“Determinants and Characteristics of temporary employment in Europe”
Anna Cristina D’Addio₤ Michael Rosholm®$£
In this study we focus on two issues. First, we investigate whether the labour market is segmented in various
categories, into different clusters. Second, we model labour market outcomes, in terms of contract types held by
workers at the different survey years’, in order to identify those characteristics that are most likely to affect them.
To tackle the first question, we use cluster analysis. To analyze empirically the second issue, we use discrete choice
models for nominal dependent variables in a panel data setting. More particularly we use a mixed logit that allows
for correlations between alternatives over time. The data used are extracted from the waves 1995-1999 of the
European Household Community panel. We distinguish three job statutes: “permanent contracts” – PE -, “fixed-
term contracts” – FC - and “other working arrangements” – OWA- (including casual jobs and the residual
category of other working arrangements). While the reference class will be represented by the most “secure”
contract (e.g. permanent job), the others will be accordingly those going from the lower level to higher level of
stability in the class of temporary contracts. To capture gender differences, the estimations have been carried out
separately on the sample of men and women. The results show that important differences exist at cluster level and
further across genders. Three clusters have been identified, and the transition rates computed for each cluster
show that workers in “low quality employment” clusters experience higher risk of labour market exclusion. The
estimation of the multinomial logit model with unobserved heterogeneity shows that the latter is very significant
for both genders and for any cluster considered. A striking result is the one associated with wages. It seems indeed
that temporary jobs always are always rewarded with lower wages. Other interesting findings are those associated
with the macro economic indicators that suggest that temporary jobs are prevalently used to face sudden change
in the workload or recession periods.
JEL codes: J24, J42, J6, C33, C35
Keywords: mixed logit, panel data, cluster analysis, temporary jobs, unobserved heterogeneity
₤ HIVA (Higher Institute for Labour Studies), Katholieke Universiteit Leuven and CAM (Center for Applied
Microeconometrics), University of Copenhagen. Corresponding author: e-mail: firstname.lastname@example.org
® Department of Economics, University of Aarhus, and IZA
$ The first author acknowledges the Department of Economics of the University of Aarhus and the Aarhus
Business School for providing research facilities while visiting them.
£ This study has been financed by the European Commission, DG Employment and Social Affairs within the
framework of the contract for the statistical support to Employment in Europe (2003), “Determinants of career
stability, job quality and labour market transitions”. We gratefully acknowledge EUROSTAT for the access to the
European Community Household Panel as well as to the New Cronos database. We would like to thank
particularly Frank Siebern-Thomas, Alfonso Arpaia, Georg Fischer, Stefano Galgliarducci as well the participants
in the various meetings at the European Commission for their helpful comments. Views expressed represent
exclusively the positions of the authors and do not necessarily correspond to those of the European Commission.
The usual disclaimer applies.
In this study we focus on two issues. First, we investigate whether the labour market is
segmented in various categories, into different clusters. Second, we model labour market
outcomes, in terms of contract types held by workers at the different survey years’, in order to
identify those characteristics that are most likely to affect them.
To tackle the first question, we use cluster analysis, i.e. a procedure consisting in searching,
among a large number of individual characteristics, those dimensions being able to be
interpreted as defining components of a dual labour market composed of “primary” and
“secondary” workers. This method allows us to include in the analysis a large number of
determinants without imposing a priori a structure on our data.
To analyze empirically the second issue, we use discrete choice models for nominal
dependent variables in a panel data setting.
The inequalities on the labour market are very significant and take many forms. Most
outstanding is income inequality. A frequent measurement of this inequality is 90-10 percentile
ratio. This measure varies rather strongly between the countries, according to the institutions
and the taste for inequality in society. It is highest in the United States where it reaches 4.5, it is
3.4 in the United Kingdom, 2.3 in Belgium and 2 in Sweden. Moreover, in the Anglo-Saxon
countries, wage inequality increased in the Eighties and Nineties. In the labour markets of
other OECD countries, the increase in wage dispersion was more limited, but sometimes
accompanied by increasing unemployment. Flexible work arrangements may also contribute to
wage inequalities, since they are often rewarded with lower wages1.
In this study we investigate whether the heterogeneity present in income and labour market
attachment can be grouped in segments, which present common determinants. Such an
approach, sometimes called “dualism”, makes it possible to refine the economic analysis of the
labour market and rests thus on the existence of a primary market, composed of stable
workers, attached to the labour market, well trained, having experience, seniority, often
receiving vocational training.
A few exceptions aside (see Dickens and Lang 1985, 1988), the traditional econometric
approaches to the labour market can only with difficulty apprehend this dimension of dualism.
We propose here a very general method of description of dualism or, more generally,
1 Attempts have been made to link the low-pay dimensions of job with the perceived low quality of work, with
one report suggesting that “policies towards low-wage jobs should centre on their quality at least as importantly as
on the level of pay which they provide” (see Salverda et al, 2001).
polarization of the labour force in more or less homogeneous segments. Our approach is
initially developed, without any a priori assumptions on the forms segmentation, to let the data
speak by extracting a certain number of common determinants among those characteristics of
the individuals in the labour force that can explain the heterogeneity. By selecting a set of
variables, we can determine a set of homogeneous sub-groups (clusters) for the initial variables.
This set of variables is selected among the usual variables affecting labour market participation
(age, sex, education) and those related to job (e.g. wages, working hours). We exclude the type
of contract that we use subsequently to characterize the clusters. Later, we describe these
dimensions while giving them a more intuitive interpretation.
The different labour market outcomes experienced by workers in flexible arrangements
relative to those in regular full-time jobs may result from the nature of the arrangements
themselves. Alternatively, they may stem from differences in the average personal and job
characteristics of individuals in those arrangements.
Various factors may lead to different labour outcomes. Besides employment protection,
which is often stronger for workers in permanent employment2, differences in abilities,
preferences and tastes may result in different behaviour for workers hired under a stable
(“permanent”) contract compared to those in flexible work agreements. For instance,
temporary workers may not be willing to invest more in specific human capital, and this may
lead to lower wages.
The nature of jobs included in the category “temporary employment” is quite heterogeneous
and may be associated with higher or lower skill qualification needs3. High turnover in a firm
may be linked to the wish of employers to adjust rapidly to labour market conditions and this
can lead to alternating paths of short jobs and unemployment spells (especially for low skilled
workers). Irrespective of their skills, individuals may be willing to accept a temporary job with
a low wage, when this would increase the probability of getting a stable job in the future in the
same firm. Still, some workers may prefer to be hired on temporary jobs in order to preserve
their independence and mobility.
Owing to the intrinsic short duration of most of temporary jobs, individuals are likely to
move quite often out and in of them. Many studies focusing on temporary jobs have ignored
the longitudinal character of the data, when modelling the different contracts hold by workers
2 See for instance Bentolila and Bertola (1990); Bentolila and Saint-Gilles (1994); Booth (1997).
3 See Farber (1999).
at the different surveys’ dates4. Using panel data it is possible to distinguish different scenarios
by exploiting the dynamics implied by temporary jobs.
Based on the groups of workers obtained through the cluster analysis, we investigate
therefore the characteristics of workers hired under different job contracts. To do this, we use
discrete choice methods for polychotomous dependent variables. Namely, we use a multi-
period multinomial logit model of job “choices” that accounts for unobserved time-invariant
individual heterogeneity. We distinguish three job statutes: “permanent contracts” – PE -,
“fixed-term contracts” – FC - and “other working arrangements” – OWA- (including casual
jobs and the residual category of other working arrangements). While the reference class will
be represented by the most “secure” contract (e.g. permanent job), the others will be
accordingly those going from the lower level to higher level of stability in the class of
For comparison purposes, the standard (pooled) multinomial logit is also presented. The
multi-period multinomial logit model (with unobserved heterogeneity normally distributed)
allows us to exploit the longitudinal component of the ECHP, and therefore studying the
subsequent labour market status of individuals holding flexible work arrangements compared
to those holding regular full-time positions at each survey date. One of the advantages of the
model, also known as “mixed logit”, is that unlike the standard multinomial logit (MNL)
model it allows for non-zero correlations of the error terms for different alternatives. If these
correlations are important, the model will yield more accurate estimates of the reduced form
determinants of labour market sector outcomes than the standard MNL.
The data used in this study are extracted from the European Community Households Panel
1995-1999. We have carried out the cluster analysis using the data about year 1995 in that it
was the first year in which people have declared the types of the contract held. For the analysis
of contracts’ determinants we have used data from 1995 onwards. Since the clusters have been
identified in 1995, Sweden, Finland and - owing to missing variables introduced in the analysis,
Luxembourg - have been excluded. Estimations have been conducted on the remaining set of
countries adopting the appropriate weights. To capture gender differences, the estimations
have been carried out separately on the sample of men and women.
4 See the vol. 112 of the Economic Journal devoted to a Symposium on Temporary work. See also Booth et al.
The results show that important differences exist at cluster level and further between
genders. Three clusters have been identified, and the transition rates computed for each
cluster show that workers in “low quality employment” clusters experience higher risk of
labour market exclusion. The estimation of the multinomial logit model with unobserved
heterogeneity shows that the latter is very significant for both genders and for any cluster
considered. A striking result is the one associated with wages. It seems indeed that temporary
jobs always are always rewarded with lower wages. This of course has an important bear on
the life of workers that hold such flexible working arrangements, not only for the present but
also for their future career prospects and their right to a “decent” life. Other interesting
findings are those associated with the macro economic indicators that suggest that temporary
jobs are prevalently used to face sudden change in the workload or recession periods.
The study is structured as follows. In the next section we describe the qualitative and
quantitative methods used. Section 3 is devoted to the description of the data used for the
cluster analysis as well as to the discussion of the results gathered through it. Transition rates
into unemployment, employment and inactivity at cluster level during the observation period
are also presented there. The data and results about the analysis of workers’ outcomes are
described in Section 4. In Section 5, we draw some conclusions.
2.1 Qualitative analysis: Clustering
It is in particular the view of temporary contracts either as “dead end” jobs or “stepping-
stones” that drives us to think of theories of segmentation in the labour market. Job
classification schemes have a long history in the literature on labour markets. Over a century
ago, Cairnes observed "What we find, in effect, is not a whole population competing
indiscriminately for all occupations, but a series of industrial layers." He identified four non-
competing groups in the English economy of the late 19th century: unskilled workers, artisans,
"producers and dealers of a higher order such as engineers and opticians," and "the learned
professions and the higher branches of mercantile business" (Dunlop, 1988). In the early post
war period, Dunlop, Ross, Livernash, Kerr and others analysed labour markets in terms of
"wage contours," "orbits of coercive comparison," and "job clusters." But since the 1960's, by
far the most influential conception has been that of a "dual economy", which distinguished
"core" from "peripheral" firms, and dual labour markets, in which "primary" jobs are
distinguished from "secondary" jobs on the basis of earnings, working conditions, job
advancement, work rules and employment stability. In the 1970's, dual labour markets were
explained by reference to a "dual economy," consisting of "core" and "periphery" sectors that
are differentiated by firm size, capital intensity, and the extent of monopoly rents (Bluestone,
1970; Harrison, 1972; Edwards, 1979). A more complex explanation for segmentation is found
in the strand of the literature that develops from Doeringer and Piore's (1971) work on internal
labour markets, which, in Rosenberg's (1989) words, locates the sources of segmentation in
"the interactions between technology, training, product demand and social class".
The micro-foundations of segmentation were advanced in the 1980's by tying dual labour
market theory to efficiency wage models, in which firms may be able to increase worker
productivity by paying high wages: primary labour markets are those in which this high wage
strategy prevails (Akerlof and Yellen, 1986; Bulow and Summers, 1986). The "primary" labour
market is often subdivided into upper and lower tiers, since, as Piore (1975) notes, "upper-tier
work seems to offer much greater variety and room for individual creativity and initiative, and
greater economic security." Piore suggested that this tripartite scheme of "independent
primary," "subordinate primary" and "secondary" segments might need to be amended to
distinguish both craft and routine white collar jobs. Gordon, Edwards and Reich (1982)
emphasized employment relations as another key source of segmentation and identify control
workers (supervisory jobs) and public sector jobs as distinct job segments.
According to job segmentation theory most flexible work arrangements are located in the
secondary segment of the labour market owing mainly to their low stability when compared to
those arising from “permanent work arrangements”.
In order to characterize individuals in different job-states or work arrangements, we have
carried out a cluster analysis using the method of k-means as illustrated below.
Cluster analysis is an appropriate method to analyse this question for two reasons. First, it
considers the full set of potentially relevant determinants without predetermining which
variables are of relevance or not. Second, it allows an easy identification of the key variables
determining the type of job: if the various clusters differ significantly with respect to some
variable, then this variable is to be considered a crucial determinant of the respective job type;
if, on the other hand, there is no significant difference in some variable across clusters, this
variable cannot be considered a distinctive job type feature.
The purpose of the methods known as “clustering methods” is mainly to group the
individuals in a restricted number of homogeneous classes. The aim is therefore that of
describing the data while carrying out a classification of the individuals. This classification is
automatic in the sense that the classes are obtained through means of a formalized algorithm
and not by subjective or visual methods. One distinguishes two wide types of classification
1. The non-hierarchical methods which directly produce a partition in a fixed number of
2. The hierarchical methods which produce continuations of partitions in increasingly vast
The method used here is that of non-hierarchical classification. It consists in grouping N
individuals in K classes so that the individuals of the same class are as similar as possible and
the classes are quite distinct. This supposes the definition of a global criterion measuring the
proximity of the individuals of the same class and thus the quality of the partition.
If one can consider the individuals as points of an Euclidean space, the problem of
“clustering” can be described as the methods that looks for a partition of a “cloud” (the data
points) formed by N points, trying to transform it in K disjoints subsets containing data
points so as to minimize the sum-of-squares criterion
j n S
n x is a vector representing the nth data point and
µ is the geometric centroid of the
data points in
j S .
Let us call I1, I2,…., IK the inertia (variance) of each cluster, calculated with respect to the
respective centre of gravity µ1, µ2,…, µK. The sum of these inertias is called intra-cluster
inertia and is noted Iw: (or within variance)
It is thus desirable that Iw is very small to have a set of very homogeneous classes. The
dispersion of the set of K gravity centers µ1, µ2,…, µK around µ, the centre of gravity of the
total cloud of N individuals, is called inter-class inertia and is noted IB, i.e. “between inertia”:
IB = ∑ P j d2(µi ; µ)
where P j is the sum of the weights of the individuals belonging to the J class, and d2(µi ; µ) is a
distance measure. A large value of IB indicates a good separation of the classes and it will thus
be appropriate that IB is as large as possible.
By the theorem of Huygens, the total inertia, I, of the cloud of N points is equal to the sum of
the intra-cluster inertia and the inter-cluster inertia, i.e.:
I = Iw + IB.
Thus, to find the maximum of IB is equivalent to find the minimum of Iw since their sum is a
constant. From the point of view of the inertia, it will be enough to characterize the best
possible partitions in K classes making the within variance (i.e. the inertia intra-class) Iw
minimal. In the extreme situation, there is of course only one (type of) individual(s) in each
cluster. In this case the within variance is zero, since each point is the centre of gravity of its
cluster. Such an extreme situation is, however, completely useless. We rather seek from now
on to obtain a partition in K clusters. The technique employed is that of the "k-means", also
called method of iterative reallocation around some mobile centers. It is a fast and powerful
iterative method. The principle is as follows (for N individuals and K desired classes): one
creates randomly K classes the center of which is formed by K individuals drawn among the N
available; one assigns an individual to the cluster to which the "central individual" is closest for
him. One calculates then the “centroids” of the K classes and reallocates the N points to the
cluster whose centroid is closest. These two steps are alternated until a stopping criterion is
met, i.e., when there is no further change in the assignment of the data points. The
disadvantage of this method is that the final distribution of the individuals depends on initial
pulling. This is why the data-processing program offers the possibility to repeat this
procedure several times5. The software will retain for the table of the results, the repetition
which maximizes the between-inertia. That criterion amounts to wanting that the classes are
as distant as possible, while ensuring an optimal homogeneity inside them.
2.2 Econometric models for nominal dependent variables
To investigate the determinants of the workers’ outcomes at each survey’s date, we use
discrete choice models for panel data. However for comparison purposes, the first model we
estimate is a standard multinomial logit pooled over time. Under assumption of random
utility, we suppose that each choice is rewarded with a specific utility as follows
V X 'βε
where i=1,…,N indexes the individuals and j=1,…,J the job types. The conditionally
independent utilities are Type I Extreme Value – distributed. An individual makes a specific
5 We have used 250 repetitions.
“choice” concerning the type of contract, by comparing the utilities with which each choice is
Prob[ individual i makes choice j] = Prob[ V(i,j) > V(i,k) for all k not j]
Defining the conditional choice probability and taking expectations of this w.r.t.
ijε yields the
unconditional choice probability
exp[ 'x( , )]
exp[ 'x( , )]
We estimate this model by maximum likelihood, pooling individuals over time. The log-
likelihood function is in this case simply equal to the sum over the individuals of the logarithm
of the unconditional choice probabilities.
The second model we estimate is a reduced form multi-period multinomial logit with random
effects. Particularly, it is called “mixed logit” (see Revelt and Train, 1998; Revelt and Train,
1999) because the choice probabilities combine the logit specification with a different
distributional assumption for the heterogeneity terms (normality in most cases).
Utility of individual i in job-type j (permanent, fixed-term, others) at time t is expressed as
where Xit is a vector of explanatory variables including individual and job-related
characteristics as well as two macro-economic indicators. The εijt are time-varying i.i.d error
terms while αij is an individual and job-type specific, time-invariant random effect. If the εijt
follow the Type I extreme value distribution, the probability of being employed in job type j at
time t (conditional on Xit and the random effects) has a multinomial logit form:
P(j t| X ,)
For identification purposes, αi1 and β β1 are normalized to zero; that is, we make permanent job
the base choice. The αij are assumed to follow a multivariate normal distribution. Namely, we
specify the vector α αi = (αi2,…,αiJ)′ to be linear combinations of J-1 independent N(0,1)
αAη , η ~ N(0,I
In the previous expression, A is a J-1 x J-1 lower triangular matrix, estimated along with the
β β’s. The covariance matrix for the heterogeneity terms α αi is then AA′ ′. If the random effects
were observed, the contribution to the likelihood of individual i employed in job-types 1 and 2
would be equal to the sequence of multinomial logit probabilities
i itii ti
L(jj P(j| X ,α )P(j | X,α )
However, the α αi s are not observed. This implies that to obtain the unconditional likelihoods,
one needs to integrate the conditional likelihoods over all possible values of η ηi (hence of α αi).
L(η )f(η )dηdη
This implies a multi-dimensional integration of dimension J-1. In our case J-1=2. In general,
the above integral has no closed form. However it can be approximated through simulation.
To do so, we need to draw R values of the η ηi for each individual. In doing this, we use Halton
draws6. The procedure allows building Halton sequences (see Train, 1999; Bhat, 1999) and has
been shown both to be able to reduce the draws needed for estimation as well as to reduce the
simulation error associated with a given number of draws (see also Greene, 2003). The
likelihood conditional on each set of values is finally calculated. We replace the integral by the
average of the R conditional likelihoods:
(6) is a smooth function of the parameters and it makes the use of simulated maximum
likelihood relatively straightforward. The simulator is consistent, however its accuracy has
been shown to depend on the number of replications and draws (Brownstone and Train
1999). The model has been estimated using 350 draws and 150 replications.
It should be noticed that the mixed logit nests the standard logit as a special case (α αi=0), so it
is possible to compare the two statistically using a likelihood ratio test. Moreover, unlike a
standard multinomial logit (whether estimated on a single or pooled cross sections), the error
terms of the utility functions for different choices are not assumed to be independent. The
6 Halton sequences are used to create a series of draws that are distributed evenly across the domain of the
distribution to be integrated. Halton sequences are created by selecting a number h that defines the sequence
(where h is a prime number) and dividing a unit interval into h equal parts. The dividing points on this unit
interval become the first (h-1) elements in the Halton sequence. Each of the h sub-portions of the unit interval is
divided as the entire unit interval was and these elements are added on to the end of the sequence. This process is
continued until the desired number of elements in the sequence is reached. Halton sequences result in a far more
even distribution of points across the unit interval than random draws.
latter assumption leads, in the standard multinomial logit model, to the property of
Independence of Irrelevant Alternatives (IIA). In the random effects model the IIA property
does not hold. If we denote the composite error term for alternative j as (suppressing the
individual i subscript) µjt = αj + εjt, we have cov(µjt, µkt) = E( [αj + εjt][αk + εkt]) = σ2αj,αk. This
implies that the unobserved portions of utility for alternatives j and k are related through the
correlation of their heterogeneity terms. The random effects specification offers therefore
advantages beyond the usual gains in efficiency associated with modelling an error
components structure over the standard MNL, by providing greater flexibility in the pattern of
3. Cluster analysis
In cluster analysis the outcome of the procedure provides the analyst with a number of classes
built around some observable variables. Specifically, we have focused on those related to job
quality and stability taking as a starting point the definition given in the report Employment in
The variables used relate both to job status and individual characteristics in 1995 (the first year
for which the question about the nature of the contract held was asked) such as:
1. occupational category
2. firm size
3. wage and working hours
4. education and training level
5. age category
Other variables like the sector of employment, the job status and the contract type have been
used subsequently to characterize the clusters. Table 1 here below reports some descriptive
statistics for the individuals in the sample. In it we have distinguished the various dimensions
of job quality according to the EU, Employment Report (2002) definition.
[Table 1 to be inserted here]
The results of the cluster analysis at EU level (as well as those concerning each Member state)
for men and women are reported in Tables 2 and 3.
[Tables 2 and 3 to be inserted here]
In Tables 4 and 5 we present the number (and shares) of individuals belonging to each cluster
at the EU level as well for each Member State.
[Tables 4 and 5 to be inserted here]
Contract type and access to training, appear to be distinctive job type features. Some further
distinctions should however be underlined, since the results for men and women differ
significantly. For both genders, we have identified three main clusters.
For men, the three clusters obtained are the following:
1. a cluster of relatively high paid and high skilled full-time employees on permanent contracts
- many of them in supervisory and intermediate positions - who are predominantly working in
high skilled, non-manual occupations or as clerks in larger firms in the public and service
sectors (this cluster will be referred to as “high quality employment cluster” in the sequel);
2. a cluster of, on average, younger and relatively low skilled full-time employees, many of
which are in low paid and temporary job arrangements (fixed-term, short-term or casual work
contracts) in non-supervisory functions without access to training, working mainly in skilled
manual or unskilled occupations – and, in the case of Belgium, Denmark and Austria also as
service workers - in small- and medium-sized private sector firms in agriculture or industry
(this cluster will be referred to as “low quality employment cluster” in the sequel); and
3. a very small, heterogeneous cluster of, in general, younger temporary and part-time
employees in manual, service or elementary occupations in small firms or self-employment,
mostly in the public and service sectors; contrary to the much larger “” cluster above,
employees in this cluster have relatively high access to training and, notably in the case of
Germany, Austria, Spain and Portugal, above average educational attainment levels (this
cluster will be referred to as “part-time employment cluster” in the sequel).
For women, the situation is somewhat different, with the following three clusters obtained:
1. a cluster of high skilled women in supervisory or intermediate positions and high paid
permanent employment with access to training, working in non-manual, skilled occupations in
the private sector; except for the UK, Ireland, France, Spain and Portugal, this cluster
comprises not only full-time employees, but also women in part-time employment with the
above characteristics (this cluster will be referred to as “high quality private sector
employment” in the sequel);
2. a cluster of relatively younger, high skilled, high paid women in non-manual skilled
occupations in the public sector with relatively high access to training; the common feature of
this cluster across all Member States is work in the public sector, while the remaining job
characteristics differ considerably: women in this cluster are somewhat more often in part-
time employment (except in Belgium, the Netherlands, Austria and Portugal), in temporary
employment and in non-supervisory positions (except in Portugal and the UK) (this cluster
will be referred to as “public sector employment” in the sequel); and
3. a large cluster of low skilled women in low paid, fixed-term contracts or casual employment
without access to training, in manual, low skilled or unskilled occupations, mainly in small
private sector firms in industry (this cluster will be referred to as “low quality private sector
employment” in the sequel).
As shown by the distinct job types for men and women described above, contract type and
access to training are important criteria of job types' classification among both men and
women. This is not the case for variables such as working time, firm size or work in the public
sector. While men in part-time employment are a category very different from full-time
employed men - independently of other personal or job characteristics - part-time work is a
common feature across all of the clusters for women. Firm size is also more important for the
classification of male employment. On the other hand, public sector employment – while
being related to job classifications for both men and women - is a much stronger determinant
of job quality for women, defining almost entirely one of the three clusters among women.
The above identification of clusters is certainly of interest in its own respect. For the purpose
of the analysis of quality in work and labour market transitions, however, it is further
important to check whether the labour market transitions also differ across the various
clusters or not. Such variation clearly is to be expected if the above dimensions of job quality,
and in particular contract type and access to training, also impact on employment stability and
career prospects in the longer run.
To be able to study the characteristics of individuals belonging to the cluster over time we
have carried out the cluster analysis for the year 1995. This allows us to study the evolution
over time of clusters in terms of transition rates between different contract types.
Transition rates at cluster level. As can be seen from the transition rates into
unemployment and inactivity by cluster at Member State level for the period 1995-1999,
reported in Table 6 following, there are in fact very significant differences in the longer term
employment performance of the above clusters.
[Table 6 to be inserted here]
In particular, transition rates into unemployment were twice as high in the “low quality
employment clusters” (cluster 2 for men and cluster 3 for women) compared with the “high
quality employment clusters” (cluster 1 for men and clusters 1 and 2 for women). For men,
the lowest employment stability was observed in the “part-time employment cluster” (cluster
3). Not only were transitions for this cluster into unemployment four times higher than for the
“high quality employment cluster”, but also transitions into inactivity were much more
The strongest impact of low quality employment on labour force attachment, though, was
found for women. Among the women in the “low quality private sector employment cluster”
(cluster 3) in 1995, almost one in five had moved into inactivity by 1999, compared to around
10% only in the two other clusters for which transitions into unemployment and inactivity
were actually similar. It should be noted, however, that the “public sector employment
cluster” includes a significantly higher share of fixed-term and temporary employees, which
seems to offset the higher employment stability of permanent employment in the public
sector, most notably in Belgium, France, Denmark, Germany, the Netherlands, Italy and
Portugal – all countries with significantly higher transition rates into unemployment for
women in cluster 2. In France, Italy and Spain, more than 20% of women in cluster 2 were
not employed five years later. Moreover, in all countries except the UK, 20% (or more) of all
women in the "low quality private sector employment cluster" (cluster 3) in 1995 had lost or
left their employment by 1999 – more than one third in Denmark, Germany and Spain.
4. Analysis of contract types’ determinants
To analyse the determinants of workers’ outcomes we have used data extracted from the
ECHP, waves 1995-1999 for the individuals identified in the cluster analysis presented in the
previous section. This means that estimations have been performed separately on each cluster
over time and further separately between men and women aged 16-64. The number of
individuals present in each cluster is reported in the tables along with the estimation results.
15 Download full-text
We have used some variables describing broadly labour market and economic conditions,
e.g. the unemployment and growth rates, besides socio-demographic and job-related
characteristics. Descriptive statistics and charts illustrating their trend over the observation
period, at the EU-level and for each Member State, are presented in Table 7 and Charts 1 and
2 here below.
[Table 7 to be inserted here]
[Charts 1 and 2 to be inserted here]
By using the occupation and industry distribution of the jobs we could have gathered some
further insights about the different contracts’ types. However, sectors of occupation as well as
characteristics describing overall job functions (supervisory, intermediate and non-
supervisory) are missing for Germany. We have therefore decided to exclude them from the
analysis. Estimations have been conducted on the set of countries remaining after dealing with
missing values and missing variables7 and the appropriate weights have been applied. Various
explanatory variables have been introduced, namely
A) Socio-demographic characteristics:
a. Age class as follows:
i. Age 16-24
ii. Age 25-34
iii. Age 35-44 (the reference)
iv. Age 45-54
v. Age 55-64
b. Education levels
i. Higher education
ii. Secondary education
iii. Primary education (the reference)
c. Cohabiting / Married
d. Children less than 12
7 Finland, Sweden and Luxembourg have been excluded from the analysis.