Appendix: Decomposing Employment Growth
The decomposition method, commonly used in labor economics, attempts to attribute the difference in the dependent variable across two groups of observations into the difference in the explanatory variables (endowments), the difference in the relationship between the endowments and the dependent variable (coefficients), and a remaining factor which is the interaction between endowments and coefficients.^{12 }particular, a differential can be decomposed as follows: In
R = y_{1}y_{2 }= (x_{1}x_{2})
2
+
x_{2}(
1

_{2}) + (x_{1 }– x_{2})(
1

_{2}) = E + C + CE
where R denotes the raw differential between the means of the dependent variable y measured for two groups of observations, x is the row vector of the means of the explanatory variables x_{1},...,x_{k}, and _{1 }and the column vectors of the coefficient for the two groups. In the final part of the expression, E=Endowments, C=Coefficient, and CE=Interaction of C & E. The question that usually comes up is how to allocate CE. In the OaxacaBlinder decomposition, it is allocated along with coefficients, so that Explained = Endowments and Unexplained = Coefficients + Interaction. However, CE can also be allocated to E, or even divided between E and C. In what follows we allocate the interaction effect along with the coefficient effect.
2
We consider 3 sources of differences in growth: ownership, sectors, and size. The decompositions are performed for the following groups of countries:
•
Cohesion group versus EU8 member states
•
EU8 member states versus SEE group
•
EU8 member states versus CIS group
•
SEE countries versus CIS economies.
There are almost no privatized and state firms in our sample of cohesion countries, and we therefore drop any remaining privatized and state firms from the Cohesion group. The TE groups retain these. The benchmark category (excluded from the decomposition regressions) is new private firms. The regressions contain 2 dummy variables for the remaining ownership categories, privatized and stateowned firms. They also include 6 dummy variables for sectors. For simplicity of estimation and interpretation, we do not interact sector and ownership dummies thus assuming that the sector growth patterns do not vary by ownership. In contrast, size effects in our specifications can vary by ownership, since we want to separate size effects from ownership effects (for example, new private firms can grow fast because they are small and/or because they are entrepreneurial). Therefore, we interact size (average employment over 200205 measured in thousands) and ownership to get sizeownership effects for the TEs.
12
We use Ian Watson’s (2005) “decomp” addin for Stata for all our decompositions.
71