As promised, this is the last edition of the “learning from last year” series (see parts I and II). Here, I’m just going to post what the two election models mentioned in Part II (the linear prediction and probit models) would have predicted, versus what actually happened. In both cases, the models got 14/15 candidates right, but they each missed a different one (and falsely predicted a different candidate as well). As a methodological note, the way that these predictions were created was by generating the probabilities predicted by each model and assigning any value over .5 as a “win” (i.e. successful election) and anything less than .5 as a “loss” (unsuccessful campaign). Without further ado, let me begin:So who are these people (unless we have a senator named “observation 1?”)?
Well, here are the results that matter (see below for the full results): both models correctly chose Zietsman, Warma, Wharton, Wanyoike, Ali, Cruz, Johnson, Limon, Sivaram, Young, Gao, Singleton, and Creasman. However, the linear model falsely predicted the election of Willmott, while missing the actual election of Jackson. The probit model falsely predicted the election of Henry while also missing the actual election of Jackson. Not too shabby.
Full results (I apologize for the awkward appearance: I blame the underlying blog platform):