MethodologyMeasuresConcept

Whip counts

How a bill's next floor vote is projected member by member, how to read the projection, and how its accuracy is checked.

A whip count is our projection of how each member of a chamber will vote on a bill that hasn't had its floor vote yet. It's a forecast, not a record. It's worked out from how those same members have voted before, and it says how often forecasts like it have turned out right.

What it predicts#

The bill's next floor vote. That's the vote in the chamber where the bill started, then the vote in the other chamber. Once both chambers have voted there's nothing left to project, and the bill shows the recorded votes instead.
The vote to pass the bill, assuming it gets that far. Most bills never reach a floor vote. A whip count doesn't say whether this one will, only how the vote would go if it happens.
Each member's vote, then a tally of the yes and no votes it adds up to.

How it's worked out#

Learn how members vote. For each chamber and session, we look at the floor votes that actually divided the chamber. Unanimous votes are left out, because they don't show who disagrees with whom. From the divided votes, each member gets a position on a single scale: members who tend to vote together sit close together. Party isn't an input. It shows up anyway, because members of a party tend to vote alike.
Place the bill. A bill that hasn't been voted on has no votes to learn from, so its likely position is estimated from what's known about it. The main clue is where its sponsors and cosponsors sit on the same scale. Its political lean, its topics, and whether it's a resolution add to that.
Project each member. A member on the bill's side of the line is projected to vote yes, and one on the other side no. The farther a member sits from the line, the stronger the call.
Attach a track record. Before a model is used, it's tested on past votes it wasn't trained on. Votes from its own session count most, and earlier sessions count for less the further back they are, so an old bad stretch fades instead of holding a model back. An earlier session where the method failed outright is left out. That test shows how often calls of each strength turned out right. A member's percentage is that record. 85% means calls like this one were right about 85 times in 100. It isn't the model grading itself.
Check whether the vote looks close. A separate check estimates whether the vote is likely to be close at all, or lopsided either way.

Each chamber's model is refitted weekly while its legislature is in session, so new votes are taken into account. At the start of a session, before the chamber has had enough divided votes for a model of its own, the previous session's model is used, and members seated since then aren't projected. The whip count says so. A refit has to pass the same tests as the original model (see below). If it doesn't, the previous model stays in use.

Reading a whip count#

The projection appears on a bill under Floor vote by party, marked as projected. See Reading a bill.

Too close to call. A member whose calls have been right less than 75% of the time is marked Uncertain, and isn't counted as a yes or a no. How many members are too close to call is the first number shown, because it says how much of the tally is open.
Coin flip. A member marked coin flip is in a range where past calls were right only about half the time. There's no percentage, because one would claim more than the record supports.
Against usual. A member marked against usual is projected to vote differently from how they usually vote. These are the calls where the projection says something about this bill in particular. When there are none, the projection says so: it reflects how these members generally vote, not anything about this bill.

Each projection also carries one overall label:

Label	What it means
Working count	Most members can be called, and the note says how often calls like the uncertain ones have been right.
Coin flip	The members who decide the vote are in a range where calls have been right about half the time. Read the tally as a range, not a number.
Not expected to be close	The vote looks lopsided. The projection is near-certain, but it says little.
No member-level read	Half the chamber or more is too close to call, so the tally can't indicate an outcome. Yes and no counts of zero mean nobody could be called, not that nobody would vote yes.

How to use it#

A whip count is most valuable as a guide to where attention should go before a vote, not as a verdict on how the vote will end. Its value is in the members it can't call and the members it expects to break from their usual vote. Those are the people whose votes are actually in play, and the ones worth a call, a meeting or a closer look. Use it to decide which members to talk to, to tell a bill with a real fight ahead from one that will pass or fail on party lines, and to check whether support is holding as a vote approaches. If most members are confidently called and none are against their usual vote, the outcome probably isn't in doubt, and the effort may be better spent elsewhere.

Don't use it to report how a legislator will vote, or to say they've taken a position. It's an estimate from their past votes, not anything they've said. And don't read the tally as a promise of the result, especially when it's marked Coin flip or No member-level read.

When there's no whip count#

Both chambers have voted. The bill shows the recorded votes.
It's a placeholder bill. Its text only states an intent to legislate later, so there's nothing yet to place. The whip count appears once its text is filled in. See Reading a bill.
There's no model for the chamber. Either the chamber hasn't had enough divided votes to learn from, or its model failed its tests. Tested on past votes it never saw, a model is used only if both of these are true:
- it does better than guessing that every vote goes the way most of that chamber's votes go;
- on the members it can't confidently call, it isn't clearly worse than giving each member the whip count's own call for their party. That's the first comparison in How its accuracy is checked.
A model that fails either test isn't offered.
The chamber's whip counts are withdrawn. Once the accuracy scoreboard is published, a chamber whose whip counts do clearly worse than that party comparison on its published results is withdrawn. Its bills show Whip count withdrawn, with the date and a link to the scoreboard, in place of a projection. Projections are still made and scored, and they come back after doing better at two publications.

Where it works best#

A whip count is most useful in chambers where votes are genuinely in doubt. Where one party holds a large majority and most bills pass with its votes, few members break from their usual vote, so the projection mostly repeats how members usually vote. In closely divided chambers, more members cross over and the projection has more to say. Even there, it can't see what happens off the record: a deal made in the hallway, a change of heart, an amendment adopted on the floor, or a member who doesn't show up.

How its accuracy is checked#

Recorded before anyone sees it. Every whip count is written to a log at the moment it's made, before it's shown to anyone. Entries can't be edited or deleted afterward, and each one is sealed to the one before it, so a change would be detectable. If a projection can't be recorded, it isn't shown.
Scored against the official roll call. Once the vote happens, each projected member's vote is compared with the recorded one. A member who didn't vote isn't counted either way. A bill that never gets its vote is set aside, not counted as right or wrong.
Compared with simple alternatives. Most members vote with their party most of the time, so raw accuracy flatters any model. Each whip count is measured against two plain guesses. The first takes the whip count's own call for each party and gives it to every member of that party. A whip count can only beat that by correctly calling the members who break from their party, which is what it's for. The second guess is that every member votes the way they usually do.
Not just the bills people looked at. We also record projections for a sample of bills nobody asked about, so the record isn't drawn only from the bills that drew attention.
Honest ranges. Members' votes on the same bill move together, so a hundred votes on one bill don't count as a hundred separate tests. The uncertainty ranges treat each bill as one unit.