And after we start dating, we may grow to be even more alike

There are at least three reasons we so often message and eventually mate with the similar. Before we even meet, myriad forces guide us away from people who are different from us – work, schooling, eHarmony’s algorithm. When we are exposed to matches, we tend to pursue people who are similar. In the face of these forces, it’s perhaps small wonder that the dimensions along which opposites attract hide in the statistical shadows.

But even believers in algorithmic approaches to love acknowledge these shadows exist. The scientists I spoke to at eHarmony and OkCupid agreed. As rich as their data sets are, the uncertainty of that first meeting remains.

Correction (April 10 6:35 p.m.): An earlier version of this article misidentified eHarmony’s website for same-sex dating; it is Compatible Partners, not Compatible Couples.

Footnotes

Because it’s extremely important to be rigorous when studying online dating, I confirmed my conclusions a few different ways. Let the man’s value of a trait be t_m and the woman’s value be t_f; let whether the man messages the woman be the binary variable y_m and whether the woman messages the man be the binary variable y_f. For each trait, I used logistic regression to regress y_m and y_f on t_f, t_m and their product, t_f*t_m. The crucial term is the product term: it’s known as an interaction term, and if it’s positive it indicates that people with similar values of t_f and t_m are more likely to message each other; if it’s negative, it indicates that opposites attract. I looked at the signs of all the product terms, as well as how statistically significant they were, and could not find any interesting cases where opposites attracted after using the Bonferroni correction for the number of traits examined.

I experimented with a few different models to ensure my basic conclusions stayed the same. I tried looking at each trait individually but controlling for obvious factors by which people choose to message mates – attractiveness, age and whether the person messaged them. I tried making the continuous variables binary (by whether they were above average). Finally, because many of these variables are correlated, I ran a giant regression including the value of every trait (along with interactions) simultaneously. None of these mathematical modifications persuaded opposites to get together, and the last one (containing 211 variables and 1 million couples) crashed my computer. I reran that regression using 200,000 couples.

Attractiveness was one trait in eHarmony’s data set, but when I asked how it was calculated, I did not get a response. The rest of the traits are self-reported by users.

This is not because men are just more willing to message everyone – I controlled for that by looking at the difference in rates at which men messaged women who were similar and women who were different.

Dan Ariely, an economist who studies online dating, compares people to wine – you may like them for reasons you can’t quantify

Race shows many interesting patterns, but they’ve been discussed in detail here and, less depressingly, here, so I do not focus on them in my analysis.

Dan Ariely, an economist who studies online dating, refers to traits where everyone prefers the same thing as examples of “vertical preferences,” as opposed to “horizontal preferences,” when people prefer those who are similar. He also finds that horizontal preferences are more important in producing the “birds of a feather” effect. For his complex but lovely discussion of the subject, see here.

These “trios” are often used in genetics to study, among other things, how genes and diseases are passed from parents to children.

Dan Ariely, an economist who studies online dating, compares people to wine – you may like them for reasons you can’t quantify

Race shows many interesting patterns, but they’ve been discussed in detail here and, less depressingly, here, so I do not focus on them beetalk in my analysis.

Dan Ariely, an economist who studies online dating, refers to traits where everyone prefers the same thing as examples of “vertical preferences,” as opposed to “horizontal preferences,” when people prefer those who are similar. He also finds that horizontal preferences are more important in producing the “birds of a feather” effect. For his complex but lovely discussion of the subject, see here.

Here, too, my 23andMe colleague Aaron Kleinman and I found that birds of a feather flock together: For 97 percent of the traits we examined, couples were positively correlated. Former smokers tended to pair with former smokers, the apologetic with the apologetic, the punctual with the punctual. It is worth noting that causality may go in both directions: Perhaps you’re attracted to your partner because he, like you, was on time for your first date; it’s also possible that he was initially incorrigibly late, but after you fell in love you trained him. (We also found some examples where opposites attracted: Morning people tended to pair with night owls, and people with a good sense of direction with those who lacked one.)