Thread by @gelliottmorris

🚨NEW from me and @TheEconomist's data team: Our interactive for the 2020 Democratic primary! We have a fancy polling average, candidate support by demographic group, data on who candidates' core voters are also considering and more. <THREAD> (No paywall!) projects.economist.com/democratic-pri…

1. Here is a collection of screenshots from the interactive. Allow me to tell you about some of the features, how they work, where we get our data and how we did our modeling. 1/13

2. The first element on the page presents our average of high-quality national polls of the Democratic primary. We include phone polls from firms that survey with live interviewers and online polls from firms that use well-documented, trusted methods (ie no IVR/Mturk combos).

3. We average the polls over time using a Bayesian implementation of a dynamic Dirichlet regression model. The model is specified to give more weight to higher quality pollsters, less weight to data collected in the past and incorporates pollster house effects.

4. The method used to calculate house effects is complicated, but know we let the model sort it out. Effects can change over time. (Something cool: the model also allows the average to learn about candidates' standings at the state level—but we're not releasing this just yet!)

5. Our 2020 site also features an array of interactive visualizations built on our weekly polling data from @YouGovUS. Breaking down support by demographic and political group is a major contribution of this project and something few other media outlets are providing.

6. We show a selection of demographics on the home page, but we make more detailed data available for each candidate on their own separate pages.

E.g. here are Warren's crosstabs: projects.economist.com/democratic-pri…

You can select candidates with the slider.

7. These stats are derived from the last month of data from YouGov. This gives us a large enough sample size to minimize sampling error while staying relevant, but also means the figures might differ from other analyses of YouGov's data. If you see small differences, fret not.

8. Though we focus quite a lot on first-choice vote intention, YouGov also provides some particularly insightful data in asking who voters are "considering" voting for. This may give us a reasonable upper bound on candidates' support in the polls.

9. One of the coolest parts of the project is this breakdown showing the share of each candidate's first-choice supporters who are considering other options. For example, Nearly 60% of Warren voters are also considering Harris, and 15% of Sanders voters consider Yang, too.

10. Because polls aren't all the information that may decide who is most likely to win the primary, and though prediction is not our ultimate goal with this project, we also show smoothed market betting odds.

11. The interactive updates at least daily, and sometimes more often when we have new polling, betting or YouGov data. We will be adding features (like state-levels polling) over time, so follow me for updates!

12. Worth giving a few hat tips. First up is @FiveThirtyEight for open-sourcing their collection of 2020 polling. @PoliSciJack also helped me implement a version of the that model we ended up using. Thanks to @PredictIt for archiving market data!

13/13. We hope you will find this as fun and insightful as we do! @martgnz and @futuraprime put their blood, sweat and tears into getting the website working and visualizations looking nice. It was a team effort.

If you have any feedback, we'd love to hear from you! Enjoy!

@martgnz @futuraprime Oh by the way, if you scroll over a candidate's face they will wiggle