What happens with multiple comparisons on same column

[brianfhead]

I'm relatively new to splink and record linkage. I was trying to see the differences in gammas when using multple comparisons on the same column, but was a little surprised to get a single gamma. I'm wondering if this is because it just uses the last comparison listed for the column or if it is somehow combining the comparisons. I have three different comparisons from the comparisons library. I'm wondering if this could be a way to build an ensemble of sorts. However, I figure if this was the case it would be mentioned in the documentation or would have been asked in the Q&A and I can find neither. Appreciate any help and the package, which so far has been great for addressing a need for me.

[RobinL]

This article should clarify, specifically look at the bit where it says:

These scenarios are mutually exclusive ...
https://www.robinlinacre.com/partial_match_weights/#remaining-columns

And see also
https://www.robinlinacre.com/computing_fellegi_sunter/
and
https://colab.research.google.com/github/moj-analytical-services/splink/blob/master/docs/demos/examples/duckdb/comparison_playground.ipynb

[brianfhead]

Thank you for the response, Robin. This is helpful, but leaves me with a follow-up question. It makes sense to me that each comparison's partial weight is included in the calculation of the final match weight. However, the table returned from the predict function includes a single gamma (comparison vector value) for each column comparison.

If I include a JaroWinklerAtThresholds and LevenshteinAtThresholds on a street address column what vector am I getting? I've been using it as a sort of simple explainability tool and want to understand what the value is when I've used a couple of different comparisons on the same column. I didn't see that covered in the links you provided, but apologies if I missed it.

Many thanks for everything.

[RobinL]

Each comparison is composed of a number of comparison levels:
https://moj-analytical-services.github.io/splink/topic_guides/comparisons/comparisons_and_comparison_levels.html#defining-similarity

They are mutually exclusive, similarity can be categorised in only one of the comparison levels. Each comparison level corresponds to a single gamma value.

If you have two comparisons that refer to the same column it might actually cause a bug or error because it shouldn't be done: it would cause that column to be double counted. If you really wanted to do it, I think you have to specify the comparison as a dict :
https://moj-analytical-services.github.io/splink/topic_guides/comparisons/customising_comparisons.html#method-3-providing-the-spec-as-a-dictionary
Ensuring you specify a different output_column_name for each

[brianfhead]

Thank you for clarifying. I was hoping that they could be combined somehow to sort of deal with trade-offs of the different comparison types. Appreciate your quick response and again thanks for building an excellent tool.