Top 7 Projection (final)

Final update 5:43 PM EST.

Updated 3:17 PM EST. No ranking changes, but Jessica and Sam are now too close to call.

Updated 11:15 AM EST. Caleb cedes the top spot to Jena.

As always, these numbers will change throughout the day. I will update. (Probabilities assume a bottom 3. If only a bottom 2 is revealed, multiply each number by 2/3).

This post content is definitely inside baseball, and some of it won’t make sense unless you follow this site or the Idolsphere in general. Sorry about that.

I’m calling it: Dialidol time of death is 4/16/2014. It was a good run, but phone voting is no longer relevant. The index is either non-responsive or anti-correlated with winner each week this entire year. A halfway decent way to guess who was eliminated would be to vote for whoever Dialidol says is best.

So far the only possibly suitable substitute I’ve found for Dialidol is a poll conducted at MJsBigBlog for “Who WILL go home?“. There are a few considerations here. If we look at the entirety of the history of this poll, it’s been relatively predictive of who’s gone home. If we do the logit of the percentage of votes that the person will go home versus whether they were not safe, we find a good fit:

Coefficients:
             Estimate Std. Error z value Pr(>|z|)    
(Intercept)  1.900010   0.625569   3.037  0.00239 ** 
MJWillInv   -0.032278   0.007569  -4.264 2.01e-05 ***

This says that each percentage point a contestant had in that poll increased her log odds of being in the bottom group by 0.03, and it’s quite statistically significant. (For more information, please see the model methodology post, which explains what these numbers mean.)

However, I’m not entirely comfortable with affirming such a high degree of effectiveness. You see, part of the time the data exists (starting in season 10) was when Dialidol was up and running. That means that likely a fair fraction of the people who voted in the poll read Dialidol and then voted in the poll. However, we can look at how the poll has done this year, when Dialidol has never functioned:

Coefficients:
            Estimate Std. Error z value Pr(>|z|)   
(Intercept)  7.80780    2.74607   2.843  0.00447 **
MJWillInv   -0.09734    0.02962  -3.286  0.00102 **

Although less statistically significant, it still certainly seems to be an ok indicator. To be conservative, I have discounted the weight of the index in the model projection relative to what this fit would say. I do this because the poll hasn’t really predicted this season’s huge misses very well. In fact, it missed as badly as the other indices when it came to Jena’s early trip to the Bottom 3, and predicted M.K. was certain to be gone when she was in fact safe in the Top 11.

I want to stress that this is partly guesswork. The safest thing any person can do is make his best guess, but hedge his bets. I know the full history of WNTS and of Votefair, so I know how seriously to take them to a high degree of certainty. But I do not yet have a good quantitative idea of well MJ’s poll does when Dialidol is gone—and neither does anyone else. That’s why I’ve reduced its contribution.