The Ins and Outs of Cyclical Unemployment in Brazil % A First

The Ins and Outs of Cyclical Unemployment
in Brazil - A First Assesment1
Fernando Siqueira2
July 22, 2009
1 The
author thanks Dieese/Seade for providing part of the data used in this paper and seminar participants at Getúlio Vargas Foundation "Brown Bag" Seminar
for their comments. All remaining errors are my own
2 Ph.D Candidate - EESP- FGV( Economic School of Sao Paulo - Getulio Vargas
Foundation)
Abstract
This paper uses di¤erent data sources to measure the cyclicality of job …nding, job separation, job creation and job destruction rates in Brazil. The
main …ndings of this research are i) job …nding and job creation present a
more cyclical dynamics than job separations or job destruction, ii) the job
…nding rate represents between 65 and 85% of unemployment variance, iii)
quits account for an important share of all separations and show a more
cyclical pattern than layo¤s. These results are all in line with the recent
international evidence. We also compare job …nding and separation rates in
Brazil with the ones available for several countries. Our estimation indicates
that worker mobility in Brazil is larger than in most countries for which we
have the same measure of worker mobility. This result is in contradiction
with the conventional notion that labor laws in Brazil made our labor market rigid. A possible explanation for these results is the large fraction of
informal workers in Brazil which could bypass the rigidities caused by labor
regulation.
1
INTRODUCTION
Unemployment dynamics is countercyclical: GDP growth requires, ceteris
paribus, more workers and thus leads to a smaller unemployment rate. Although this is a conventional wisdom, the way it happens is controversial.
The variation in total employment is a result of ‡ows into and out of employment. Some argue that the ins are more important, while others favor
the outs.
Up to 2002 the conventional way of modeling the cyclicality of labor dynamics was to assume that job …nding is relatively stable, while job loss
moves countercyclically with GDP: contractions in GDP leads to an increase
in unemployment due to the increase in the job separation rate. An adverse
shock reduces the pro…tability of jobs and since wages are sticky, the only
way to adjust that is to lay o¤ some workers. This view on labor market
adjustments has been proposed by Davis, Haltiwanger and Schuh (1996),
Blanchard and Diamond (1990), among others. This conventional wisdom
was challenged by earlier versions of research papers like those of Hall (2005)
and Shimer (2007). Using a detailed database on the transition of workers
between employment and unemployment, these papers reached the opposite
conclusion: the driving force in unemployment dynamics is the job …nding
rate instead of job separation. Job loss rates are relatively stable or acyclical.
An adverse shock lowers the pro…tability of jobs and leads to a decrease in
hiring intentions, causing unemployment to increase. The results obtained
by Hall (2005) and Shimer (2007) do not demonstrate a consensus yet. Fujita and Ramey (2009), Elsby, Michaels and Solon (2009), among others,
reach somewhat di¤erent conclusions using the same database. In particular, Fujita and Ramey (2009) and Elsby et al. (2009) …nd job separation to
be countercyclical (rather than acyclical, as pointed out by Hall (2005) and
Shimer (2007)). Petrongolo and Pissarides (2009), analyzing three European
countries (UK, France and Spain), …nd that the importance of job …nding
and job creation varies across countries and that both variables are important
for unemployment dynamics.
In Brazil, the analysis of labor market dynamics at a macroeconomic level
is relatively scarce. In the last few years, most works on labor markets at
meetings such as Anpec and SBE1 have focused on microeconomic aspects,
including wage di¤erentials and convergence, demand for skilled workers,
among other themes. Macroeconomic aspects of the labor market, for in1
ANPEC and SBE are Brazilian annual meetings on economics and econometrics research respectively. Papers presented at those meetings can be found at www.anpec.org.br
and www.sbe.org.br.
1
stance, gross job creation and job destruction, job …nding and separation
rate, have been ignored so far, at least up to the time this study was being
written. Even in the publications by the Central Bank of Brazil (BCB) on
applied macroeconomics, there has been too little discussion on labor market
dynamics, with more emphasis being placed on a few variables. The discussion reported in monetary meeting minutes and in‡ation reports is restricted
to the evolution of the unemployment rate, a subject that plays a more important role in in‡ation dynamics and job creation, especially in formal job
creation, a variable thought to have a positive correlation with the business
cycle.
The aim of this paper is to contribute to the analysis of labor market
dynamics in Brazil. This research study is not intended to be exhaustive, but
to bring some more detailed statistics about the labor market into discussion
and thus motivate new research on the subject. The methodology used in
this paper closely follows that of Shimer (2007), Petrongolo and Pissarides
(2008), Elsby, Hobijn and Sahin (2008), among others. Some parts of the
paper also rely on Blanchard and Diamond (1990) and Davis and Haltiwanger
(1999).
To address the evolution of unemployment dynamics in Brazil based
on job …nding and separation rates, we use three di¤erent data sources:
Dieese/Seade, IBGE and CAGED. Seade/Dieee has conducted a monthly
employment survey in the metropolitan area of Sao Paulo since 1989. IBGE
also performs an employment survey in six metropolitan areas in Brazil, but
due to methodological changes, the data we use start in 2002. CAGED traces
down admissions and separations regarding formal jobs. The main advantage of using CAGED data is that they provide details on the reasons for job
separation.
Our results indicate that job …nding has a more cyclical dynamics than
job separations. These results are shown by the three databases used: the job
…nding rate accounts for 65 to 85% of unemployment variance, according to
Dieese/Seade data and IBGE data, respectively whereas admission is found
to be more cyclical than layo¤s, according to CAGED data. Quits are also
found to present strong cyclicality. This result is in line with the international
literature. Our results indicate that layo¤s have a somewhat countercyclical
pattern, but some estimates also show an acyclical pattern.
This paper is organized as follows. Section 2 presents a simple model of
unemployment dynamics based on job …nding and job separation rates and a
way to decompose unemployment variation into these two variables. Section
3 o¤ers more details on the data we use and Section 4 presents our estimation
results for the impact of GDP growth on admissions (job creation) and layo¤s
(job destruction) and the results for unemployment decomposition into job
2
…nding and job separation rates. In Section 5 we present our conclusions and
o¤er directions for future research.
2
JOB-FINDING AND JOB-SEPARATION
MODEL OF UNEMPLOYMENT DYNAMICS
2.1
SIMPLE MODEL2
This section presents a simple model of unemployment dynamics based on
job …nding and job separation rates. The model assumes a constant labor
force and is formulated in continuous time. Since data are only available in
discrete time, some adjustments must be made.
For t 2 f0; 1; 2; :::g we refer to the interval [t; t + 1) as the time t. Let
Ft 2 [0; 1] be the job …nding probability and Xt 2 [0; 1] the employment exit
probability (the probability of losing the job) at time t and let us assume that
unemployed workers …nd a job and employed workers lose their jobs according
to Poisson process with arrival rates ft = log(1 Ft ) and xt = log(1 Xt );
respectively.
Now let 2 [0; 1] be the time elapsed since the last measurement date,
et+ the number of employed workers at time t + and ut+ the number of
employed workers that are unemployed at time t + but were employed at
some time t0 2 [t; t + ]. It is convenient to de…ne uSt+1
uSt (1), the total
amount of short-term unemployment at the end of period t.
Using the variables just mentioned we can write:
ut+ = et+ xt
ut+t ft
(1)
ust ( ) = et+ xt
ust ( )ft
(2)
These equations represent the law of motion for unemployment and shortterm unemployment, respectively. Note that we can rewrite equation (2) as:
et+ xt = ust ( )ft + ust ( )
Substituting (3) into (1) we get:
2
This section is based in Shimer (2007).
3
(3)
ut+ = ust ( )
ust ( ))ft
(ut+t
(4)
By construction, uSt (0) = 0 and given an initial condition for ut this
di¤erential equation can be solved for ut+1 and ust+1 ust+1 (1) :
Ft )ut + ust+1
ut+1 = (1
(5)
Rearranging equation (5) we get the probability of …nding a job:
"
#
ut+1 ust+1
Ft = 1
ut
(6)
This expression connects the job …nding probability as a function of unemployment and short-term unemployment, two variables usually available
from employment surveys. To get the job separation rate we …rst solve equation (1) forward to get:
ut+1 =
(1
e (ft +xt ) xt
lt + e
ft + xt
(ft +xt )
ut
(7)
where lt = et + ut , which is assumed to be constant. This is not a good
assumption in some cases, and we will return to this issue when we report our
results in Section 4. The only unknown in equation (7) is the job separation
rate xt , which can be easily solved numerically.
2.1.1
Unemployment decomposition
Given that we have already calculated ft and st we can use these two variables
to infer what is the steady-state unemployment rate, u
et , and decompose the
its variation into the variations of job-…nding and job-separation rates.
The standard formula for steady-state unemployment rate in search and
matching models are given by:
u
et =
st
st + ft
(8)
Using this equation, the standard approach to compute the contributions
of ft and st to unemployment variation is to assume that u
et ' ut and compute
the time variation in ut to obtain:
ut = (1
ut )ut
st
1
st
4
1
ut (1
ut 1 )
ft
ft
(9)
For notation convenience, let us rewrite equation (8) as:
du = dus + duf
(10)
Note that by doing so we are using the approximation u
et ' ut . Equation (9) decomposes unemployment variation into a component attributed
to variations in the job …nding rate and another one attributed to the job
separation rate. By using this equation, we can also compute the percentage
contribution of each factor as:
j
=
cov(du; duj )
;
var(du)
j = s; f
(11)
Given equation (9), we know that s + f = 1. With f ; s at hand we
can infer which factor, job-…nding or job-separation, accounts for the most
part of unemployment variation.
But before computing the 0 s some adjustments are usually necessary.
The …rst one refers to the calculated statatistics of st and ft : Fujita and
Ramey (2009) use an HP …lter to smooth ft and st to capture the "trend"
of each variable and the corresponding steady-state unemployment. Shimer
(2007) also uses an HP …lter but with a much lower smoothing parameter
than the = 1600 usually used for quarterly series. His motivation for this
is that the usual parameter removes much of the cyclical variation from the
variables. Using a lower smooth parameter, Shimer (2007) removes the excess
variation in ft and st attributed to measurement errors without taking o¤ all
cyclical behavior.
The second adjustment refers not to the way the variables enter into the
decomposition equation, but to the way the composition equation is computed given that the assumption u
et ' ut may not be a good one. Petrongolo
and Pissarides (2009) remove from their sample observations where di¤erences between the implied (given st and ft ) unemployment rate and the
actual rate is more than 10%. Elsby et al. (2008) use a more elaborated approach to deal with these misalignments. Instead of assuming that the actual
unemployment rate is always equal or close to steady-state values, Elsby et
al. (2008) assume that actual unemployment rate tends to converge to the
steady-state equilibrium implied by the calculated st and ft , but the rate of
convergence is not large enough to imply that u
et ' ut for all t. This case
can be important in countries where the values of st and ft are small (this
is often used as a proxy for rigid labor markets). When Shimer (2007) …rst
used the hypothesis of full convergence, implied in the u
et ' ut hypothesis, he
correctly pointed out that the U.S. labor market is one of the most ‡exible
in the world and that the values of st and ft were large enough to support
5
this assumption. Adjusting equation (8) to allow for rates of convergence
less than one, we need to rewrite it as:
ln ut '
t 1
(1
ut 1 )[ ln st
ln ft ] +
1
t 2
ln ut
1
;
(12)
t 2
where t 1 e 12(st +ft ) re‡ects the rate of convergence to steady-state
unemployment for monthly job-…nding and job-separation rates. For U.S.
data,
1 given that both f and s are high. On the other hand, for
some OECD countries is much lower than one (indeed, in some cases it is
lower than 0.50). This supports the analyisis of both the Shirmer (2007) and
Elsby et al. (2008): for U.S., where the labor market is highly ‡exible and
transition rates are large, the convergence to the steady-state unemployment
is close to complete in a period of one quarter. For countries with more
rigid labor markets, i.e., small transition rates, the convergence is far from
complete and some adjustment must be made. In Brazil, as we show with
more detail in the Appendix, the data support the notion of fast convergence
and this kind of adjustment is less important: the averages of s, f and are
2.2%, 11.0% and 0.80 respectively using Dieese/Seade data and 22.8%, 2.8%
and 0.95 respectively using IBGE data.
3
DATA
We use data on job …nding and job separation rates from the Dieese/Seade
and IBGE databases. We also present worker ‡ow data based on CAGED.
3.1
DIEESE/SEADE
Dieese/Seade has conducted an unemployment survey for the Metropolitan
Area of Sao Paulo since 1985. It also has entered into an agreement with
other regional institutions for the calculation of unemployment rates, among
other statistics, for other …ve metropolitan areas since 1998. Due to data
availability, we use only data for the Metropolitan Area of São Paulo. Data
for other regions are also available from Dieese, but not with all the details
needed for this study. The other source of data for unemployment is IBGE
(further information on this data source will be given later). The problem
with using IBGE data is that they cover a short time period: the methodology
for calculating the unemployment rate changed in 2003 and data for previous
years can only be traced back to 2002. If we want to correlate the evolution of
6
labor market variables with the business cycle, this small sample size makes
the exercise simply impossible.
Although Dieese data relate to the Metropolitan Area of Sao Paulo only,
the region is big enough to make the sample representative of the Brazilian labor market. Figure 1 shows the evolution of the two unemployment
measures since the beginning of each sample.
Unemployment Rate From January 1985 to December 2008
Figure 1: Unemployment rate: Dieese-Seade and IBGE
Dieese usually separates unemployed workers into non-employed and marginally employed. We will treat both as the same thing. Despite wasting
part of the information, the analytical solutions for the model we will use
are clearer when the labor market has only two states (employed and nonemployed). Additional states can be addressed but at the cost of intractability of the model (Shimer, 2007). We use data for short-term unemployment
and short-term employment (unemployed or employed up to 1 month) to
construct the job separation and job …nding rates. Figure 2 depicts the evolution of these two variables since 1994. As can be seen, the job separation
7
rate is more volatile than the job …nding rate.
From 1Q94 to 4Q08
Figure 2: Dieese-Seade: job …nding and separation rates
3.2
IBGE
IBGE (the Brazilian Institute of Geography and Statistics) computes most
of the economic data for Brazil, including, GDP, industrial production, retail
sales, consumer prices and unemployment rate, among others. The advantage
of using IBGE data is that they have a wide coverage. Most of the data
computed by IBGE use samples that allow tracing the pattern for the whole
Brazilian economy, while most of the "shadow indicators" usually use a small
number of regions in their sample. The "Monthly Employment Survey"
conducted by IBGE is based on a sample of six metropolitan areas in Brazil,
including Sao Paulo and, therefore, it is more representative of the country’s
employment situation than the Dieese / Seade survey. The problem with
IBGE data is that there are recurrent changes in methodology. The current
Employment Survey methodology is based on a redesign performed in 2001.
The details we use are only available from 1Q02 onwards. We use data up
to 1Q09.
8
From 1Q02 to 1Q09
Figure 3: IBGE: job-…nding and separation rates
3.3
BRAZILIAN MINISTRY OF LABOR
Data from the Brazilian Ministry of Labor (MTE, in Portuguese) are also
used as a source. MTE computes formal3 gross job creation and job destruction every month. The disadvantage of this source of data is that it does
not include informal workers, which account for roughly 40% of the Brazilian
workforce. The main advantage is the accuracy with which each movement is
computed. Since MTE has the records of every formal worker, gross ‡ows are
readily available. MTE also computes layo¤s and quits. Since workers who
quit the job cannot make use of the FGTS deposit and are not eligible for
unemployment insurance, the reason for job destruction must be disclosed.
A possible drawback of this segmentation is a tacit agreement between employee and employer: the employer agrees to dismiss the worker even if the
worker is asking to quit the job. This may happen due to some kind of good
3
A formal job is one registered at MTE. Formal job positions must comply with all
brazilian labor laws. In particular, "formally" employed workers have several rights and
bene…ts beyond regular salary including, among others, remunerated vacation, FGTS deposits and thirteenth salary.
9
relationship between workers and employers. But even such cases can be
addressed: since most workers who are quitting a job already have another
job opportunity open, they will not …le for unemployment insurance. So we
can use …ling processes for unemployment insurance to estimate the total
number of layo¤s. A possible drawback of CAGED data is that any trend
towards formalization of workers will create an overestimation bias towards
job creation. Given the limitation of this source of data, we cannot address
this bias. In performing the calculation in the next section, we will be making the assumption that formalization is a slow movement with little variance
explained by the business cycles. Job creation and destruction, on the other
hand, are assumed to be cyclical. This implies that most of the variation in
job …nding rates is related to the cyclicality of job …nding. Besides, formalization is empirically found to be procyclical. This does not eliminate the
problem because the cyclical components may not be perfectly correlated,
but it at least reduces the upward bias in job …nding rate due to a trend toward formalization. In addition, we will use the "detrended" CAGED ‡ows
for most of the calculations. Assuming that the trends towards formalization
or "unformalization" are slow movements, we can remove part of this e¤ect
by using smoothing techniques such as the HP …lter.
10
In percentage of total
From 1Q98 to 4Q08
Figure 4: Formal jobs
11
Major ‡ows
Other ‡ows
From 1Q96 to 1Q09
Figure 5: Worker ‡ows
12
Our sample period for CAGED data goes from 1Q96 to 1Q09. The data
are quarterly averages of monthly data. All data were seasonally adjusted.
4
4.1
EMPIRICAL ANALYSIS
ECONOMIC ACTIVITY, GROSS FLOWS AND
JOB-TO-JOB TRANSITION
This section presents a brief detour from our objective of measuring job
…nding and job separation rates. Instead, we will use gross ‡ows to infer the
sensitivity of job creation and destruction to the economic cycle. As we have
already mentioned, the advantage of using worker ‡ows is the accuracy with
which the variables are calculated. Also, the separation between admissions,
layo¤s and quits provides better information regarding the kind of ‡ows we
can expect based on the information we have or expect about the economic
cycle.
To estimate the impact of output gap on these ‡ows we use a dynamic
model of labor market ‡ows similar to the one estimated by Blanchard and
Diamond (1990) and Davis and Haltiwanger (1999). Let ct ; dt and qt represent job creation (admissions), job destruction (layo¤s) and quits, respectively (all variables measured as a percentage of total employment) and be
other reasons for separation. We know that E = ct dt qt zt . Let
us assume that labor demand is a function of output, yt , and that there is
no labor supply shortage, i.e., we assume that total employment depends on
output but that output does not depend on labor supply. This hypothesis
embeds the notion that unemployment is always above its steady-state level.
Even though this is a strong hypothesis, it is not totally inconsistent. In our
sample period, unemployment has been very high for international standards
and grew during the 1990s. Only in the …nal sample period, particularly in
some moment between 2004/05 and 2007/08, were there some signs of lack
of workers, mainly of more educated and experienced ones. In our model,
we also assume that worker ‡ows are endogenous. Using the language of
Blanchard and Diamond (1990), all worker ‡ows can be seen as reallocation
shocks and the output gap is the only aggregate activity shock. Table 1
shows the statistical description of these data. Some variables labeled u in
the equation above deserve some comments. The …rst one is termination for
cause. The Brazilian labor laws give the employer the right to lay o¤ workers
with some kind of bad behavior. This cause represents a very small share
of total layo¤s and we will treat this variable as exogenous in the estimated
13
model. Other reasons are death and retirement. Death is close to constant
in our sample and will not be included in the analysis since its e¤ects can be
captured by a constant in the estimation. Retirement was high at the beginning of the sample and is close to constant at the end of the sample. This
re‡ects the large in‡ows of young workers in the labor force in the 1990s: the
average age of workers in Brazil decreased from 40 years in 1995 to 36 years
in 2008.
0
Let Yt = ct dt qt
denote the endogenous vector of employed pop0
ulation movements and Xt = yt zt
be a vector of exogenous variables,
including the measure of economic activity and other employed population
movements. Our empirical methodology consists in estimating a VAR model
of Y and X. With this model it is possible to estimate the impact of admissions shocks on quits, a measure we will use as a proxy for job-to-job
transition. This can be represented by:
Yt = A(L)Yt + B(L)Xt +
(13)
t
Table 1 indicates that most of the movements in labor force are accounted
for by admissions, layo¤s and quits. Termination for cause (df ), retirement
(r) and deaths (de) are many times smaller than the other variables. Despite
the smaller magnitude, some movements in these variables could originate
some rearrangement among the variables in Y , at least in some moments in
our sample. For instance, retirements were greater in the 1990s, and recently,
there has been a large increase in terminations for cause. Table 3 also reports
ADF unit root tests. The results indicate that it is not possible to reject the
unit root hypothesis in most of the series. Due to this fact, we will estimate
the VAR model with all variables represented by di¤erence from a time trend.
Our proxy for economic activity is the real GDP, also measured as a deviation
from trend.
Panel A
Average
St.Dev.
Max
Min
c
d
df
3:58% 2:66% 0:04%
0:50% 0:22% 0:01%
4:78% 3:37% 0:06%
2:77% 2:33% 0:04%
q
0:64%
0:14%
1:08%
0:45%
Table 1: Summary Statistics
14
r
0:01%
0:01%
0:03%
0:00%
de
0:01%
0:00%
0:01%
0:01%
Panel B
c d
df
c
1 0:86
0:12
d
1
0:06
1
df
q
r
de
q
r
de
0:91
0:81
0:61
0:86
0:73
0:64
0:25 0:36
0:58
1
0:63
0:39
1
0:76
1
Table 2: Cross-Correlations
Panel C
Level
ADF stat.
p-Value
Deviation from Trend
ADF stat.
p-Value
c
d
df
q
r
de
3:038
0:132
2:040
0:566
2:180
0:490
2:215
0:471
2:909
0:168
3:664
0:034
3:196
0:002
3:602
0:001
4:859
0:000
3:862
0:000
3:946
0:000
3:664
0:034
Table 3: Unit Root Tests
An important issue in analyzing impulse responses in VAR models of the
labor market as the one we are estimating is the identi…cation assumption.
Davis and Haltiwanger (1999) (henceforth DH) o¤er a detailed analysis of
identi…cation assumptions and their e¤ects on the impact of each shock in
the model. A di¤erence between our approach and that of DH is that we do
not treat any particular shock to our endogenous variables as an allocation or
aggregate shock. DH separate the evolution of job ‡ows between a measure
of job creation (POS) and job destruction (NEG). POS is constructed using
data on admissions and quits assuming a constant rate of quits replacement
of 85%, and NEG is based on layo¤s and quits. The VAR model of DH is
composed of these two variables and they postulate that particular terms of
the VAR innovations can be de…ned as aggregate shocks and other allocation
shocks. With these interpretations of the shocks, they analyze several kinds
of restrictions.
15
Figure 6: Impulse response: VAR(1) model
We will not follow this approach in our VAR model. Instead, we will
make use of minimal identi…cation assumptions. In particular, we will not
impose any kind of restriction on the way admissions and layo¤s are a¤ected
by quits and vice versa. Instead, we will include quits in the VAR model and
check the responses of these variables to shocks in admissions and layo¤s.
We also include a GDP growth in the VAR model as an exogenous variable.
In doing so, we use this variable as a measure of aggregate shock. We also
include a dummy variable for the 1996:1-1999:2 period. This period was
marked by a sensitive trend towards unformalization of workers8 and anemic
job creation. The reasons for this include increasing trade openness, which
induces the use of capital-intensive techniques, and also job losses in the
industrial sector due to external competition. Chahad et al. (1999) claim
that the appreciated exchange rate that prevailed during most of this period
contributed to a process of inverse import substitution, which a¤ected more
the formal jobs than the total employed population, as the industrial sector
is where formalization reaches its highest in the country.
16
To account for the e¤ect of formalization and also to bypass the unit root
present in our worker ‡ow measures, we estimated the model in di¤erence
from a time trend (using an HP …lter series for each variable). In this model,
we do not use the dummy variable we have just mentioned. Given the structure of our VAR model, we are prone to interpret the shocks to each variable
as allocation shocks. Given the small number of terminations for cause and
the constant number of dead workers in our sample period, we did not include
them in our exogenous vector . Another aspect of VAR model estimation is
the choice of lag length. Since di¤erent criteria (AIC, SBC, H-Q) lead to
di¤erent choices, we estimated the model with one and two lags. The results
are very similar and our choice is to deal only with the VAR(1) model. The
impulse response functions are presented in Figure 6. The inclusion of the
dummy variable and the retirement variable did not change the results in a
signi…cant manner and we opted to exclude them from the model.
The main …nding shown in Figure 6 is that admission shocks lead to a
signi…cant and positive impact on quits. Also from Figure 6 we can see that
all shocks are highly persistent. All impulse responses approach zero after
no more than three quarters. There are other signi…cant responses in Figure
6: the response of layo¤s to admissions, the response of layo¤s to quits and
the response of quits to layo¤s. Despite signi…cance in some quarters, each of
these responses seems small compared to the others we mentioned. Although
some of these responses can be of economic signi…cance, we can also think of
these results as indications of model misspeci…cation.
The VAR model also permits us to analyze the impact of GDP growth
on worker ‡ows. Since GDP growth was treated as an exogenous variable,
we can perform simulations of the impact of a 1% increase in GDP growth
without caring about the covariance matrix of residuals.
17
Impulse response: VAR(1) model
Accum. Impulse response: VAR(1) model
Figure 7: Output gap impact on worker ‡ows
The impact of a shock in GDP gap leads to an immediate increase in both
18
admissions and quits and also lead to a decrease in layo¤s. The accumulated
responses show an increase in layo¤s. This result is due to the net increase in
employment. This can be seen using the long-term impact or the static-long
run solution of the model. The long-term impact for any single dynamic
equation in the form A(L)zt = + B(L)yt + "t is given by = 1 +B(L)
. For
A(L)
the model we estimated, the long-term impact of changes in the output gap,
y, in the gross ‡ows rates, c, d and q are 0.120, -0.013 and 0.013 respectively.
This implies that job creation is much more sensitive to changes in GDP
than layo¤s. The static long-run solution for each variable is given by:
c = 0; 00008 + 0; 064y
(0;00025)
d=
0; 00012 + 0; 021y
(0;00011)
q=
(15)
(0;014)
0; 00005 + 0; 030y
(0;00009)
(14)
(0;033)
(16)
(0;010)
From the equations above, we can see that both admissions and quits are
procyclical while separations are acyclical. Note that none of the coe¢ cients
on layo¤s is signi…cant at reasonable levels.
A possible drawback of these results is the way the variables are measured.
All variables are measured as deviation from an HP …lter trend. This data
transformation is helpful once worker ‡ows seem to have a unit root and
also because there were some trends in the labor market that could not be
captured by the GDP growth alone. The recent trend towards formalization
and the "unformalization" in the 1990s are the most important examples
of such trends. Another possible drawback refers to the naive assumptions
used to estimate the VAR model. Davis and Haltiwanger (1999) and also
Blanchard and Diamond (1990) use a more "structured" VAR model to infer
the impact of di¤erent job market forces in modeling the dynamics of worker
‡ows. This could be a good re…nement of the present paper we expect to
address in future research. As we mentioned in the introduction, job market
analysis at a macroeconomic level is in its infancy in Brazil and we expect
some re…nements will arise after the release of the present paper.
19
4.2
UNEMPLOYMENT DECOMPOSITION AND THE
CYCLICALITY OF JOB-FINDING AND JOBSEPARATION RATES
In this section, we try to infer the importance of job …nding and job separation to determine unemployment dynamics. With the results from the
previous section at hand we could expect the major would come from the
job …nding rate. As we mentioned in the introduction, the major role of
job …nding rates is also the common …nding, documented in Shimer (2007),
Hall (2006), Petrongolo and Pisssarides (2009), Elsby et al. (2009), among
others. Before going into the unemployment variance decomposition results
we will make a cautionary note on the results. Other authors have already
pointed out that the di¤erence between actual unemployment rate and the
steady-state rate implied by job …nding and separation rates could be high.
This could arise, for instance, because job …nding and separation rates are
small and the convergence rate is slow. Another possible reason is measurement error. A …nal possibility is the failure of one of the hypotheses: the
constant labor force. In our data, this is particularly important in the 1990s:
labor force growth was close to 2.5% per year between 1994 and 2004 and
close to 0.5% from 2004 to 2008. Taking this into account, we also measure unemployment variation decomposition starting in 2000 for the longer
Dieese/Seade data. For IBGE data, we do not need to do this since the data
start in 2002. The di¤erence between actual and steady-state unemployment
are illustrated in Figure below. The measurement error was smoothed using
a four-quarter moving average of the series.
20
Absolute values
Percentage of s.s. unemployment
Figure 8: Deviation of unemployment from steady-state
The di¤erences between actual unemployment and implicit steady-state
21
are clearly larger at the beginning of the sample. Besides, Seade data have
a much larger deviation than IBGE data. Petrongolo and Pissarides (2009)
exclude from their sample observations where this di¤erence was larger than
10%. We will not do this here. Instead, we broke the larger Seade sample into
two and we warn the reader that the deviation is actually large in some parts
of each sample and the results should be read with caution. It is also worth
mentioning that the misalignments in both series are not much larger than
those described in the international literature. Shimer (2007), Petrongolo
and Pissarides (2009) and Elsby et al. (2008) made their data on job …nding,
job separation and actual unemployment available. Using these data, it is
possible to replicate the …ndings of these authors and calculate misalignment.
Performing such calculations for the data used by Shimer (2007) for the USA
leads to an average misalignment of around 4%, and for the data obtained by
Petrongolo and Pissarides (2009), we found an average misalignment of 6%
for the UK, 4% for France and 14% for Spain. These numbers are comparable
with average misalignments of 7% for Seade and 4% for IBGE. So, relative
to the average unemployment rate, misalignment in Brazil can be said to be
as small or large as that in other countries. Given that the unemployment
rate in Brazil was high in the period we cover, in absolute terms, Brazilian
misalignment is larger. To be more precise, average misalignment is 0.4%
and 1.2% for IBGE and Seade data, respectively. In the USA, the average
deviation in absolute terms is just 0.2%.
The decomposition of unemployment variance is depicted in Table 4. As
expected, the job …nding rate explains much of the cyclical variation in unemployment. The international evidence presented in Petrongolo and Pissarides
(2009) and Elsby et al. (2008) indicates a range from 0.5 to 0.9 for this parameter. The results we found also fall within this range. The size of the
error term is usually not reported and no comparison can be made. The
decomposition of IBGE unemployment data leads to a greater role for the
job …nding rate. This result is not due to the di¤erence in the sample period:
using the same period for Seade data would lead to values pretty close to the
ones already reported.
Period
Data Source
f
1Q94 4Q08 Dieese-Seade 68%
1Q00 4Q08 Dieese-Seade 65%
1Q02 1Q09
IBGE
85%
s
23%
33%
9%
Error
9%
2%
6%
Table 4: Unemployment variation decomposition
Another possible explanation is that in the period covered by IBGE
22
the unemployment rare presented a clear downward trend (steeper than the
Dieese trend). This explanation is not very compelling since several papers
already measured the contribution of these two variables in di¤erent economic situations: Elsby et al. (2009) analyze the behavior of job-…nding and
job-separation in the last ten U.S. recessions and …nds that the job-…nding
rate decreased relatively more than the job-separation increased in all cases;
Petrongolo and Pissarides (2009) computes the contribution of job-…nding
and job-separation for the same country in di¤erent moments and do not
…nd any indication that the contribution of job-separation increases when
the unemployment rate was moving up. In our opinion the most serious
problem with IBGE data is its short sample period and in this case the only
remedy is time.
5
CONCLUSION AND DIRECTION FOR
FUTURE RESEARCH
This paper assesses the determinants of unemployment ‡uctuation in Brazil
using three di¤erent datasets. The main …nding of this research is that job
…nding variation or admissions account for the majority of cyclical unemployment or worker ‡ow ‡uctuation in Brazil. Admissions are signi…cantly procyclical. Quits represent a large fraction of job separations and also present a
signi…cant procyclical behavior. Job separation is found to be countercyclical
but with less cyclical dynamics than admissions and layo¤s. The job …nding
rate accounts for 65 to 85% of unemployment variation while job separation
rate accounts for 9 to 33%.
Admissions present a strong procyclical pattern as do quits. Layo¤s, on
the other hand, seem to present a weak countercyclical pattern, and by some
measures, they could be considered acyclical. These results are based on a
vector autoregression (VAR) with naive assumptions. A possible direction
for future research would be to use more structural VAR as the ones presented in Davis and Haltiwanger (1999). Another possible re…nement of this
work would be to further look into the relation between admissions, layo¤s
and quits. Our estimation results indicate a signi…cant response of quits to
admission shocks. These results implicitly assume a constant relation to admissions and quits. Other authors claim that this replacement ratio would be
also procyclical. We did not address this possibility in this paper. Another
line of research using worker ‡ows could use the gross data instead of the
data as a percentage of employed workers. The use of percentage ‡ows is
attractive since it gives the elasticity of worker ‡ows the output gap, . The
23
disadvantage of using these variables is that they are much less volatile than
the output gap, which could render the standard errors of the estimates less
precise.
24
References
[1] Aguas, Marina; Valeria Pero and Eduardo P. Ribeiro (2008). Heterogeneidade no Mercado de Trabalho no Brasil: Deseprego e Inatividade
no Brasil. Encontro Nacional de Economia, 20
[2] Corseuil, Carlos and Luciana Servo (2006). Criação, Destruição e Realocação de Empregos no Brasil. IPEA, Rio de Janeiro, 2006.
[3] Davis, Steven and John Haltiwanger (1999). “On the Driving Forces behind Cyclical Movements in Employment and Job Reallocation”. American Economic Review, vol. 89 (5),1234 –1258.
[4] Davis, S.; J. Haltiwanger and S. Schuh (1996). Job Creation and Destruction. MIT Press, Cambridge, MA.
[5] Elsby, M., B. Hobijn and A. Sahin (2008). “Unemployment Dynamics
in the OECD”. Mimeo.
[6] Elsby, M.; R. Michaels and G. Solon (2009). “The Ins and Outs of Ciclycal Unemployment”. American Economic Journal: Macroeconomics,
vol. 1.
[7] Fallick, Bruce and Fleischman, Charles. “Employerto-Employer Flows in
the U.S. Labor Market: The Complete Picture of Gross Worker Flows.”
Finance and Economics Discussion Series Working Paper 2004-34, Board
of Governors of the Federal Reserve System, 2004.
[8] Fujita, S. and G. Ramey (2009). “The Cyclicality of Separation and Job
Finding Rate”. International Economy Review, may/09.
[9] Hall, R. (2005a). “Job Loss, Job Finding and Unemployment in US
Economy over the Past Fifty Years”. NBER Annual 2005.
[10] Hall, R. (2005b). “Employment E¢ ciency and Sticky Wages: Evidence
from Flows in the Labor Market”. Review of Economics and Statistics,
vol. 87 (3), 397 –407.
[11] Nagypál, Eva (2005). “Worker Reallocation over the Business Cycle:
The Importance of Job-to-Job Transitions.” Unpublished manuscript,
Northwestern University, 2005.
[12] Pentragolo, B. and C. Pissarides (2008). “The Ins and Outs of European
Unemployment”. American Economic Review, Papers and Proceedings,
2009.
25
[13] Shimer, R. (2007). “Reassessing the Ins and Outs of Unemployment”.
Mimeo, University of Chicago.
[14] Shimer, R. (2005). “The Cyclicality of Hires, Separations, and Job-toJob Transitions”. Federal Reserve Bank of St. Louis Review vol. 87 (4),
493 –507.
APPENDIX - International Comparison of
Job Transition Rates
This section presents sample statistics for job …nding, separations, quits and
other measures used in this paper for several countries. The aim of this
section is merely illustrative. As we will see, job transitions in Brazil are
larger than in most of OECD countries where the same statistics are available.
One possible explanation for this high mobility in Brazilian labor force is the
presence of informal jobs and self-employed workers, classes which are widely
known to be the most ‡exible. As mentioned earlier, informal jobs accounted
for nearly half of the jobs in Brazil in our sample period. Although this
feature can account for some part of the elevated worker mobility, it cannot
account for all of it. Corseuil and Servo (2006) compare worker mobility
using only the formal sector and also …nd Brazilian worker mobility to be
among the largest in a large sample of countries.
Figure A1 shows the dispersion of job …nding and separation rates for
several countries. The source of data for countries other than Brazil is Elsby
et al. (2008). It is clear from Figure A1 that Brazilian transition rates are
above the median, particularly for separation rates. This result is astonishing
given the several penalties a company must pay upon laying o¤ a worker in
Brazil.
26
Figure 9: Job …nding and job separation accross countries
A second feature of these data is the implicit rate of convergence. The
Elsby et al. (2008) measure of unemployment convergence to steady-state
equilibrium is given by t = 1 e 12(st +ft ) . Using the sample averages of
ft and st we can have a proxy for the rate of convergence across countries
and infer the importance of taking into account the slow convergence in
calculating the contribution of job …nding and job separation rates as drivers
of unemployment rate. Figure A2 shows the sample averages of convergence
rates calculated as t 1 e 12(st +f t ) :
27
Average speedy of convergence ( )
Figure 10: Average rate of convergence (theta)
28