Thoughts on Gdp Up, Happiness Down
In the past I've talked about decision markets, more on the Futarchy discussion group I created than on this blog. One idea in Futarchy is that GDP or some extension thereof is a reasonable proxy for happiness. (There's much more. I'm greatly short-changing it)
"After a gradual upward trend [in happiness] that ran from January to April, 2009, the overall time series has shown a gradual downward trend, accelerating somewhat over the first half of 2011."
Is that comparison right?
The article didn't say what they denominated GDP in, or I missed it. It's surely in the paper but I just read the article 2 minutes before starting this blog post, so I haven't checked.
But if their GDP is denominated in dollars, the apparent paradox may just mean that the dollar is worth less now. No surprise there.
How they measured happiness
It was interesting how they tried to measure happiness: by analyzing Tweets. They had a database from Amazon's Mechanical Turk of the emotional positiveness of English words. They graded 46 billion words worth of Tweets on emotional positiveness.
Is it representative? Not a lot
The researchers admit that their happiness metric is less than representative: "It does skew toward younger people and people with smartphones and so on - but Twitter is nearly universal now," Dodds said, "Every demographic is represented."
Could that metric work for futarchy? Not like that.
We might want to for two reasons:
- We'd rather measure happiness than dollars, which we were treating as a proxy for something like happiness anyways.
- Having more good metrics is always better.
But that's about the most spoofable metric ever. If I had money riding on "Happiness will go down", I would have an army of webbots out there tweeting variations of "I feel suicidal" over and over and over.
The researchers would catch on, you say? Sure, and then there'd be an arms race. I gave you the cheap quip version, but off the top of my head I can think of half a dozen ways to fake more human-like miserable tweets. Ultimately I am confident that the fakes would win the arms race.
So it's another example of Goodhart's Law that proxy measures only work OK when nothing important depends on them.
Could that sort of idea work at all?
Well, can we design around Goodhart's Law with this? Perhaps only choose the proxy after the fact, and reject ones that look like they've gotten distorted.
It might work in some situations:
- If there so many chooseable proxies that beforehand, the expected cost of distorting them, spread over many proxies, are more than the expected gain of distorting the right one(s).
- If the procedure for rejecting distorted ones is not itself gameable.
So the idea isn't that promising, but is worth a mention.
1 University of Vermont (2011, December 16). GDP up, happiness down: From Twitter, scientists measure mood. ScienceDaily. Retrieved January 1, 2012, from http://www.sciencedaily.com/releases/2011/12/111216174440.htm