There have been several attempts to estimate the validity of such models and determine all true sources of economic growth. The model of extreme boundaries sensitivity analysis (EBSA) , Bayesian model of averaging (BMA) [2; 3], the Bayesian averaging of classical estimates (BACE) approach , and general-to-specific methodology [5; 6] have been worked out. All the studies were conducted on cross-section data and investigated over 50 factor variables that had ever been claimed significant in cross-countries growth regressions. In [5; 6] the EBSA model was shown to be too strict, the BMA, and BACE — too loose. General-to-specific methodology happened to have the desired properties in the part of keeping the balance between the type 1 and type 2 errors.
The research method advocated in this work presumes the logic of the ideas stated in [5; 6]. Though the model to be built should be adjusted in the part of the variables to be included, and transmitted to panel data.
The countries, which compose the data set serving the base for the final specification, can be regarded population. So it is appropriate to contract the basic model applying fixed effects estimation procedure. At the end, however, to distinguish between various initial conditions’ impact reestimation of the final model could be provided with random effects.
It is reasonable to assume there were no shocks during transition common for all the countries under investigation. So it is appropriate to consider only specific effects.
Let’s assume that the impact the lagged variables have on the current value of growth is realized completely through more investment (and not resource reallocation). So we include among regressors a lag for investment and no lags for other variables, no lagged growth either. This enables to escape finite sample bias all existing estimators would have.
Estimation can be potentially biased at any step of the procedure because of two reasons. The first one is the omitted variables bias. This, however, is not a crucial thing, as all the wrong final specifications would be eliminated at the later stage of performing encompassing tests. Redundant regressors at the first stages do not create bias. Let’s assume that the number of true regressors is not larger than 15. The other problem is with choosing the appropriate estimator. The panel, which can be obtained for transition countries, would be neither long nor broad. LSDV (FGLS for the random effects model) and Maximum Likelihood estimators, which are proposed as estimators for panel models without the dynamic component, are sensitive to violating assumptions imposed on the error term. We are going to employ LS estimators, so it is necessary to control the autocorrelation and heteroscedasticity.
There are a number of search paths considered. The general guideline is as follows.
1. Choose any 15 series among the potential explanatory variables.
2. Create 2 sub-samples: one including the observations for all but the last year, the other — for all but the first year. Estimate the general specification, using the full set of candidate variables and applying LSDV.
3. Run a bunch of tests: LM test for the first-order serial correlation; F-test of the hypothesis that the current specification is a valid restriction of the general specification. The number of tests failed is recorded.
4. If there are variables insignificant at more than 5 percent level, eliminate the variable with the lowest t-statistic and go to 3. The statistics is calculated using White’s heteroscedasticity corrected standard errors.
5. If the current specification with all the variables significant at 5 percent level can pass all the tests as in 3, record it as the terminal specification and go to 10.
6. If any of the tests fails, return to the last specification, for which all the tests are passed.
Чтобы распечатать файл, скачайте его (в формате Word).
Ссылка на скачивание - внизу страницы.