Revising Revisionist Polling History

The polling industry took a lot of heat—some of it deserved—in the wake of the 2016 presidential election. To be sure, several polls in key swing states (like Wisconsin, Michigan, Ohio, and Pennsylvania) under-sampled key constituencies that ended up voting for Donald Trump, and not (as some of those polls suggested) Hillary Clinton. But the vast majority of polls during that election cycle measured national, not state-wide, sentiment. So, how did those polls fare in projecting the national popular vote?

According to the Real Clear Politics average of national polls conducted between November 1 and November 7, 2016, Clinton averaged 46.8%—in fact, she ended up getting 48.2% of the actual vote. That means the polls were off by 1.4 points. The same RCP average showed Trump with 43.6% in the polls—and he got 46.1% on Election Day. So, the polls for him were off by 2.5 points. 

And, if you look at the projected margin of popular vote victory from these same polls, Clinton was predicted to win by 3.2 points—in fact, her popular vote margin was 2.1 (a difference of 1.1 points). Not exactly a catastrophic failure of the polling industry.

Many pollsters re-calibrated their techniques for the 2020 election cycle and did much better on state-based surveys. Of course, the only true way to project an Electoral College win is to cobble together 50 state polls and do the math. The expense for such an effort would be prohibitive—so the media continues to focus on national averages. To be fair, the polling industry should be judged on what it (and its media sponsors) can realistically conduct: national samples that project popular vote. On that score, we have been (and continue to be) pretty darned good.