Methods for Custom Polling


We generally measured public opinion using Random Device Engagement (RDE). RDE means targeting Ad IDs, device identifiers, randomly on the platform unique users behind the Ad ID spend their time.  That is to say, we reach potential respondents where they spend time organically: when they engage in their quotidian tasks at home, get information at home, and interact with friends and family. Coverage is reasonably high, insofar as smart phones (as a key device type today) have very high penetration ~ 70%), and quality of data is good because surveys  and incentive to participate are embedded into the feel and design of the application where we pick up respondents.

We then create estimates of public opinion using the most bleeding edge analytics of modeling and post-stratification (affectionately known as MRP+). Compared to other polling companies, our approach is differentiated—and superior—in three important ways:

  1. Depth: We can present movements with unprecedented demographic and geographic granularity. The data file includes the combination of two demographic (e.g., 18-24 year olds with a "some college" education). This type of depth will be crucial in online targeted campaigns moving forward.
  2. Speed: We can assess opinions on political matters within approximately three hours from customer request to result presentation. This also allows us to assess public opinion to crises relating to core democratic institutions, protests, etc.
  3. Cost-Effective: We can do a lot of polling, because our method costs less money.
  4. Accurate: It is also accurate. Our state-of-the-art method uses machine learning to convert whatever polling data we collect into results that are at least as accurate as results from traditional random-digit dialing polls.

First, we model the raw responses to each question, given each respondent's age, gender, location, education level, race, marital status, party identification, income, family size, and urbanicity. This information divides the population into thousands of demographic categories. For each sub-group and poll question, we predict the percent of people that would provide each answer if the entire country showed up to the poll.

Each of these predictions is informed by all responses, including responses received in previous polls. To achieve this, we have developed a complex dynamic model that allows us to parse out variance in sample composition from true swings over time. This is crucial for assessing trends in polling data, as illustrated by this cautionary tale from the world of political polling.

Second, we project our estimates for each sub-group onto our best estimate of the likely voting population, a process known as post-stratification. Specifically, we weight our predictions by the fraction of that sub-group in the overall target population. We derive the target population from a Big Data combination of population-level census data, and proprietary financial and political data sets, including background information on all registered Americans.

The transformed data provides meaningful information about many segments of the population. Our approach has been validated in peer-reviewed articles published in leading academic journals (see here and here) and real-world event predictions (see here and here). Our polling has lead to major publications in the US’ leading news platforms, including but not exclusive to The New York Times (see here and here), the Washington Post (see here and here), or Slate (see here). For a full white paper, click here.