1 Background

Going into the 2016 Presidential Election, most pollsters were confident of a Clinton win. The aftermath of Trump’s win resulted in many questions being asked of pollsters and speculation as to how so many got it wrong. I won’t get into the reasons why; here are some articles with coverage on that. Instead, I want to focus on quantifying and visualizing the amount of error in the polls – where were they wrong, and how were they wrong? I used polling data collected by FiveThirtyEight and scraped the final results from David Wasserman.

Note: FiveThirtyEight’s dataset of poll results includes both raw and adjusted results. Adjusted results correct for historical biases in different pollsters, and can be useful for eliminating some noise from our visualizations (assuming we trust FiveThirtyEight’s adjustments). However, because I’m primarily interested in visualizing how the polls actually did, my analyses make use of the raw data, unless stated otherwise.

5 Errors in state-level polls

While national-level polling doesn’t seem to have done too badly in predicting the popular-vote margin, errors in state-level polls can be more costly because of how the Electoral College works. Ultimately, whether a candidate wins the popular vote within a state determines whether they receive the electoral votes they need (with the exception of a couple of states). How well did state-level polls do in estimating the margins in the state-level popular vote? Were polls in some states especially biased?

With any poll, there will be some over or underestimation of the margin. If the actual margin in a given state is a Trump win by ~2%, a poll biased towards Trump (projecting a win of > 2%) may have a projected margin that’s too high, but it would still project the right outcome (Trump wins). A poll biased towards Clinton could project the right outcome if the bias is small enough (if the bias is 1% towards Clinton, then Trump shoulds win by ~1%), or it could get the outcome wrong if the bias is large enough (if the bias is 3% towards Clinton, then Clinton should win by ~1%).

If the polls were all equally accurate, we would expect that the distribution of bias to be symmetrical around 0 (though no single poll would have exactly 0 bias, the average bias of all the polls across all the states should come close). We woud also expect that across states won by each candidate, the projected margin will sometimes be overestimated, and sometimes be underestimated.

The following plot displays the errors between the actual state-level popular vote margin and the margin projected from polls in the last two weeks of the campaign. The separate plots are for states won by either candidate, and they’re arranged from states where Trump was most underestimated to states where Clinton was most underestimated.

A few things jump out from this plot:

In states where polls showed a Trump victory, the actual margin would be larger than projected. In states where polls showed a narrow Clinton victory, the amount of bias played a big role – if the polling was biased against Trump sufficiently, the actual outcome would be flipped. This mattered most of all in several close swing states where Clinton was expected to win but ended up losing.

6 Conclusion

As stated above, there’s been a ton of analysis about how the polls missed Trump’s win. This post has avoided those issues, except to show that there are specific ways that the polls did and did not get things right. I’m planning to write more on the data presented here – I would love to talk if you have any ideas for interesting analyses!