Math

April 25, 2003 1:42 PM

Every Unhappy Family Has Its Own Bilinear Influence Function.
posted by semmi (19 comments total)

In longitudinal studies of couples who have passed through the Love Lab, Mr. Gottman and his colleagues have successfully predicted which ones will divorce with greater than 90-percent accuracy.

If peer review supports this statement, that's simply amazing.
posted by ZenMasterThis at 2:03 PM on April 25, 2003

I would definitely be interested in reading more about this. Anyone happen to read the book by any chance? (I'm not even totally sure it's been released yet)
posted by ookamaka at 2:22 PM on April 25, 2003

...equations that described a dynamic system in which a snarky comment by one spouse would result in negative emotions in the partner, sometimes resulting in a downward spiral.

They call it "Bilinear Influence Function", I call it Metafilter.
posted by m@ at 2:23 PM on April 25, 2003

[insert snarky comment here]
posted by ZenMasterThis at 2:31 PM on April 25, 2003

Interesting link, semmi, thanks. It would be interesting to see some of that data translated into Tufte's lingo.
posted by yoga at 2:36 PM on April 25, 2003

He has argued, for example, that a crucial predictor of divorce is a husband's inability or unwillingness to be influenced by his wife's suggestions and emotional expressions.

So, in essence, if a husband doesn't do what his wife tells him to do, they'll end up getting divorced?

Sounds like his research is dead spot on.
posted by hurkle at 2:37 PM on April 25, 2003

MarriageFilter.

(you knew someone would say it sooner or later)
posted by ZenMasterThis at 2:40 PM on April 25, 2003

Slate had a piece on this a bit ago, if anyone's interested.
posted by Skot at 2:45 PM on April 25, 2003

"So, in essence, if a husband doesn't do what his wife tells him to do, they'll end up getting divorced?"

I could have told them that without the fancy calculus.

As for the "Love Lab", show me the way
posted by Outlawyr at 2:53 PM on April 25, 2003

[insert negative emotions here]
posted by blue_beetle at 2:55 PM on April 25, 2003

Fuck you all, you slimy pigs! I'm divorcing the lot of you, contemptuous fools.
posted by PigAlien at 3:07 PM on April 25, 2003

Every Unhappy Family Has Its Own Bilinear Influence Function.

I guess Tolstoy was wrong. Huh.
posted by sonofsamiam at 3:16 PM on April 25, 2003

Please professor, a mathematical model of this marriage.
posted by jfuller at 3:43 PM on April 25, 2003

"In longitudinal studies of couples who have passed through the Love Lab, Mr. Gottman and his colleagues have successfully predicted which ones will divorce with greater than 90-percent accuracy."

If peer review supports this statement, that's simply amazing.

My guess is that their model had an R^2 of 0.9, or that it "explained" 90% of the variation in the data, and not that there were any out-of-sample predictions actually being made. Sloppy journalism and technical terms never mix very well.

"Statistics will just give you the bare facts. If you want to understand why a dynamic system behaves as it does, then you need to use nonlinear tools."

This guy claims to understand catastrophe theory, but he thinks that statistics are all linear? Watch the leg, pal...another yank like that and it may just fall right off...

I'm all for rating facial expressions on a scale from -4 to 4, and drawing inferences therefrom, but this article was painful on the brain.
posted by dilettanti at 4:06 PM on April 25, 2003

"Unhappy Family" is the name of an entree at the Chinese Restaurant across from the Bronx County Courthouse.
posted by ParisParamus at 5:10 PM on April 25, 2003

Given that it's a nonlinear model, it probably won't have a R2. I think they meant that they assigned a probability-of-divorce (within the time frame) to every couple, and 90% of the time, couples that divorced had PODs over 0.5, and couples that didn't had PODs under 0.5.

I wonder how they dealt with the censoring problems? Some couples that they met surely just aren't divorced *yet*, and there may be cases where the marriage ended with someone's death, etc etc.

I'm not really impressed with a 90% success rate. We don't know what kind of success rate you get with a simple null model. I've run probits/gompits with a predictive success of about 85%, where the model was predicting that *everyone* would get a 1, and 85% did. Also, I bet it's really rilly easy to pick out half the divorces or so from other variables (age at marriage, or conversational correlates with that, etc).
posted by ROU_Xenophobe at 5:30 PM on April 25, 2003

He has argued, for example, that a crucial predictor of divorce is a husband's inability or unwillingness to be influenced by his wife's suggestions and emotional expressions.

So, in essence, if a husband doesn't do what his wife tells him to do, they'll end up getting divorced?

I can vouch for that - happened to me.

Men: you are either blessed with luck and the woman of your dreams doesn't go all weird after you marry

OR

Men: obey your wives or face divorce
posted by SpaceCadet at 7:02 PM on April 25, 2003

Personal finance teacher at college gave this advice: "Guys, marry a smart woman, and do everything she says." His rationale had a lot to do with divorces wreaking havoc on one's fincancial security, but it tied in with one's own happiness too. A good wife will make you work harder on things you would never do otherwise, but it'll pay off in all sorts of ways.

It seemed like good advice to me and ever since then I've been keeping my eye open for a woman who's seems to be going somewhere where I would like to go too. Still looking.
posted by wobh at 9:04 PM on April 25, 2003

Given that it's a nonlinear model, it probably won't have a R2.

Actually, an R2 measure is frequently calculated for nonlinear binary dependent variable models (e.g., logit and probit models) [see, for instance, pp. 37-39 of Maddala's Limited-Dependent and Qualitative Variables in Econometrics]. The standard R2 is just the correlation between "predicted" (model) and observed values. My point, though, however obscured it might have been, was that the 90% was likely some sort of "goodness of fit" measure, and not the prediction accuracy of the model on some independent (later) sample. Of course, without reading the book, I can't be sure.
posted by dilettanti at 10:59 AM on April 28, 2003