Closed light-and-salt closed 8 years ago
I think this is an overstatement. The real issue here is the plot's density. https://github.com/uwdata/compassql/issues/277
I think density is related, but clusters also matter. Even if the data is dense, if there are clear-cut color clusters, user can still reason about clusters.
For example, in the chart below, the purple and orange clusters tell something: Dramatization and Factual types are based on real life events... (whether that's truthful or not is a different story)
When there is less color clustering, then you don't know what to say about a colored chart. Is there a relationship? Hard to say!
Then it starts coming to a point where you can apply statistics, which will in turn introduce inconsistencies.
Interesting research to be done though
When you have a QxQ scatterplot, adding categorical field as color isn't generally helpful unless the categories can cluster in the QxQ scatterplot (like the
x:Horsepower, y:Miles_per_Gallon, color:Origin
example in our tutorial).