About · Contact · Home
  People Matching Project

Help needed!
If you are interested in helping out with the goals of the People Matching Project, we need volunteers to take our survey.

[Updated 17 April 2017]


The People Matching Project has conducted what is probably the most in depth investigation into what individual differences (personality, etc.) can be used to predict if two individuals will get along. This has been done with the goal of developing free and open source tools for matching together compatible individuals for a variety of purposes (e.g. friends, dating, room-mates). This page summarizes the research program. If you want to participate in the research program (please do!), see the banner above.

Theoretical foundation

The main approach of the research program has been looking at real world friendships and surveying real world pairs of friends and couples about their personalities and seeing on what traits the pairs are similar. If it is found that for any given trait X, people high on the trait are more likely to be paired with people high on the trait and people low on that trait are more likely to be paired with people low on that trait then, the assumption that we are working with is that, the similarity was probably a partial cause of that pair's formation. In some cases this assumption may be a bit spurious, for example students at Harvard are much more likely to be friends with other students at Harvard than they are to be friends with students at Yale — even though they probably have just as much potential to be friends if stuffed together in the same room, but for general personality this assumption looks good. Of course, the gold standard would be to randomly assign people to pairs and see how they ended up getting along — no one has the resources to do this.

The research program has sought to find questions that correlate between pairs of people, that can be then combined to produce scales that correlate between pairs of people.

Screening of personality test items

It is important that a large and diverse set of possible test questions be tested to make sure that no domain of predictive power is left untapped. In total, nearly 2,000 questions have been evaluated for their power to predict friendship.

The data used as the final evaluation metric comes from pairs of individuals where both have answered the same sets of questions. Recruitment of volunteers for these surveys is slow and a major roadblock, so questions first went through a pre-screening survey where the survey respondent answered a personality survey about themselves and then rated a friend of theirs from the observer perspective. This procedure may draw immediate objections -assertions that this is not valid and that correlating self with observer rating is subject to too much bias to be useful -, but analysis of this issue indicates that these self-rating / observered rating correlations do somewhat predict self-rating / self-rating correlations. See a further breakdown of this issue here. Nearly 2,000 questions were evaluated this way using 35,000 respondents.

Questions that passed the initial screening then were put in paired surveys distributed to recruited pairs of individuals with a real life life connection (recruitment advertized for best friends). In total 601 questions were evaluated this way using four surveys, sample sizes on these surveys ranged from between 350 to 700 pairs. The summary data for all these questions can be downloaded at items-screening2-table.csv.

The ten best items found so far in our research are in the table below along with the correlation between friend pairs (the metric of quality we are using). See the linked file for the full list.

ITEM TEXTCorrelation between friends
I am a feminist.0.449
I have smoked a lot of marijuana in my life.0.443
I support the United States.0.43
I would be open to dating a transsexual person.0.416
I like tattoos.0.387
I use facebook every day.0.386
I believe in evolution.0.382
I make jokes about autism.0.379
I believe in a universal power or God.0.37
I would not want my children to be gay/lesbian.0.365

As can be seen, the best questions all seem to be more about values than things usually considered part of personality (e.g. the big five personality traits: extroversion, neuroticism, agreeableness, conscientiousness, and openness), although some lifestyle questions also do well. In the table below some questions taken from the Big Five Aspect Scales are excerpted from the same data so a comparison can be made.

ITEM TEXTAssociated traitCorrelation between friends
I get angry easily.Neuroticism0.042
I am filled with doubts about things.Neuroticism0.134
I feel others' emotions.Agreeableness0.091
I respect authority.Agreeableness0.197
I carry out my plans.Conscientiousness0.054
I like order.Conscientiousness0.102
I make friends easily.Extroversion0.041
I take charge.Extroversion 0.038
I am quick to understand things.Openness-0.003
I enjoy the beauty of nature.Openness0.161

For more detailed statistical results from each survey see:

  • Data from the First People Matching Project Friend Pairs Survey (n=2,005; 160 items)
  • Data from the Second People Matching Project Friend Pairs Survey (n=1,229; 160 items)
  • Data from the Third People Matching Project Friend Pairs Survey (n=1,412; 160 items)
  • Data from the Fourth People Matching Project Friend Pairs Survey (n=1,164; 160 items)
  • Data from the Fifth People Matching Project Friend Pairs Survey (n=3,727; 215 items)
  • Other item formats evaluated

    Two other item content formats were tested and included in the paired surveys. In one format, respondents rated paintings (15 items) and the second they rated the attractiveness of individuals (15 items). These two item formats were not found to have enough predictive power to currently warrant pursuing.

    Also, it was investigated if facial recognition could be used to predict compatability based on the theory that people may subconciously get along with people who share facial traits (one dating app, FaceMate does just this). To do this photos of about 250 couples were cropped from announcement wedding websites and their similairty was evaluted. Couples had slightly more similar facial features than would be expected by chance, but only trivially - this was judged to be a dead end.

    Currently, the ratings of song clips are being investigated for potential - based on the observation that personality test items about generes and artists do decently well.