David M Rothschild on Posted on

Democratic presidential nominee Hillary Clinton is 81% likely to win the presidential election on PredictWise‘s market-based forecast. And, poll-based forecasts have started to converge with PredictWise. As we roll past major events: the conventions, the debates, “October surprises”, we run out of events that can shake up the voting population and swing the few remaining persuadable voters. Finally, Election Day becomes our last major source of uncertainty: how closely will the voting match the polling? Could Clinton outperform polling due to strong Get-out-the-Vote (GOTV) operations? Could Donald Trump outperform the polling due to social desirability bias (i.e., the Bradley Effect)?

While individual polls have a lot of error, the average of all polling is very accurate year-after-year. Sharad Goel, Andrew Gelman, Houshmand Shirani-Mehr, and I have a paper (digested in the New York Times) that shows that while the true margin of error of any individual poll us upwards of 7 percentage points, the average of presidential polls in any given contested state is off by about 2 percentage points. The largest error, in any contested state, from 2000 to 2012, was just 5 percentage points. Currently, Clinton is winning in over 270 Electoral Votes by over 4 percentage points.

Clinton will likely benefit from an advantage in GOTV operations, we just have no great way to estimate it. We do not know the true impact of GOTV, because we have never seen such an imbalance. Research can estimate the impact of GOTV in a world in which both sides have robust GOTV in competitive states, but that does not necessarily apply to a world in which one candidate has a huge advantage in resources. What we can say is that it should help her, or campaigns have been wasting tens of millions of dollars each election cycle!

Could this advantage be off-set or “trumped” by Trump benefiting on Election Day, versus the polls, from voters being too embarrassed to tell pollster they support Trump? The social desirability bias occurs when people say they will do something, because it is socially desirable. It is nicknamed the Bradley Effect after Tom Bradley, an African-American, was leading the polls for California governor, but lost. The theory is that people said they would vote for him, to sound post-racial, but actually voted for his white opponent. Applied to 2016, the theory is that Trump supporters are too embarrassed to say so, but will come out to vote for him.

To investigate social desirability, I divided Huffington Post’s Pollster’s aggregation of national polls into four parts: either including Johnson or not including Johnson, and live telephone or IVR (automated calling)/online/internet polling.

Hypothesis: people will be more comfortable telling a machine they support Trump than a live person.


1) Clinton is winning no matter how the data is sliced: Clinton’s support is 45.5%, 43.1%, 46.4%, and 44.8%, while Trump’s support is: 41.7%, 41.1%, 42.0%, and 40.8%. By slicing the data into four parts, we are actually lowering Clinton’s true support a little, because there are already so few post-debate polls in the aggregation. So, this is generous to Trump and he is losing no matter how you look at it.

2) Trump is relatively stable between live person and machine: He is slightly lower with machines: -0.6 pp and -1.2 pp. He is not doing better with a machine he is doing worse. Of course, this has to do with a small mode effect, live people get less Johnson/Undecided/Other than machines. But, I think it is safe to say that Trump is basically flat with between 41-42 percent support.

3) Clinton drops more between live person and machine: -2.4 pp and -1.8 pp, and this is actually pretty good for Clinton: What this means is that when not explicitly given the choice of Undecided/Other the respondents answer Clinton (i.e., on the phone with a live interviewer). In the voting booth, Undecided/Other is not an option, so these voters are more likely leaning towards Clinton when forced to make a choice.

In short, there is no indication that social desirability is artificially deflating Trump support. I feel very confident that Trump’s support is mired around 41-42 percent, while Clinton is somewhere between 45-47 percent depending on the variables in play. And, Clinton has more room for growth.