Avoiding the pitfalls when economics shifts from science to engineering.
New York Times; New York, N.Y.; Aug 29, 2002; Hal R. Varian

ECONOMISTS think of themselves as scientists; their primary goal is to understand how the economy works. But scientific knowledge is not their only goal; as a famous economist once remarked, ''The point is not to understand the world, but to change it.''

Economists are increasingly being called on to give advice about how to design new economic institutions. They have been consultants in the design of auctions, power exchanges, financial exchanges and a variety of other market and market-like mechanisms.

In these applications, economics looks more like engineering than it does pure science. Just as a civil engineer applies principles of physics and mechanics to design bridges, economists apply principles of economic analysis to design exchange mechanisms.

Al Roth, an economist at Harvard, recently described an interesting case study of ''economist as engineer.'' In the mid-1990's, Mr. Roth worked with the National Resident Matching Program to design a new system for matching residents and hospitals. Documents that describe his experience are available at www.economics.harvard.edu/aroth/alroth.html.

As with civil engineering, economic engineering at its best involves an understanding of the historical and institutional environment, coupled with theoretical models, computational models and real-life experiments.

Start with the institutional background. Each year this program matches about 20,000 medical students to hospitals. Students interview with potential employers and then submit a ranking indicating their preferences. Similarly, hospitals submit a ranking of residents. The program then uses a computer algorithm to match residents with hospitals.

The clearinghouse started in the early 1950's and worked reasonably well for a while, but by the 1980's there was considerable dissatisfaction. The chief issues had to do with whether the system was fair to students, whether there were ways to ''game the system'' and whether there were better ways to treat married students who wanted jobs near their spouses.

As it happens, there is a large theoretical literature on ''matching problems'' of this sort. In this literature, a match is ''unstable'' if there is some student and some hospital that prefer each other to their current matches. If there are no such pairs, the matching is ''stable.''

Stability is obviously a desirable property. It turns out that in the simple cases studied in theory there are usually many stable matches. There are relatively simple algorithms that can find the ''best'' stable match for each side of the market, but these are typically different.

So far so good -- there are easy ways to find good matches, at least in theory. What about gaming the system? Here the news is not so good. It turns out that there is no algorithm that cannot be manipulated in matching students and hospitals: there will always be cases in which applicants or hospitals will want to change their stated ranking to improve their prospects.

But the theory only says there are some cases in which participants want to misrepresent their preferences; it does not give guidance about how often this occurs.

This observation led Mr. Roth to do some computational experiments to determine just how likely such misrepresentation would be. It turned out that it was not a common problem for realistic distributions of preferences. Even though manipulation was always possible, it was not much of a problem in practice, if an appropriate algorithm was used.

What about the couples problem? Here theory was even more pessimistic: there might be no stable matching if couples were involved. Still, simulations showed that this was also relatively rare.

The next stage was to experiment with real-life data. Experiments are important in economics for the same reason they are important in engineering or medicine: something may work well in a controlled environment and fail miserably in real life.

Some experiments are computational: applying various algorithms to actual data, to see how they would fare. This was particularly easy in this case, since the data from other years was readily available. One could also use laboratory experiments with human subjects, but that did not have a big role to play here. (In other work Mr. Roth reports subsequent experiments designed to examine various alternative matching models.)

Theoretical and computational models are important in coming up with designs. But they should be subject to experiment to really test them out, since effects that aren't in the models could be critical in practice. In civil engineering these might be things like wind and snow; in economics they might be things like psychological bias and social norms.

In some cases, market participants can figure out loopholes the designer never thought of. This has been a particular problem in the design of spectrum auctions and electricity exchanges.

Of course, as we learn more about these issues they can be incorporated into the theoretical and computational models. But even when you think you have all the pieces in the model, it is good practice to try something out on a small scale first.

The matching program adopted the Roth algorithm in 1997. ''The first few years of operation in the new match design seem to have been extremely smooth,'' he reports, ''and while many issues related to the medical labor market seem to be of lively concern, the organization of the match no longer appears to be a source of significant controversy.''

Other economic engineering projects have not fared so well, with the California electricity market being a notorious example. The ''famous economist'' quoted above was Karl Marx, who also had ideas about economic design that ended disastrously.

These examples should be sobering lessons to economic engineers. An understanding and appreciation of existing institutions, good theory, good computational modeling and well-designed experiments are critical ingredients to a successful design.