The basic job of an economic theorist is to write papers, usually with at most one or two co-authors, that develop new models of some phenomenon of interest. A reasonably successful theorist writes about one such paper per year. Each paper contains a brand-new model, which while often similar to other models that have come before, has to be built up from scratch. The fact that you have to build up a new model with each paper, combined with the fact that you have to write lots of papers, means that the models can’t be too complicated, or at least can’t be complicated in ways other than the specific ways that you want them to be. They have to be tricked out in just such a way as to allow you to get at the question of interest, while leaving a whole bunch of other (important) stuff out.

This is not the only way that economic modelling could be done. You could imagine an alternative in which teams of modellers work for a long time on developing much more complicated models, running different versions of them (using different assumptions or different values for the various parameters of the model), seeing what pops out, and then writing a string of papers describing the ways in which the model has been tweaked and reporting the results that have been obtained. There are models like this floating around (I think the Federal Reserve has a big one) but they are rare and almost completely absent from the academic literature. Years ago they were somewhat common in Macroeconomics, but they seem to have fallen out of fashion. There might be good reasons not to use these kinds of models. One theorist friend suggested that working through mathematical proofs of relatively simple models provides more intuition and insight than just cranking away on a huge model that’s too complicated for anyone to really understand. I think he has a good point, and there may be others. But I suspect that these kinds of models are rarer than they should be, and I think the reason is because they are not fun for theorists to build. Building a nice model from scratch feels like a creative act, one with a lot of aesthetic appeal, almost more art than science. Building a huge model with a bunch of other geeks and sticking ugly numbers into it feels much different, and the kind of people who like that feeling are the kind who would have become empirical or experimental economists, and not theorists, in the first place.

Sorry, to be clear, I'm referring to microeconomic models like Arrow-Debreu equilibrium or game theory; e.g. the sort of stuff that appears in http://econtheory.org/ . Looks like the other commenters have macro models in mind.

Making big complex models sounds like a great idea if you can fit them to real data (or attempt to do so). But theoretical economics is so theoretical, the models are related to reality only though analogies. Thus the only value you can get out of them is to explore their mathematical properties and hope those properties are analogous to properties in the real world. Since it's easier to explore mathematical properties of simpler models, simple models are better for pure-theoretical modeling.

Now, of course, that doesn't address the value of the entire enterprise of theoretical modeling... I still like it, but I admit that's more due to aesthetics than a conviction it's good for the advancement of scientific understanding. (I confess to having written both an honors and master's thesis that were purely theoretical...)