Are low-skill public sector workers really overpaid? A quasi-differenced panel data analysis

Abstract

Public–private sectoral wage differentials have been studied extensively using quantile regression techniques. These typically find large public sector premiums at the bottom of the wage distribution. This may imply that low skill workers are ‘overpaid’, prompting concerns over efficiency. We note several other potential explanations for this result and explicitly test whether the premium varies with skill, using Australian data. We use a quasi-differenced Generalized Method of Moments (GMM) panel data model which has not been previously applied to this topic, internationally. Unlike other available methods, this technique identifies sectoral differences in returns to unobserved skill. It also facilitates a decomposition of the wage gap into components explained by differences in returns to all (observed and unobserved) skills and by differences in their stock. We find no evidence to suggest that the premium varies with skill. One interpretation is that the compressed wage profile of the public sector induces the best workers (on unobserved skills) to join the public sector in low wage occupations, vice versa in high wage occupations. We also estimate the average public sector premium to be 6% for women and statistically insignificant (4%) for men.

Keywords:

• public sector
• wages
• quasi-differenced panel data
• GMM
• Australia

JEL Classification::

• J45
• J31
• J38

Notes

¹ Some have used father's sector of employment as an exclusion restriction (e.g. Hartog and Oosterbeek, 1993; Hou, 1993; Terrell, 1993; Dustmann and van Soest, 1998; Bender, 2003; Melly, 2006). Such estimates are biased if intergenerationally transmitted attitudes to public sector employment are correlated with intergenerationally transmitted (unobserved) skills. Others have used attitudes towards unions (e.g. Bender, 2003; Heitmeuller, 2006; Melly, 2006), which are likely to be endogenous to working in a unionized environment. Some use parent's education (Hartog and Oosterbeek, 1993; Hou, 1993), which is also likely to be correlated with unobserved skills. Others have used age (Kanellopoulos, 1997; Borland et al., 1998). But age is correlated with risk aversion (Pålsson, 1996; Halek and Eisenhauer, 2001), which may be rewarded differently in the two sectors (Gregory and Borland, 1999).

² See also Gibbons et al. (2005) who use a similar approach in the context of industry wage models.

³ Analysis was conducted using SAS V9 and Stata V11.

⁴ In linear IV, instrument relevance can be determined by testing the significance of the instrument(s) in the first stage regression. This is not the case for nonlinear GMM (see Stock et al., 2002).

⁵ It is acknowledged that the effect of a 1 year increase in experience may differ across the experience distribution, as reflected by the standard practice of including experience in quadratic form in wage equations (Mincer, 1974; Preston, 1997). It would be possible to include experience in quadratic form in the wage equation here. This is not pursued for a number of reasons. First, such an inclusion would make the interpretation of ψ more difficult. In the preferred model, ψ facilitates a simple assessment of whether differences in returns to skills differ between sectors. Second, the nonlinearity in returns to experience would only be identified through an increase in 1 year of experience for each employee. To reiterate the nature of this restriction, it assumes that returns to the last single year of experience do not vary across experience levels. However, there is no restriction to the functional form of returns to all previous years of experience. This restriction is thus unlikely to be of any substantive consequence.

⁶ Author's calculations are from the SEAS Expanded Confidentialised Unit Record File. The percentage contribution was calculated by the author for each employee as total employer contributions divided by usual weekly income from main job. The sample was restricted to employees, excluding employees of own business. People with more than one job were excluded as the employer contribution variable does not differentiate between jobs. At the time of the survey, the minimum legislated employer contribution was 8%. Employees with monthly income below $450 per month are exempt, as are those under 18 years of age working less than 30 hours per week. Thus it is reasonable for the average contribution to be less than 8%.

⁷ Employees employed by their own business at either observation were excluded.

⁸ Public sector employees are those who identified their employer as a ‘Government business enterprise or commercial statutory authority’ or ‘Other governmental organization’.

⁹ There are, however, a number of other possible explanations. It may result from reporting errors in the change in employer variable, since this relies on retrospective recall. It is also possible for employees to change sector without changing employer. This is the case when a public corporation is privatized. In any case, the conservative approach is taken here, by limiting the sample to employees who reported a change in employer.

¹⁰ Current work schedule is self-reported. Shift work is defined as any schedule other than a ‘regular daytime schedule’. Most employees classified as shift workers reported ‘a rotating shift (changes from days to evenings to nights)’; an ‘irregular schedule’; or a ‘regular evening schedule’.

¹¹ The results of the decomposition are a function of the estimated coefficients and the sample means. The SEs of the decomposition take account of the variance–covariance matrix of the estimated parameter vector. They also take account of the SEs on the sample means. They also account for the fact that the estimated mean time invariant characteristics of workers in each sector ( and ) are functions of the estimated parameters and the sample means.

¹² The decompositions were estimated using the user-written Stata module –Oaxaca– (Jann, 2008).

¹³ Industry dummies are not included due to the heavy industrial segregation of public sector employment.

¹⁴ The fixed effects models were estimated using the user-written Stata module –xtivreg2– (Schaffer, 2005).