lukestein’s avatarlukestein’s Twitter Archive—№ 7,281

  1. …in reply to @paulgp
    @paulgp I don’t think either helps with the aggregate. But to totally abuse terminology, I think any model evaluation should pick something up about what the model adds *incrementally* to e.g., just using state populations. 1/
    1. …in reply to @lukestein
      @paulgp Reductio ad absurdum: Consider a model which just forecasts claims = 𝛽*population for every state. claims predicted vs actual will have the SAME POSITIVE R^2 for ANY choice of 𝛽>0, even though nailing the aggregate is all about getting 𝛽 right? R2 totally uninformative. 2/
      1. …in reply to @lukestein
        @paulgp Using per capita claims predicted vs actual will have the same ZERO R^2 for any choice of 𝛽>0, even though nailing the aggregate is all about getting 𝛽 right. Again, R2 totally uninformative. 2/