A low R2 might well indicate that variables are poorly measured, that important
variables have been excluded, or that the model has been miss-specified in other ways
(e.g. effects are non-linear or non-additive). But, this does suggest that R2 should
generally be of only secondary interest to us. If a correctly specified model with wellmeasured
variables produces a small R2, then so be it. We should be much more
interested in the determinants of R2 than in R2 itself. And, if we are going to make
comparisons of R2, we should make sure we are doing so correctly. Rather than just
saying R2 differs across groups, times, or variables, we should try to explain why it
differs (and we should definitely avoid misleading statements, such as those which
erroneously imply that a larger R2 is the result of larger structural effects.)