What a Target Model Actually Is

Most campaigns meet their target model as a spreadsheet. One row per registered voter, a few columns of numbers between 0 and 100, and a vendor on the phone explaining which column to call for persuasion and which to call for turnout. The file goes to the field director, the walk lists get cut, and nobody asks what the numbers are estimates of.

They are estimates of specific quantities, and the quantities are different from each other. A support score is an estimate of the probability that a named voter, if they cast a ballot, votes for your candidate. A turnout score is an estimate of the probability that they cast one at all. Those two numbers answer different questions, they are built from different evidence, and a campaign that averages them into one priority ranking will knock the wrong doors for eight weeks.

What the model is built from

Three inputs go in. The first is the voter file: the registration record, with vote history, party registration in states that record it, address, age, and whatever commercial and census data the vendor has appended to it. The second is responses from real people, collected for this race or a comparable one, in which somebody actually stated a preference and an intention to vote. The third is a statistical method that learns the relationship between the first two and then applies it to everybody on the file, including the overwhelming majority nobody ever contacted.

That third input is where the word model earns its place and where the honest uncertainty lives. A model trained on 1,200 completed interviews is producing estimates for 400,000 people. For almost all of those people the estimate is an extrapolation from somebody else who resembles them on a handful of recorded variables. That is a reasonable thing to do, and it is also the ceiling on how good the file can be, because no method repairs a response set that over-represents whoever was easiest to reach.

So the first question about any target model concerns what it learned from rather than how it was fitted: how many people, recruited how, when, in what geography, answering what questions. A vendor who will discuss the method in detail and the response collection only in general terms has told you where the weakness is.

Two universes that contain different people

A persuadable universe is the set of voters whose choice is genuinely open and who will be there to make it. A turnout universe is the set of voters whose choice is not in doubt and whose attendance is. Both are smaller than campaigns expect, they overlap very little, and a single score per voter collapses them. The collapse is invisible on the spreadsheet.

Four cells a single score collapses

An illustrative district of 100,000 registered voters, split by whether the voter is reliable and whether the choice is open. The two cells a campaign can move are marked.

Look at the two marked cells. Persuasion contact can move 9,000 voters, because those are the people who are both undecided and reliably present. Turnout contact can move 17,000, because those are the people who will vote your way if they vote at all. Different households, different scripts, different mail, and in a close race different budgets.

The ratio between those two cells is among the most consequential facts about a district, and it varies enormously. A low-turnout municipal race is almost entirely a turnout problem: the committed are the electorate, and the question is which of them shows up in June. A high-salience federal race with near-universal engagement sits closer to the persuasion end, because the people who will vote are already going to vote and the contest is over a thin band in the middle. School Board, County Commission, Mayor, Governor and President do not sit at the same point on that scale, and a model built for one of them answers the wrong question for another.

A deliverable should state the size of both universes in plain counts, by precinct, with the assumptions that produced them. If it states one number per voter and leaves you to work out what it means, the vendor has kept the easy part and handed you the hard one.

Why precinct level matters

District-wide support estimates are averages, and averages hide exactly the variation a field program exists to exploit.

What a district average hides

Modeled support in twelve precincts of one illustrative district, sorted low to high, against the district-wide average of 48 percent.

Forty four points of range sit underneath that one average. A campaign working from the average treats the 27 percent precinct and the 71 percent precinct the same way, which wastes persuasion money in the first and turnout money in the second.

Precinct-level estimates do three things a district number cannot. They route a canvass, because a walk list is a geographic object. They can be checked, because precinct returns are published and a model's precinct aggregates can be held against them afterward. And they show you where the model is confidently wrong, which is where the money leaks and the only place to look if you want the next model to be better.

Ask for estimates at the lowest geography the vendor will stand behind, then ask separately what the error looks like at that level. Precinct estimates built out of 1,200 interviews across a whole state carry real uncertainty, and a vendor who reports them to the decimal without a word about it has told you something about how the shop works.

The assumptions that should be written down

Every target model rests on assumptions that are choices rather than findings. Five of them belong in the document, in sentences a campaign manager can read, and the first one moves the answer more than the other four combined.

• The electorate the model assumes: how many people vote, and with what age, geographic and partisan composition.
• Where the training responses came from: which race, which dates, which collection method, how many people.
• What happens to voters with no vote history, who are a large group in any state with recent registration growth.
• How people who did not answer were handled, and which variables the weighting used.
• Which variables the model was allowed to use, and which were deliberately excluded.

The same model under three turnout assumptions

One voter file, one set of responses, one method. Only the assumed shape of the electorate changes.

None of those three assumptions is wrong. They are answers to three different questions about who shows up. A deliverable that reports one projected margin without naming which electorate it assumed has not given you an estimate. It has given you somebody else's judgment about turnout, wearing the clothes of a measurement.

Questions to ask a vendor before you accept the file

Seven questions, and the one about the data file matters more than it sounds. A model you cannot interrogate is a model you have to trust, and trust is an expensive substitute for a holdout.

• How many people did the model learn from, and how were they recruited?
• What is the support score the probability of? One sentence.
• How does the model score on respondents it never saw during training?
• In the last comparable race, how far off were your precinct-level aggregates, and in which direction?
• What electorate size and composition does this file assume, and where is that written down?
• Which voters does the model decline to score, and why?
• Do I get the response-level data, and in what format?

How to judge the file you were handed

Three checks are worth making, in order of how much they tell you.

The first is calibration on held-out respondents. Any model looks accurate when it is scored on the people it trained on. Ask for performance on a random slice held back from training, and ask for it as calibration rather than accuracy: among the voters this model scored at 70 percent support, did roughly 70 percent support you? A model that ranks voters in the right order but is systematically overconfident will spend a persuasion budget on people who were never movable.

The second is the precinct backtest. Aggregate the model's scores to precinct level in a past race and hold them against the published returns. The errors will not be zero and should not be. What you are looking for is whether they are random or whether they lean one direction in one kind of place, because a lean is a bias the next race inherits.

The third is stability. Re-score the file with a different random seed and a different training slice, then count how many voters cross the thresholds you actually use. If the top persuasion decile reshuffles by a third, the deciles are not real and the field plan should not be built on them.

All three are routine requests. A firm that treats them as adversarial has answered the question you were really asking.