ches Bundesamt). The survey consists of all the large German manufacturing firms that have
500 or more employees over the entire period. To limit the reporting burden for smaller firms,
firms with 20 499 employees are included only as a random sample that can be assumed as
being representative for this size category as a whole. Firms with less than 20 employees are
not included.8 As a rule, the smaller firms report for four consecutive years and then are substi-
tuted by other small firms (rotating panel).9 Because the estimation of firm-specific fixed effects
requires at least two observations, firms with only one observation are excluded, thus leaving
approximately 39,000 firms in the sample.
The Cost Structure Census contains information for a number of input categories, including
payroll; employer contributions to the social security system; fringe benefits; and expenditures
for material inputs, self-provided equipment, goods for resale, and for energy. Also included
is information on expenditures for external wage-work, external maintenance and repair, tax
depreciation of fixed assets, subsidies, rents and leases, insurance costs, sales tax, other taxes,
public fees, and interest on outside capital, as well as “other” costs such as license fees, bank
charges, and postage or expenses for marketing and transport. Further information available in
the Cost Structure Census includes industry affiliation; type of business (craft or manufactur-
ing); location of headquarters; value of the stocks of raw materials, goods for resale, and final
output; and the amount of R&D expenditure as well as the number of R&D employees.10 The
information on employment comprises the number of owners actively working in the firm and
the number of full-time, part-time, home-based, and temporary workers.
3.3 Estimation results of the production frontier
Table 1 displays the parameter estimates of a translog production function according to Equa-
tion (1) based on the micro data of the individual firms.11 We include dummy variables for the
different years of the observation period, with 2005 being the year of reference to account for
yearly shifts in the frontier. The fit of the regression (R2) is remarkably high (0.995) and the
fixed firm effects as well as the year effects are highly significant.12
Several specification tests were performed to see whether our estimated technology is con-
sistent with predictions from neoclassical production theory. First, we investigated whether
the translog specification is superior to a simple Cobb-Douglas specification by testing the null
8Since 2001 the statistics also contain firms with 1-19 employees. However, these firms are not included in our
analysis due to a rotating sampling scheme; only one observation is available for most of these small firms.
9Due to mergers or insolvencies, some firms have less than four observations. Note, however, that firms are
legally obligated to respond to the Cost Structure Census; thus, there are actually almost no missing observations
due to nonresponse.
10Information on resources devoted to R&D has been gathered in the Cost Structure Census since 1999.
11Least squares dummy variables (LSDV) method for panel data; see Baltagi (2001) and Coelli, Rao and Battese
(2002) for details on this approach.
12The results of a Hausman-Wu test indicate correlation between fixed effects and the other explanatory vari-
ables. Thus, a random effects model or a stochastic frontier framework is not appropriate in this case.