"[T]he purpose of a model is not to be realistic. After all, we already possess a model that is completely realistic—the world itself. The problem with that “model” is that it is too complicated to understand. A model’s purpose is to provide insights about particular features of the world. If a simplifying assumption causes a model to give incorrect answers to the questions it is being used to address, then that lack of realism may be a defect. . . . If the simplification does not cause the model to provide incorrect answers to the questions it is being used to address, however, then the lack of realism is a virtue: by isolating the effect of interest more clearly, the simplification makes it easier to understand."Models are never perfectly realistic. I don't think I've ever suggested that perfect realism is necessary or even good, so I should hope Bob doesn't think that's my standard. But realism of assumptions matters so I will never say something like "all that matters is whether your predictions are accurate". It's not all that matters. One can plausibly sacrifice predictive power for realistic assumptions just as one can plausibly sacrifice realism in assumptions for predictive power.
--David Romer, Advanced Macroeconomics (New York: McGraw-Hill, 1996), pp. 11–12.
The scientist's utility function is quite well behaved. When we are looking for assumption realism/empirical validity bundles, our utility functions are concave and we have diminishing marginal rates of substitution between the two. The real world and the critiques of our peers present us with the constraints that we're optimizing against.
Every time an economist needs to explain these epistemological things on the realism of his model is because he went too far off the mark. Did you notice that no other sciences has similar complains about this issue?
ReplyDeleteBest,
Pablo Mira (economist)
Argentina
No, Daniel, I'm sorry but you are simply mistaken on this. Look again at what Romer wrote:
ReplyDeleteIf the simplification does not cause the model to provide incorrect answers to the questions it is being used to address, however, then the lack of realism is a virtue:
Look at how strong that statement is. Romer isn't saying that he's willing to tolerate a lower degree of realism in his model, so long as it allows him better accuracy.
No, Romer is saying that he WANTS the model to be simpler--i.e. less realistic--so long as that doesn't hurt its explanatory accuracy. This is the opposite of what you are saying.
Romer doesn't think there's a tradeoff between realism and predictive accuracy (or explanatory power, if you prefer). He actively dislikes "realism" if it makes the model more complicated. The only reason he would accept a more realistic model, is if it achieved more explanatory power for the phenomena under consideration.
Now maybe he's got a weird set of priorities, and yours are better, but you are simply wrong if you think Romer is agreeing with you on this.
Of course we want the model to be simpler and less realistic - that's what models are. We have to simplify. It's a virtue that we have a simpler picture than reality.
DeleteI can't comprehend how this is not coming through. I am having another Mugatu moment.
Romer likes realism, non-complication, and explanatory power. I do too.
DeleteRomer recognizes that there is a very real trade-off between realism and non-complication and a very plausible (although not necessary) trade-off between realism and explanatory power.
He does not think we should be at the corner solution of any of those.
Neither do I.
Given the earlier comment perhaps I should be more realistic in my point about interior solutions: WE LIKE TO SEE non-realism with some explanatory power and some non-complication. That is our optimum. When we start to lose explanatory power we start to consider the extent to which non-realism is too high a cost.
Delete*be more specific in my point about interior solutions.
DeleteI read it exactly like Bob. It sounds like Romer is saying simple is the preference, add complexity only if necessary.
ReplyDeleteThis reminds me of a discussion about modelling sigmoidal curves at a biological assay conference. A 5 parameter logistic fit the data best and for good reasons. There was quite a discussion about whether 4 parameter, 3 or even simple linear regression would be the preferred approach because they are simpler and an FDA guideline require justification for not using the simplest model.
re: "It sounds like Romer is saying simple is the preference, add complexity only if necessary."
DeleteYep!