That was my first thought as well. Also, traditional R squared with binary variables, or maybe categorical variables, never made much sense to me. The "meaning

That was my first thought as well. Also, traditional R squared with binary variables, or maybe categorical variables, never made much sense to me. The "meaning" of the variance I don't think is quite the same. You generally have to nonlinearly transform (e.g., logit) any linear model quantities into something else to put it on an observed variable scale.