and Arora, (1972) and Wansbeek and Kapteyn (1982) proposed alternative FGLS approaches. Wallace and Hussain (1969) use OLS residuals; Swamy and Arora (1972) use residuals from Within and Between estimators; and Wansbeek and Kapteyn (1982) use Within estimator residuals. Since our model includes time-invariant variables which can not be estimated by Within estimator, Wallace and Hussain’s approach is an appropriate method. The standard error component model given by equation (3.1) assumes that the regression disturbances are homoscedastic with the same variance across time and individuals. When heteroscedasticity is present assuming homoscedastic disturbances will result in consistent but inefficient coefficient of the variables and biased standard errors (Baltagi, 2001). The heteroscedastic models assume that the variances of the μitand/or the uitchange between cross-sectional units and this difference is not accounted for by variables. LMhet.statistic could be used under the null hypothesis of homoscedasticity to test the heteroscedasticity in uit. This LM test uses the residuals from the OLS estimation. If we expect heteroscedasticity to exist in the uitin the REM case, then the statistic based on the OLS estimation may also be used in the REM. Once the heteroscedasticity is detected, one possible way to correct heteroscedastic bias in the variance-covariance matrix is to use the White’s estimator (Erlat, 2006; Greene, 2003). International Trade and Finance Association: International Trade and Finance Association Working Papers 2008
5 2122.1ˆˆ2∑=−=NiihetTLMεεσσ∼21−NχSimilar to the problem of heteroscedasticity, the presence of the serial correlation results in consistent but inefficient estimates of the regression coefficients and biased standard errors (Baltagi, 2001). Wooldridge (2002) proposes an AR(1) serial correlation test under the null hypothesis of no serial correlation for one-way panel data models. The test is applied by regressing the residuals from the OLS estimation of first-differenced variables on the lagged residuals. If the residuals from this estimation have an autocorrelation coefficient of -0.5, then the null hypothesis can not be rejected. Drukker (2003) developed a Wald test (FAR(1)) for this testing approach that is used here. 5. Empirical Findings The gravity model estimation and the tests results are reported in Table 1. The LM and Honda statistics show that both random individual and time effects are significant. The LMgroupand Hondagroupshow that individual random effects are significant. On the other hand, time effects are not significant according to the LMtime and Hondatime. The Hausman test confirms that there is no correlation between individual random effects and explanatory variables, indicating that the REM is consistent and efficient. Tests results verify our model selection and refer to the one-way REM including only individual effects.