Latent Class Modeling: I'm pretty sure this is the future. It's a much more powerful, higher resolution method than linear correlation and its derivatives. The whole issue of frog jumping I talked about in an earlier post is eliminated by attending to different profiles of participants, which LCM does.
Non-narcissistic self-esteem: It would be extremely cool if someone built a reliable self-esteem measure that doesn't also carry narcissism. RSE is going to carry both high self-esteem and high narcissism (or narcissistic self-esteem and non-narcissistic self-esteem.) So you have to use supplemental measures, like NPI-16, to tease them apart, or cross explicit and implicit self-esteem (which isn't as valid, probably.)
Kernis was heading in that general direction before he passed away. Implicit self-esteem is tough. I know that Diener and colleagues used initial letter liking as an implicit measure. I think it might have done some work in predicting life satisfaction. (Letter liking is presenting people with their own initials and capturing their attitudes, like J for me, or D. I'm not sure how much validation we have on that.)
Rosenberg also bothers me a bit because is wording is dated, and probably comes off as awkward to a lot of contemporary participants. This is likely to be another source of noise.
I'll probably try out some new measures. I also love Simine Vazire's method of repeating questions on personality measures, basically Are you sure this is the right answer? It correlates better with peer ratings, thus seems more accurate.