about me
game theory
... more
Subscribe Weblog

The recent empirical literature on economic growth has identified a substantial number of variables that are partially correlated with the rate of economic growth. The basic methodology consists of running cross-country regressions of the form
extrem01where γ is the vector of rates of economic growth and x1, x2, ..., x2 are vectors of explanatory variables which vary across researchers and across papers.

Since the multiplicity of possible regressors is a pain in the neck, researchers soon started to ask themselves what variables are "truly" correlated with growth. An initial answer was given by a version of Edward Leamer's extreme bounds analysis since it promised to be capable of identifying "robust" empirical relations for economic growth. In short, the extreme bounds test works as follows: imagine that you have a pool of K variables previously identified as related to growth and you are interested in knowing whether variable z is "robust". Then you would estimate regressions of the form
where y is a vector of fixed variables that always appear in the regression (since one knows that those variables really^2 make sense), z is the variable of interest and xj is a vector of up to (say) three variables taken from the pool of the K variables available. One needs to estimate this regression or model for all possible xj combinations. For each model, j, one finds an estimateextrem03

Since this test was too strong for most variables to pass (all variables are fragile), Sala-i-Martin (1997) proposed to depart from this "extreme" test and decided to assign some "level of confidence" to each variable. To this end, he constructed weighted averages of all the estimates of βzj and its corresponding standard deviations, using weights proportional to the likelihoods (somebody has just opened Pandora's box) of each of the models. If you think that's interesting you might wan't to download Determinants of Long-Term Growth: A Bayesian Averaging of Classical Estimates (BACE) Approach (Sept. 2004) Here's the free version (Feb. 2003). There is an even older version available at the NBER (2000, uses a different dataset but the theoretical part hasn't been altered (except typos)). Sala-i-Martin et al. were even able to come up with posterior inclusion probabilites - the posterior probability that a particular variable is in the regression.

His results? Variables that are strongly or robustly related to growth (don't accuse me of doing sign econometrics, take a look at the paper for further information):
  • dummy for East Asian countries (+)
  • primary schooling enrolment rate in 1960 (proxy for human capital) (+)
  • average price of investment goods between 1960 and 1964 (-)
  • initial level of per capita GDP(-)
  • fraction of tropical area (-)
  • population density of coastatal aereas (+)
  • Malaria prevalence (-)
  • life expectancy (+)
  • fraction Confucian (+)
  • African dummy (-)
  • Latin American dummy (-)
  • fraction of GDP in mining (+)
  • former Spanish colony (-)
  • number of years an economy has been open (+)
  • fraction Muslim (+)
  • fraction Buddhist (+)
  • ethnolinguistic fractionalization (-)
and last but not least the share of government consumption in GDP is also robustly estimated and its sign is negative. "This could be expected because public consumption does not contribute to growth directly, but it needs to be financed with distortionary taxes which hurt the growth rate."

stolen from Determinants of Long-Term Growth: A Bayesian Averaging of Classical Estimates (BACE) Approach, 2003

via AdamSmithee

related stuff:
Economic Growth & Parameter Heterogeneity