another attribute to a model, for example, “the use of proportional voting”, it can have two effects on model performance. Firstly, it could change the coverage (/recall) of the model. This is highly likely because the more highly specified a model is the more likely it only fits a smaller proportion of cases. Secondly, it could change the consistency (/precision) of the model, for better or worse.

When”the use of proportional voting” is added to the two attribute model described above this does reduce the coverage of the model (from 67% to 44%) but it has no effect on the consistency of the model, which remains at 100%. So, in a sense, it is not a scope condition of much interest. If an additional attribute did reduce the consistency of the model it would be more important, because of potential practical implications for efforts to influence or engineer the presence of model attributes.  In the Krook dataset, none of the five attributes functioned as scope limiting attributes. But it is very conceivable that in a larger data set of African countries, some attributes in this set or otherwise would so so. They would be worth identifying.