Naming is a social problem, and there's a strong risk of correlation with hidden social variables in any empirical study, while any non-empirical study is suspect because humans are what counts.

In particular the most likely explanation for correlation between bad names and bad code is unskilled or inexperienced programmers. This needs to be controlled for.

My personal observation is that well-factored code can use shorter identifiers because it deals with fewer concerns. Long identifiers frequently indicate muddy thinking that mixes concepts, where the identifiers are burdened with disambiguation.

Further, in languages with particularly verbose identifiers (Java especially), it can get hard to see the wood for the trees: the density of text obscures code and data flow in the program. There's a similar argument to be made about the tradeoff between symbolic operators and textual operators: the more frequently they're used in a domain, the better they're represented by a symbol. Symbols, of course, are especially short identifiers.