Science tells us that fast food lovers are more likely to marry other fast food lovers

Emma Pierson writes:

I’m a statistician working at the genetics company 23andMe before pursuing a master’s in statistics at Oxford on a Rhodes scholarship. I’ve really enjoyed reading your blog, and we’ve been doing some social science research at 23andMe which I thought might be of interest. We have about half a million customers answering thousands of survey questions on everything from homosexuality to extroversion to infidelity, which as you can imagine produces an interesting dataset.

1. We found that customers who answer our survey questions in the middle of the night are significantly less happy and significantly more likely to be manic. See here and here.

abc

2. Using genetic data, we identified 15,000 couples along with the child they had had together. We showed that couples tended to be similar — 97% of traits showed a positive correlation between woman and man, even when we controlled for race and age — although there were often intriguing exceptions.

At this point Pierson shows an info graphic that says that “punctual people,” “skiers,” “hikers,” “non-smokers,” “fast food lovers,” and “apology prone people” are more likely to marry each other, while, in contrast, “early birds” are more likely to marry “night owls,” and “human GPSs” are more likely to marry “constant wrong turners.”

She continues:

We also showed that couples who were dissimilar in terms of BMI or age tended to be less happy (even when we controlled for individual BMI + age). See here and here.

What I’d really like to see is the full list. I’m not so interested in learning that skiers and hikers are likely to marry each other, but if I could see an (organized) list of all the traits they look at, this could be interesting.

14 thoughts on “Science tells us that fast food lovers are more likely to marry other fast food lovers

  1. This is also a good survey to make into a 2-way street. That is ask the question why people with different age and BMI who are relatively unhappy together stay together. Or even better why couples which are unhappy, but have similar age/BMI, have less chance to stay together.

  2. Of course, there’s a massive self-selection bias here. Who is not choosing to complete these surveys? How do people who pay for such a service compare to those who do not? Who exactly is being represented here? And what analytic choices were made before they arrived at this point?

    I see very little merit to this research. This looks like someone running a bunch of random correlations in search of an interesting story. It’s exactly the same garbage that is regularly slammed on this blog, just with a genetic company’s data instead of a floundering psychologist’s.

    • Picking up on DO’s point, the inference is wrong because the data are not longitudinal. If your spouse is punctual, a skier, a hiker, disapproving of smokers, loving of fast food,… etc., you might more easily pick up the trait. Apologising, when sincere, might be infectious. Also, if there were gender differences for navigational abilities or body clocks you’d get “significant” correlations for heterosexual couples. Just too easy too pick holes here. More positively, its easy to see that to address questions about mate preferences you need to assess people before they pair up. Considering Brandon’s point, you’d want the sample to be as close to random as possible, and you’d still need to guard against fishing (Rahul’s point). It’s obvious randomised longitudinal studies are vital for addressing such basic questions. And yet such such studies are rare, because they’re very difficult.

    • Rahul:

      The particular issue of multiple comparisons should be resolved by analyzing all possible outcomes at once, ideally in a multilevel model, as discussed in my paper with Hill and Yajima. That’s one reason I wrote above that I’d like to see the full list, not just a few selected comparisons.

    • Yes, it would be a huge concern if the primary metric you’re going by is a p-value. But what’s wrong with sidestepping hypothesis testing altogether, simply estimating the effect sizes (i.e., correlations), and then making guarded inferences on its plausibility as a population effect based on the size of the interval and point estimates? I see no a priori reason why inductive/exploratory approaches such as this are inherently bad; it’s all in how it’s carried out and how “support” is defined.

  3. 1. “… more likely to marry each other”? This suggests causality. Do they claim that their subjects had these traits *before* they became members of a couple?

    2. “… controlled for race and age”. They should control for income, number of children and geographical region. My intuition is that once you group by these 3, you get very similar values for things like “skiing” and “fast food”. And then, of course, couples tend to do things like that together.

  4. This gets at something that may be useful, which would be a look at the traits which require more shared time and traits which require more individual time. Speaking from anecdote, I know many couples – including the one I’m in – which divide activities into those we share and those we prefer to do on our own. Driving is an example: though we do it together, I let my wife go her way and she lets me go mine, meaning I like to navigate my way and drive my way and she prefers to do that herself too. By not doing the navigation together, we instead trust the competence of the other and share the space with something else. Night versus morning gets at the same thing: time to get done what you alone need or want to do, even if that’s just letting your mind wake or whatever. Cooking can be a shared activity for couples but in many houses, like ours, we take turns and/or share aspects.

  5. > We have about half a million customers answering thousands of survey questions on everything from homosexuality to extroversion to infidelity, which as you can imagine produces an interesting dataset.

    Since this is 23andMe, you also have (Wikipedia estimates) something like 700,000 sets of SNP data, which is enough to produce a lot of GWAS hits for many traits. These correlations seem rather boring, why don’t you look instead at GWAS hits for homosexuality, extroversion, education level (I know you have that since I just bought 23andMe and the site immediately asked me what degree I held), infidelity – the most interesting part of your dataset!

  6. Well, sometimes people are compatible when they like the same sorts of things, eg, hiking or the same kind of movie. Other times, people with complimentary likes are compatible, eg, For a package of mixed flavored jellybeans, someone who likes black jellybeans is compatible for someone who hates them.

Leave a Reply

Your email address will not be published. Required fields are marked *