Backend

• update to multivariate normal rng in generated quantities
• update to 10,000 iterations (1,250 per chain, 8 chains)
• re-run from 5/1 onwards
• Add all missing function documentation
• Update repository README
• Add sub-dir level READMEs
  • R (add a README with a section for each of the sub-dirs of this folder, i.e., ## data, ## model ...)
  • data
  • out
  • stan

Frontend Support

• Finalize text
  • Heading
  • EVS blurb
  • Probability blurb
  • Electoral map blurb
  • Voteshare blurb
  • Similarity blurb
• Formatting
  • Manually set tooltip font in CSS
  • Fix footer (remove attribution for me / D)
  • Add final linebreak
  • Set font color for text (maybe headings?)
  • Adjust manual margin to match {.column-margin} formatting
• Functionality
  • update generate_links() function to perform internal site linking
  • Add link to model code and output to footer

Site

Tracking in this repository just to keep checklists in one place

• Write How this works blog
• National Page
  • Finish qmd body content
  • Add execute param for pulling from main branch
• State Pages
  • Write qmd body content using workable format
  • Add execute params for pulling from main and specifying state
• Site-level support
  • Add rendering script
• Publicize
  • test test test test on dev
  • launch on main
  • draft comms for
    • email
    • linkedin
    • twitter (thread)
    • tr slack wins
    • dsh discord

Bugs

• Fix issue of missing state-similarity plots on deployed pages with spaces (e.g., New Hampshire, Nebraska CD-1, etc.)

i.e., "if biden/trump win in x state, the conditional probability of them winning the electoral college increases from a -> b"

not sure if this should be for all states & plopped into the state pages, or just for competitive states and plopped onto the national page

Currently says "as of ... Joe biden is all but guaranteed to beat donald trump in District of Columbia." Should instead read as the District of Columbia

I've been mashing together GDP growth using a combination of CPI and GDP, but it may be better to just pull from the real GDP YOY growth table directly, which also goes back to 1947 and exactly matches BEA estimates.

I incorrectly labeled files "abramovitz" instead of "abramowitz"

Need to investigate if the difference in ordering is intentional, but if not, need to coalesce poll modes that are just rearrangements of each other. For example:

Live Phone/Online Panel/Text
Live Phone/Text/Online Panel

This is the only example, so even if they do need to be collapsed, only 4/294 polls use either of these patterns, and only 1 uses Live Phone/Text/Online Panel. Would warrant a minor version update only but not a full historical update.

Part of 2.0, but currently the model runs in prior mode from 5/1 -> 5/7. This is just because the fixed effect variables only have one factor for a bit (e.g., only LV polls). The solution is to change the fixed effect variables to not have variable lengths & just map the ids in directly.

render a daily quarto doc that shows:

• top-level probability change over the past week
• any states that have changed ratings
• runtime plot
• diagnostics plot
• count of polls plot
• state probability of winning vs. number of polls (will need to add the state abbr here)
• biden's forecasted approval
• the nifty ratings over time plot
• others (?)

• update model to accommodate 2 democratic candidates
• model still uses biden v trump polls to inform poll biases, but only gq for harris is used for democratic ticket forecast
• update pipeline to use a poll aggregation function based on candidate name
• (possibly) update model to estimate aggregates on the outcome scale
  • see, for example, the dc page. The voteshare credible intervals are a little wonky mostly due to how aggregation works.
  • This is a low priority issue
• add reference point to time series plots indicating date of switch (i.e., a copy of the 'annotate end date')
• rerun from 5/1 onwards (?)

possibly filter out polls by the created_at column. Using end_date in the model, but polls that end on a given day aren't necessarily available on that day (would be available by the created_at date, so swap out the filter for end_date with created_at)

Basically just need to write the draws somewhere.

On election night, when any competitive states are called, filter to the subset of draws that match reality. Only needs to be run interactively during election night, so should be a separate script. Likely will want to amp up the number of draws for that run specifically. Will also want to communicate uncertainty around the probability.

This can be a 1.2 level effort that gets done after #4