.

that is a terribly broad question ...

my two cents (maybe less ...) is that the term "model" will have two different meanings there :

• statistical models are there to reveal the correlation structure in the data, possibly adding any prior domain-knowledge available
• economic models are there to reveal the causal structure in the phenomena underlying (or producing) the data and (possibly) to allow careful design of experiments so as to test such causal structures (the same would be true for "physical models" if we were talking about physics instead of economics, and indeed, in physics, the design of experiments is a mandatory part of the job !)

models (either statistical or economic) should be chosen according these different goals ...

and of course, there must come a time where both kinds of models have to be compared and discussed !

for a minimalist approach of the link between "Probability, Causality, and the Empirical World", you might be interested in reading the excellent paper below :

http://projecteuclid.org/download/pdfview_1/euclid.ss/1089808272

.

Try and make sure the model satisfies three criteria: Theory, Analogy and Model. Ultimately will help to draw a balance between the absolute (economic theory criteria) and the inadequate (empirical evidence without a theory) (Derman 2011, Chapter 5 & 6). This I gleaned from an interesting reflection on model building by Emanuel Derman, a former quant. You may want to temper this view with a more scholarly perspective from Roman Fryman and Michael D. Goldberg. The main thesis being that models are reflexive because they are a human artefact- they are not permanently infallible. Or another way to put it: once a model is created it begins to sow its own seeds of destruction especially if it becomes popular to the point of being the canonical one for the job.

Derman, E. (2011, October 25). Models. Behaving. Badly.: Why Confusing Illusion with Reality Can Lead to Disaster, on Wall Street and in Life. Simon and Schuster.

Frydman, R., & Goldberg, M. D. (2011, February 7). Beyond mechanical markets: Asset price swings, risk, and the role of the state. Princeton University Press.

To summarize:

1.Choose a model that is relatively new so that it does not become "part of the statistic" noting its theoretical pedigree. (taking Fryman & Goldberg 2011's advice)

2.Choose a model that is relatively simple so that its complexity does not fall prey to inadequacies of reality (economically fanciful but of no practical use since one might need to fit it to reality through many rounds of statistical estimations. Scaling downwards by making things simple is harder than scaling upwards by complicating the equation.