This new paper responds to the recent CSO study and also tackles several new dimensions of this question.
Paper here.
Abstract:
This paper provides a sub-sectoral analysis of changes in the public-private sector pay gap in Ireland between 2003 and 2006. We find that between March 2003 and October 2006 the public sector pay premium increased from 14 to 26 per cent and that there was substantial variation between subsectors of the public service. Within the public service the premium in 2006 was highest in Education and Security Services and lowest in the Civil Service and Local Authorities. In the private sector the pay penalty in 2006, relative to the public sector, was most severe in Hotels & Restaurants and in Wholesale & Retail and least severe in Financial Intermediation and Construction. The paper tests for the sensitivity of the pay gap estimates using a matching framework, which provides a stronger emphasis on job content, and finds the results to be broadly comparable to OLS. Finally, the study highlights the problems associated with controlling for organisational size in any study of the public-private pay gap in Ireland.
8 replies on “The Public-Private Sector Pay Gap in Ireland: What Lies Beneath?”
Interesting paper but this data is now three years old. I wonder where the trend has gone now post pension-levy. Second, I’m not entirely convinced of their OLS model. They acknowledge the difference in education between the public and private sectors and yet while the raw pay gap moved from 16 to 25%, the adjusted gap went from 14 to 26%. Something doesn’t seem right.
I wonder should a Dublin factor be included among the explanatory variables? Presumably the public sector is disproportionately Dublin based, and presumably there is a higher cost of living/wage equilibrium in Dublin too?
Another factor is the gender gap in the private sector which is not as big in the public sector. I think the headline figures hide a lot of important detail. The biggest gap was in education. I fail to see how they accounted for the fact that there is no private sector equivalent of a professor. The statistical model is cute but not validated properly.
James Conran says ‘….presumably the public sector is disproportionately Dublin-based…’.
This is often presumed, but does not appear to be true – the civil service proper tends to be Dublin-based, although less so than used to be the case, but the civil service is only about 10% of the total public service, which is mostly in Health and Education, and well spread around the country. Ditto Gardai, army, prisons, Dept Ag staff, Rev Commissioners, local authorities, many quangoes.
Table 12, Volume 7 of the 2006 Census gives the following figs (ILO definitions) for the sum of Public Administration, Health, Education, as % of total employed, by region.
State 21.8%
Border 23.7%
Dublin 22.0%
MidEast 20.6%
Midlands 22.3%
MidWest 21.1%
SouthEast 20.8%
SouthWest 21.0%
West 22.9%
There is very little variation really, and Dub close to State, but the three categories given are not precisely co-terminous with public sector – they exceed it in aggregate. So far as I know, there are no published regional breakdowns available for ‘public servants’, but I reckon there cannot be much regional variation.
As noted by the authors themselves, the primary application of propensity score modeling within labour economics, is the evaluation of labour market programmes. These are typically pseudo-experimental (observational) studies exploiting the power of longitudinal microdata. Through necessity, this paper must rely on two cross sectional waves of the National Employment Survey.
The motivation for matching arises from the absence of a valid counterfactual against which to evaluate the outcomes of those who receive a “treatment” (in the example of labour market programmes retraining is a treatment). The goal of a matching exercise is to artificially construct a control group who are identical in all significant respects to the treated group – achieving conditional mean independence (CIA).
To meet the assumptions required for matching, the outcome must be conditionally mean independent of treatment, conditional on the propensity score P(X). The vector of covariates used to estimate propensity scores should include all variables jointly affecting both participation and outcome.
In the programme evaluation literature a correctly specified selection model (the determinants of participation) is critical, as a number of factors may influence peoples’ decision to accept or shirk a treatment. The extent of covariates which ought to be included in modelling selection is not universally agreed upon – “Imbens”:http://www.barcelonagse.eu/Causal_Inference.html advised a room of aspiring labour economists at last year’s Barcelona GSE to put the kitchen sink [my words…] into selection in a stepwise process, whereas “Pearl”:http://ftp.cs.ucla.edu/pub/stat_ser/r350.pdf appears to favour rigid parsimony (ht: Brendan Halpin). No mechanical algorithm exists for choosing these covariates, though evidence has shown the choice of covariates makes a substantial difference to the performance of the propensity score[1].
Pairing cases on their estimated propensity scores is relatively easy — but matching algorithms also seek to achieve balance in the underlying covariates that informed the selection model.
bq. “Thus, where important occupational differences exist within a particular component of the public sector, this will be accounted for within the PSM framework and a like-with-like comparison will be made.” p. 18
p. Putatively yes, but how professions that are almost exclusively associated with certain sectors can be matched[2] across treatment and control groups is not clear to me. The authors refer to this problem on p. 7, but only in the context of achieving balance within “security services” employment across groups. I would expect very weak common support in all occupational codes where a single sector dominates employment. I am curious how the authors met balancing conditions for occupations with no obvious correspondence between sectors?
Without the benefit of microdata, I can only make some assumptions from the frequencies reported in Table A2. Of these, 15 of the 26 occupational categories could potentially suffer from weak common support:
* Legislators & Senior Officials
* Managers of Small Enterprise
* Engineering & Science Professionals
* Life Science & Health Professionals
* Engineering & Science Associate Professionals
* Life Science & Health Associate Professionals
* Customer Service Clerks
* Skilled Agricultural & Fishery Workers
* Extraction & Building Trades Workers
* Precision, Handicraft & Related Trades Workers
* Stationary Plant & Related Operators
* Machine Operators & Assemblers
* Sales & Services Elementary Occupations
* Agricultural, Fishery & Related Labourers
* Mining, Construction, Manufacturing & Transport Labourers
Hopefully the authors will elaborate on common support in a subsequent draft.
In any case, matching on cross-sectional data is not without pitfalls and estimates of propensity scores are likely to suffer from greater bias than those derived from longitudinal data. Systematic differences may remain between treatment and control groups even after conditioning on observables, see Smith & Todd.
As difference-in-differences matching estimators generally perform better than their cross-sectional cousins, it would be interesting to see if a longitudinal panel such as EU-SILC could contribute anything to the current -debate- hysteria surrounding public sector pay.
fn1. Smith, J. and P. Todd (2005), “Does matching overcome LaLonde’s critique of nonexperimental estimators?”, Journal of Econometrics 125, 305-353.
fn2. Without replacement.
Two things stand out about the public sector – there are a lot of women working in it and there are a whole lot of people working in education and health. The paper finds a male premium and a premium for education and health. However we’re left a bit in the dark about the interaction of variables. The paper looks at a gender variable and then goes on to do male and female regressions separately. I think we’d learn a whole lot more by looking at the interactions. For instance in table 3 I’d like to see not just intercept but also slope estimates for gender. This is such a simple thing to do – why haven’t they done it?
Another thing that I find interesting is that premium in education is greater than the premium in health – a priori I’d have expected the opposite as medical doctors are really well paid in Ireland. This makes me suspect that the matching is not ideal. I’d love to see what the premium for public sector workers looks like if doctors are excluded.
@juggernaut
“a priori I’d have expected the opposite as medical doctors are really well paid in Ireland”
And surely those doctors in private practice tend to earn even more? Thus reducing the premium over the entire health sector.
FYI, the dominant providers of private specialist care in the Ireland are Consultants in the public not the private sector. For example, 90% of cardiologists in Ireland are on the public payroll but earn the mega bucks from private practice. Part of the reason is that not only are do public consultants control the bulk of private practice but also also that you (the tax payer) subsidise that private income in a myriad of ways including by paying medical insurance (through the State Claims Agency which insures the private practice of a ‘public’ consultant on a public hospital site) as opposed to a Consultant in private practice who has to pay all their all own costs.
Although the Revenue Commissioners figures show that this profiteering in hospital medicine in Ireland comes from the public sector doctors, the actions of the HSE, the DOHC, and even the Competition Authority over decades have revealed no interest in addressing this cartel. Disappointing but hardly surprising given the recurrent and systematic failure of Official Ireland since Independence most recently outlined by Dan O’Brien.