1. If the driver has as much information as the econometrician, setting both parameters on slope and gas equal to 0 will fit the data equally well.
2. If the econometrician observes slope and gas and speed with error (as will almost always be the case), then GMM will estimate both parameters on slope and gas as zero (if those 3 errors are independent of each other), even if the true parameters aren't.
3. If the driver observes gas or slope with error, and the econometrician doesn't, then the econometrician can indeed actually estimate the parameters.
How likely is 3? Only a stupidly arrogant econometrician would assume a priori that he knows better how to drive the car than the guy actually driving it. (OK, maybe the econometrician has final revised data, and the driver has only real time data, and the final revised data is better than the real time data. But even then the driver will be observing other indicators that the econometrician doesn't have data on, so the econometrician will interpret the driver's response to those omitted variables as "gas pedal shocks", and will screw up the estimation royally.
Yep, all that Sims VAR stuff is wrong. Here, read this.
I should have been a little more specific; the approach I described assumes that the driver makes some errors, so the speed is not actually constant in practice. That can be because the driver slightly errs in seeing the slope or gas, or because of other sources of error (gusts of wind, maybe). Without that variation, you're right, the system is unidentified. The larger the variation, the better the identification. But none of this is news to economists, and none fundamentally requires the econometrician to have better information than the driver (although measurement error can matter in some circumstances, it matters in a different way).
As I'm sure you know, these arguments (i.e. mainstream Economics is wrong and ignorant!) are more convincing when they're accompanied by a model. DSGE models have a lot of limitations, but they're pretty good for demonstrating failure of identification. Monetary policy is tricky to identify, and I'd be sympathetic to an argument showing that deviations from a Taylor-rule are bad for identifying the effects of monetary policy shocks (lots of people would agree with this, the interesting question is whether they're bad in empirically important ways or just conceptually bad) but your car+driver analogy seems like it's aiming to be broader than that.
Thanks psuet. I'm not one of those guys who goes around saying "mainstream economics is wrong and ignorant". Well, OK, we are often wrong and ignorant, but people who aren't in mainstream economics are usually even more wrong and ignorant!
This is just a particular beef I have. Because I keep on seeing examples where economists make mistakes by missing this point. Like when economists try to test whether headline or core inflation is better at forecasting future inflation, and use those results to give policy advice on whether inflation-targeting central banks should respond to core, headline, or both indicators. All they are modelling is the central bank's mistakes, in responding too strongly or too weakly to those indicators.
I think proper identification does require the econometrician be (in some sense) a "better" driver than the driver. If you can see the hills better than the driver can, then you can see the effects of a hidden hill that you know he doesn't see.
\hat parameters = argmin (average_over_t speed(gas_t, slope_t, parameters) - const)^2"
No you can't.
Think about it.
1. If the driver has as much information as the econometrician, setting both parameters on slope and gas equal to 0 will fit the data equally well.
2. If the econometrician observes slope and gas and speed with error (as will almost always be the case), then GMM will estimate both parameters on slope and gas as zero (if those 3 errors are independent of each other), even if the true parameters aren't.
3. If the driver observes gas or slope with error, and the econometrician doesn't, then the econometrician can indeed actually estimate the parameters.
How likely is 3? Only a stupidly arrogant econometrician would assume a priori that he knows better how to drive the car than the guy actually driving it. (OK, maybe the econometrician has final revised data, and the driver has only real time data, and the final revised data is better than the real time data. But even then the driver will be observing other indicators that the econometrician doesn't have data on, so the econometrician will interpret the driver's response to those omitted variables as "gas pedal shocks", and will screw up the estimation royally.
Yep, all that Sims VAR stuff is wrong. Here, read this.
http://worthwhile.typepad.com/worthwhile_canadian_initi/2011...
You have just proved my point: economists don't understand Milton Friedman's thermostat.
Forget your fancy stuff. Just STOP AND THINK about what you are really doing when you try to estimate parameters.