Name of "association measures"

[Datseris]

As I am updating my tutorials and everything else, I am also taking the change to be very careful about the naming of things, as well as using the same name for the same concept.

One thing I have been going back and forth at is how we name "association measures".

Originally, I had the tutorial "causal timeseries analysis". We already discussed why its good to avoid using the word "causal", because often the "physical causality" is not what a method infers, or what we are after anyways. Hence, I decided to start naming "relational timeseries analysis".

In the docs of CausalityTools.jl we have "Association measure". However, after googling the term shows that it is used exclusively for correlation (or linear regression coefficients). I am wondering, is there any other term that we can use that is not already plagued by a specific concept?

RelationalMeasure would be something, but I think @kahaaga you have read more of the literature to know if a generic name exists.

If you argue that "AssociationMeasure" is the best name, then I will change the workshop to "associative timeseries analysis".

[haagadev-kristian]

Hey @Datseris! Thanks for bringing this up.

I've actually been pondering for some time whether this package should be renamed to something like AssociationMeasures.jl. Or, perhaps, inspired by your suggestion: RelationalMeasures.jl. Having "causal" in the name of a package that is dedicated to computing many different measures, some related to causal inference methods and some not, isn't the best idea, I think.

Modern "causal inference" inference methods are typically graph-based and use (conditional) independence testing, and takes some (conditional) association measure as input. One could in principle argue that these graph-based methods belong in a separate package, because they are "meta". However, at this stage, we've implemented too few of them to rationalize having their own package, so they should stay. Anyways, the term "causal inference" is used explicitly in the graph-causality frameworks, with good theoretical justification, but otherwise we should be careful. As you say, the use of the word "causal" and "causality" really necessitates a precise definition of what the word means in the specific context it is applied. This again comes down to subjective choices, definitions and assumptions.

As far as I can see, there no good consensus in the literature on an umbrella term for the methods offered here. One fact is that we offer statistics (functions of input data) for two two or more variables, that just so happens to quantify some interesting aspect of the data, some of which just so happens to be useful for independence testing and "causal inference".

We want something that explicitly captures that the measures quantify something between two or more variables. But it must also be generic enough to account for conditioning. I think "association" works, and "relational" does too, so it comes down to preference.

I personally favor "association", because it is the most generic and captures the notion of "statistical association", which may or may not be physical. "Statistical association" is at least terminology I've seen before and regularly use myself. On the other hand, "relation" seem to suggest some sort of physical relation. Moreover, "statistical relation" sort of leaves me hanging - I immediately want to ask "what sort of relation, then? Please specify" (just from common use of the word "relation" in regular language).

[Datseris]

okay, sure, that makes sense. I will be using "associative timeseries analysis" in my lectures from now on.

Regarding renaming the package. I am not sure. The name is catchy and has been registered already. The name AssociationMeasures.jl is of course "more correct" from a techinical view point. There is an argument for both sides. The goal of the package name is to attract attention. The goal of the package description is to be more accurate. The package description can be "Generic and extendable framework offering a multitude of definitions and computational tools for inferring associations (linear, dynamic, causal, or other) between data." which is both very accurate description and very formal. For the name, I will leave the decision to you. I am fine either way but I don't find CausalityTools.jl a bad name either. You should strive for the name that can bring a newcomer in more.

[Datseris]

BTW, I am assuming that I am talking to Kristian A Haaga, right? What has happened to @kahaaga ?