Making the Connection: What 947 Recorded Speed Dates Reveal About Why People Click

By Noam Shemla · The Studies

TL;DR Three researchers recorded 947 four-minute speed dates and asked, after each one, how well the two people had clicked. Who people were (height, weight, dating experience) predicted clicking most. But the way they talked added about half as much again, inside four minutes, and it followed a clear pattern: clicking went with animated voices, with the woman as the subject of the conversation and the man visibly following and supporting her, and with few hedges and few filler questions.

• Paper: McFarland, D. A., Jurafsky, D., & Rawlings, C. (2013). Making the Connection: Social Bonding in Courtship Situations. American Journal of Sociology, 118(6), 1596–1649.
• Setting: Speed-dating evenings run for graduate students at a private US university in 2005. Dates lasted four minutes; each person had about 20 in an evening.
• Sample: 110 daters (56 men, 54 women) and 947 dates with usable audio and transcripts, giving 1,883 “conversation sides” (one person in one date).
• What was recorded: Both people wore an audio recorder on a shoulder sash. Every date was transcribed and time-stamped turn by turn.
• Outcome: Right after each date: “How well did you click with this person?” on a 1–10 scale, and yes or no to seeing them again.
• Method: An actor–partner interdependence model: a multilevel regression that separates the effect of your own speech from the effect of your partner’s.
• Headline: Speech added about 7.5% of explained variance on top of traits’ 15%, and clicking, not the speech itself, carried speech’s effect on who wanted a second date.

Most research on speed dating asks a selection question: who gets chosen? It finds what you would expect. Men pick women who are slim, women pick men who are tall, and everybody leans towards people like themselves.1 What that research mostly leaves out is the four minutes in between, the conversation itself.

This paper, by Daniel McFarland and Dan Jurafsky of Stanford and Craig Rawlings of the University of California, Santa Barbara, is about those four minutes. Its question is simple to state and hard to answer: when two strangers say afterwards that they "clicked", was there something about how they talked that made the difference, and if so, what?

The authors start from a theory in sociology: a bond forms when two people share a focus of attention and a mood, and each builds up the other's emotional energy, the way a good conversation seems to feed itself.2 If that is right, the conversation should leave traces that a machine can count.

Why speed dating

A speed-dating evening is close to a laboratory built by accident. Strangers are paired at random, so who meets whom is not chosen. Every date is the same length. And straight after each one, both people fill in a card. That combination is what lets the authors ask a question about conversation rather than about the people having it.

The data come from events the team ran for graduate students in 2005. Everyone wore a recorder, so each date exists as two audio files and a transcript marked with who spoke when, who laughed, and where speech overlapped. After losses to equipment and operator errors, 947 dates had both audio and a transcript: on average 812 words and 93 turns each.

What they measured in the talk

The authors built two kinds of measures, one for how people sounded and one for what they did with words.

For sound, they extracted 18 acoustic features from each person's side of each date: the floor, peak and average of their pitch, how much it varied, their loudness and its variation, and how long their turns were. Many of those are near-duplicates, so a factor analysis reduced them to six independent dimensions that together capture 85% of their variation: maximum pitch, loudness, minimum pitch, variable loudness, turn length and variable pitch. Two more measures of excitement sit beside them: speaking rate and laughter.

For words, they counted the moves a conversation is made of.

The conversational moves the study counted

Eleven of the speech measures, described in our words. Each was counted separately for each person in each date.

The authors also recorded the obvious rivals: each person's height and body-mass index, whether they were born abroad, how often they dated, whether they were looking for something serious, how far into the evening the date fell, whether the two already knew each other, and how different their hobbies were. That matters. A speech feature that only mattered because tall men also talk a certain way would vanish once height is in the model.

The model: your speech and theirs

A date has two people in it, and each one's sense of clicking can respond to their own behaviour and to their partner's. The paper uses the standard tool for that situation, the actor–partner interdependence model.3 In each date it links each person's speech (X) to both their own rating (Y) and their partner's.

Written out, the first level of the model predicts one person's click score like this:

A second level gives every date its own baseline, so that a date where both people happened to be in a good mood does not masquerade as an effect of speech. Every predictor is standardised, so each coefficient reads the same way: a one-standard-deviation increase in the feature is associated with this many points of change on the 1–10 click scale.

What they found

Who you are matters more, and how you talk still matters

Start with the uncomfortable part. Traits were the bigger story. In supplementary models, people's traits explained about 15% of the variance in clicking; features of their speech explained a further 7.5%. Men reported more connection when they were tall and when their partner had a lower body-mass index. Both sexes clicked more when they were looking to date, experienced at it, and native-born, and with people whose hobbies matched theirs.

But 7.5% from four minutes of conversation, measured on top of everything the people brought with them, is not small. And the authors found it grew with time: in models using only the first minute of each date, the speech effects were much weaker. The longer people talked, the more the talk mattered.

One more number points the same way. The longer someone took to decide whether they wanted to see the other person again, the more they reported clicking: a one-standard-deviation increase in decision time went with 0.44 more points on the click scale for men and 0.60 for women. Connection, in other words, was something that happened in the conversation rather than a verdict delivered at hello.

The pattern in the talk

Here is the full model, with every trait control in it, for all nineteen speech features. The two columns for each sex are the two arrows in the diagram above: how your own speech related to your rating, and how your partner's speech related to it.

How each speech feature related to the sense of having clicked (full model)

The pattern is easier to see drawn out. The authors plotted every effect that reached significance by its size (up and down) and its strength (left to right). We have redrawn their figure below.

Read the two panels together and a single shape comes out of them, which the authors call a reciprocal asymmetrical performance. The two people are not mirroring each other. They are playing complementary parts.

Both became more animated when they clicked, in different registers. Men laughed more and varied their volume, and their pitch actually became less variable, closer to a conventionally masculine delivery. Women raised and varied their pitch, varied their volume, spoke more softly and took shorter turns. Variable loudness is the one feature that predicted clicking strongly for both.

The woman was the subject. Women clicked when they talked about themselves ("I") and used "you know" and "I mean", the markers of someone absorbed in their own story. Men clicked with women who did the same: her "I" is one of the three partner effects on the men's side. A look back through the transcripts found that this kind of talk came when people were telling stories with some passion, not answering questions about their hobbies.

The man followed. Women clicked with men who said "you", who reacted with appreciation and sympathy, who laughed right after they did, and, most surprisingly, who interrupted them.

The interruptions that helped Interruptions are usually read as a bid for control, so the authors went back to the transcripts. From the dates women rated highest, they took the first 100 of the 327 times the man interrupted, and read each one. Between 80 and 90 were supportive: finishing her sentence, offering a matching experience, extending her idea, or saying he understood. That is the same collaborative overlap researchers have described among close friends.

Questions and hedges went the other way. Every question a woman asked, and to a lesser degree every question a man asked her, went with less connection. The authors found that questions in the flat dates were doing a specific job: keeping a stalling conversation alive ("Where are you from?", "What do you study?"). The dates that clicked had long runs of shared stories and few questions. A later study of the same recordings found that the kind of question is what matters: follow-up questions, the ones that build on what the other person just said, went with more second-date yeses. Hedges ("I guess", "sort of") predicted less clicking for women's own speech and, for men, in the woman's. This matches the team's related finding that hedging is a mark of an awkward date.4

From clicking to a second date

The card also asked a blunter question: do you want to see this person again? When the authors added the click scores to that model, the actor's own sense of clicking dominated it. A one-standard-deviation increase in how much you felt you had clicked went with 5.74 times the odds of saying yes for men and 4.06 times for women, about five times the size of any other feature. And once clicking was in the model, the speech features dropped out of significance: they worked through the feeling of connection rather than beside it.

• 947 Dates analysed. 1,883 conversation sides, 110 daters
• +7.5% Variance from speech. on top of 15% from traits
• 5.74× Odds of a yes, men. per SD of clicking
• 4.06× Odds of a yes, women. per SD of clicking

The study in four numbers.

What this study cannot tell you

• It is correlational. Nobody was told to laugh more or hedge less. The paper shows what clicking dates sounded like, not that changing your speech would make a date click. Some of the arrows may point backwards: a date that is going well makes people animated.
• It is one kind of room. Graduate students at one elite university, in 2005, in heterosexual four-minute dates. The gender pattern it describes is the norm of that setting at that time. The authors say plainly that such asymmetric roles may reflect social inequalities and are not a recipe.
• The effects are modest. A coefficient of 0.2 means a fifth of a point on a ten-point scale per standard deviation of the feature. Traits still did twice as much work.
• Many measures, one sample. The authors tried a wide range of word lists and features and report that most had no relationship with clicking. Testing many features on one data set raises the chance that a few of the weaker effects are noise, which is why the strongest results (variable loudness, laughter, maximum pitch, self-reference) deserve more trust than the marginal ones.
• The outcome is a feeling. "Clicked" is a self-report straight after the date. The authors note that only about half of the matched pairs followed up by email.

What a speaker can take from it

The study is about dates, and it would be easy to over-read. What transfers is not the gender script, which belongs to its setting, but three observations about conversation that recur elsewhere in the research.

Animation reads as engagement. Variable loudness was the most consistent feature in the study: when people clicked, their volume moved. It is close kin to what Exprea reports as vocal energy, and it is one of the easiest things to hear in your own recording.

Hedging costs connection, not just authority. The paper found hedges on the wrong side of clicking for both sexes, which fits a wider body of work on powerless speech. It is one reason hedging rate is part of how we score a take.

A good question is a follow-through, not a filler. The questions that hurt were the ones that kept a flat exchange limping along. The moves that helped were the ones that showed the other person they had been heard: appreciation, sympathy, a matching story, finishing their thought. That is a skill, and like any skill it gets better with practice and feedback.

This article summarises a published paper. Every number above is the authors’ own, taken from the paper’s text, Tables 5, 7 and 8, and Figures 2 and 3.

Measured

• Acoustic features extracted with Praat from each person’s recording, reduced to six factors
• Word and turn features counted from time-stamped transcripts
• Click scores (1–10) and second-date decisions from scorecards filled in after every date

Inferred, not measured

• The link from speech to clicking is statistical association, controlled for traits, not an experiment
• The |t| positions in the effects figure were read off the published chart and are approximate

References

1. McFarland, D. A., Jurafsky, D., & Rawlings, C. (2013). Making the connection: social bonding in courtship situations. American Journal of Sociology, 118(6), 1596–1649. https://doi.org/10.1086/670240
2. Collins, R. (2004). Interaction Ritual Chains. Princeton University Press. https://openlibrary.org/works/OL28947W
3. Kenny, D. A., Kashy, D. A., & Cook, W. L. (2006). Dyadic Data Analysis. The Guilford Press. https://openlibrary.org/works/OL18353687W
4. Ranganath, R., Jurafsky, D., & McFarland, D. A. (2013). Detecting friendly, flirtatious, awkward, and assertive speech in speed-dates. Computer Speech & Language, 27(1), 89–115. https://doi.org/10.1016/j.csl.2012.01.005