Tuesday, 16 April 2013

Social science statistics: of muscle cars and macro-level statistics.



AMC Rebel 1970, a typical American muscle car
Trying to think through another worry about social science practice, and in particular how we do macro-level empirics - comparing countries, polities or regions. Warning: I strongly suspect that this will mostly prove that I should read more Barbara Geddes. (And who shouldn't?)

Suppose you were a Chrysler executive in 1970, and were evaluating the potential threat from Japanese cars.

Here is an approach you could take. Divide cars into American-made and Japanese-made. Examine the average performance of each. Find that American cars are superior to Japanese cars. Go to sleep contentedly. Inferior cars will never take over our market.
http://upload.wikimedia.org/wikipedia/commons/0/0d/Toyota_Corolla_First-generation_001.jpg
First generation Toyota Corolla, 1966

This approach is obviously flawed. "American cars" and "Japanese cars" are not natural kinds, like "killer whale" and "grey whale". They refer to the deliberate creations of human intelligence. The average Japanese car of the 1960s was not a useful predictor for the average Japanese car of the 1970s. The best car would probably have been a more useful predictor, because Japanese car industry executives were capable of learning.

Now consider how political scientists evaluate the relative performance of democracies and authoritarian regimes. (Or parliamentary versus presidential regimes, or whatever else.) Assume away all the problems of causal identification, unobserved heterogeneity and so on. Even beside all of this: they are basically taking the same approach as above. They look at the average performance of democracies versus dictatorships in the past. To get a big sample, they might go quite far back - to 1950, or even beyond to the 19th century.

But political regimes are also the products of human learning and intelligence. "Democracy" and "dictatorship" are not natural kinds either. They are systems put in place and altered by people. For example, British democracy of 2013 -- with its legal checks on the executive, its relationship to the EU, its devolved governments in the nations, and its quangos and bureaucrats -- is quite different from British democracy in the 1970s -- with its corporatist structure, industrial policy and so forth. Bluntly, both democratic and non-democratic regimes learn, and are constantly rebuilding themselves over time.

For this reason, average past democratic or authoritarian performance may well not be the quantity of interest. Democrats (dictators) should worry about the best, most successful authoritarian (democratic) regimes. Data from the 1950s are likely to have limited relevance.

File:2011 Toyota Corolla -- NHTSA.jpg
Toyota Corolla 2011