I’m particularly impressed by Joel et al 2017, where participants spent half an hour filling out several hundred of the most useful survey/psych questions (many extremely similar to what ‘date documents’ record) first, and past the global ‘hot or not’ ratings everyone could agree on, they throw high-powered random forests at trying to predict pairs of men/women, and the pairwise random forests do not merely fail to add much predictive power, they actually make the predictions worse! I, uh, did not predict that.
As far as I can tell, the outcome that the study was trying to predict was perceived compatibility on a four-minute speed date? If that’s the amount of time you have to get to know a person, it doesn’t sound too surprising if the global ‘hot or not’ ratings are the only useful predictor. Many people even reserve deal-breaker questions like “kids or no kids” until the second full date or later.
Given that squidious was talking about cases where people jump into a relationship and might find out about serious problems only much later, e.g. at a point where they might already have kids, it seems that the kinds of long-term issues she was talking about would also go unnoticed in a situation where you only had four minutes to assess the other person.
The authors note this limitation themselves, and seem to say that the actual question squidious is referencing hasn’t even been studied, since it’s methodologically too hard:
The present findings address only obliquely the predictability of long-term romantic compatibility. Even if unique desire in initial interactions is not predictable a priori, a matching algorithm could serve a useful function by surrounding users with partners with whom they would ultimately enjoy long-term compatibility should a relationship develop. Building and validating such an algorithm would require that researchers collect background measures before two partners have met and follow them over time as they become an established couple. To our knowledge, relationship science has yet to accomplish this methodological feat; even the commonly assessed individual-difference predictors of relationship satisfaction and breakup (e.g., neuroticism, attachment insecurity; Karney & Bradbury, 1995; Le et al., 2010) have never been assessed before the formation of a relationship. For these variables to be useful in a long-term compatibility algorithm that also separates actor, partner, and relationship variance, researchers would need to predict relationship dynamics across participants’ multiple romantic relationships over time (Eastwick et al., 2017). Predicting long-term compatibility may be more challenging than predicting initial romantic desire.
Great point. Also, going on speed dates is already a selection effect, for example I consider speed dates to be a waste of time, especially with random people (I think they are usually not preselected?)
As far as I can tell, the outcome that the study was trying to predict was perceived compatibility on a four-minute speed date? If that’s the amount of time you have to get to know a person, it doesn’t sound too surprising if the global ‘hot or not’ ratings are the only useful predictor. Many people even reserve deal-breaker questions like “kids or no kids” until the second full date or later.
Given that squidious was talking about cases where people jump into a relationship and might find out about serious problems only much later, e.g. at a point where they might already have kids, it seems that the kinds of long-term issues she was talking about would also go unnoticed in a situation where you only had four minutes to assess the other person.
The authors note this limitation themselves, and seem to say that the actual question squidious is referencing hasn’t even been studied, since it’s methodologically too hard:
Great point. Also, going on speed dates is already a selection effect, for example I consider speed dates to be a waste of time, especially with random people (I think they are usually not preselected?)