La Unidad de Caja

To investigate the theoretical hypothesis we define two measures which reflect the degree of wage inequality within an establishment. First, we use the observed wage gap: 1 1 1 , , g=m,f g j N m f g g j j j j g ij j j Gap w w w w N = = − =

∑

where wij denotes the log earnings for individual i at establishment j superscripts m and

f refer to male and female observations. g j

N indicated the number of male and female employees, respectively, in establishment j. Since the wage information in our data set is right-censored (see Section 2.4 for more details), the observed wage gap defined in equation underestimates the actual raw wage differential. In order to determine the actual observed wage gap we apply a simple Tobit model. By estimating the following equation for each establishment, we can directly derive the wage differential between male and female employees:

,

ij j j ij ij

w =α +γ fem +μ (2.2)

where α is an absolute term measuring the average wage rate in establishment j, fem is a dummy variable reflecting the gender of individual i and μ_ij denotes the error term. The estimated coefficient γˆ then represents the raw GWG in establishment _j j (Gap1j) taking into account that wij is censored from above.

From an economic viewpoint the wage gap which is due to differences in occupational skills shall be deemed to be justified and comprehensible. Therefore, we calculate a second measure of the gender pay differential which is adjusted by the difference in human capital of employees:

(

)

2 1 m m m f

j j j ij j ij

Gap =Gap − β X −β X (2.3)

ij

X includes mean characteristics of the individuals i at establishment j and m j βˆ is a vector of estimated coefficients – derived from wage regressions – of the individual characteristics Xij of male employees in establishment j. Hence, Gap2 reflects the difference in the rewards for individual human capital characteristics and unobserved wage effects between male and female employees within each establishment j. The calculation of this measure requires the estimation of wage equations for male

employees only.22 In order to allow for the heterogeneity and complexity of the wage setting process we estimate – as far as possible – separate wage equations for each establishment:

100 .

m m m m

ij j ij ij

w =β X +ε for establishments with at least male employees (2.4) The dependent variable describes the daily log wage rate. We restrict the wage equation to a standard Mincer equation aiming to adjust the observed wage rate by differences in human capital endowments between men and women. Since other possible wage determinants, such as the occupational status and the occupational group are determined by human capital and presumably also by other firm characteristics, we exclude them from our wage equation. Hence, m

ij

X includes potential experience (squares), dummy variables for different education levels and job tenure.23 The right- censoring of the dependent variable again requires the estimation of a Tobit model. In order to make sure that our firm-specific wage estimations yield statistically meaningful results, we only take into account establishments with at least 100 male employees. This procedure is most suitable to take into account the heterogeneity among establishments. The benefit of this approach is only feasible, however, at the expense of the number of considered establishments. To exploit information from establishments with less than 100 male employees, we run pooled regressions for all establishments with between twenty and ninety-nine male employees:

100 .

m m m m

ij ij ij

w =β X +ε for establishments with less than employees (2.5) Given the results of equation (2.4) and (2.5) respectively, we can calculate Gap2 which describes the GWG within establishments assuming that men had the same human capital endowment as women within an establishment. Note, however, that part of the differences in characteristics may be explained by inequality in access to and encouragement of education. Furthermore, there might be a discriminating element in the selection of employees such that observed characteristics of employees as well as estimated coefficients are not distributed randomly across establishments.24

22 _{Alternatively, one may use female wages and characteristics to determine the remuneration of human}

capital. Given that the regression of male wages are unlikely to be biased due to selection problems and that men are less concerned with discrimination, we argue that male wage coefficients better represent the market value of selected qualification characteristics.

23 _{Note that the inclusion of firm effects or industry-level variables is not required in this specification}

because we run firm-specific wage regressions and hence identification would not be feasible.

24 _{In order to correct for this selection we would have to estimate employment probabilities (Datta Gupta}

1993). Due to the lack of information on the household context and the individual background, it is difficult to implement this procedure which requires convincing exclusion restrictions.

Using these two measures for the firm-specific wage differential as dependent variables allows us to analyze the effect of firm characteristics and institutional framework on wage inequality within establishments. To ensure the significance of our results, both indicators are only calculated for establishments with at least 20 male and 20 female employees.

, 1,2.

j j j

GapK =δZ +ε K = (2.6)

The observed wage gap (Gap1) as well as the GWG which is adjusted for the difference in human capital characteristics (Gap2) is assumed to depend on the vector Zj including firm characteristics and information on the institutional framework of establishment j. δ captures the impact of the corresponding explanatory variables, derived from the theories expounded in Section 2.2. To investigate the hypotheses based on Becker’s discrimination model, we use the establishment size, the relative establishment size within the sector and the export quota. Implications from the bargaining model are tested by variables such as “application of collective wage agreements” and “existence of a works council”. To determine whether the naive notion of collective bargaining holds, i.e. that unions aim to increase wages at the lower tail of the wage distribution irrespective of sex, we add the female share of union members in the relevant union to vector Zj in equation (2.6). In order to test whether the works council acts in favor of the majority of the workforce, we interact the existence of a works council with the female share in the establishment. As well as the variables attributable to we also use specific control variables such as region and industry.

In this second estimation step we can exploit the panel structure of the data by applying a random effects model. In the first estimation step, that is the wage estimation, it is not possible to apply fixed-effects panel estimation in a Tobit model, because most of our human capital variables are time invariant. Even if it would be straightforward to apply a random effects Tobit model, we currently refrain from this approach because of computer time restrictions. Since the variance of the calculated gender wage gap varies by establishment size per definition, we calculate robust standard errors accommodating heteroscedasty in the dependent variable.

A general issue in assessing the effect of firm characteristics is the firm self-selection into works council status or the adoption of collective agreements. In our setting, this problem becomes relevant if employees of establishments with high gender wage gaps were more or less likely to implement works council or follow collective agreements.

According to Addison et al. (2006), however, endogeneity of works councils does not seem to be important for Germany.