You may have noticed an “election composite” mentioned in the Data Selector or below the ratings diagram in the Analytics view . For example:
This post explains what an election composite is and why we use one, identifies some other notable composites, and describes the specific composite that we use.
To evaluate the likely future election outcomes for a set of districts, you need actual past election results.
One’s first instinct for congressional districts is to simply use past US House elections. Those “endogenous” election data are distorted in several ways though, including the impacts of specific candidates, incumbency effects, and district-by-district spending levels, etc. To remove such distortions, the accepted practice is to use statewide election results — e.g., President, US Senate, and races for state office — which are “exogenous.”
With that in mind, one’s next instinct is to simply pick one election, like the most recent Presidential election. There are three problems with this approach though:
- What you want is statewide voting patterns, i.e., what is the typical two-party vote split precinct by precinct? Any single election — even a statewide one — is subject to some variation from the underlying pattern.
- One prominent example of that is when a third-party candidate gets a significant percentages of the vote — that distorts the two-party vote shares.
- Presidential elections can vary significantly from typical statewide voting patterns (especially in highly contentious elections, like 2016).
The solution to these problems is to combine the results from several elections into what we call a “composite.” A composite of elections is a voter preference index (VPI), a measure of the two-party partisan lean of each precinct .
You may already be familiar with one: Cook’s Partisan Voting Index (PVI). It combines the last two Presidential elections. Another index, the Hofeller Formula, was made famous in a North Carolina gerrymandering case. There Republicans used results from seven recent elections to gauge partisan lean:
- Two Gubernatorial elections for 2008 and 2012
- Three Senate elections for 2008, 2010, and 2014
- The 2008 Commissioner of Insurance election, and
- The 2012 Commissioner of Labor election
Our VPI combines the results of three types statewide general elections:
- President — If only one is available, we use that. If two are available, we average them (like PVI).
- US Senate — If only one is available, we use that. If two are available, we average them.
- State office— The most recent elections for Governor and Attorney General. If only one or the other is available, we use that. If both are available, we average them.
We average these three subtotals to create precinct totals. Note: We exclude uncontested elections and elections that a 10% or more 3rd-party / independent votes. We include lopsided elections that are contested, because they help flesh out the seats-votes curve. The partisan metrics in Analytics and Advanced views use two-party vote shares, so you want to use an election (or composite) where there the “Other” vote is relatively small.
We took this approach for several reasons:
- With a few exceptions, all states have all of these elections so we can both get the election data and the index will work consistently across states.
- Presidential elections only have a one-third weight in the composite, so they are included but don’t dominate the results.
- The maximum time between elections is short enough that discounting older elections is not worth the added complexity. We weight each type of election equally regardless of when it occurred.
- We can create this composite with now using the 2012–2018 election data that we have, and as we get more 2018 election results over the course of the coming year we can seamlessly update the composites state by state. Similarly, we can slide 2020 election results in as we get them.
To sum up, we use an election composite in District Statistics, because it is a better gauge of typical statewide voting patterns than any single election.