12

presence of this country effect, LSDV is the main technique used for analysis of panel

data when time is dominant over units,

16

producing unbiased and also efficient ß (given

the small number of units in the analysis). The estimated models assume that the

disturbances across panel are heteroscedastic. I also assume first-order autocorrelation

within panels (an AR1 process).

17

In the presence of unit effects, the alternatives to a

LSDV model are two: to lag the dependent variable and estimate an OLS regression with

panel corrected standard errors,

18

or estimate the GLSE ‘random effect’ model.

19

In spite of the fact that a shortcoming of LSDV ‘fixed effects’ for time dominant designs,

in comparison to GLSE ‘random effect,’ is that in the presence of time invariant

regressors, these will be collinear with LSDV dummies (Beck, 2001). In the concrete

case of my models, the slow moving regressors included vary more across time than

across units. LSDV model results should be considered conservative in the sense that low

significance relationships among regressors and the inequality measure could be masked.

Nevertheless, in the world of model alternatives and assumption violations, a type II error

is preferable to model misspecification.

Table 3 displays the analysis results for the *Step 1 models*. Model 1 summarizes the

analysis of the food manufacturing sector FDI effect on employment in manufacturing,

and includes the major macroeconomic and domestic economic factors and welfare

variables emphasized in previous studies. As expected, *Food manufacturing* FDI has a

significant (p ≤ .05) negative effect on *manufacturing employment*. This result supports

16

One of the most important critiques to LSDV model is that it removes degrees of freedom. In this case,

however, the analysis of seven countries (units) will have a cost of 6 degrees of freedom.

17

Heteroskedasticity and AR1 are not tested given the presence of unit effect and the choice for LSDV

model.

18

I opted not to lag the dependent variable precisely because, as Achen (2000) and others warn against,

LDVs tend to dominate the regression equation, generating downwardly biased coefficient estimates on the

explanatory variables. When lagging, both in step 1 models (employment by sector as dependent variable)

and in step 2 models (Gini as dependent variable) the LDV included explains most of the variance of the

former, affecting the significance of the other ß estimates. From a substantive point of view, this leads to a

useless auto-deterministic explanation of inequality.

19

The decision to use fixed effects over random effects laid both on theoretical and methodological

considerations. With respect to the substance, random-effects would be more appropriate if the sample is

drawn from a substantially larger population, factor that in the case of cross-country studies favors the

fixed-effects model. On top of this, the ‘random effects’ model assumption of no correlation of unit specific

errors with model’s independent variables is a too strong one for country panel data.