David Greenwald, writer, photographer and cat owner from Portland, Oregon, tweeted on Nov. 11 that there could have been some discrepancies in swing states between counties using electronic and paper ballots. This is captured on today's web edition of The Guardian, together with a link on some tweets by Nate Silver, including his regression analysis as follows:

Run a regression on Wisc. counties with >=50K people, and you find that Clinton improved more in counties with only paper ballots. HOWEVER:

enter image description here

...the effect COMPLETELY DISAPPEARS once you control for race and education levels, the key factors in predicting vote shifts this year.

enter image description here

Nobel laureate Paul Krugman seemingly closed the loop... on tweeter... by saying:

The response from Nate was exactly what we needed. I personally wanted real reasons to dismiss this scare, not "don't be silly" 2/

It seems silly, almost, now, and in retrospect, but there is already a letter to Congress signed by the who's who of higher education.

So perhaps it is not so self-evident after all, and I would like to ask for some senior comment as to when is it necessary to consider controlling for confounding variables, and how a seasoned statistician goes about it. In the end it is really impressive how a quick OLS is likely to have put an end to a looming conspiracy theory.

  • $\begingroup$ Please note that this is meant as a CW post intended to deepen the understanding of a current event likely being followed by millions of people, and finding its resolution in the understanding of basic statistical concepts. $\endgroup$ Commented Nov 23, 2016 at 15:12
  • $\begingroup$ Please search our site for "Simpson's Paradox." $\endgroup$
    – whuber
    Commented Nov 23, 2016 at 18:58
  • $\begingroup$ @whuber I'll take advantage of this interchange to wish you a Happy TxGiving tomorrow (and will erase this promptly). You know, I have even posted some entries of my own doing about Simpson's paradox (probably too scrappy to mention, but just to say that I actually have searched CV for some pearls on the topic - a very good one recently by Pere off the top of my head), but especially when there is no contingency table (proportions or counts) I still have nagging doubts as to whether it's confounding or Simpson's, as in this post. $\endgroup$ Commented Nov 23, 2016 at 19:07
  • 1
    $\begingroup$ Thank you @Antoni for your kind thoughts. I believe Simpson's Paradox can be construed in general as a confounding phenomenon. Even if you wish to maintain a strict construction of it as concerning contingency tables only, it occurred to me that nevertheless the same considerations apply: the decision to adjust for confounding variables seems to be the same as the decision to break a table down by the individual values of some factor. In both situations, the results and the interpretation can change profoundly. That's why I was hoping a search might turn up directly useful posts. $\endgroup$
    – whuber
    Commented Nov 23, 2016 at 20:22
  • $\begingroup$ @whuber Great, thank you. I'll keep the question open, then, since it seems to me after reading your comment, that there is room for someone to elaborate on these points. I realize that it is a very soft question, and it's very general, but it blew my mind seeing such a simple OLS dispel concerns that were taking a somber tone, and with so many "macro" considerations. $\endgroup$ Commented Nov 23, 2016 at 20:27


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.