The Needle Returns for New Hampshire. Here’s How It Works.

Once polls close in the New Hampshire primary tonight, The Times will start publishing a live estimate of the final result, better known as the Needle.

How to watch the Needle tonight

Our results pages feature graphics designed to help you understand how each candidate is faring.

The hypothetical chart below shows how our live estimates work. Our best estimate for each candidate’s final vote share is shown along with a range of estimates for where things might end up.

And our results pages also include county-by-county maps showing where votes remain to be counted and which candidate our model thinks stands to benefit from that outstanding vote.

Last week, we published estimates for the Iowa caucus, which featured a look at an area of unusual interest: who was likeliest to take second place. This time around, now that Ron DeSantis has dropped out of the race, we are producing individual estimates only for Donald Trump and Nikki Haley.

How the Needle works

The fundamentals behind the Needle are surprisingly simple. In a way, it considers two big questions: Where is the vote that remains to be counted, and which candidate is faring better than expected?

As results start coming in, the Needle compares what is reported with pre-election expectations, county by county and precinct by precinct. It then estimates who will win the remaining vote based on the patterns it has seen in the results so far.

All of this, of course, is easier said than done.

The pre-election expectations are created by combining data from Times/Siena College polling, other public polls, voter registration files, the census and historical election results.

Once it receives results, the Needle uses a statistical model to spot demographic patterns that help it understand how the vote varies in different types of counties or precincts. It then blends its pre-election expectations, the output of the statistical model and the actual tabulated results into a single estimate.

Like any statistical model, the Needle improves as it sees more and better data. It fares best when an election has predictable partisan and demographic divides, which allow it to quickly judge whether a candidate is on track for victory. It benefits greatly when election officials break out returns by different methods of voting, like in-person or mail. And it benefits when it gets granular results by precinct, not just the typical results by county.

In New Hampshire, the Needle has some, but not every one, of those advantages.

Most votes will be cast in-person, which avoids the challenges posed by the use of multiple methods of voting. The Needle also benefits from a quirk of election administration in New England: Results will be reported by township, rather than by county, as is customary in the rest of the nation. Townships, which are smaller geographical areas than counties, provide the model with more detailed information to base its estimates on.

The Needle also gains an advantage from last week’s Iowa caucus. Granular, precinct-level data from that race collected by The Times allows our model to better understand how Mr. Trump and Ms. Haley’s support breaks down along demographic lines.

But the Needle also faces a big disadvantage: The race has changed since then, now that Mr. DeSantis has dropped out. And a presidential primary or caucus is not nearly as predictable as a partisan contest, when, for example, the Needle can confidently assume that New York City will be great for a Democrat.

Why you should watch the Needle

We think the Needle is the premier tool available for making sense of incomplete election results. It can offer a much clearer picture of the outcome of an election than just looking at the early returns, which are often biased toward a particular candidate based on which counties or precincts happen to report first.

In 2016, the Needle foresaw Mr. Trump’s surprising victory hours before the television networks. In 2020, it anticipated that Joe Biden would win Georgia even though it was days before he took the lead in the tabulated vote. In 2022, it made clear that the race for control of the House was close and that there was no red wave.

Tonight may not be so dramatic, given that Mr. Trump’s support exceeds 50 percent in many polls. But the Needle will be learning, and telling us about important aspects of the race in real time.

When to trust the Needle

The Needle does more than just create an estimate for a candidate’s final share of the vote. It also looks at a range of possibilities for how the race could go. As more votes come in, the Needle will become more certain of the final outcome.

One way the Needle conveys uncertainty is through confidence intervals, which quantify a range of reasonably plausible final outcomes. The Needle simulates many random scenarios that diverge from our best estimate at each point in the night. The extent of the randomness is informed by the model’s accuracy at comparable stages of vote counting in historical elections. The displayed confidence intervals — shown by the colored bands on the charts — contain 95 percent of those simulated results; outcomes beyond these bands aren’t impossible, but the Needle considers them unlikely.

Probabilities are worth taking seriously: A candidate with a 75 percent chance of winning, in the view of the Needle, really might win only 75 percent of the time. One in four times, that candidate will lose. Over a long primary season, there will be many times when a candidate with that chance of winning goes on to lose. That’s what we want to see: 75 percent doesn’t mean 100 percent.

In fact, the Needle by itself is never enough to say that there’s a 100 percent chance a candidate will win. The Needle does not make race calls, as there are countless ways for our model to be led astray. In rare instances, the vote totals reported by election officials and tabulated by The Associated Press contain errors, which the Needle may assume to be factual before they are corrected. And in the past, we’ve sometimes had to pause the Needle to deal with unanticipated bugs or quirks in data feeds.

In the end, it still takes a human being to figure out when a race is really over.