2.2·Change the model

Does the finding change when the model changes?

To keep the nuance across countries, we ran one topic model for each country. An alternative would be to run a single pooled model across all three, but the largest collection might then dominate it. This section shows what we lose when we switch from a country model to a shared one.

How the model test worked

The main analysis used a separate topic model for England, Scotland and the Republic of Ireland. This allowed each collection to be organised around its own language and patterns.

We also ran one pooled model across all three collections to compare what was gained and lost when the same evidence was organised in a shared model rather than in three country-specific models.

Both approaches analysed the same 60,678 passages. No evidence was added or removed. We traced the same passages through both approaches to identify which topics merged across countries, which remained distinct, what new groupings appeared and which passages were left without a topic.

This isolates one modelling choice: three country-specific models or one pooled model. It shows what each structure preserves, combines or leaves out. It does not test whether the document collection was representative, whether the topic labels were correct or whether topics grouped together by the pooled model carried the same meaning in each country.

What changes when the model changes?

England Scotland Republic of Ireland merged: more than one country arrives outlier, or scattered

What do you make of this comparison?

What to take from this

Changing the model does not erase the broad differences, but it changes how the evidence is organised and how much of each collection receives a topic.

What this shows

Where the same passages go when three country-specific models are replaced by one pooled model.

How we read it

Country-specific patterns remain visible, but pooling affects subjects differently and leaves more Scottish and Irish material unassigned.

What it does not tell us

Which model structure is correct, or whether topics that merge use the same language for the same institutional meaning.

Before you use this

Treat agreement between the models as evidence of robustness, not proof of validity. Any result from the pooled model must carry its lower coverage of the Scottish and Irish collections.