How Analytical Models Failed Clinton

This story was originally published on nationaljournal.com on December 27, 2016 The Novem­ber elec­tions pit­ted Demo­crats against Re­pub­lic­ans, con­ser­vat­ives against lib­er­als, Trump-style pop­u­lists and tea parti­ers against the es­tab­lish­ment and con­ven­tion­al politi­cians. An­oth­er con­test, fol­lowed mainly by polit­ic­al afi­cion­ados, matched tra­di­tion­al poll­sters against newly fash­ion­able ana­lyt­ics wiz­ards, some of whom—pre­ten­tiously in my opin­ion—called them­selves “data sci­ent­ists.” It was well known that tra­di­tion­al polling was hav­ing prob­lems. The numb­ing ef­fect of bil­lions of tele­market­ing calls and the ad­vent of caller ID and voice mail had re­duced re­sponse rates (the per­cent­age of com­pleted in­ter­views for every hun­dred at­tempts) from the 40s a couple of dec­ades ago to the high single di­gits. As they struggled to get truly rep­res­ent­at­ive samples, poll­sters “weighted” their data more than ever be­fore, mak­ing as­sump­tions of what the elect­or­ate would look like on elec­tion days that were weeks, months, or even a year or more away. Prob­lems with tra­di­tion­al, live-tele­phone polling led to ex­per­i­ment­a­tion and more re­cently a grow­ing ac­cept­ance of new meth­ods like In­ter­act­ive Voice Re­sponse, pop­ularly known as robo-polls, and on-line polling. Each new meth­od brings