This section is all about our own polling. We collect data from either online or mobile-based panels and then model and post-stratify the data with bleeding edge technology. Example of work include the PredictWise/Pollfish (with Wisconsin and Pennsylvania going to Trump) and MSN (with Wisconsin and Michigan going to Trump) 2016 election polling. Our methodology also showed Trump by 1 percentage point in Florida, where everyone else had Clinton from 1 to 4 percentage points.
Blog Posts on Findings from PredictWise/Pollfish Poll:
Trump Voters: extra surveillance for Muslims. PredictWise January 8, 2017.
Trump Voters: Abandon NATO for Russia?. PredictWise January 7, 2017
Trump voters support blocking immigration, but not that into deportation. PredictWise December 21, 2016.
Trump on Trade: Good Politics Bad Governance. PredictWise December 18, 2016.
Trump tax plan would shock Trump voters. PredictWise December 11, 2016.
Obamacare is very popular. PredictWise December 5, 2016.
Trump and the GOP Have Massively Unpopular Tax Policies. Slate November 30, 2016.
Pollfish Election 2016. PredictWise November 10, 2016.
New poll: 30% of Trump supporters say they’ll protest the election if he loses. Washington Post November 8, 2016
‘It Really Does Get Into Your Head.’ The Election Through the Eyes of Teenage Girls. New York Times November 4, 2016
Preparing for Future: Smartphone Only Polling. Huffington Post October 30, 2016
Poll: 60% of Republicans believe illegal immigrants vote; 43% believe people vote using dead people’s names. Washington Post October 24, 2016
Will Republican voters flee Trump, post-video? Here’s what we found. Washington Post October 14, 2016
Tax Increases for High Income Really Popular PredictWise September 29, 2016
Many Voters Want to Decrease Flow of Immigrants PredictWise September 29, 2016
Strong Support for Raising Taxes on Income Over 250k PredictWise September 5, 2016
Guns, Massacres, and Ephemeral Shifts in Public Policy Positions. PredictWise August 22, 2016
Quick Take: solid convention bounce for Clinton. Huffington Post July 30, 2016
Who cares about free trade? Not many Americans, it turns out.. Washington Post July 29, 2016
Guns And Immigration: Critical Wedge Issues Of 2016. Huffington Post July 9, 2016
There is no “Trump Bump” in the polls — just a growing lead for the Democrats. Washington Post, June 7, 2016
New polls show that more Americans prefer Democrats’ policies. Washington Post, May 27, 2016
Will it be Clinton or Trump in November? Here’s what’s wrong with most predictions. Washington Post, May 3, 2016
Trump is closer to what most Republicans believe than the ‘establishment’ candidates are. Washington Post, March 6, 2016
Reports from PredictWise/Pollfish Poll:
Civil Rights Poll on January 5, 2016: PredictWisePollfish_CivilRights_20170105
National Defense Poll on January 3, 2016: PredictWisePollfish_NationalDefense_20170103.
Trade & Immigration Poll on December 13, 2016: PredictWisePollfish_TradeImmigration_20161213.
Taxes & Regulations Poll on December 5, 2016: PredictWise&Pollfish_Taxes&Regulation_20161205.
Healthcare Poll on November 28, 2016: PredictWise&Pollfish_Healthcare_20161128.
Technical Details of PredictWise/Pollfish Poll:
Pollfish runs in-application polling on third party mobile applications. We are reporting the modeled and post-stratified results. First, we run hierarchical logit regressions of the answers 1-7 on age, gender, geography, education, race, and party identification. This is run in Stan on R. Second, we estimate the population of likely voters using TargetSmart‘s voter file. We determine the percentage of the voters that are in any demographic cell (i.e., combination of demographics). Third, we post-stratify the results of the model onto the target populations of likely voters in the 2016 presidential election. This is experimental work, but we want to highlight three things:
Cheap: the retail cost of running each survey is $1,000 versus $20,000 retail for a comparable sized RDD telephone poll.
Fast: while we are trying to be consistent with the questions, we can add questions and have results in 1 or 2 hours. Further, the lower cost allows us to have high time granularity, which is prohibitively expensive with RDD data collection.
Accurate: our research shows that this method is accurate. But, we are also going to emphasize shifts in sentiment as much as levels of sentiment, where shifts in sentiment reflect the within survey design changes. As long we are surveying from a similar sample frame each week, the shifts are real within that sample frame, regardless of any concerns over the exact level of the sentiment.
There are two really topical papers to reference for this work. First, David and Sharad (along with Adam Obeng) have a paper called Non-Representative Surveys: Fast, Cheap, and Mostly Accurate which talks about the spectrum of data collection. Second, David and Sharad (along with Wei Wang and Andrew Gelman) have a paper called Forecasting elections with non-representative polls which talks about the analytics of modelling and post-stratifying in more depth.