TOWARDS A FORMAL PROFILING MODEL TO FOSTER ACTIVE LABOUR James Obben

Document technical information

Format pdf
Size 210.8 kB
First found May 22, 2018

Document content analysis

Category Also themed
Language
English
Type
not defined
Concepts
no text concepts found

Persons

Organizations

Places

Transcript

TOWARDS A FORMAL PROFILING MODEL TO FOSTER ACTIVE LABOUR
MARKET POLICIES IN NEW ZEALAND
James Obben1
Department of Applied and International Economics
Massey University
Palmerston North
New Zealand
ABSTRACT
Following the unemployment hysteresis of the 1980s, discussions of
methods for reducing the natural rate of unemployment tend to focus
on long-term unemployment (LTU). A broad consensus exists among
economists for efforts to move towards active labour market policies
or ALMPs that increase employment opportunities. The premium is
on accurate and early identification of jobseekers at risk of LTU (an
activity known as profiling) so that action on referral to an
appropriate ALMP can be taken as early as possible in the
unemployment spell. Evidence shows that statistical-model based
profiling is more accurate than profiling based on other methods. To
advance a statistical profiling model for New Zealand, this paper has
attempted to fit a logit model to more than 2.2 million completed
unemployment spells. Despite the prevailing data limitations, the
estimated coefficients have the expected signs and form the most
plausible basis yet for estimating the probability an identified
jobseeker has of suffering LTU.
JEL Classification: J64
1. INTRODUCTION
Much as it is decried, some unemployment is necessary for any market economy to function
properly (Layard et al. 1991, p. xiii). Thus a lot of attention in the literature has been directed at
measuring, describing and explaining unemployment, especially its equilibrium level – the level at
which inflation stabilises.2 Ever since the introduction of the natural rate hypothesis (NRH) and the
expectations-augmented Phillips curve in seminal papers by Phelps (1967) and Friedman (1968),
the concept of the ‘natural’ rate of unemployment or the NAIRU (the non-accelerating inflation rate
of unemployment) has dominated the discussions of equilibrium unemployment.3 In its simplest
version, the NRH posits that in the absence of supply shocks unemployment will deviate from the
1
I wish to thank Andy Reynolds who made it possible for me to obtain the data. Ted Drawneek provided invaluable
computing assistance in editing the data for analysis. Comments on an earlier draft by Jim Alvey and S. Shakur were
helpful and are appreciated. Any remaining weaknesses or errors are my responsibility
2
Nickell (1990) surveys the vast literature on unemployment theories.
3
See the volume edited by Cross (1995).
1
natural or equilibrium rate only if expectations about inflation are wrong (i.e., inflation expectations
error has a negative effect on unemployment levels). The natural rate or NAIRU is an attractor (but
might also be an “attractee”) of the actual rate of unemployment. The main policy implication is
that if unemployment is pushed below the NAIRU, inflation pressure builds up; and if
unemployment stays above the NAIRU, disinflation ensues. It is only at the NAIRU that there is
no tendency for inflation to either increase or decrease. The problem is, estimation of the NAIRU
is quite imprecise (Staiger et al., 1997) and there is controversy as to its practical usefulness
(Galbraith, 1997). That notwithstanding, estimates of the time-varying NAIRU (Gordon, 1997) are
watched with keen interest by country administrations and central banks (Stiglitz, 1997 and
Richardson et al., 2001). If estimates of the NAIRU are lower than the prevailing unemployment
rate, it may be inferred that expansionary policies may be pursued without igniting inflation; if the
natural rate rises and it is used as a serious “speed limit” to stop the reduction of unemployment in
an anti-inflation policy, then relatively high unemployment rates may be interpreted as being
normal or natural.
Unemployment in all the major industrial economies rose steadily from the mid-1970’s, stayed high
in the 1980’s and 1990’s in most European countries (even when inflation stabilised) but fell in the
US, Japan and a few other countries (Martin, 1994). Since the NRH argues that the unemployment
rate over long periods cannot move too far from the natural rate, the persistently high observed
unemployment meant the natural rate rose. To help explain this phenomenon, some economists
have resorted to the theory of “hysteresis” – the notion that long-lasting recessions leave relatively
permanent effects on the time path of the economy.
There are a number of ways in which unemployment hysteresis might occur. In the view of Layard
et al. (1991) a key to understanding this is the emergence of long-term unemployment4 (LTU) that
accompanied the high unemployment in Europe. A deep recession displacing a lot of workers for
extended periods means the proportion of long-term unemployed workers among the unemployed
increases. If the unemployment benefit scheme is quite generous, there is little incentive for
jobseekers to accept low-paying jobs and might prefer remaining unemployed for a long time. If
the long-term unemployed become demoralised and search less effectively or lose their skills,
employers may not find them employable and may in fact discriminate against hiring the long-term
unemployed applicants (Nickell, 1990, p. 420). If wage setting is largely through union bargaining,
then outsiders cannot exert any downward pressure on wages and thereby bring about an increase
in employment: insiders will pursue high real wages for themselves leaving outsiders to remain
unemployed. Unemployment (which is the same as excess supply of labour) exerts downward
pressure on wages. However, the higher the proportion of long-term unemployed, the less effective
is unemployment in holding down wages and helping to lift employment. Thus discussions of
methods for reducing the natural rate tend to focus on LTU.
In focussing on LTU to reduce the natural rate, there is a broad consensus among economists for
efforts to move away from the “passive” help given through unemployment benefits to “active
labour market policies” or ALMPs that increase employment opportunities. Three mechanisms are
identified here: (i) job-search assistance or placement and counselling services, (ii) education and
training, and (iii) direct job creation and recruitment subsidies (Layard et al., 1991; Katz, 1994).
4
In New Zealand, long-term unemployment is defined as continuous unemployment spell lasting 27 weeks or more in
contradistinction to the OECD definition that uses 52 weeks.
2
The details and relative merits of these are discussed in the references above. If these policies are
targeted successfully and in a cost-effective manner at the long-term unemployed or those
jobseekers at risk of being long-term unemployed, it should be possible then to remove more
people from the unemployment register sooner rather than later. Certainly the premium is on
accurate and early identification of jobseekers at risk of LTU (an activity known as “profiling”) so
that action on referral to an appropriate ALMP can be taken as early as possible in the
unemployment spell.
Profiling may be based on either (a) statistical models estimated to determine the probability of an
unemployment insurance claimant exhausting their benefit [as in the US] or to identify those
jobseekers who have the highest risk of LTU [as in Australia], or (b) “characteristics screening” –
sorting jobseekers according to whether or not they belong to groups with known unemployment
handicaps, or (c) the judgement of the public-employment-service staff.
Even though the
predictive power of statistical models is relatively low, evidence from Australia and the United
States show that model-based profiling is more accurate than profiling based on judgement of
public-employment-service staff and on characteristics screening (OECD Proceedings, 1998, p.16).
New Zealand, along with the other OECD countries, experienced increases in unemployment and
LTU in the 1980’s and 1990’s (see Table 1). Indeed, during the early 1990’s New Zealand policy
makers (Employment Task Force, 1994) identified the growth of LTU as the single most worrying
feature of the country’s recent economic and social performance (see the incidence of LTU in New
Zealand during the 1987-99 period summarised in Tables 2 and 3). In response to these
developments some New Zealand economists directed research effort towards analysing the New
Zealand Employment Service (NZES) dataset. Using survival analysis techniques and censored
data (i.e., a mixture of completed and uncompleted unemployment spells), the objectives of those
studies have been to determine the extent to which the length of unemployment duration can be
predicted from the characteristics of the unemployed individuals (e.g., Watson et al., 1997) and to
determine the impact that certain characteristics and unemployment duration have on the
probability of someone exiting the unemployment register (e.g., Gardiner, 1995). Unlike in
Australia where formal profiling using a statistical model to identify at-risk jobseekers for early
intervention was initiated in 1994, in New Zealand assistance is provided to registered individuals
when they are clearly identified as long-term unemployed. Some concerns about the timing of
assistance and early identification of at-risk jobseekers had been raised by the Prime Ministerial
Task Force on Unemployment (1994) but to date no formal profiling is done.
To advance a statistical profiling model for New Zealand, this paper extends a previous work
(Obben et al., 2002) that estimated a logit model based on a “smaller” random sample of 100,000
completed unemployment spells (i.e., uncensored data) from the NZES database traversing the
1988-97 period to cover more than 2.2 million observations available for the period. Earlier studies
modelling unemployment duration in New Zealand had used censored data but, as Salant (1977)
has noted, parameters based on incomplete spells suffer from interruption bias and length bias (or
the censoring problem). Utilising completed spells or uncensored data thus skirts the objection
raised by Salant and also allows all the cases to be neatly categorised either as being long-term
unemployed or otherwise, thus making it possible to apply binary choice models as an alternative
approach. Barron and Mellow (1981) and Heikensten (1984) are two examples of studies that used
the binary choice approach to estimate re-employment probability.
3
Table 1: Incidence of Long-Term Unemployment (as Percentage of Total Unemployment) in Surveyed Selected OECD
Countries 1988-99
Country
Australia
≥6mths
≥12mths
Austria
≥6mths
≥12mths
Belgium
≥6mths
≥12mths
Canada
≥6mths
≥12mths
Czech Rep
≥6mths
≥12mths
Denmark
≥6mths
≥12mths
Finland
≥6mths
≥12mths
France
≥6mths
≥12mths
Germany
≥6mths
≥12mths
Greece
≥6mths
≥12mths
Hungary
≥6mths
≥12mths
Iceland
≥6mths
≥12mths
Ireland
≥6mths
≥12mths
Italy
≥6mths
≥12mths
Japan
≥6mths
≥12mths
Korea
≥6mths
≥12mths
Luxembourg
≥6mths
≥12mths
Mexico
≥6mths
≥12mths
Netherlands
≥6mths
≥12mths
New Zealand
≥6mths
≥12mths
Norway
≥6mths
≥12mths
Poland
≥6mths
≥12mths
Portugal
≥6mths
≥12mths
Spain
≥6mths
≥12mths
Sweden
≥6mths
≥12mths
Switzerland
≥6mths
≥12mths
Turkey
≥6mths
≥12mths
United Kingdom ≥6mths
≥12mths
United States
≥6mths
≥12mths
Total OECD
≥6mths
≥12mths
1988
47.4
28.4
1989
40.6
23.0
1990
41.0
21.6
1991
49.6
24.9
1992
58.7
34.5
1993
57.1
36.5
1994
41.0
36.4
90.0
77.5
20.7
7.4
87.5
76.3
20.8
6.8
81.4
68.7
20.2
7.2
77.9
62.9
22.5
6.5
74.7
59.0
26.1
9.4
70.4
52.9
28.6
11.4
51.7
28.7
…
…
64.6
44.8
65.0
46.7
71.3
48.1
50.8
25.9
23.6
6.9
63.7
43.9
66.7
49.0
73.5
52.4
53.2
29.9
32.6
9.2
55.5
38.0
64.7
46.8
71.9
49.8
54.4
31.9
32.6
9.2
58.0
37.2
54.1
31.5
71.5
47.6
49.9
26.9
…
…
58.1
36.1
55.4
33.5
70.4
49.7
45.5
25.2
52.8
30.6
58.2
34.2
60.1
40.3
71.0
50.9
77.6
61.6
84.3
68.1
38.1
17.9
77.4
58.9
69.8
58.2
36.2
15.9
76.9
59.1
76.5
57.7
34.4
17.2
47.4
26.3
36.7
13.3
62.2
32.4
75.2
58.3
28.4
12.5
40.9
21.6
54.0
32.1
52.8
30.6
61.7
38.3
63.8
44.3
72.8
50.5
62.6
41.3
31.4
14.3
80.7
64.3
79.5
61.5
36.1
17.5
20.6
5.4
54.7
29.6
60.3
46.1
39.0
21.2
39.1
20.2
76.8
43.9
53.2
31.9
41.1
23.5
79.1
52.3
52.5
33.2
45.6
27.2
58.4
38.7
68.4
51.1
17.6
4.2
27.5
17.0
37.5
31.0
66.1
47.4
25.9
8.3
37.8
19.6
45.2
43.4
69.6
50.1
32.0
10.9
47.5
20.2
47.2
28.8
13.0
6.3
57.3
35.4
20.6
11.2
62.9
42.5
20.4
11.7
82.2
66.0
85.9
69.0
40.5
20.2
82.4
67.3
58.7
70.4
37.3
18.7
57.7
42.3
57.1
38.1
65.6
50.0
29.4
11.3
15.9
6.3
66.1
49.9
34.5
14.7
29.5
11.6
67.0
51.2
75.0
61.5
21.2
8.2
66.6
48.3
72.7
58.5
18.4
6.5
61.5
44.7
12.1
7.4
57.2
40.8
9.9
5.7
13.6
6.7
81.0
66.0
85.2
69.8
39.0
19.1
13.9
2.6
66.7
42.9
63.6
49.3
39.5
20.9
40.8
20.4
62.8
39.0
62.4
44.8
70.2
54.0
22.2
12.1
26.2
16.4
72.6
47.0
50.3
34.4
10.0
5.5
44.6
30.9
Source: OECD Employment Outlook, various issues.
4
77.5
49.4
50.0
32.2
43.8
28.9
65.1
40.3
57.2
43.4
73.4
56.1
38.7
17.3
49.3
27.6
68.5
45.4
63.4
45.4
20.3
12.2
52.4
35.2
1995
51.3
30.8
42.8
25.6
77.7
62.4
27.8
14.1
52.5
30.6
46.6
27.9
56.6
37.6
64.0
42.3
65.9
48.7
72.7
51.4
73.0
50.6
33.3
16.8
77.9
61.4
80.2
63.6
37.7
18.1
16.9
3.4
49.5
23.8
7.9
1.5
80.4
46.8
43.3
25.5
39.1
24.1
63.1
40.0
65.1
50.9
72.8
56.9
45.6
27.8
50.4
33.6
60.3
36.3
60.8
43.6
17.3
9.7
49.8
33.9
1996
48.5
28.4
42.5
25.6
77.3
61.3
29.3
16.7
52.3
31.3
44.4
26.5
55.5
34.5
61.5
39.5
65.3
47.8
74.7
56.7
75.2
54.4
31.2
19.8
75.7
59.5
80.8
65.6
40.4
20.2
16.0
3.9
44.6
27.6
9.8
2.2
81.8
50.0
36.6
20.7
31.1
16.0
62.2
38.0
66.7
53.1
72.2
55.7
48.4
30.1
52.1
25.0
65.9
43.5
58.1
39.8
17.5
9.5
50.3
34.2
1997
51.4
30.7
47.7
28.7
77.2
60.5
26.9
16.1
53.0
30.5
45.7
27.2
48.6
29.8
63.7
41.2
68.5
50.1
76.5
55.7
73.5
51.3
27.0
16.3
73.6
57.0
81.8
66.3
41.3
21.8
15.8
2.6
61.1
34.6
6.9
1.8
80.4
49.1
36.4
19.4
26.1
12.0
60.4
37.4
66.7
55.6
71.8
55.5
50.8
33.4
49.4
28.5
62.7
41.6
54.8
38.6
15.9
8.7
50.8
34.9
1998
52.2
33.6
43.3
29.2
76.3
61.7
24.1
13.7
54.6
31.2
41.4
26.9
42.2
27.5
64.2
44.1
69.6
52.6
74.8
54.9
71.0
49.8
22.9
16.1
…
…
77.3
59.6
39.3
20.9
14.7
1.6
55.2
31.3
3.3
0.9
83.5
47.9
37.9
19.4
20.5
8.2
…
…
64.5
44.7
70.4
54.1
49.2
33.5
48.9
34.8
60.3
40.0
47.3
32.7
14.1
8.0
48.3
33.1
1999
48.4
29.4
47.6
31.7
73.5
60.5
21.4
11.6
61.9
37.1
38.5
20.5
46.4
29.6
55.5
40.3
67.2
51.7
…
…
70.4
49.5
20.2
11.7
…
…
77.2
61.4
44.5
22.4
18.6
3.8
53.8
32.3
6.8
1.7
80.7
43.5
39.0
20.8
16.2
6.8
…
…
63.8
41.2
67.9
51.3
…
…
61.0
39.8
60.6
33.7
45.7
29.8
12.3
6.8
46.2
31.2
Table 2: Percentage Share of Long-term Unemployment1 in the Total Unemployment of Various
Demographic Groups in New Zealand, 1987-1999
Group
Age Group
15-19
20-24
25-34
35-44
45-54
+55
Gender
Male
Female
Ethnic Group
European
Maori
Pacific Isld
Other
Education Quals.
No quals.
Sch. quals.
Post-sch.
Other quals
Occupation
Sought 2
MP&T3
Clerks
S&SW 4
A&FW 5
Trades 6
P-MO&A 7
Elem. Occ8
Not specifd
Overall
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
17.4
14.7
18.2
21.4
30.0
…
19.6
21.2
20.6
25.3
30.4
39.3
21.0
31.2
31.0
30.2
32.6
41.3
24.2
30.6
33.1
35.1
45.5
43.4
26.2
32.2
37.3
41.2
45.1
46.9
34.8
42.0
44.3
46.9
52.8
56.9
37.4
46.9
50.4
54.7
60.0
65.9
34.1
45.0
49.0
52.7
60.8
65.8
28.2
39.0
43.3
48.8
58.8
62.7
23.7
27.4
35.7
43.5
50.4
60.0
20.2
30.2
31.4
37.3
48.3
47.1
20.2
27.8
32.0
40.5
46.6
47.9
23.1
29.4
31.0
42.4
45.9
52.2
27.1
22.2
31.5
23.7
37.7
28.7
42.7
30.4
43.8
33.9
52.1
40.1
58.2
48.3
57.0
45.8
53.8
40.7
44.7
35.0
40.9
31.6
39.4
31.9
42.3
33.3
17.6
23.4
…
…
20.8
26.8
…
…
27.6
32.7
32.4
28.2
31.9
31.7
41.3
31.1
31.8
43.7
46.2
34.3
39.4
51.9
57.4
41.2
46.6
53.5
68.6
45.1
45.3
52.8
58.6
52.2
40.1
48.3
56.7
41.5
33.7
37.0
50.5
39.7
30.7
36.8
43.3
33.3
29.7
36.7
39.8
40.0
32.0
40.2
39.1
40.2
22.7
13.6
22.0
13.6
28.0
16.1
24.1
16.3
33.4
24.1
31.5
23.2
39.5
26.7
36.5
18.5
45.1
26.8
41.9
25.7
54.1
34.5
47.3
35.1
58.4
41.9
58.3
40.9
56.7
39.1
58.0
39.9
52.9
33.9
52.7
35.3
44.5
28.4
42.6
28.9
41.9
23.0
36.6
29.3
43.4
23.8
35.8
26.8
42.7
29.4
41.1
28.3
35.7
36.6
39.1
39.6
46.9
47.3
52.4
57.5
41.5
43.9
44.1
50.6
44.9
50.0
58.6
59.4
60.3
43.8
49.9
43.5
46.9
41.9
49.5
54.8
55.0
57.3
42.5
48.5
36.4
45.4
34.9
44.9
60.3
53.4
51.7
38.9
43.7
34.1
38.2
27.6
44.0
41.8
40.6
48.8
33.8
36.3
28.1
30.8
29.4
41.3
37.5
42.9
43.5
31.1
33.3
29.8
28.6
29.8
40.7
33.3
43.1
40.9
31.8
33.2
32.4
33.1
28.8
41.8
38.9
43.1
44.0
34.5
35.1
19.1
22.3
29.0
32.7
1
The long-term unemployed are those people who have been unemployed for 27 weeks or longer.
NZSCO: The New Zealand Standard Classification of Occupations of 1968 was revised in 1990, causing a
discontinuity in the series in 1990/91.
3
MP&T = Managers, Professionals and Technicians.
4
S&SW = Services and Sales Workers.
5
A&FW = Agriculture and Fishery Workers.
6
Trades = Trade Workers.
7
P-MO&A = Plant and Machine Operators and Assemblers.
8
Elem. Occ. = Elementary Occupations.
2
Source: Statistics New Zealand, Labour Market Statistics 1999.
5
Table 3: Percentage Decomposition of Those Who Constitute the Long-term Unemployed1 in New
Zealand, 1987-1999
Variable
Age Group
15-19
20-24
25-34
35-44
45-54
+55
Gender
Male
Female
Ethnic
Group
European
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
30.3
13.1
23.8
14.8
12.3
…
25.8
20.0
22.6
14.8
11.0
7.1
18.3
22.2
26.1
16.9
9.9
6.7
17.2
19.3
26.2
16.1
15.0
6.3
16.3
17.8
27.3
19.4
13.2
6.2
15.4
20.8
26.9
18.1
13.4
5.4
13.7
19.1
27.0
19.6
14.0
6.7
13.0
18.2
25.7
20.2
16.3
6.6
12.1
17.0
24.4
21.4
17.6
7.4
14.2
13.4
21.6
24.9
17.7
8.2
12.6
15.2
22.8
21.3
19.2
8.7
11.9
16.2
22.6
23.8
17.8
8.1
11.9
14.7
20.4
24.9
18.2
9.8
57.4
42.6
63.9
36.1
66.2
33.8
66.5
33.5
66.0
34.0
66.4
33.6
64.6
35.4
64.0
36.0
65.2
34.8
61.4
38.6
61.7
38.3
61.5
38.5
62.5
37.5
62.3
63.2
64.4
65.4
57.7
61.3
59.8
56.9
51.8
54.6
29.5
25.8
22.9
19.9
24.9
22.0
23.2
25.0
24.1
25.7
25.6
Pacific Isld
…
…
8.1
10.4
12.8
12.8
12.5
13.4
11.8
11.2
9.2
Other
…
…
3.9
3.8
4.6
3.8
4.6
4.7
57.7
(69.0)
22.6
(14.5)
11.7
(5.6)
7.7
(10.9)
56.4
Maori
58.4
(79.1)
25.6
(9.7)
11.1
(3.8)
4.7
(7.4)
7.6
11.4
10.8
Education
Quals.
No quals.
64.8
63.9
59.5
59.4
52.9
50.3
47.9
48.2
48.5
48.8
48.6
48.9
44.0
Sch. quals. 15.6
17.4
19.7
19.3
19.8
20.0
20.4
20.2
19.5
18.9
17.8
18.1
20.7
Post-sch.
9.0
9.0
9.9
11.4
11.8
11.3
11.8
13.3
13.5
12.2
11.8
11.4
12.7
Other quals
9.8
11.0
11.3
9.0
15.1
17.5
19.2
18.1
17.9
19.9
21.5
21.6
22.7
Occupation
Sought 2
MP&T3
8.4
9.4
9.2
8.3
10.9
10.0
11.9
11.9
Clerks
8.7
10.0
9.8
9.5
9.7
8.9
7.6
8.2
S&SW 4
10.8
12.1
11.8
9.3
10.4
11.8
12.1
12.1
A&FW 5
5.0
5.4
6.7
5.4
5.5
5.0
5.2
4.7
Trades 6
10.5
10.2
8.9
8.3
5.7
5.5
5.2
5.7
P-MO&A 7
7.1
6.8
7.9
6.9
6.5
7.1
6.7
6.3
Elem. Occ8
15.2
20.4
25.3
23.7
14.7
15.0
15.9
12.1
Not specifd
34.3
25.3
20.2
28.6
36.6
37.0
35.4
39.3
Note: Figures may not add to 100.0 because of rounding. Figures in parentheses for ethnic groups show percentage
shares in total population in the respective census years.
1
The long-term unemployed are those people who have been unemployed for 27 weeks or longer.
2
NZSCO: The New Zealand Standard Classification of Occupations of 1968 was revised in 1990 causing a
discontinuity in the series in 1990/91.
3
MP&T = Managers, Professionals and Technicians.
4
S&SW = Services and Sales Workers.
5
A&FW = Agriculture and Fishery Workers.
6
Trades = Trade Workers.
7
P-MO&A = Plant and Machine Operators and Assemblers.
8
Elem. Occ. = Elementary Occupations.
Source: Statistics New Zealand, Labour Market Statistics 1999.
6
Important modifications have been made in the empirical logit model in this extended work. For
example, age is modelled as a polychotomous rather than a continuous variable; the categories for
ethnicity and educational attainment have been expanded; annual dummy variables have been
utilised so that the full dataset can be analysed in one step; the geographical locations have been rearranged to coincide with political regional council areas; and distributed lags of the growth rate of
the economy have been successfully incorporated. Constrained by the set of available variables on
which information is gathered by the NZES and the characteristically low predictive power of the
statistical models, the parameters obtained in this study form a highly credible objective basis for
estimating the probability an identified jobseeker has of suffering LTU. The a posteriori success
rate of about 55% of the model in identifying LTU sufferers bodes well for the time when currently
unavailable but relevant information on jobseekers and local labour markets can be integrated in the
model.
The paper contains five sections. A short review of pertinent earlier studies is provided in Section
2. The dataset used and analytical method utilised are described in Section 3. Section 4 presents
and discusses the empirical results. The last section provides a summary and outlines the
conclusions from the study.
2. A REVIEW OF EMPIRICAL STUDIES ON UNEMPLOYMENT DURATION
2.1 General Introduction
In the literature, most empirical studies modelling the determinants of the length of out-of-work
duration time for unemployed persons are based on survival analysis. The data for survival
analysis arise in investigations where the subjects of study are followed until they reach a specified
endpoint which is often, but by no means always, death (Everitt and Der 1996 p. 93). In
unemployment duration models, survival time is the spell of unemployment, i.e., the time it takes
for a job search to end in a successful [re-]employment. Because survival times are restricted to
being positive, and because the data often contain censored observations (i.e., observations for
which at the end of the study the event has not occurred) the familiar multiple regression techniques
may not be justifiable. Central to the analysis of survival times are two functions that can be used
to describe their distribution, namely the survival function and the hazard function.
The survival function S(t) is defined as the probability that an individual survives longer than t, i.e.,
S(t) = Pr(T > t) where T denotes survival time. The graph of S(t) versus t is known as the survival
curve and it is generally negatively sloped. From such fitted curves, researchers can estimate the
probability that an individual with a given set of characteristics will survive past a given length of
time. On the other hand, the hazard function h(t) is defined as the probability that an individual
experiences the event of interest in a small time interval s, given that the individual has survived up
to the beginning of this interval. The hazard function is also known as the instantaneous failure
rate; it is a measure of how likely an individual is to experience an event at a particular point in
time. Empirically, it is assumed that the survival and hazard functions are adequately captured by
the Weibull distribution.
7
2.2 International Studies
Predicated on job search theory, most empirical research has been on the estimation of the hazard
function in which the probability is made a function of different exogenous variables that usually
include personal characteristics of those studied, conditions in the local labour market, the length of
spell of unemployment and economic incentives such as the magnitude of the unemployment
benefits (i.e., the replacement ratio). The parameters are usually estimated via the maximum
likelihood approach and, even though the particular functional form may vary, the most frequently
used specification is the proportional Weibull hazard function:
λ(t) = αtα-1 exβ ⋅ν
where t represents the length of the unemployment spell; α is the parameter used to capture the
effect of possible duration dependence; ν is a stochastic error component that is there to pick up
heterogeneity from omitted variables, x and β are vectors of explanatory/independent variables and
their coefficients, respectively (Holmlund et al., 1989, p. 175). Values of 0 < α < 1 would suggest
negative duration dependence; α = 1, no duration dependence; and α > 1, positive duration
dependence. Variance of the error term σ2 ≠ 0 would suggest a heterogeneous sample, whilst σ2 = 0
would suggest a fairly homogeneous sample.
Kaitz (1970) is one of the seminal papers on unemployment duration in the US. Lancaster (1979)
was the first to suggest and empirically use the Weibull hazard function in a UK study. Other early
studies include Nickell (1979a), Kiefer and Neuman (1979, 1981), Narenranathan et al. (1985) and
Albrecht et al. (1986). Some of the important country studies include Clarke and Summers (1979)
and Akerlof and Main (1980) for the US; Björklund (1981) for Sweden; Main (1981) and Nickell
(1979b) for the UK; Hassan and De Broucker (1982) for Canada; and Trivedi and Baker (1983) for
Australia. Good reviews of those studies have been done by Kiefer (1988), Holmlund et al. (1989)
and Layard et al. (1991, Ch. 5).
Almost all the studies report negative duration dependence (i.e., exit probability or the probability
of re-employment diminishes with length of unemployment spell) but the evidence on the effect of
benefits on unemployment duration is mixed. In their survey of empirical studies, Holmlund et al.
(1989) surmise that estimations of the benefit elasticity of unemployment duration based on time
series data are not robust; those based on cross sectional data report values ranging from 0.3 to 0.6.
Persons in age group 15-24 years are reported to have higher rates but shorter spells of
unemployment than other groups in the population. The European studies show that personal
characteristics (e.g., age, health, gender, marital status, and previous employment record) are more
important than the labour market situation and the replacement ratio for expected spell lengths
(Holmlund et al. 1989, pp. 209, 210 216).
The standard specification of the Weibull hazard function is not without flaws. For one thing,
theoretical and empirical problems with the sensitivity of ad hoc parameterisations have led some
writers to ponder what features of duration models can be identified non-parametrically. The
seminal paper on “non-parametric techniques” is Cox (1972) that shows how to recover β without
specifying the form of the baseline hazard, a process that does not allow the estimation of the
duration dependence. To address that weakness and other data problems, Heckman and Singer
8
(1984) and Han and Hausman (1990), among others, have proposed various alternative extensions
of Cox’s model. A drawback of Han and Hausman’s approach is the assumption of a prespecified
distribution for the error term about which economic theory offers little guidance. Anton et al.
(2001) have further extended Han and Hausman’s approach by adopting an alternative
semiparametric estimator of a proportional hazard model that has been proposed by Lewbel (2001),
albeit in a different context. Using duration data constructed from a panel sample of unemployed
men in Spain and focusing on the issues of heterogeneity and censoring, Anton et al. (2001) report
that (i) the effect of censoring in a parametric model is to change the signs and magnitudes of the
coefficients; (ii) heterogeneity is detected by using a censored ordered logit model with the
modification proposed by Han and Hausman (1990) but the signs and magnitudes of the
coefficients are not significantly affected; and (iii) if allowance is made for a wider class of
heterogeneity, as done in their semiparametric model, then changes in both signs and sizes of the
coefficients are observed.
2.3 New Zealand Studies
Among the empirical studies on New Zealand unemployment duration, Gardiner (1995) calculates
the exit probability function (EPF) using the NZES data covering the period July 1989 to June 1993
and analyses the EPF to see if its shape and/or level changed over time and whether individuals
with different characteristics have different exit probabilities. The EPF for 1993 is presented as a
representative EPF and examined in detail to ascertain the different exit rates for the different
demographic groups. The data contained censored observations and the approach taken in the exit
probability analysis is tantamount to estimating a hazard function to generate the probability that an
individual who has reached a specified duration will leave the unemployment register within the
following seven days. It is reported that the EPF declines as duration increases, implying that the
longer a person remains unemployed the less chance he or she has of leaving unemployment in the
following week. Or it could also be that certain individual characteristics underlie a lower exit
probability. The paper goes on to outline the disaggregated EPF according to gender, age group,
ethnicity, educational qualifications and regional location (the variables on which NZES collects
information). The conclusion, however, is that the model has low predictive power because of
omitted relevant explanatory variables such as income (both from a previous job and from sources
whilst unemployed), the number of dependents and local demand conditions.
Watson et al. (1997) take the “continuance curve” (or the survival curve) approach to investigate
the relationship between jobseeker characteristics and the duration of unemployment spells in New
Zealand. The dataset used comprised a little over 300,000 cases covering the period January 1995
to February 1997 and contained censored observations for which survival analysis is well-suited.
The authors of that paper report that the continuance curve approach does not allow accurate
identification of NZES jobseekers who are likely to become long-term unemployed on the basis of
the few key demographic characteristics that are measured when the jobseekers enroll with the
NZES. In predicting cases of LTU, the model always included a high proportion of short-term
spells. The paper recommends that a broader range of jobseeker and local labour market
information may improve the ability to model the duration of unemployment spells and to more
accurately identify jobseekers who are most at risk of LTU.
9
Moving away from survival analysis, Obben et al. (2002) fitted a binary choice model to a random
sample of 100,00 observations of completed unemployment spells covering a much longer period
(1988-97) thereby covering different phases of the business cycle. The cases for individual years
were analysed as well as data for the whole decade. The objective was to determine the probability
that an unemployed individual with a given set of attributes will be long-term unemployed. The
regression results made sense but the predictive power of the model was low.
3. ANALYTICAL METHOD AND DATA
The research objective is to estimate the extent to which the variables on which NZES collects
information about jobseekers – age, gender, ethnicity, educational qualifications and regional
location – are predictive of whether or not an unemployed individual would be long-term
unemployed. The model of choice is the logit model5 in which the dependent variable is a dummy
variable taking the value 1 if the spell of unemployment is 27 weeks or longer (i.e., a case of New
Zealand-defined LTU) or 52 weeks or longer (i.e., a case of OECD-defined LTU) and 0 otherwise.
Using the maximum likelihood estimation procedure, the discrete and dichotomous dependent
variable Yi is regressed on a set of discrete and/or continuous explanatory variables yielding a
model that can be written as
(1)
Yi = ln [Pi / (1 – Pi)] = â1 + â2 Xi2 + … + âk Xik
i = 1, …., N
where ln is natural logarithm, Pi is the probability that individual i is a long-term unemployed
person, the â’s are the regression coefficients and the X’s are the explanatory variables. Since the
logit model assumes that Yi is a logistic random variable, the dependent variable is the log of the
odds that an individual will be a long-term unemployed. A given slope coefficient shows how the
log of the odds (that an individual will be a long-term unemployed) changes as the corresponding
explanatory variable changes by one unit (in the case of a continuous or discrete variable), or as an
attribute different from that of the base category is considered (in the case of a dummy variable).
The statistical significance of the slope coefficients may be assessed from their respective standard
errors, t-ratios or p-values. A test of the hypothesis that all of the regression coefficients in the
model are zero can be done via the likelihood ratio test where the chi-square test statistic has k-1
degrees of freedom. In dichotomous dependent variable models, the conventionally computed
coefficient of determination (the R-square) is of questionable value as a measure of goodness of fit
(Gujarati, 1995, pp. 545-546, 561, 579). Consequently, several alternatives are suggested in the
literature. A number of these are reported with the analytical results in Section 4.
Furthermore, when the regression coefficients are exponentiated, the derived values or the antilogs
indicate the effect of each explanatory variable directly on the odds of being a long-term
unemployed rather than on the log-odds. Subtracting 1 from the antilogs and multiplying the
results by 100 would give the percentage changes in the odds corresponding to one-unit change in
the explanatory variables.
5
For the econometric theory of binary choice models see, for example, Intriligator (1978, pp. 173-176); and for the
rationale for the preference of the logit over the probit model see, for example, Pindyck and Rubinfeld (1991, p. 256),
Griffiths et al. (1993, p. 751) and Gujarati (1995, Ch. 16).
10
The probability that individual i would be long-term unemployed can be estimated from the
antilogs of Equation (1):
(2)
where
Pi / (1 – Pi) = antilog i
Pi = (1 – Pi) antilog i
= antilog i / (1 + antilog
i
i)
is the estimated value of the response variable from the regression for individual i.
The data for this study comprised all of the 2,476,898 cases of completed unemployment spells in
the NZES database covering the period from 1988 to 1997. For each spell (determined from the
start date and end date) the client number, age, gender, ethnicity, educational attainment and office
of registration were noted. In preparing the data for analysis, the observations were sorted
according to start date in order to note the number of cases in each year. Since there was no
interest in the identities of the registrants, the client numbers were ignored. Cases containing no
entry or meaningless entry or out-of-range values for any variable were deleted. All of the cases
for the year 1997 were deleted because they were found to be a sub-selection of spells that ended in
the same year; these were clearly incompatible with the others. After such corrections, the usable
dataset came to 2,229,125 cases, implying about 10% of the original data were lost due to mistakes
in them. Apart from the year 1988 which had 63,466 observations the other years had observations
ranging from 252,152 to 285,272. The following statistics in weeks were noted about the
unemployment spells: range: 0.14 to 477.14; mean: 31.03; median: 17.43; mode: 9.00; and standard
deviation: 40.79. The distribution of the unemployment spells is certainly skewed to the right.
For estimation purposes the values of the various variables were divided into different categories so
that dummy variables could be used. Unemployment spells 27 weeks or longer were given the
dummy variable LTU27 to reflect the New Zealand definition of LTU; those 52 weeks or longer
were given the dummy variable LTU52 to reflect the OECD definition of LTU. About 36% of the
total cases were LTU27s, and about 16% were LTU52s.
The age variable was divided into nine classifications: TEENAGER (15-19 age group);
EARLY20S (20-24 age group); LATE20S (25-29 age group); EARLY30S (30-34 age group);
LATE30S (35-39 age group); EARLY40S (40-44 age group); LATE40S (45-49 age group);
EARLY50S (50-54 age group); and BTWN5565 (55-65 age group). The base category for age was
TEENAGER.
For obvious reasons, there were only two classifications for gender: MALE and FEMALE; the
latter was used as the base category. The ratio of males to females in the dataset was 3:2.
The study employed six variables to match the six classifications of ethnicity reported in the NZES
database: EUROPEAN (European or white); MAORI (Maori); MAO_EURO (mixed MaoriEuropean); PACIFIC (Pacific Islander); MXDMAORI (mixed Maori-Pacific Islander); and
OTHETHNC (all other ethnic groups). EUROPEAN was used as the base category for ethnicity.
There were eight educational attainment categories that the jobseekers are classified into; the
corresponding variables are: LT3YRSCH (no formal school or less than three years of secondary
11
schooling); LT3SCERT (less than three School Certificate passes or equivalent); GE3SCERT
(three or more School Certificate passes or equivalent); SXTHFORM (Sixth Form Certificate,
University Entrance or equivalent); BURS_HSC (Bursary, High School Certificate); OSCHQUAL
(other school or trade qualifications); POSTSECQ (post-secondary qualifications); and
DEGPROFQ (university degree or professional qualifications). DEGPROFQ was utilized as the
base category of educational attainment.
To reduce the number of points of geographical location, the study identified the jobseeker as being
located in the regional council of the unemployment office or the city the jobseeker registered at.
The variables used to represent the thirteen regional councils are as follows: NRTHLAND
(Northland); AUCKLAND (Auckland); WAIKATO (Waikato); BOPLENTY (Bay of Plenty);
HAWKSBAY (Hawke’s Bay); TARANAKI (Taranaki); MANAWANG (Manawatu-Wanganui);
WELLNTON (Wellington); NELSON (Nelson); WESCOAST (West Coast); CANTBURY
(Canterbury); OTAGO (Otago); SOUTHLND (Southland). Wellington, the national capital, was
used as the base category.
4. EMPIRICAL RESULTS
Two different sets of the logit model were estimated. In the first set, Model 1, the NZES-based
variables plus dummy variables for the various years were regressed on LTU27 and on LTU52.
The annual dummy variables are meant to capture all the year-specific factors not reflected in the
other variables. This model is principally to furnish a posteriori probabilities for the registered
jobseekers and thereby provide a measure of the predictive power of the logit model. However, for
ex ante probabilities, which are more relevant for ALMPs, the results of Model 1 cannot be used to
estimate probabilities for either the current year or future years. To get around that problem, the
model was re-estimated with distributed lags of the growth rate of the economy replacing the
dummy variables for the individual years. Up to four lags were deemed adequate. The thinking is
that when appropriate sets of growth figures (actual or predicted) are substituted along with the
attributes of an identified jobseeker, the probability of LTU for that individual can be calculated for
out-of-sample years. The regression results and the figures for percentage change in odds obtained
for Model 1 and for Model 2 are reported in Table 4 and Table 5, respectively. The econometric
software SHAZAM (White, 1997) was utilised for the computations. Since the main interest is in
the implications of the New Zealand definition of LTU, the discussion will concentrate on the
LTU27 results; the implications of the OECD definition of LTU may be inferred from an
examination of the LTU52 figures and the discussion of the LTU27 results.
The LTU27 results in Table 4 show that only four (OTHRSCHQ, WAIKATO, OTAGO and DV94)
out of the 43 regressors are not statistically significant. Even though the sampled R-squares are
expectedly low, the chi-square score for the likelihood ratio test clearly leads to the rejection of the
joint test that all the regression coefficients are equal to zero.
12
Table 4
Regression Results of Model 1
LTU27
Variable
MALE
EARLY20S
LATE20S
EARLY30S
LATE30S
EARLY40S
LATE40S
EARLY50S
BTWN5565
MAORI
MAO_EURO
PACIFIC
MXDMAORI
OTHETHNC
LT3YRSCH
LT3SCERT
GE3SCERT
SXTHFORM
BURS_HSC
OTHRSCHQ
POSTSECQ
NRTHLAND
AUCKLAND
WAIKATO
BOPLENTY
HAWKSBAY
TARANAKI
MANAWANG
NELSON
WESCOAST
CANTBURY
OTAGO
SOUTHLND
DV89
DV90
DV91
DV92
DV93
DV94
DV95
DV96
CONSTANT
Summary statistics
Likelihood Ratio Test
2
Maddala R
2
Cragg-Uhler R
2
McFadden R
Prediction success rate
Coefficient
0.37645
0.21385
0.14525
0.14811
0.15264
0.16478
0.23701
0.32966
0.22267
0.26754
0.13096
0.21919
0.14785
0.10361
0.32183
0.16582
0.07469
0.05372
0.12773
0.00639
-0.03655
0.10653
-0.01561
-0.00393
0.01501
-0.07113
-0.06848
-0.04802
-0.07587
-0.07637
-0.10799
-0.00801
-0.19683
-0.03797
0.18174
0.44961
0.35578
0.18816
-0.01287
-0.08876
-0.21609
-1.27976
T-ratio
127.32
51.58
30.97
28.46
26.41
25.87
33.23
40.06
21.25
67.87
21.83
36.27
10.45
14.91
49.93
23.15
10.29
7.27
13.64
0.49
-4.43
12.99
-2.81
-0.58
2.16
-10.43
-7.51
-6.23
-9.49
-5.21
-16.93
-0.98
-20.85
-4.01
19.42
48.29
38.12
20.11
-1.36
-9.33
-22.55
-109.29
LTU52
% Change
in Odds
45.71
23.84
15.63
15.96
16.49
17.91
26.74
39.05
24.94
30.67
13.99
24.51
15.93
10.91
37.96
18.03
7.75
5.52
13.62
0.64
-3.58
11.24
-1.54
-0.39
1.51
-6.86
-6.62
-4.68
-7.31
-7.35
-10.23
-0.79
-17.86
-3.72
19.93
56.77
42.73
20.71
-1.28
-8.49
-19.43
-72.19
65991 with 41 d.f.
0.029
0.040
0.023
55.1%
Coefficient
0.43425
0.15259
0.16972
0.22495
0.25736
0.32356
0.43297
0.55622
0.06746
0.34095
0.16219
0.29207
0.20376
0.12418
0.50594
0.29341
0.16425
0.10006
0.16361
0.05336
0.03177
0.17303
0.03259
0.00864
0.01902
-0.24599
-0.09321
0.01249
-0.19707
-0.15785
-0.17273
-0.06451
-0.35755
-0.00114
0.07888
0.35641
0.09454
-0.14276
-0.38393
-0.47085
-0.97374
-2.40323
T-ratio
107.13
27.05
27.01
32.68
33.83
39.17
47.72
55.01
4.57
67.94
20.41
38.24
10.98
12.99
54.06
28.47
15.61
9.23
11.89
2.86
2.67
16.67
4.457
0.98
2.11
-26.72
-7.63
1.23
-17.91
-7.86
-19.93
-5.85
-27.26
-0.09
6.72
30.79
8.05
-12.01
-31.49
-38.19
-74.29
-153.56
81272 with 41 d.f.
0.036
0.062
0.042
74.5%
13
% Change
in Odds
54.38
16.48
18.49
25.22
29.35
38.2
54.18
74.41
6.97
40.62
17.61
33.92
22.6
13.22
65.85
34.1
17.85
10.52
17.77
5.48
3.23
18.89
3.31
0.87
1.92
-21.81
-8.9
1.25
-17.88
-14.6
-15.86
-6.25
-30.06
-0.11
8.21
42.82
9.91
-13.3
-31.88
-37.55
-62.23
-90.95
Table 5
Regression Results of Model 2
Variable
MALE
EARLY20S
LATE20S
EARLY30S
LATE30S
EARLY40S
LATE40S
EARLY50S
BTWN5565
MAORI
MAO_EURO
PACIFIC
MXDMAORI
OTHETHNC
LT3YRSCH
LT3SCERT
GE3SCERT
SXTHFORM
BURS_HSC
OTHRSCHQ
POSTSECQ
NRTHLAND
AUCKLAND
WAIKATO
BOPLENTY
HAWKSBAY
TARANAKI
MANAWANG
NELSON
WESCOAST
CANTBURY
OTAGO
SOUTHLND
GRTHRATE
GRATEL1
GRATEL2
GRATEL3
GRATEL4
CONSTANT
Summary statistics
Likelihood Ratio Test
2
Maddala R
2
Cragg-Uhler R
2
McFadden R
Prediction success rate
Dependent variable is LTU27
% Change
Coefficient T-ratio
in Odds
0.37663
127.46
45.73
0.21572
52.07
24.07
0.14805
31.59
15.95
0.15123
29.09
16.32
0.15595
26.99
16.87
0.16844
26.46
18.34
0.24091
33.79
27.24
0.33387
40.61
39.63
0.22831
21.79
25.64
0.26626
67.59
30.51
0.13509
22.53
14.46
0.21805
36.11
24.36
0.14815
10.47
15.96
0.10486
15.09
11.05
0.32246
50.05
38.05
0.16789
23.45
18.28
0.07821
10.78
8.13
0.05734
7.76
5.91
0.13132
14.03
14.03
0.00867
0.68
0.87
-0.03593
-4.36
-3.53
0.10719
13.08
11.31
-0.01418
-2.55
-1.41
-0.00493
-0.74
-0.49
0.01353
1.95
1.36
-0.07288
-10.69
-7.03
-0.07038
-7.72
-6.79
-0.04807
-6.24
-4.69
-0.07678
-9.61
-7.39
-0.07478
-5.11
-7.21
-0.10632
-16.68
-10.08
-0.00916
-1.13
-0.91
-0.19522
-20.1
-17.73
-0.07957
-90.84
-7.65
0.01738
15.96
1.75
-0.06977
-65.29
-6.74
-0.02269
-17.05
-2.24
-0.03878
-41.16
-3.81
-0.94152 -110.84
-60.99
Dependent Variable is LTU52
% Change
Coefficient T-ratio
in Odds
0.43581
107.57
54.62
0.15384
27.29
16.63
0.17186
27.37
18.75
0.22597
32.85
25.35
0.25815
33.95
29.45
0.32471
39.33
38.36
0.43371
47.82
54.29
0.55704
55.12
74.55
0.06347
4.3
6.55
0.33832
67.45
40.26
0.16454
20.71
17.88
0.28986
37.97
33.62
0.19953
10.76
22.08
0.12145
12.71
12.91
0.50915
54.43
66.38
0.29706
28.84
34.59
0.16971
16.13
18.49
0.10417
9.61
10.98
0.16641
12.11
18.11
0.05711
3.06
5.87
0.03332
2.81
3.39
0.17326
16.7
18.91
0.03439
4.71
3.5
0.00899
1.02
0.9
0.01986
2.2
2.01
-0.24662
-26.8
-21.85
-0.09436
-7.73
-9
0.01135
1.12
1.14
-0.19583
-17.8
-17.78
-0.15719
-7.83
-14.54
-0.17174
-19.83
-15.78
-0.06404
-5.81
-6.2
-0.35379
-26.99
-29.79
-0.11634
-99.39
-10.98
0.04529
31.24
4.63
-0.13754
-91.37
-12.85
0.00741
4.44
0.74
0.03357
27.62
3.41
-2.32328 -193.94
-90.2
63776 with 38 d.f.
0.028
0.039
0.022
55.0%
79945 with 38 d.f.
0.035
0.061
0.041
74.5%
14
The figures for percentage change in odds suggest that males are 46% more likely than females to
suffer LTU. This result is in consonance with the finding by the Employment Task Force which
reported that women are more likely to “leave the labour force” on becoming unemployed (1994, p.
14). All the age groups are more likely than teenagers to suffer LTU. Ranking them, the logit
results suggest that those in the early 50s are the worst off, followed by those in their late 40s, 55
and older, early 20s, early 40s, late 30s, early 30s and late 20s. This is consistent with much of
what is reported in the literature.
Compared to those identified as Europeans/whites, all the other races are more likely to suffer
LTU. The Maori are 31% more likely than Europeans to suffer LTU, followed by Pacific Islanders
(25%), mixed Maori-Pacific Islanders (16%), mixed Maori-European (14%) and other ethnic
groups (11%). These figures tell the same qualitative story as the figures in Tables 2 and 3.
With respect to educational attainment, the only group that does better than university graduates are
those with post-secondary qualifications – they are 4% less likely than graduates to suffer LTU.
Those with “other school qualifications” are nearly at par with graduates. As is to be expected,
those with no schooling or less than three years of schooling are the worst off, being 38% more
likely than graduates to suffer LTU. They are followed by those with less than three school
certificates (18%), those with bursary or high school certificate (14%) and those with sixth form
qualifications (6%).
Being the most economically depressed part of the country, Northland unsurprisingly turns out the
highest positive percentage change in odds among the regional locations. The only other place to
turn out a positive figure is Bay of Plenty. Whereas registrants in Northland are 11% more likely to
suffer LTU than those in Wellington, the figure for Bay of Plenty is only 2%. The rest of the
regions range from slightly better to significantly better than Wellington. The region least likely to
experience LTU is Southland, followed by Canterbury, West Coast, Nelson, Hawke’s Bay,
Taranaki, Manawatu-Wanganui and Auckland.
Using 1988 as the base year, the logit results indicate that everyone’s odds of experiencing LTU
decreased by 4% in 1989 but increased by 20%, 57%, 43% and 21%, respectively, in the next four
years. Subsequently, the odds decreased by 1%, 8% and 19% in 1994, 1995 and 1996,
respectively. It may be noted that New Zealand experienced acute recession in both 1988 and
1989.
When the annual dummy variables are replaced by the distributed lags of the growth rate of the
economy, the LTU27-results of Model 2 in Table 5 show that the coefficients for the demographic
and regional variables are comparable to those in Model 1, suggesting that the parameters from the
logit model are quite robust. Turning attention to the growth variables, it can be seen that all of
them have highly significant coefficients. The configuration of their signs suggests that an increase
in the growth rate of the economy may decrease the log of the odds of LTU in the current year,
raise it in the second year and lower it continually thereafter. The second-year effect is puzzling
and would require investigation.
Finally, the parameters of Model 2 will be used to illustrate the estimation of out-of-sample
probabilities of LTU for a few representative jobseekers. The March 2002 publication of Key
Statistics [for New Zealand] shows the March-ending annual growth rates to be 3.06%, 1.85%,
15
0.41%, 4.68% and 2.67% for the years 1997 through 2001. These growth rates have been used in
conjunction with the LTU27-parameters of Model 2 to predict the values of the dependent variable
for eight random individuals and to calculate their respective probabilities of suffering LTU in the
year 2001. Individual ‘A’ is a male, mixed Maori-European in his late 20s having other school
qualifications and living in Auckland.
When the location of individual ‘A’ is changed to
Northland, he is identified as individual ‘B’. Individuals ‘C’ and ‘D’ are the European/white
counterparts of ‘A’ and ‘B’, respectively. The female equivalents of ‘A’, ‘B’, ‘C’ and ‘D’ are
named ‘E’, ‘F’, ‘G’ and ‘H’. The values of the response variable for these individuals and the
calculated probabilities of LTU are summarized in Table 6.
Table 6
Estimating Out-of-Sample Probability of LTU for Representative Jobseekers
Individual
A
B
C
D
E
F
G
H
Estimated Value of the Response Variable (= Log of the
Odds of Experiencing LTU)
–0.94152 – (0.07957*2.67) + (0.01738*4.68) – (0.06977*0.41) –
(0.02269*1.85) – (0.03878*3.06) + 0.37663 + 0.14805 + 0.13509 +
0.00867 – 0.01418 = –0.60762
–0.94152 – (0.07957*2.67) + (0.01738*4.68) – (0.06977*0.41) –
(0.02269*1.85) – (0.03878*3.06) + 0.37663 + 0.14805 + 0.13509 +
0.00867 + 0.10719 = –0.48625
–0.94152 – (0.07957*2.67) + (0.01738*4.68) – (0.06977*0.41) –
(0.02269*1.85) – (0.03878*3.06) + 0.37663 + 0.14805 + 0.00867 –
0.01418 = –0.74271
–0.94152 – (0.07957*2.67) + (0.01738*4.68) – (0.06977*0.41) –
(0.02269*1.85) – (0.03878*3.06) + 0.37663 + 0.14805 + 0.00867 +
0.10719 = –0.62134
–0.94152 – (0.07957*2.67) + (0.01738*4.68) – (0.06977*0.41) –
(0.02269*1.85) – (0.03878*3.06) + 0.14805 + 0.13509 + 0.00867 –
0.01418 = –0.98425
–0.94152 – (0.07957*2.67) + (0.01738*4.68) – (0.06977*0.41) –
(0.02269*1.85) – (0.03878*3.06) + 0.14805 + 0.13509 + 0.00867 +
0.10719 = –0.86288
–0.94152 – (0.07957*2.67) + (0.01738*4.68) – (0.06977*0.41) –
(0.02269*1.85) – (0.03878*3.06) + 0.14805 + 0.00867 – 0.01418 = –
1.11934
–0.94152 – (0.07957*2.67) + (0.01738*4.68) – (0.06977*0.41)
(0.02269*1.85) – (0.03878*3.06) + 0.14805 + 0.00867 + 0.10719 =
0.99797
Probability
LTU in 2001
35.26%
of
38.08%
32.24%
34.95%
27.20%
29.67%
24.61%
– 26.93%
–
Note: definitions of individuals A to H are given in the text.
As it will seen in the last column of Table 6, changing the location of individual ‘A’ from Auckland
to Northland increases his probability of experiencing LTU from 35% to 38%. If the person’s
ethnicity was not mixed Maori-European but European/white, his probability of experiencing LTU
16
would be three percentage points lower at both places. Changing the gender to female reduces the
respective probabilities by eight percentage points but the relative ethnic and locational differences
are maintained. It would seem from this illustration that plausible measures of the probability of
suffering LTU can be obtained for all identifiable jobseekers. Given that there are 2*8*6*13*9 =
11,232 possible combinations of the attributes of registrants from the classifications of gender,
educational attainment, ethnicity, regional office and age group, the illustration above should
suffice. The probabilities calculated this way can constitute an objective basis which can be
augmented with relevant supplementary information to decide on the suitability of a particular
ALMP for various jobseekers.
5. SUMMARY AND CONCLUSION
To advance a statistical profiling model for New Zealand, this paper has extended a previous work
based on a random sample of 100,000 to cover more than 2.2 million cases of completed
unemployment spells over the 1988-97 period. Different variants of a logit model were estimated
to find the relationship between a set of recorded attributes describing a registered jobseeker and
the probability that the individual would be long-term unemployed. The signs and magnitudes of
the regression coefficients and other measures calculated from them lead to the endorsement of
what is qualitatively known about the demographic and locational characteristics of the
unemployed in the country. The fit of the model is satisfactory but the predictive power is
characteristically low. However, the incorporation of the distributed lags of the growth rate of the
economy is an innovation that can allow the calculation of out-of-sample probability of suffering
LTU for any identifiable jobseeker. Consequently, this study advocates the adoption of the
parameters obtained herein as the first step in the construction of a formal profiling model for New
Zealand.
17
REFERENCES
Akerlof, G. A. and Main, B. G. (1980) Unemployment Spells and Unemployment Experience,
American Economic Review, Vol 70, No. 5, pp. 885-893.
Albrecht, J. W., Holmlund, B., and Lang, H. (1986) Job Search and the Transition to Employment:
Theory, FIEF Working Paper No. 26, Stockholm.
Anton, A. A., Sainz, A. F. and Rodriguez-Poo, J. (2001) Semiparametric Estimation of a Duration
Model, Oxford Bulletin of Economics and Statistics, Vol 63, No. 5, pp. 517-533.
Barron, J. M. and Mellow, W. (1981) Changes in Labor Force Status Among the Unemployed,
Journal of Human Resources, Vol 16.
Björklund, A. (1981) Studies in the Dynamics of Unemployment, EFI (Economic Research
Institute), Stockholm.
Clark, K. B. and Summers, L. H. (1979) Labor Market Dynamics and Unemployment: A
Reconsideration, Brookings Papers on Economic Activity, Vol 1, pp.13-60.
Cox, D. R. (1972) Regression Models and Life-Tables, Journal of Royal Statistical Society, Ser. B.,
Vol 26, pp. 186-220.
Cross, R. (1995) The Natural Rate of Unemployment: Reflections on 25 Years of the Hypothesis
(ed), Cambridge University Press, Cambridge.
Employment Task Force (1994) Employment: The Issues, Prime Ministerial Task Force on
Employment, Office of the Prime Minister, Wellington.
Everitt, B. S. and Der, G. (1996) A Handbook of Statistical Analyses Using SAS, Chapman & Hall,
London.
Friedman, M (1968) The Role of Monetary Policy, American Economic Review, Vol 58, No. 1, pp.
1-17.
Galbraith, J. K. (1997) Time to Ditch the NAIRU, Journal of Economic Perspectives, Vol 11, No.
1, pp. 93-108.
Gardiner, P. (1995) “An Analysis of Exit Rates and Duration Dependence in Registered
Unemployment” in Morrison, P. S. (ed.) (1995) Labour, Employment and Work in New
Zealand 1994, Proceedings of the Sixth Conference 24-25 November 1994, Victoria
University of Wellington.
Gordon, R. (1997) The Time-Varying NAIRU and its Implications for Economic Policy, Journal of
Economic Perspectives, Vol 11, No. 1, pp. 11-32.
Griffiths, W. E., Hill, R. C. and Judge, G. G. (1993) Learning and Practicing Econometrics, John
Wiley & Sons, New York.
Gujarati, D. N. (1995) Basic Econometrics, 3rd edition, McGraw-Hill, Inc., New York.
Hall, R. E. (1979) A Theory of the Natural Rate of Unemployment and the Duration of
Unemployment, Journal of Monetary Economics, Vol 5, pp. 153-169.
Han, A. and Hausman, J. (1990) Flexible Parametric Estimation of Duration and Competing Risk
Models, Journal of Applied Econometrics, Vol 5, pp. 1-28.
Hasan, A. and De Broucker, P. (1982) Duration and Concentration of Unemployment, Canadian
Journal of Economics, Vol 15 No. 4, pp. 735-756.
Heckman, J. and Singer, B. (1984) A Method for Minimizing the Distributional Assumptions in
Econometric Models for Duration Data, Econometrica, Vol 52, pp. 271-320.
Heikensten, L. (1984) Studies in Structural Change and Labor Market Adjustment, EFI (Economic
Research Institute) Stockholm.
18
Holmlund, B., Löfgren, K. G. and Engström, L. (1989) Trade Unions, Employment and
Unemployment Duration, Clarendon Press, Oxford.
Intriligator, M. D. (1978) Econometric Models, Techniques and Applications, Prentice-Hall, Inc.,
Englewood Cliffs, N.J., U.S.A.
Kaitz, H. B. (1970) Analyzing the Length of Spells of Unemployment, Monthly Labor Review, Vol
93, pp. 11-20.
Katz, L. F. (1994) Active Labour Market Policies to Expand Employment and Opportunity, Federal
Reserve Bank of Kansas City, Symposium on Reducing Unemployment: Current Issues and
Policy Options, pp. 239-290.
Kiefer, N. M. (1988) Economic Duration Data and Hazard Functions, Journal of Economic
Literature, Vol 26.
Kiefer, N. M. and Neuman, G. R. (1979) An Empirical Job Search Model With a Test of the
Constant Reservation Wage Hypothesis, Journal of Political Economy, Vol 87, pp. 89-107.
Kiefer, N. M. and Neuman, G. R. (1981) Individual Effects in a Nonlinear Model: Explicit
Treatment of Heterogeneity in the Empirical Job-Search Model, Econometrica, Vol 49 No.
4, pp. 965-979.
Key Statistics (various issues), Statistics New Zealand, Wellington, New Zealand.
Lancaster, T. (1979) Econometric Methods for the Duration of Unemployment, Econometrica, 47,
pp. 939-956.
Layard, R., Nickell, S. and Jackman R. (1991) Unemployment: Macroeconomic Performance and
the Labour Market, Oxford University Press, Oxford.
Lewbel, A. (2000) Semiparametric Qualitative Response Model Estimation with Unknown
Heteroskedasticity or Instrumental Variables, Journal of Econometrics, Vol 97, pp. 145177.
Main, B. G. (1981) The Length of Employment and Unemployment in Great Britain, Scottish
Journal of Political Economy, Vol 28, pp. 14-164.
Martin, J. P. (1994) The Extent of High Unemployment in OECD Countries, Federal Reserve Bank
of Kansas City, Symposium on Reducing Unemployment: Current Issues and Policy
Options, pp. 5-40.
Narendranathan, W., Nickell, S. and Stern, J. (1985) Unemployment Benefits Revisited, Economic
Journal, Vol 95, pp. 307-329.
Nickell, S. (1979a) Estimating the Probability of Leaving Unemployment, Econometrica, Vol 47
No. 5, pp.1249-1266.
Nickell, S. (1979b) The Effect of Unemployment and Related Benefits on the Duration of
Unemployment, Economic Journal, Vol 89, pp. 34-49.
Nickell, S. (1990) Unemployment: A Survey, Economic Journal, Vol 100, No. 401, pp. 391-439.
Obben, J., Engelbrecht, H-J. and Thompson, V. W. (2002) A Logit Model of the Incidence of
Long-Term Unemployment, Applied Economics Letters, Vol 9, pp. 43-46.
OECD Proceedings (1998) Early Identification of Jobseekers at Risk of Long-Term
Unemployment: The Role of Profiling, OECD Secretariat, Paris.
Phelps, E. S. (1967) Phillips Curves, Expectations of Inflation and Optimal Unemployment Over
Time, Economica, Vol 34, No. 3, pp. 254-281.
Pindyck, R. S. and Rubinfeld, D. L. (1991) Econometric Models and Economic Forecasts, 3rd
edition, McGraw-Hill, Inc., New York.
19
Richardson, P., Boone, L., Giorno, C., Meacci, M., Rae, D. and Turner, D. (2000) The Concept,
Policy Use and the Measurement of Structural Unemployment: Estimating a Time-Varying
NAIRU Across 21 OECD Countries, OECD Economics Department Working Papers No.
250.
Salant, S. W. (1977) Search Theory and Duration Data: A Theory of Sorts, Quarterly Journal of
Economics, Vol 91, pp. 38-57.
Staiger, D., Stock, J. H. and Watson, M. W. (1997) The NAIRU, Unemployment and Monetary
Policy, Journal of Economic Perspectives, Vol 11, No. 1, pp. 33-49.
Stiglitz, J. (1997) Reflections on the Natural Rate Hypothesis, Journal of Economic Perspectives,
Vol 11, No. 1, pp. 3-10.
Trivedi, P. K. and Baker, G. M. (1983) Unemployment in Australia: Duration and Recurrent Spells,
Economic Record, Vol 59, pp. 132-148.
Watson, R., Maré, D. and Gardiner, P. (1997) Predicting the Duration of Unemployment Spells,
Labour Market Bulletin (New Zealand Dept of Labour) Vol 2, pp. 51-65.
White, K. (1997) SHAZAM: User’s Reference Manual Version 8, McGraw-Hill Book Company,
New York.
20

Similar documents

×

Report this document