full pape

INCOMPLETE DRAFT.
The dimensions of youth unemployment in South Africa
Gareth Roberts
Abstract
The government of South Africa has initiated a number of programmes that are intended to help young
people transition into employment, although they have had only limited success (Bernstein, 2008). One
feature of the problem in South Africa is that the National Youth Policy (2009) refers to young people as
“those falling within the age group of 14 to 35 years”, and the National Youth Development Agency
(2011) points out that “[t]hose aged 20-35 who are not employed and not at school become the key
focus for youth development efforts.” Both the National Youth Policy and the National Youth
Development Agency do, however, recognise that “the 14 to 35 age range is by no means a blanket
general standard, but within the parameters of this age range, young people can be disaggregated by
race, age, gender, social class, geographic location, etc.” (NYP, 2009). This raises an important question
and may provide a possible explanation for the limited success of the youth development programmes
in South Africa: to what extent are these young people disadvantaged because of their age, and not just
by these and other characteristics? Since the presence of omitted characteristics associated with being
employed and unemployed at a particular age is, however, likely to bias any estimates of the causal
relationship between age and employment, we specifiy this relationship as a dynamic correlated
random effects model that considers the conditional distribution of these age-independent
characteristics associated with the initial state of the respondent. This paper finds that those less than
25 who are looking for work are consistently less likely to be in employment , because of their age.
Furthermore, we argue that there is no reason to extend the definition of youth beyond age 30.
Introduction
The problem of youth unemployment is not unique to South Africa, nor is it a new problem. This
country, however, has one of the highest rates of unemployment among young workers in the world
(Treasury, 2011). In the first quarter of 2011, four out of every five workers in the labour force aged 1519 were not employed and wanted to work. A little over 60% of workers aged 20-25, 50% aged 25-29,
and a third of those aged 30-34 were not employed but wanted a job. More importantly, despite
modest growth in GDP, these rates have increased each year since the South African economy went into
recession at the beginning of 2009. And while the official unemployment rate appears to have stabilized
at just over 25%, these figures mask higher levels of discouragement, particularly among younger
workers (Verick, 2011).
The government of South Africa has initiated a number of programmes that are intended to help young
people transition into employment, although they have had only limited success (Bernstein, 2008). An
important feature of the problem in South Africa is that the National Youth Policy (2009) refers to young
people as “those falling within the age group of 14 to 35 years”, and the National Youth Development
Agency (2011) points out that “[t]hose aged 20-35 who are not employed and not at school become the
key focus for youth development efforts.” Both the National Youth Policy and the National Youth
Development Agency do, however, recognise that “the 14 to 35 age range is by no means a blanket
general standard, but within the parameters of this age range, young people can be disaggregated by
race, age, gender, social class, geographic location, etc.” (NYP, 2009). This is also apprent when one
considers that the range of ‘youth’ in this definition means that in some cases it covers parents and their
children, and children receiving Government support (the national Child Support Grant programme was
extended to children aged 17 in 2010). It also raises an important question and may provide a possible
explanation for the limited success of the youth development programmes in South Africa: to what
extent are these young people disadvantaged because of their age, and not just by these and other
characteristics?
This paper attempts to answer this question by investigating the relationship between the age and
employment-state levels and transitions of a sample of workers. Descriptions of the levels of
unemployment show that there is no clearly defined break in this relationship. Rather, they show that
joblessness is highest among young people aged 23 to 27. The presence of omitted characteristics
associated with being employed and unemployed at a particular age is, however, likely to bias any
estimates of the causal relationship between age and employment. For example, young people who
have left school and have no yet had the opportunity to acquire job-specific skills are likely to remain
unemployed for longer when compared to their older peers, regardless of any barriers to employment
that are associated with their age. They may also find it easier to transition into employment because
they have acquired skills that are better suited to vacancies in the labour market. In contrast, younger
workers entering the labour force may always be more productive as complements to newer
technologies, or they may, on average, always be less suited to the emotional demands of work and so
on. Indeed, the descriptions of the age-employment relationship show that the proportion of younger
employed-workers increases, with age, at a faster rate than for older workers.
The paper therefore distinguishes between those unobservable characteristics that are, technically,
independent from the worker’s age, and those that cannot be separately identified. To do this, we
specifiy this relationship as a dynamic correlated random effects model that considers the conditional
distribution of these age-independent characteristics associated with the initial state of the respondent.
The estimation is, however, limited to workers who stayed in the same dwelling over the course of one
year. These workers were inteviewed four times in the year. There are 13 such groups, with the first of
these sampled in the first quarter of 2008 and the last in the first quarter of 2011. This narrows the
population to people that do not move while their dwelling is in the frame. Nevertheless, we find that in
this group those younger than 25 are significantly less likely to be employed, and that after 25 the
average partial effect (APE) on employment of an additional year in age for male workers is increasing
but at a decreasing rate. For females this APE even decreases after age 27. The effect of age on the
probability of being employed plateaus at age 30 for African males. Furthermore, these trajectories are
similar for workers who have completed school and those that have not – although the former are more
likely to be employed at any given age.
We argue that these results, together with the descriptions of the age-employment relationship, suggest
that there is no reason to extend the definition of youth beyond age 30. Those less than 25 who are
looking for work are consistently less likely to be in employment , because of their age. Furthemore, we
argue that the available evidence suggests that the only way to reduce the number of unemployed
youth, in the absecence of exogenous increases in aggregate demand, is to target those age-indepdent
unobservablce characteristics. One way to do this is to, for a short period, facilitate transitions into
employment. This will allow us to better test the extent to which the positive relationship between
one’s lagged employment status operates through behavioural channels or is instead contigent on the
equilibrium we observe.
The youth unemployment problem
In one of the first1 comprehensive economic studies to examine youth unemployment, Freeman and
Wise (1982) identify several dimensions related to the youth unemployment problem that distinguish it
from the generally problem of unemployment. Younger workers are more likely to switch between
searching for work and `non-economic’ activities such as education, and they are prone to being
discouraged2 or less active job seekers. Furthermore, in their study, youth unemployment is generally
concentrated among “a small group who lack work for extended periods of time,” which has
distinguishing characteristics. They also provide several explanations for the causes of youth
unemployment, including the general level of aggregate demand in the economy and the proportion of
young people in the population. There is positive correlation between education and both employment
and wages, and they find that there are characteristics associated with youth unemployment that are
not related to wages, including evidence which suggests that young workers from poor families
experience higher rates of unemployment. They believe that the youth unemployment problem is a
concern not only because of the immediate social and psychological effects of inactivity but also
because, while a long spell of unemployment following the completion of school has no effect on
employment more than three years later, such unemployment is associated with a sizable negative
effect on the wages.
This study, however, deals exclusively with the problem in a developed economy context. There is a
relative paucity of research on the issue in developing countries, particularly in Africa and Asia - due
mainly to a general “lack of good data” (Blanchflower, 1999). Notable exceptions include O’Higgins
(2003) and the International Labour Office (ILO, 2010) annual “Global Employment Trends for Youth”
which, despite this constraint, plot aggregate trends associated with the problem in developing
countries and attempt to draw common inferences within regions. A key finding is that both the level
and rate of youth unemployment are high and may be growing in most developing countries. O’Higgins
(2003) points out that in developing countries the youth unemployment rate is between two and eight
times higher than the corresponding rate for adults. Together, these findings suggest that the youth
unemployment problem in developing economies may not necessarily be confined to a small group.
1
2
Others include Rees (1986) etc. etc.
They want to work but do not actively search for work
In Africa, with data from 13 countries, Guarcello et al. (2005) show that the average duration of the
transition from school to work is very long, “suggesting young people in these countries are faced with
substantial labour market entry problems upon leaving the school system.” Garcia, M. and Fares, J.
(2008) argue that most African youth generally “start working too early and are unprepared to meet the
demands of the labor markets.” Leibbrandt and Mlatsheni (2004), however, show that there is
“considerable variation from country to country which indicates that a number of country specific
circumstances.”
In South Afric, Mlatsheni and Rospabe (2002), define young people as those aged 15 to 30 since entry in
the labour market in South Africa is thought to occur later than in developed countries (Mlatsheni and
Rospabe, 2002), and find that “large amounts of the differences in employment of youth and older
participants are attributable to disparities in observable characteristics such as experience and
education in the case of wage employment and family characteristics in the case of self-employment.
The latter is also likely to be greatly influenced by differences in access to credit.” Furthermore,
unemployment is highest among African youth, young females and those with less education.
Lam, Leibbrandt, and Mlatsheni (2007) extend the definition even further to 35. However they
acknowledge that the different groups in this range are not homogenous, and therefore propose that
there are three distinct cohorts within this broader classification of you: 15-19, 20-24, and 25 to 35.
Using data from the Cape Area Panel Study (CAPS), they find that “by age 20, only 20% of African
females and 31% of African males have ever done any paid work, using a very broad definition. In
contrast, 86% of white females and 90% of white males have done paid work, with only slightly lower
percentages for coloured youth.” They also find that while African and Coloured youth experience a
sharp jump in labour force participation immediately after leaving school, Coloured youth find work
much more quickly. Among African youth, there is a “steady increase in the percentage searching for
work during the first 20 months after leaving school.” However, “by the 20th month after leaving school,
only about 30% of African males and 20% of African females are working.” Lam at al. (2007) also find
that, while there is a high correlation between completed Grade 12 or higher education and the
probability of finding employment in the first 20 months after leaving school, this impact is halved when
they include scores from a literacy and numeracy exam that was administered to the CAPS respondents.
This, they argue, may “indicate that employers do not use schooling alone as a signal, but are also able
to discriminate on the basis of ability.”
In most of the literature on the subject in developed countries, youth unemployment refers to
unemployment among workers aged 15 to 24 or 20 to 24. However, in South Africa even the official
definition of youth extends this age-group by a decade, despite an average life expectancy that is lower
than in these developed economies3. While there appears to be no official explanation, the 15-24
definition that is generally used in the literature on youth unemployment is itself largely arbitrary.
Even though this definition does not constrain investigation of unemployment, it is likely to have
implications for the efficiency of the allocation of resources that are intended to address the issue,
particularly if the impact of homogenous active labour market interventions is heterogeneous within the
group. Their efficiency also depends on the diagnosis of the problem. Is high youth unemployment the
result of cohort crowding, are firms less likely to employ young people for some reason, or are young
people more likely to be voluntarily unemployed etc.? If increased levels of youth unemployment are
caused by population growth (and/or entry into the labour market) that outpaces economic growth,
particularly in labour absorbing sectors and where there is a wage floor, there is no need for a definition
of youth in this context. Changes in the level of youth unemployment will be highly correlated with the
adult rate, and the ratio of unemployed younger workers to unemployed workers (in general) will
remain approximately constant. The prescription is straightforward: develop the demand for labour
absorbing goods that can be produced with the stock of human capital that is available e.g. by removing
barriers to exporting such as labour regulation; or develop the stock of human capital that is needed to
meet the demand for goods that cannot be produced with existing levels.
However if firms favour older workers to younger workers with identical skills, perhaps because they
cannot observe productivity, the effectiveness of policies that facilitate matching will depend on the
functional form of this discrimination in relation to the age of the worker. It makes little sense to treat
workers aged 34 in the same way to those aged 24 if the risk associated with employing an `identical’
but younger worker is significantly higher. In this case the definition of youth will depend on the
intervention. Identifying the particular relationship between the age of the worker and their risk is,
however, very difficult because the absence of any difference between workers aged 24 and 34 requires
a very good explanation for the intervening 10 years. In other words, it may be even riskier to employ
an older unemployed worker unless she is able to signal her productivity.
3
In 2012, life expectancy in South Africa is 53 years
There are many potential sources of such heterogenity that is associated with how firms view younger
workers including, amongst others, how they behave in the workplace – in teams or with clients, the
probability of the worker voluntarily leaving employment (particularly after the fixed costs associated
with training), and even their propensity to take disruptive action in wage disputes. Similarly there may
be certain contexts where firms believe that being young is advantageous e.g. in jobs that require a
more `flexible’ working hours or that are unlikely to appeal to older workers, and so on.
Similarly if, for example, the younger of these otherwise identically-skilled workers have systematically
different unobservable characteristics (such as their discount factor), the effect on each of a
homogenous intervention is likely to differ. The problem here, again, is that is very difficult to identify
these preferences since these are determined endogenously and it is almost impossible to separately
identify their impact on aggregate youth unemployment. Indeed this is generally the case for many
other charactertics that are correlated with being young – such as the the networks these young people
have access to, the skills they have acquired at differrent levels of education (since these are likely to
differ across generations), and the differential support they receive from their environment (including
their support structures), and so on.
Given the large number of potential causes of the heterogenity in the employment and unemployment
rates of younger and older workers, it is not surprising that numerous interventions have been proposed
to mitigate the problem. In the most extensive investigation into the impact of different interventions
that attempt to address the youth unemployment, Betcherman, Gofrey, Puerto, Rother, and Stravreska
(2007) examine evidence relating to 289 interventions from 84 countries (not including South Africa).
The interventions they examine include those that make the labour market work better for young
people - such as improving information (counseling and job search skills), those that increase labour
demand such including wage subsidies and public works programmes, and those that remove
discrimination. They also extend their investigation to interventions that are intended to promote
entrepreneurship among young people, those that attempt to resolve post-school training problems and
training market failures, mobility barriers, and regulatory reforms such as changes in labour law.
The results from their investigation show that there are, on average, no differences across the different
interventions in terms of their impact on employment. Nevertheless, while most had a positive impact,
less than half were efficient. They find that the impact of the programmes on employment was generally
greater in developing countries, but also find that the impact was lower in countries with less flexible
labor markets. Betcherman et al. (2007) do, however, find that the highest returns for disadvantaged
youths come from early and sustained interventions. They therefore argue that “any policy advice on
addressing youth unemployment problems should emphasize that prevention is more effective than
curing.” Furthermore, they point out that the level of evaluation of programmes is generally weak and
conclude that there is a “need for major improvements in the quality of evidence available for youth
employment interventions” and that “the absence of rigorous evaluations almost certainly leads to an
overestimation of program impacts by policy-makers.”
In a study that attempts to replicate the Betcherman et al. (2007) inventory for South Africa, Bernstein
(2008) draws attention to the absence of rigorous evaluation of the impact of such programmes in this
country. Bernstein (2008) points out that less than a third of the 114 interventions identified “are
externally assessed in any way,” that with a few exceptions there was no financial data that would
permit an assessment of the cost to benefit ratio of these programmes, and that “no data appears to be
collected on the experiences of young people in the months and years after their contact with an
employment creation intervention.” The study finds that approximately 40% of the governmentsponsored programmes involve direct employment creation (e.g. public works programmes), 30%
promote skills development, and the remaining 30% support business development. Furthermore,
despite the breadth of access to these programmes, the available evidence suggests that together they
have facilitated a transition into employment for only a small proportion of the unemployed youth.
Data
This paper uses a panel of data that was constructed from Statistics South Africa’s (StatsSA) Quarterly
Labour Force Survey (QLS). The survey, which was first introduced in 2008, is used to calculate the
official unemployment rate. Respondents are sampled in dwellings that, when weighted, are intended
to be representative of the national population. After every quarter, a quarter of these dwellings are
rotated out of the sample-frame. The survey collects data about the individuals in each dwelling a
maximum four times over the course of one year. It is released as a set of independent cross-sections
for the quarter. There is, however, a unique identifier for a significant proportion of the respondents4,
and an algorithm is used to match those people within dwellings where this was not the case. The
following table presents descriptions of the number of working-age individuals and the number of
periods data was collected for unique individuals.
Table 1: Sample
Individual observation
Number of period observations
Year
Quarter
1
2
3
4
Total
1
2
3
4
Total
2008
1
59,025
0
0
0
59,025
25,050
15,287
10,860
7,828
59,025
2008
2
24,919
33,148
0
0
58,067
9,650
3,994
3,222
8,053
24,919
2008
3
23,766
15,800
18,306
0
57,872
8,672
3,772
3,146
8,176
23,766
2008
4
22,788
15,213
11,531
7,976
57,508
8,377
3,200
3,120
8,091
22,788
2009
1
23,022
14,526
11,507
8,213
57,268
7,967
3,795
2,993
8,267
23,022
2009
2
21,353
15,032
11,291
8,338
56,014
7,256
2,970
2,892
8,235
21,353
2009
3
20,777
14,183
11,334
8,243
54,537
6,985
3,092
2,498
8,202
20,777
2009
4
20,733
13,990
11,324
8,448
54,495
6,878
2,419
2,736
8,700
20,733
2010
1
21,156
14,177
10,960
8,414
54,707
5,035
3,024
3,268
9,829
21,156
2010
2
18,255
16,017
11,667
8,365
54,304
3,588
2,436
2,815
9,416
18,255
2010
3
16,327
14,679
13,065
8,873
52,944
3,190
2,148
2,319
8,670
16,327
2010
4
15,902
13,307
12,435
10,034
51,678
3,235
1,816
2,401
8,450
15,902
2011
1
17,536
12,797
11,180
9,580
51,093
3,526
2,486
2,705
8,819
17,536
2011
2
17,001
13,957
11,006
8,841
50,805
3,623
2,610
10,768
0
17,001
2011
3
18,397
13,767
11,598
8,630
52,392
4,667
13,730
0
0
18,397
2011
4
18,062
14,243
11,567
8,961
52,833
18,062
0
0
0
18,062
359,019
234,836
168,771
112,916
875,542
125,761
66,779
55,743
110,736
359,019
Total
4
See Verick (2011)
Descriptions of the labour market in South Africa
The figure5 below presents the proportions in 2008, 2009, 2010, and 2011 of each age-cohort that is
studying, unemployed, not economically active, in private wage-employment, public employment and
self-employed. The majority of young people aged less than 20 are students. The unemployment rate
increases as these young people leave education and is highest at age 25. The proportion of employed
workers then increases until age 45, when increasing numbers of workers transition out of the labour
market. There is no discernable change in these trajectories after age 35. Similarly, there is no point at
which the relationship between age and employment flattens out6. Table x, however, shows that after
age 27, the proportion of workers, those that have been unemployed for less than a year, and the
proportions of those that stay employed from quarter to quarter or move into employment start to
converge. The table also shows that workers older than 30 are more likely to be underemployed.
Figure 1
5 Following Soyer, Emre and Hogarth (2011), the following and subsequent sections rely extensively on grpahs to
describe the labour market for young people in South Africa.
6
The proportions presented are smoothed using a cubic spline of age.
Table 2: Descriptions
Employed
Employment
transitions
Unemployed
No matric
Labour
Full-time
Under
Job
duration
in Q
Never
worked
<6
months
6 months
to 1 year
>1
year
Stay
Into
18
83.0
14.0
3.2
0.5
3.8
96.4
2.0
0.9
0.7
67.9
1.4
19
71.4
25.0
7.0
0.8
4.0
93.8
3.1
1.6
1.4
74.8
2.4
20
63.9
39.2
11.5
1.3
4.7
89.8
4.8
2.7
2.7
77.4
3.6
21
58.4
50.1
17.3
1.7
5.4
85.0
6.4
3.7
4.9
81.0
5.3
22
56.5
60.3
23.6
2.4
6.1
79.6
8.3
4.8
7.2
83.3
6.4
23
55.2
67.7
28.7
2.9
6.7
74.8
8.9
6.5
9.8
85.1
7.9
24
54.1
73.4
34.0
3.0
7.7
68.0
11.4
7.4
13.3
87.1
9.3
25
55.0
76.7
38.1
3.4
8.7
64.1
11.3
8.1
16.4
87.1
9.5
26
55.2
77.9
40.6
3.3
9.6
59.8
12.2
7.8
20.2
89.0
9.7
27
54.6
80.4
43.8
3.7
10.5
55.6
12.6
9.2
22.6
88.8
11.3
28
55.3
80.2
45.8
3.7
11.3
52.0
13.1
9.0
25.8
89.3
11.7
29
54.8
80.3
47.2
3.8
12.4
51.7
12.7
9.1
26.5
90.2
11.0
30
55.1
80.8
48.9
4.1
13.8
48.5
13.1
8.7
29.8
90.3
11.8
31
55.5
81.1
50.7
3.7
14.5
46.8
13.2
8.2
31.7
91.2
11.5
32
55.1
81.2
52.3
4.0
15.4
43.7
13.0
8.9
34.3
90.8
12.1
33
55.5
81.0
53.3
4.3
16.9
42.5
13.1
8.7
35.7
91.4
12.7
34
55.1
81.3
54.5
4.1
18.3
37.4
13.0
8.7
40.9
91.2
12.6
35
55.8
81.1
54.4
4.6
19.6
37.2
12.4
8.3
42.1
91.6
13.1
36
57.0
80.7
56.2
4.2
21.4
37.2
13.1
7.7
42.0
91.6
13.0
37
58.9
80.8
55.9
4.3
22.9
33.3
13.4
7.6
45.7
91.5
12.3
38
58.2
80.8
56.3
4.2
24.2
31.6
13.4
8.1
47.0
92.2
12.1
39
60.6
79.9
56.6
4.5
26.1
30.6
12.4
7.5
49.5
91.6
13.0
Age
Figure 2 presents the estimates of the contribution (average) of each age-cohort to the working-age
population, labour force, the unemployed (defined broadly), and those employed – again in 2008, 2009,
2010, and 2011. The distributions each year are remarkably similar. In this sample the 18-19 cohort of
unemployed and wage-employed represent a disproportionately smaller share of these statepopulations. The 19-24 cohorts, in contrast, represent disproportionally more of the unemployed and
disproportionally fewer of the wage-employed. Young employed workers older than 27 contribute more
than their relative share of the employed. However, they contribute an even higher proportion of the
unemployed. Indeed, workers aged 25 represent the highest share of the unemployed across all agecohorts, after which this share starts to decrease until age 30 when the relative share of employed and
unemployed workers appears to converge. The share of unemployed drops below the share of the
working-age population after 34.
The consistency of these distributions over these periods (before, during, and after a recession) support
the argument that there is a youth unemployment problem, in the sense that a higher proportion of
younger workers are unemployed when compared to their older counterparts. However, these
distributions also suggest that there are several sub-groups within the 18-34 cohort in this sample: 1819, 20-24, 25-34. The first is characterised by lower levels of participation, the second by significant
increases in participation and search frictions, and the third by increasing levels of employment and
decreasing levels of unemployment. Similarly, those aged 25 30 represent the highest proportion of the
labour force. This poses several questions: Are young people in cohorts with disproportionally more
wage-employed workers disadvantaged, and what explains the convergence we observe from age 25?
Figure 3 adds an important additional perspective: youth ‘discrimination’ in employment is far greater in
public (which includes NGOs and parastatals) and self-employment than it is in private-firm
employment. While the proportionate-contribution of those aged less than 24 to the number of wageemployed is lower than their corresponding contribution to the labour force, the difference is
substantially smaller than for the latter two. Indeed, convergence occurs much later for these two
employment-states: at age 34. The official definition may not, therefore, be merely coincidental.
Finally, Figure 4 ssuggests that one of the reasons why both unemployment and labour force
participation are higher after 24 may be because, after 24 the relative share of the population in
education is disproportionally lower than that cohort’s share of the population, and that from age 23
the contribution of the labour force tracks the age-cohort’s contribution to the population of workers
aged 18 to 65.
Figure 2:
Figure 3
Figure 4
The causal effect of age on the probability of being employed
While these figures show us that younger workers are (at any given time) less likely to be employed than
older workers, they are less informative about the extent to which unemployed younger workers are at
a ‘general’ disadvantage when it comes to finding employment. These levels are cumulative in that all of
those who are employed would have found this work when they were younger. Furthermore, even
though a person’s age is independent, the distributions of the unobservable characteristics of workers
who are in any given state are not likely to be independent. For example, younger workers (in general)
are likely to have less work experience simply because they have had less time to accumulate this
experience. One could also reasonably expect that those with less experience (particularly job-specific
skill) would be less likely to be employed, regardless of the selection mechanism that leads to these
skills. Similarly, older unemployed workers may differ in ability to younger workers that have
transitioned into the labour force from school. Neglecting these omitted characteristics that are
associated with being employed or unemployed at a given age may bias any estimates of changes in
transition probabilities that may be attributed to age.
There are a number of parametric approaches that attempt to address the initial conditions problem,
particularly in linear models e.g. Anderson and Hsiao (1982), Arellano and Bond (1991), Arellano and
Bover (1995), and Ahn and Schmidt (1995). There are also ways to estimate the parameters in non-linear
models, including non-parametric specifications e.g. Chamberlain (1992), Honore (1993), Wooldridge
(1997), Honore and Kyriazidou (2000), Altonji and Matzkin (2003), and Chernozhukov, Fernández-Val,
Hahn, and Newey (2009). With these models, however, it is either not possible to include lagged
dependent variables or fixed covariates, or one is unable to distinguish point-estimates of the average
partial effects (APE) of any strictly exogenous covariates. Heckman (1981) instead proposes a dynamic
probit model with covariates that approximates the conditional distribution of the initial condition
(Wooldridge, 2005), and Orme (2001) proposes a model that includes a Heckman-type (1979) selection
correction term. Both models, though, require the inclusion of exclusion restrictions in the specification
used to identify this distribution. Wooldridge (2005), in contrast, suggests that one could instead model
the distribution of the unobserved effect conditional on the initial value and the distributions of strictly
exogenous explanatory variables for all time periods. This approach requires a balanced panel, and while
Wooldridge (2010) makes suggestions on how to deal with non-random sample selection in Correlated
Random Effect models, these methods do not extend to those with a lagged dependent variable on the
right hand side. Wooldridge (2005) does, however, point out that the density obtained using his simple
solution for a lagged dependent variable has the advantage that it would not only be conditional on the
exogenous explanatory variables, but also depend on the initial state “in an arbitraty way.”
When studying the performance of this estimator, Akay (2009) finds that in the case of unbalanced
panels the Wooldridge (2005) estimator performs better than estimators which assume the initial
condition is exogenous. However, Akay also finds that Heckman’s (1981) estimator is not affected “by
the unbalancedness of panel data”7, and that this estimator performs better than the Wooldridge (2005)
estimator in panels with fewer than five periods. Since the data available to this paper only contains a
maximum of four waves for any given individual, the Heckman (1981) is the most appealing. However,
the QLFS data includes no additional pre-sample data for the individual that may be regarded as
exogenous (other than age, race and gender, e.g. place of birth, or a relative’s level of education). Race,
gender, or even interactions between these and the time period could be used as exclusion restrictions,
although these are also likely to be significantly correlated with the subsequent dynamics.
This consequently makes the Wooldridge estimator the only non-linear alternative available to this
paper, even if as Wooldridge (2005) points out it is not possible to estimate the ‘causal’ effect of the
respondent’s gender and race on their employment dynamics – since these cannot be separately
identified from the unobservable component. There is nevertheless reason to believe that any bias dues
to non-random sample selection will be smaller than one might expect in panels following only one set
of workers over several periods since, as mentioned the QLFS follows a rotating-dwelling sample
scheme. This means that every quarter, a quarter of the sample consists of new dwellings which should,
since the rotation is random, contain individuals similar to those who left their dwellings in the
preceding period. Of course this does not include information about the dynamics associated with
employment state transitions when these individuals move dwellings, which is clearly an important
concern - particularly for young people who are more likely to move around. However, this is true for
every initial state of the individuals in the sample. Since the sample includes people that moved into the
dwelling just before it was sampled for the first time, and who do not move out of the dwelling in the
subsequent three periods, it may be plausible to assume the duration (when less than four quarters)
spent in a given dwelling is orthogonal to the relationship between age and employment, conditional on
the distribution of unobservable characteristics. In other words, it is possible that the distributions of
the unobservable characteristics associated with the initial state are equivalent for those workers that
7
Although this does not imply that the estimates are not biased as a result of non-random attrition
we observe for at least four periods and those that we observe for less than four periods, and that the
difference is captured by the individual time-specific error.
Figure <> presents the proportion of each age-chort that is employed in this population (over four
periods), and these proportions for those we observe for less than four periods. It shows that those less
than 25 in the latter population were less likely to be employed, although these proportions follow
parrallel trends until age 25. It follows then that our estimates of the age-employment relationship are
likely to over-estimate the probability of being employed for those younger than 25, if these differences
are not captured by the conditional distribution of the age-independent unobservables.
Perhaps the greatest threat to the validity of the estimates then is the distribution of these conditional
unobservables. This would seem a far greater threat, even in longer balanced panel, than differences in
the dynamics that can be attributed to the people we do not observe. The validity of this assumption,
however, is not testable and as such remains a mix of speculation and the desire to find an answer to
the type of question that an experiment would not be able to provide a consistent estimate (as the
number of experiments and time both increase to infinity).
Figure 5:
Consequently, to answer the question of where youth ends in the South African labour market, this
paper specifies the causal effect of the respondents’ age, as opposed to the effect of her unobservable
characteristics that are associated with her age, by controlling for the effect this selection on
unobservables and the duration-dependence in the respective employment states (y) – employed or
unemployed as:
P(yi,t = 1 | yi,t-1 , yi,t-2 , … , yi,0, zi, ci) = Φ(αzait + ρyi,t-1 + α 0zi + ρ0yi,0 + ci )
where αzait and ρyi,t-1 include the age of the respondents in full time employment or who are
unemployed in t-1, α0zi the mean age of the respondent over the four periods, and ρ0yi,0 her
employment state in the first period.
y is a binary variable since there is, as yet, no equivalent estimator for more than two states (e.g.
employed, unemployed, and not economically active). This, consequently, makes defining an
appropriate measure of unemployment in this context very important. The QLFS questionnaire is
structured so that it is only possible to recover whether the respondent is currently engaged in any
education activities if the respondent has not worked or has not searched a job. It follows that a
proportion of those who are working may be full-time students who e.g. work over weekends or are
looking for casual once-off jobs. Similarly, a proportion of those searching for work or even those who
claim to be discouraged may be studying at the same time or waiting for the school term to start and so
on. The sample also includes those workers who transition from school to work, or from a full-time job
or from searching for work back into education. Furthermore, a proportion of those who were employed
in the week prior to the survey were employed once-off for only a few hours, or they are engaged in
‘sporadic’ self-employment. In both cases these people are, effectively, in a general state of
unemployment. Then, while many of the discouraged job-seekers who want to work but have given up
searching for work because they can’t afford to or because they do not believe they’ll find a job, it is not
clear if all of these people should be regarded as being part of the labour force if there is a wage-offer
distribution for these workers.
In addition to restricting the sample to the respondents who happened to be in all four periods in
dwellings that were sampled over four periods, it also excludes those respondents who were studying
and were not looking for work in all four periods, and those who were not retired and not looking for
work during this time. Restricting the sample to four-period observations permits a more precise
measure of age (in quarters)8. This is important because the four periods span over only one year, which
means that a respondent will only experience a single discrete change in her age when this is measured
in years.
The following table and figures presents the results from the estimation, and plots of the APE of an
additional year of age on being employed. They show that those workers aged less than 25 are
consistenly less likely to be employment after we consider the distribution of age-independent
unobservable characteristics associated with being in employment. There is some heterogeneity in the
this relationship after age 25. For male workers the probability of being employed continues to increase
with age. However for females it decreases, and for African male workers there is no change in this
probability after age 30.
More importantly, as figure <6> shows, the significance of the age-independent unobservable
characteristics explaing the difference between the APE of age and the observed proportions of workers
in employment at any given age increases with age This, and the signficance of the lagged employment
8
Age in years is captured as an integer in the QLFS
state in explaining the probability of being employed, seems to suggest that the only way to change the
age-employment relationship we observe would be through aggregate demand or by targeting these
unobservable characteristics, perhaps through active programmes that facilitate transitions in
employment for those younger than 25. This paper cannot confirm whether the latter would sustain
employment or have an effect on the outcomes of older workers since it does not identify (other than
assuming that conditional on age and the lagged employment state they are normally distributed) what
these unobservables characteristics are. These can only be identified through experiments.
Table 3: Regression estimates of probability of being in employment
Age
Age squared
Employed in t-1
Mean age
Mean age squared
Employed in t=1
Constant
Ln(Sigma)
Observations
Number of respondents
All
Male
Female
Employed
African male
African female
No matric
Matric
0.471***
(0.0787)
-0.00577***
(0.000972)
0.944***
(0.0218)
-0.312***
(0.0785)
0.00378***
(0.000971)
3.707***
(0.0591)
-5.122***
(0.102)
0.648***
(0.0325)
0.517***
(0.113)
-0.00695***
(0.00142)
0.862***
(0.0324)
-0.337***
(0.113)
0.00474***
(0.00142)
3.650***
(0.0849)
-5.264***
(0.150)
0.638***
(0.0478)
0.443***
(0.110)
-0.00500***
(0.00133)
1.007***
(0.0295)
-0.286***
(0.109)
0.00302**
(0.00133)
3.640***
(0.0798)
-5.225***
(0.140)
0.631***
(0.0443)
0.540***
(0.126)
-0.00713***
(0.00160)
0.837***
(0.0353)
-0.358***
(0.125)
0.00488***
(0.00160)
3.534***
(0.0906)
-5.317***
(0.168)
0.646***
(0.0523)
0.465***
(0.121)
-0.00493***
(0.00147)
0.983***
(0.0321)
-0.286**
(0.120)
0.00273*
(0.00147)
3.454***
(0.0841)
-5.638***
(0.160)
0.601***
(0.0491)
0.431***
(0.0969)
-0.00539***
(0.00114)
0.863***
(0.0255)
-0.280***
(0.0966)
0.00350***
(0.00114)
3.483***
(0.0653)
-4.993***
(0.122)
0.609***
(0.0384)
0.493***
(0.146)
-0.00586***
(0.00203)
1.162***
(0.0431)
-0.336**
(0.145)
0.00405**
(0.00203)
3.762***
(0.124)
-5.140***
(0.188)
0.613***
(0.0658)
252,492
84,396
110,327
36,865
142,165
47,531
84,489
28,229
112,731
37,682
160,530
53,707
91,962
30,689
Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1
Figure 5: Average Partial Effect of an additional year in age
Figure 6: Predicted level of employment
Figure 7: Average Partial Effect of an additional year in age for male workers
Figure 8: Average Partial Effect of an additional year in age for female workers
Figure 9: Average Partial Effect of an additional year in age for African male workers
Figure 10: Average Partial Effect of an additional year in age for African female workers
Figure 11: Average Partial Effect of an additional year in age for workers without matric
Figure 12: Average Partial Effect of an additional year in age for workers with matric
Conclusion
…
References
Soyer, Emre and Hogarth, Robin M. (2011). The Illusion of Predictability: How Regression Statistics
Mislead Experts. International Journal of Forecasting, Forthcoming. http://ssrn.com/abstract=1996568
Wooldridge, J. (2010). Correlated Random Effects Models with Unbalanced Panels. Mimeo, Michigan
State University; accessed Septermber 2012. http://econ.msu.edu/faculty/wooldridge/docs/cre1_r4.pdf
...
Appendix
Figure 13: Official unemployment rate
Figure 14: Official GDP growth rate