we are able to combine the effects of both internal and environmental factors and also control
for a number of other variables.16 Further details about the data can be found in the Appendix.
With the standard errors of efficiencies estimated in the first step, it is possible to apply
the more efficient weighted least squares method, instead of OLS, in the second step, using the
reciprocals of the standard errors of efficiency as weights. However, the results changed so little
that we decided to report only the OLS results in the following sections.17
4.2 Empirical results
Table 5 displays the partial R2 values, which indicate the relative importance of a variable
for the entire observation period, 1992-2005 (Model I), or for the last six years, 1999-2005
(Model II). Conducting the analyses for the subperiod of 1999 to 2005 allows the inclusion
of information on R&D intensity and temporarily employed (subcontracted) labor, which is
only available for from 1999 onward. Table 6 provides the signs, magnitudes, and t-values for
all continuous and some selected categorical variables. We include the number of observation
periods as a control variable for sample selection. Of potential concern in these estimates is that
some inefficient firms exit the market and are consequently not included in the sample in later
years, a situation known as panel attrition. This could lead to an attrition bias since efficiency
is the dependent variable of the analysis. If this is the case, we should find a significantly
positive relationship between a firm’s observation periods and its efficiency. However, we find
that the number of observation periods is negatively correlated with efficiency, although with
low explanatory power measured in terms of partial R2. Hence, we cannot preclude that there is
a sample selection bias, but in the opposite direction of attrition - firms that stay in the sample
longer, presumably the larger ones, tend to be less efficient. An indication of an attrition bias
is found only for the subgroup of least efficient firms (Table 8), which is probably due to a
moderate survivor bias for this group of firms.
Several conclusions can be drawn from the results in Tables 5 and 6. First, in both models,
for the 1992-2005 and 1999-2005 period, all included independent variables - except the year
effects - have significant explanatory power at the 1 percent level. This might in part be driven
by the huge size of the dataset. However, with regard to the magnitudes of partial R2s, we can
state that industry affiliation, firm size, and location have by far the most important effects on
productive efficiency. Jointly, the effects adds up to 84 percent (Model I) and 82 percent (Model
II) of the models’ explanatory power.
16Note that the industry classification changed in 1995 from WZ1979 to WZ1993, the latter corresponding to
the international NACE classification. We kept only those firms in the sample for which an industry affiliation
according to WZ1995 is available, i.e. which have at least one observation after the year 1994. Furthermore, in the
second step of our analysis of the determinants of efficiency, we excluded all firms that changed industry affiliation,
location, or legal form during the observation period.
17The WLS results can be obtained from the authors upon request.
12