Risk Attitudes and Occupations - Department of Economics

University of Zurich
Department of Economics
Center for Institutions, Policy and Culture in the Development Process
Working Paper Series
Working Paper No. 420
Risk Attitudes and Occupations: Self-Selection
or Adapting Preferences?
Sebastian Findeisen
April 2013
Risk Attitudes and Occupations: Self-Selection
or Adapting Preferences?
Sebastian Findeisen∗
University of Zurich, IZA
April 19, 2013
Abstract
This paper investigates the link between occupation and risk preferences. I
use data from the German Socio-Economic Panel to estimate a panel model of
income risk, controlling for unobserved heterogeneity and distinguishing permanent and transitory wage risk across occupations. Using a sample of young labor
market entrants who make occupational choices after or at the time risk preferences are elicited, I find no evidence for selection on risk attitudes. For a sample
of experienced workers, I show that the idiosyncratic history of exposure to past
income shocks negatively affects the willingness to take risks. Importantly this
remains robust to controlling for wealth and current exposure to risk, ruling out
differences in background risk as a potential explanation. Together with the evidence of no systematic selection, this suggests that idiosyncratic risk exposure
shapes preferences.
JEL-classification: J24, J31, J62
Keywords: Occupational Choice, Risk Preferences, Preference Adaption, Self-Selection
∗
Contact: [email protected]. Sebastian Findeisen acknowledges financial support from
the ERC Advanced Grant IPCDP-229883.
1
Introduction
The degree of risk-aversion is a pivotal parameter for answering a range of normative and positive questions in many fields such as macroeconomics, public finance, or
labor economics. For tractability and the lack of reliable evidence on the distribution
of risk-attitudes, the standard approach is the use of one common parameter, which
is the same across agents. Recently, economists have made progress to get a sense of
the degree and importance of heterogeneity in risk preferences. A popular method
is to try to elicit risk preferences directly with surveys.1 . Barsky, Juster, Kimball, and
Shapiro (1997) and Dohmen, Falk, Huffman, Sunde, Schupp, and Wagner (2011) are
two important studies in this regard using large scale representative surveys. Both papers uncover substantial heterogeneity in risk preferences and document correlations
with personal attributes like gender, age and height.
The aim of this paper is to empirically investigate how heterogeneous risk preferences are linked to individual occupations and, more broadly, labor market histories.
Using the German Socioeconomic Panel (GSOEP), I start by estimating a dynamic
panel model of income risk across a broad set of 28 occupations. The method allows
to distinguish between permanent and transitory shocks across occupations and provides a way to rank occupations with respect to their wage riskiness.
I focus on two margins. The first is on occupational sorting with respect to risk
preferences. Can differences in risk attitudes explain occupational choices? Do the
more risk-averse sort into low-risk jobs? A simple model with incomplete insurance
and heterogeneous risk preferences would predict these patterns, conditional on ability and no other sources of heterogeneity. A key challenge to answer these questions
is that risk attitudes are usually elicited only years after people have sorted into occupations. This puts a threat to identifying an effect from risk attitudes on occupation
choices to the extent that preferences are endogenous to one’s environment. I circumvent this by focusing on a sample of labor market entrants for which one observes risk
preferences before occupations are chosen. I find no evidence for the sorting hypothesis. Conditional on gender and education, the influence of risk attitudes on sorting
into occupations is estimated around zero.
I also examine how risk preferences and the riskiness of occupations is related
among experienced workers with labor market histories of over 20 years. Working in
occupations with large short-term income fluctuations is robustly associated with a
lower willingness to take risks, controlling for standard individual characteristics, importantly also including net-wealth. Recently, economists have become increasingly
1
Another approach is to use revealed preferences and exploit choices in markets, for example, on
insurance purchases. The distribution of risk-aversion can be inferred from structurally estimating
preference parameters, which are consistent with observed choices under an expected-utility framework (Cohen and Einav (2007)).
2
interested in how preferences might be shaped by an individual’s cultural and political environment or historical experiences, respective widely-known examples are
Guiso, Sapienza, and Zingales (2004), Fuchs-Schundeln and Alesina (2007) or Malmendier and Nagel (2011). I test whether an individual’s idiosyncratic history of exposure to wage risk influences the degree of risk-taking later in life. For the measure
of the individual history of exposure to wage risk, I use an individual’s occupation
history before risk-preferences are elicited. I document a pronounced negative effect
of past exposure on the current willingness to take risks, also controlling for wealth
and the current exposure to income risk. Taken together with the result that riskpreferences are no predictor of occupational choices, this suggests that risk-attitudes
are adapting in the direction that more exposure increases risk-aversion. Allowing for
varying weights on past risk exposure suggests that individuals place more weight on
early idiosyncratic experiences.
Related Literature. An important predecessor is the already mentioned article by
Dohmen et al. (2011). Using a field-experiment they show that the measure of risk
preferences I employ in this paper, namely the answers to the general risk question
in the GSOEP are a good predictor of risks taken in lotteries. Additionally, the same
risk questions also predicts stock-holdings, smoking and engagement in other risky
activities fairly well.2
Previous work has addressed the self-selection of individuals on risk attitudes
into occupations using quite different approaches. Fuchs-Schündeln and Schündeln
(2005) use a difference-in-difference design to exploit that occupational choice was
determined by largely political reasons and, therefore, was likely orthogonal to riskpreferences in the Eastern Part, but not so in Western Germany before reunification.
They distinguish between two types of occupations: public or private sector employment. Guiso and Paiella (2004) use Italian data on declared risk-attitudes and distinguish between being self-employed or not, and working in the public or private sector. Bonin, Dohmen, Falk, Huffman, and Sunde (2007) also use the GSOEP and distinguish between occupations at the three and two digit ISCO-88 classification level, as in
this paper. One of the main advantages of the approach I employ to test sorting is that
I use a sample of labor market entrants, who declare preferences before occupational
choices are made. This addresses endogeneity concerns, i.e. an effect of occupational
risk on risk-attitudes, and is, arguably, a more direct test of the sorting hypothesis.
Bonin et al. (2007) use the residual from Mincerian cross-sectional wage regressions
from the year 2004. I instead estimate a dynamic model of wage risk covering up to
25 years, controlling for unobserved individual heterogeneity and time-occupation effects. This follows the common approach to estimate income risk in macro and labor
2
In a related paper Jaeger, Dohmen, Falk, Huffman, Sunde, and Bonin (2010) show that individuals
with a higher willingness to take risks are also more likely to be mobile and change residence.
3
economics following, for example, Carroll and Samwick (1997), Meghir and Pistaferri (2004), Storesletten, Telmer, and Yaron (2004), and Guvenen (2009), and allows to
distinguish transitory and permanent risk. In addition, a dynamic model allows for
unemployment risk, potentially a major source for unpredictable income fluctuations.
This paper also contributes evidence how exposure to risk might shape risk preferences. Malmendier and Nagel (2011) show that macroeconomic experiences very
robustly explain differences in individual risk preferences and stock market participation across cohorts. The results here are complementary and indicate that idiosyncratic
experiences are also very likely to play a role in preference formation.
2
Occupational Income Risk
I define and estimate income risk as the variance of unpredictable changes in individual income. I disentangle permanent from transitory innovations to income, building
upon various contributions estimating US labor income risk using the PSID, among
them Carrol and Samwick (1997), Meghir and Pistaferri (2004), Storesletten, Telmer,
and Yaron (2004), and Guvenen (2009). The estimation procedure in this paper is
adopted from Carrol and Samwick (1997) and slightly expanded to attribute income
risk to a set of occupations.
2.1
Specification
Income Definition. The dependent variable is the log of yearly labor income of an
individual. About 30% of all individuals in the sample, report at least one unemployment spell in their working history. This indicates that a riskiness measure for
occupations should explicitly take unemployment risk into account. I, therefore, attribute UI benefits and assistance payments to the labor income of the individual in
the given year. During an unemployment spell, the last occupation held is assigned
to the individual, motivated by the fact that benefits are linked to past wages.
Income Process. The log of individual i’s labor income working in occupation j at
time t is assumed to be governed by the following process:
wijt = λi + αjt + βj · xijt + uijt ,
where λi is an individual fixed-effect, αjt is an occupation-time trend effect, and xijt
age and age squared, which are the only time-varying individual effects. Note that βj
is free to vary across occupations, implying that returns to seniority (experience) are
heterogeneous. Income variation in the data caused by different life-cycle patterns of
income across occupations are, hence, not contained in the error uijt . The variation
4
which is left to identify βj comes from time-varying occupations of given individuals
(occupation switchers) and variation within individuals over time.
Income Risk: Identification. I decompose the error into the sum of two components:
uijt = ωijt + ijt ,
(1)
where ωijt is the permanent component. ωijt is referred to as permanent, since innovations to its process have lasting effects on labor income, as captured by the random
walk:
ωijt = ωijt−1 + ηijt ,
where innovations ηijt are normal and uncorrelated over time and individuals, thus
ηijt ∼ N (0, ση2j ). Relatedly, the transitory shocks have no persistence and are drawn
from a normal distribution, so ijt ∼ N (0, σ2j ). Importantly for the purpose of this
paper, it is assumed that ση2j and σ2j are occupation specific. This specification and
the decomposition into permanent and transitory components follows previous work
estimating income risk using the PSID, adopted to the purpose of distinguishing wage
risk across occupations.
2.2
Estimation
Taking the n-th difference between two residuals uijt in (1Specificationequation.2.1)
yields:
∆n uijt = uijt+n − uijt = ηijt+1 + . . . + ηijt+n + ijt+n − ijt .
To fix ideas, suppose an individual never switches occupations. Then applying the
variance operator gives:
(2)
var [∆n uijt ] = nση2j + 2σ2j .
In their influential paper Carrol and Samwick (1997) propose to estimate a similar
specification like (2Estimationequation.2.2) by OLS and I follow their lead here. The
LHS variable is conveniently obtained by taking the (estimated) squared differences
(∆n ûijt )2 . ση2j and σ2j are then estimated as coefficients on the regressors n and 2.
Measurement error in the LHS will add noise and inflate standard errors, but will
not attenuate the estimates. To account for occupational switching, I estimate a more
general version of (2Estimationequation.2.2):
var [∆n uijt ] =
ση21
t+n
X
I1k + · · · +
k=t+1
=
J
X
j=1
"
ση2j
ση2J
t+n
X
IJk + σ21 (IJt+n + IJt ) + · · · + σ2J (IJt+n + IJt )
k=t+1
n
X
#
Ijk=1 + σ2j (IJt+n + IJt ) ,
k=1
5
Table 1: Occupations
Occupation Code
11
12
13
21
22
23
24
31
32
33
34
41
42
51
52
61
71
72
73
74
81
82
83
91
92
93
110
Legislators and Senior Officials (inc. senior gov. and party officials)
Corporate Managers (incl. responsibilities for certain divisions/departments)
General Mangers (responsibilities for whole companies, persons who manage enterprises on their own behalf)
Science Professionals
Life Science and Health Professionals
Teaching Professionals (includes professors, school teaching which requires high educational background)
Other Professionals (inc. accountants, lawyers and others)
Physical and Engineering Professionals
Life Science and Health Associate Professionals (includes medical assistance and nursing)
Teaching Associate Professionals (mostly primary education teaching)
Other Associate Professionals (includes estate agents and travel consultants)
Office Clerks
Customer Service Clerks
Personal Services Workers (includes personal care workers and barbers)
Shop Salespersons/Models
Agricultural and Fishery Workers
Extraction and Building Trade Workers
Metal, Machinery, and Related Trades Workers
Precision, Handicraft, Printing and Related Trades Workers
Other Craft and Related Trades Workers (includes textile, wood and processing)
Stationary-Plant and Related Operators
Machine Operators and Assemblers
Drivers and Mobile-Plant Operators
Sales and Services Elementary Occupations (incl. street vendors and door-to-door salesmen)
Laborers in Agriculture and Fishery
Laborers in Mining, Construction, Manufacturing and Transport
Armed Forces
where Ijk=1 is an indicator function, picking up the number of times a permanent
innovation from distribution j is drawn in between two periods with distance n.
2.3
Data
The (G)SOEP is a representative panel survey of the adult population in Germany.
The SOEP came into life 1984, and for this study all waves from 1984 to 2008 are used.
It surveys all members of a household on a wide range of economic and non-economic
topics. Importantly, information on the occupation held by an individual is included,
according to the International Standard Classification of Occupations (ISCO88) by the
International Labour Organization. Throughout the analysis, the two digit occupations code (plus the armed forces) will be used, differentiating between 28 occupations. Table 1 lists the different occupations. The earnings variable is the annual labor
income, derived from the main occupation of an individual, which includes all compensation received, i.e. wages, bonuses, commissions etc. from that source. Like
motivated above, if during any year an individual goes through a spell of unemployment, earnings include all benefits and government transfers linked to the labor
market status of the individual. The sample is restricted to males working full-time,
for which I observe a labor market history of at least 15 years (the maximum being
6
Table 2: Risk Across Occupations
Transitory Income Variance
Permanent Income Variance
1.
High Level Mangers
1.
2.
Sales and Services Elementary Occupations
(incl. cleaning personnel, garbage collectors)
Extraction and Construction Workers
2.
3.
26.
27.
28.
..
.
Customer Service Clerks
(incl. cashiers and receptionist)
Senior Government Officials Professional Politicians
(incl. Professional Politicians )
Teaching Professionals
(includes professors, high school teaching )
3.
26.
27.
28.
Drivers and Mobile-Plant Operators
(incl. Taxi Drivers)
Personal Services Workers
(includes restaurant services)
Teaching Professionals
(includes professors, high school teaching )
..
.
Customer Service Clerks
(incl. cashiers and receptionist)
Sales and Services Elementary Occupations
(incl. cleaning personnel, garbage collectors)
Primary Education Teachers
25). In particular, only those years are kept for which an individual is either working
full-time or unemployed. The analysis is, moreover, restricted to Western Germans,
since occupational choices in the former German Democratic Republic (GDR) were
severely restricted. This creates a non-balanced panel of 1838 individuals, with complete information for 19 years on average, and 34,831 observations in total. In the next
Section, I will link occupations to declared risk preferences. The Appendix contains
some more details on sample selection and variables.
2.4
Results
Table 2Risk Across Occupationstable.caption.2 ranks occupations according to their
estimated amount of permanent and transitory risk. The estimates reflect several intuitive patterns. High level managers are subject to large transitory income uncertainty, which might be explained by incentive based contracts. Another example is
the case of construction workers, who are often employed on short term contracts.
In contrast many senior government officials and teachers are tenured for life, minimizing transitory income fluctuations in these professions. Permanent shocks are
typically associated with layoffs, job mobility (within occupations), or health shocks.
Drivers and workers in gastronomy seem to be strongly affected by those permanent
shocks. In contrast and perhaps surprisingly, a few low-skilled service occupations
like garbage collecting are associated with low transitory and permanent uncertainty.
This may be explained by the fact that I focus on males staying attached to the labor market the whole time, so conditional on constant labor force participation these
7
occupations bring relatively speaking little risk.3 In general and consistent with previous estimates of permanent and transitory shocks, the magnitude of the transitory
variances is much larger; the mean of σ2j is 0.06 versus 0.006 for ση2j .4 Note that because labor income as it is defined here already includes some degree of insurance
against unemployment and short-term health risks, the estimated parameters capture
residual risk not covered by government transfer programs.
3
Risk Attitudes and Occupational Sorting
3.1
Occupational Choices and Risk Attitudes For Labor Market Entrants
In this Section I test if risk preferences are a good predictor of occupational choices.
Simple models featuring imperfect insurance and preference heterogeneity would
predict this kind of sorting as an equilibrium outcome when holding the skill level of
workers constant. Suggestive evidence for a positive sorting pattern has been found,
for example, by Schulhofer-Wohl (2011) for the case of aggregate shocks to the US
economy and by Bonin et.al. (2007), who use the cross-sectional residuals from Mincerian regressions as a measure for occupational earnings risk. Importantly and in
contrast to other papers, my focus is on young individuals who made first-time occupational choices after or in the same year risk preferences were elicited. This addresses the potential endogeneity of risk preferences, which is dealt with in the next
Section. I use two different strategies to estimate the relationship between preferences
and occupations. First, I use the unpredictability of income streams across chosen occupations as the dependent variable, and risk attitudes among with education and
other variables as controls. Second, a multinomial logit model of occupational choice
is estimated. I then use the predicted probabilities to generate a predicted occupational income risk profile for each individual. Finally, I test whether these individual
profiles are correlated with risk attitudes in a regression framework.
3
Also maybe surprisingly, teaching professional seem to associated with low transitory but high
permanent risk. This may be due to tenure decisions, which have long-lasting effects wage effects.
4
The welfare consequences of transitory and permanent shocks depend, of course, to what extent
they can be insured. For example, in a well-known paper Levine and Zame (2002) show that selfinsurance with a riskless bond is a very effective device to smooth transitory shocks in an incomplete
markets world. Empirically, Blundell, Pistaferri, and Preston (2008) find in a recent influential paper
that transitory shocks seem to be very well absorbed and do not translate into significant fluctuations
in consumption for the US.
8
Table 3: Risk Attitudes and Riskiness of Occupation Among Labor Market Entrants
Dependent Variable:
Transitory Shocks Occupation
(1)
(2)
Permanent Shocks Occupation
(3)
(4)
Willingness To Take Risks
0.81
(0.70)
-0.09
(0.09)
0.12
(0.72)
-0.01
(0.09)
Sex
-1.45***
(0.28)
-0.01
(0.04)
Education (Degree)
-0.46***
(0.13)
-0.05***
(0.02)
0.55
(0.33)
Yes
786
-0.07
(0.05)
Yes
786
Age
Region Fixed Effects
Observations
No
803
No
803
OLS Estimates. Robust Errors. ***, **, * indicate significance at 1%, 5%, and 10% level. Coefficients
for the willingness to take risks and age as well as corresponding standard errors multiplied by 103 ,
coefficients for sex and degree as well as corresponding standard errors multiplied by 102 . Constant
included in regressions.
3.1.1
Risk Preferences and Risk in Chosen Occupations
The simplest test for sorting is to look at the correlation between occupational earnings risk and risk preferences. Table 3Risk Attitudes and Riskiness of Occupation
Among Labor Market Entrantstable.caption.3 presents the results of regressing σ2j
and ση2j on the measure of risk attitudes and controls.5 Risk-attitudes are measured
on a discrete 0-10 scale, with higher values indication a higher risk tolerance. More
details are provided in the appendix and Dohmen et al. (2011).
The first column shows a positive but insignificant correlation between the attitude
to take risks on the probability of choosing an occupation with higher transitory risk.
Controlling for education, age, gender and region drives the estimate towards zero.6
The results show that male, less educated, and older labor market entrants sort into
high transitory risk environments. The marginal effect of being male corresponds to
5
Estimating ordered multinomial models or generalized Tobit models (interval regressions), using
the measures of income risk in the chosen occupation as censored dependent variable gives the same
patterns; results are available on request.
6
Region here is the at the state level (Bundeslaender). Education is measured by the highest degree obtained consistent with the ISCED-1997 definition. The classes are: (0)’none yet’ (1)’general
elementary’ (2)’middle vocational’ (3)’vocational + German Abitur’ (4)’higher vocational’ (5)’higher
education’.
9
Table 4: Multinomial Logit Model of Occupational Choice
Variable:
χ2 (23)
P-Value
Willingness To Take Risks
Sex
Education (Degree)
Age
Region Fixed Effects
28.47
397.37
1019.22
57.09
22.30
0.20
0.00
0.00
0.00
0.50
moving from the 25th to the 50th percent quantile, obtaining a higher degree reduces
transitory occupational income risk by moving from the 40th to 30th percent quantile.
Also for permanent occupational income risk there exists no evidence in favor of
sorting on risk preferences. The coefficients are estimated negatively and imprecisely.
Again, education can be seen as a good insurance device against labor market risk:
an additional degree is associated with a reduction in permanent income risk, going
from roughly the 40th to the 30th percent quantile of the occupation risk distribution.
In unreported results, I aggregate permanent and transitory risk by adding them with
varying weights. Also in this exercises there is no robust relationship between income
risk and risk attitudes.
3.1.2
Risk Preferences and Predicted Income Risk
I next use a more indirect approach to estimate the preference-occupation risk relationship. Table 3Risk Attitudes and Riskiness of Occupation Among Labor Market
Entrantstable.caption.3 summarizes the main outcomes of estimating a multinomial
logit model for the choice stage.7 Displayed are the critical and p-values for the hypothesis that all alternative specific coefficients are zero. Age, degree and gender are
robust predictors of occupational choices. Next, with the predicted probabilities I
construct expected occupation risk income measures for every individual:
E[ση2i ] =
X
p̂ij ση2j
j
and analogously for transitory risk σ2j . I then test, if estimated parameters are robustly related to risk attitudes by regressing E[ση2i ] and E[σ2i ] on measured risk attitudes and controls, as displayed in Table 5Risk Attitudes and Predicted Riskiness of
Occupation Among Labor Market Entrantstable.caption.5. Across the board the co7
To gain precision, three occupations chosen by less than 10 individuals are omitted in the model.
10
Table 5: Risk Attitudes and Predicted Riskiness of Occupation Among Labor Market
Entrants
Dependent Variable:
Transitory Shocks Occupation
(1)
(2)
Permanent Shocks Occupation
(3)
(4)
Willingness To Take Risks
0.82***
(0.02)
-0.08***
(0.02)
0.17*
(0.09)
-0.11***
(0.01)
Sex
-1.50***
(0.04)
-0.01***
(0.00)
Education (Degree)
-0.47***
(0.13)
-0.05***
(0.00)
Age
0.59***
(0.04)
Yes
786
-0.07***
(0.01)
Yes
786
Region Fixed Effects
Observations
No
803
No
803
OLS Estimates. Robust Errors. ***, **, * indicate significance at 1%, 5%, and 10% level. Coefficients
for the willingness to take risks and age as well as corresponding standard errors multiplied by 103 ,
coefficients for sex and degree as well as corresponding standard errors multiplied by 102 . Constant
included in regressions.
efficients remain very similar to the ones in Table 3Risk Attitudes and Riskiness of
Occupation Among Labor Market Entrantstable.caption.3. Higher risk tolerance is
now weakly significantly associated with higher expected transitory income risk. In
contrast, for expected permanent occupational risk the correlation with risk tolerance
is negative, contradicting sorting on risk preferences.
4
Risk Attitudes and Occupations Among Experienced
Workers
In summary the evidence from the previous Section provides no support for the hypothesis that risk preferences predict sorting into occupations with respect to wage
riskiness. I now turn to the empirical relationship of occupational wage risk and risk
attitudes among experienced labor market participants; in particular, I use the same sample of workers, which was used to estimate income risk across occupations. Risk
preferences were first elicited in 2004, implying an average of 48 years and about 25
years of labor market experience. In contrast to the sample of entrants, preferences for
11
the experienced may have been shaped by past risk exposure.8 For this group, it is not
clear, whether any correlation between risk attitudes and held occupations is driven
by self-selection or adapting preferences. Given the evidence from the previous Section that the data seems to reject sorting on preferences for labor market entrants, this
Section puts the adapting preference channel into focus.
4.1
Risk in Main Occupation
I start by assigning one main occupation to each individual of the sample. For occupational switchers, I use the mode of occupations held over the working life. In the
next Section a finer measure of exposure to income risk is employed, taking switching into account. Table 6Risk Attitudes and Riskiness of Occupation Among Old
Workerstable.caption.6 displays the result of an ordered Probit model, where the dependent variable is the willingness to take risks by the individual.9 Displayed are
marginal effects on the probability of declaring the median answer to the risk question, evaluated at sample means. Significant levels and signs are the same as for the
untransformed coefficients of the ordered Probit specification.
The first column shows that working in an occupation with higher transitory income risk is very robustly associated with a lower willingness to take risks. This is not
true for permanent occupational income risk, in contrast, as becomes obvious from
column two. The coefficient is positive but extremely imprecisely estimated. Column
3 confirms these pattern in a specification including both shocks. The coefficient on
transitory risk is hardly affected, whereas the coefficient on permanent risk switches
sign and is very imprecisely estimated. The specification in Column four controls for
wealth, age and education leaving the conclusions regarding the role of transitory
shocks unaffected. The signs for wealth and age are expected and consistent with
previous results by Dohmen et al. (2011).
The finding that people who have worked in a risky occupation for most of their
life are less willing to take on risks may be driven by the notion of background risk.
Having worked in risky occupation is correlated with currently or in the future working in a risky occupation. This may lower the willingness to take on risks even in a
completely neo-classical model with stable preferences, as long as insurance markets
are incomplete or there are liquidity constraints. Guiso and Paiella (2008) use Italian
survey data to make this point empirically; Eeckhoudt, Gollier, and Schlesinger (1996)
investigate theoretically under what kind of changes in background risk one should
8
In recent work, Malmendier and Nagel (2011) use the Survey of Consumer Finances and establish
a sound link between experienced macroeconomic history and financial risk taking, as well as risk
attitudes.
9
All results are robust to estimating an Ordered Logit instead.
12
Table 6: Risk Attitudes and Riskiness of Occupation Among Old Workers
Dependent Variable: Willingness To Take Risks
(1)
(2)
(3)
(4)
(5)
Transitory Shocks in Main Occupation
-30.99***
(10.92)
-32.20***
(11.69)
-31.70***
(12.34)
-43.01**
(18.12)
-18.97
(64.85)
9.37
(70.13)
104.54
(100.17)
Wealth
5.67
(21.64)
6.52
(21.61)
Age
-0.20***
(0.04)
-0.20***
(0.04)
Education (Degree)
0.62***
(0.24)
0.64***
(0.24)
14.32
(16.18)
Permanent Shocks in Main Occupation
44.46
(60.64)
Transitory Shocks
Current Occupation
Permanent Shocks
-123.84
Current Occupation
(99.58)
Observations
1463
1463
1463
1285
1285
Ordered Probit Estimates. Marginal Effects (times 100) on answering 5 to risk question, evaluated at
sample mean. Robust Errors based on Huber/White/sandwich estimator. ***, **, * indicate significance
at 1%, 5%, and 10% level. Wealth scaled by 107 .
expect changes in risk aversion. Column five adds wage risk in the current occupation
at the time when risk preferences were elicited. Jointly with wealth and the human
capital level (degree and age), which controls for permanent income, this specification
takes into account, the ability of an individual to insure themselves against additional
risk. The main conclusions regarding the influence of past exposure on current risk
attitudes remain unaffected. By the high correlation between main occupation and
current occupation, standard errors increase significantly because of collinearity.10
Finally, in unreported results, the data shows no relationship between the riskiness
of the first observed occupation for old workers and risk preferences.11 This provides
evidence that it is not different selection patterns of old cohorts into occupations,
which drive the correlations but rather the risk exposure channel.
10
In unreported result, I exclude riskiness in one’s main occupation from the model. Higher risk
in one’s current occupation predicts that risk taking decreases, consistent with theories of background
risk.
11
The first observed occupation does not necessarily coincide with an individual’s real first occupation since some individuals enter the SOEP after their labor market entry.
13
4.2
Exposure Measures Accounting For Occupation Switching
Average Exposure. The previous specifications associated the exposure to income risk
with the income risk in one’s main occupation, which was defined as the mode occupation. A different approach is to take occupational switching into account. Indeed,
the median number of occupations an individual holds is three and at the 75% quantile it is five. To get a better measure of risk exposure over the working life, I start by
computing wage risk averages as:
Ti
1 X
σ2 ,
Ei =
Ti t=1 j(t,i)
where Ti is the number of periods an individual has worked up to answering the risk
2
question and σj(t,i)
is the variance of innovation in occupation j the individual i holds
in t. This is done separately for permanent and transitory innovations.
Table 7Risk Attitudes and Riskiness of Occupation Among Old Workerstable.caption.7
shows the results of estimating the same Ordered Probit models using this measure.
The patterns are almost identical. Workers with histories of holding occupations with
larger transitory income shocks are less risk taking, controlling for wealth, age and
education. Consistent with the hypothesis that the history of idiosyncratic risk exposure shapes preferences, the effects for transitory risk gets stronger, as the measure
of exposure gets better vis-a-vis Table 6Risk Attitudes and Riskiness of Occupation
Among Old Workerstable.caption.6. The estimates for the exposure to permanent occupational income risk suffer from high standard errors and show no robust pattern.
Flexible Weighted Average of Exposure. Next, I allow for the possibility that the
timing of risk exposure may influence declared risk attitudes. For example, early
experiences on the job may carry an especially important role in risk preference formation. On the other early labor market shocks might be forgotten and more recent
events might have a bigger influence on declared risk preferences. The histories of
transitory and permanent wage risk exposure are now calculated as weighted averages:
Ti
X
∗
2
Ei =
ωit σj(t,i)
,
t=1
where the weights ωit =
Lλ (t)
PT i
λ
t=1 L (t)
are a function of the time lag relative to when the
risk question was answered and λ is a free parameter, controlling if early or more recent wage riskiness carries more weight. λ > 0 implies that early exposure to income
risk matters comparatively more, whereas λ < 0 implies the opposite. Note the specification nests the previous case with average exposure (λ = 0). A similar specification
14
Table 7: Risk Attitudes and Riskiness of Occupation Among Old Workers
(1)
History Exposure To Transitory Shocks
Dependent Variable: Willingness To Take Risks
(2)
(3)
(4)
(5)
(6)
-35.24***
(13.52)
-36.30**
(14.48)
-31.72**
(16.55)
-48.90*
(26.73)
-42.19**
(18.50)
-16.90
(80.90)
26.30
(89.98)
210.87
(151.48)
97.08
(104.09)
Wealth
8.05
(22.59)
10.00
(22.87)
8.19
(22.84)
Age
-0.20***
(0.04)
-0.21***
(0.04)
-0.21***
(0.04)
0.62**
(0.25)
0.63**
(0.25)
0.56**
(0.25)
Transitory Shocks
Current Occupation
16.78
(18.31)
4.47
(12.38)
Permanent Shocks
Current Occupation
-183.15
(118.16)
-93.76
(84.17)
0
2.55
(fitted)
History Exposure To Permanent Shocks
57.56
(75.60)
Education (Degree)
0
λ
0
0
0
Observations
1463
1463
1463
1285
1285
1285
Ordered Probit Estimates. Marginal Effects (times 100) on answering 5 to risk question, evaluated at
sample mean. Robust Errors based on Huber/White/sandwich estimator. ***, **, * indicate significance
at 1%, 5%, and 10% level. Wealth scaled by 107 .
is employed by Malmendier and Nagel (2011), whose lead I follow here. λ is chosen
such that the likelihood function implied by the specification involving all controls is
maximized.12
Column six shows that the best model fit is obtained for λ = 2.55, implying that
labor income risk beard at the beginning of one’s career seems to be of more importance in preference formation than more recent exposure. The estimated coefficients
remain stable for the exposure to transitory risk, education, age, and wealth.13
12
I use a two step procedure. I maximize the likelihood functions over a λ grid and then pick the λ
giving the best fit.
13
A question I do not adress in this paper, is why transitory risk seems to be more important in
shaping preferences. A potential explanation is that although permanent shocks have a bigger long run
impact and are less insurable, they hit less individuals. For example, layoff risk may only be perceived
as such by an individual, when that individual or someone close to him in the same occupation is
actually laid off.
15
5
Conclusion
Heterogeneity in risk preferences is a prevalent feature of survey data. This paper
has two main contributions. First, I investigate, if the observed cross-sectional variation in risk attitudes can explain selection into occupations with respect to wage risk;
importantly I use a sample of young workers who enter the market after or in the
same year they declare their preferences. No robust sorting effect can be detected.
Second, I test if personal experiences or idiosyncratic past exposure to labor market
risk, has an influence on risk attitudes, using a sample of experienced workers, who
declare preferences after having participated in the labor market for over 20 years on
average. The data show a robust negative relation between idiosyncratic exposure to
transitory wage risk on risk preferences, conditional on demographics and ruling out
background risk by controlling for wealth as a potential explanation.
This paper conducted an indirect test for the reaction of preferences to variation
in an individual’s economic environment. An alternative approach would be to use
repeated survey exploiting variation in risk attitudes within individuals over time.
This is left as a further area for research, when such panels become available.14
Acknowledgments.
I am grateful to Matthias Doepke, Fabian Krueger, Uwe Sunde, Rainer Winkelmann, Christoph Winter, Eva Berger and seminar participants in Ammersee (IZA
European Summer School in Labor Economics) and Zurich for comments, and especially to Fabrizio Zilibotti for constant advice and support. The data used in this
paper was extracted using the Add-On package PanelWhiz for StataÆ. PanelWhiz
(http://www.PanelWhiz.eu) was written by Dr. John P. Haisken-DeNew ([email protected]).
See Haisken-DeNew and Hahn (2006) for details. The PanelWhiz generated DO file
to retrieve the data used here is available from me upon request. Any data or computational errors in this paper are my own.
References
A NDERSEN , S., G. H ARRISON , M. L AU , AND E. R UTSTRÖM (2008): “Lost in State
Space: Are Preferences Stable?,” International Economic Review, 49(3), 1091–1112.
14
Direct evidence on the stability of risk preferences is very scarce. Over a 17-month time span,
Andersen, Harrison, Lau, and Rutström (2008) use a field experiment and report very little variation
in revealed attitudes to risk over time, finding both increases and decreases.Also the SOEP provides
a small panel dimension for risk preferences, as some people also answered the question in 2006 and
2008. I find very tiny variation in stated risk preferences over this time frame.
16
B ARSKY, R., F. J USTER , M. K IMBALL , AND M. S HAPIRO (1997): “Preference parameters and behavioral heterogeneity: An experimental approach in the health and
retirement study,” Quarterly Journal of Economics, 112(2), 537–579.
B LUNDELL , R., L. P ISTAFERRI , AND I. P RESTON (2008): “Consumption inequality and
partial insurance,” The American Economic Review, pp. 1887–1921.
B ONIN , H., T. D OHMEN , A. FALK , D. H UFFMAN , AND U. S UNDE (2007): “Crosssectional earnings risk and occupational sorting: The role of risk attitudes,” Labour
Economics, 14(6), 926–937.
C ARROLL , C., AND A. S AMWICK (1997): “The nature of precautionary wealth,” Journal of monetary Economics, 40(1), 41–71.
C OHEN , A., AND L. E INAV (2007): “Estimating Risk Preferences from Deductible
Choice,” The American Economic Review, 97(3), 745–788.
D OHMEN , T., A. FALK , D. H UFFMAN , U. S UNDE , J. S CHUPP, AND G. WAGNER
(2011): “Individual Risk Attitudes: New Evidence from a Large, Representative,
Experimentally-Validated Survey,” Journal of the European Economic Association.
E ECKHOUDT, L., C. G OLLIER , AND H. S CHLESINGER (1996): “Changes in background
risk and risk taking behavior,” Econometrica, pp. 683–689.
F UCHS -S CHUNDELN , N., AND A. A LESINA (2007): “Good-Bye Lenin (Or Not?): The
Effect of Communism on People’s Preferences,” American Economic Review, 97(4),
1507–1528.
F UCHS -S CHÜNDELN , N., AND M. S CHÜNDELN (2005): “Precautionary Savings and
Self-Selection: Evidence from the German Reunification Experiment,” The Quarterly Journal of Economics, 120(3), 1085–1120.
G UISO , L., AND M. PAIELLA (2004): “The Role of Risk Aversion in Predicting Individual Behaviors,” CEPR Discussion Paper No. 4591.
(2008): “Risk aversion, wealth, and background risk,” Journal of the European
Economic Association, 6(6), 1109–1150.
G UISO , L., P. S APIENZA , AND L. Z INGALES (2004): “Does Local Financial Development Matter?,” Quarterly journal of Economics, 119(3), 929–969.
G UVENEN , F. (2009): “An empirical investigation of labor income processes,” Review
of Economic Dynamics, 12(1), 58–79.
17
J AEGER , D., T. D OHMEN , A. FALK , D. H UFFMAN , U. S UNDE , AND H. B ONIN (2010):
“Direct evidence on risk attitudes and migration,” The Review of Economics and
Statistics, 92(3), 684–689.
L EVINE , D., AND W. Z AME (2002): “Does market incompleteness matter?,” Econometrica, 70(5), 1805–1839.
M ALMENDIER , U., AND S. N AGEL (2011): “Depression Babies: Do Macroeconomic
Experiences Affect Risk Taking?,” The Quarterly Journal of Economics, 126(1), 373.
M EGHIR , C., AND L. P ISTAFERRI (2004): “Income variance dynamics and heterogeneity,” Econometrica, pp. 1–32.
S CHULHOFER -W OHL , S. (2011): “Heterogeneity, risk sharing and the welfare costs of
idiosyncratic risk,” Journal of Policitcal Economy, pp. 925–958.
S TORESLETTEN , K., C. T ELMER , AND A. YARON (2004): “Cyclical dynamics in idiosyncratic labor market risk,” Journal of Political Economy, pp. 695–717.
Appendix
Table 8: Estimates of Persistent and Transitory Income Risk
Occupation Code
11
12
13
21
22
23
24
31
32
33
34
41
42
51
Permanent Shock
0.011
0.0051
0.0132
0.0034
0.0023
0.0164
0.0020
0.0080
0.0064
0.000
0.0026
0.0046
0.0006
0.0173
Transitory Shock
0.0035
0.0430
0.2648
0.0456
0.0078
0.0000
0.0442
0.0759
0.0652
0.0571
0.0665
0.0620
0.0071
0.0278
Occupation Code
52
61
71
72
73
74
81
82
83
91
92
93
99
110
Permanent Shock
0.0030
0.0083
0.0031
0.0042
0.0037
0.0073
0.0049
0.0091
0.0183
0.0000
0.0007
0.0064
0.0000
0.0066
Transitory Shock
0.0571
0.0343
0.0972
0.0952
0.0676
0.0887
0.0581
0.0833
0.0471
0.1145
0.0237
0.0896
0.0874
0.0097
Based on pooled regression with 322677 observations.
Variables
Measurement of Risk Preferences
The risk question was asked in 2004, 2006 and 2008. For every individual, I use the
first non-missing answer. For the sample of old workers out of initially 1832 per18
sons, 1634 answer the risk question at least once. Answers are measured on a elven
point scale from 0-10. From Dohmen et a. (2011) the wording of the question is:
ìHowdoyouseeyourself ?Areyougenerallyapersonwhoisf ullypreparedtoakerisksordoyoutrytoavoid
R heirpaperalsoprovidesbehavioralvalidationof thequ
andthevalue10means´verywillingtotakerisks.T
experiment.
Education
Education is measured by the highest degree obtained consistent with the ISCED-1997
definition. The classes are: (0)’none yet’ (1)’general elementary’ (2)’middle vocational’
(3)’vocational + German Abitur’ (4)’higher vocational’ (5)’higher education’
Wealth
Wealth data is available for years 2002 and 2007. For individuals with data available
for two years, I use the average. Otherwise, I use the period which is non-missing.
Wealth measures net wealth and includes the value of own property (housing) minus
the debt on own property, financial assets/debt, business assets and private pension
claims.
Sample of Labor Market Entrants
I keep all individuals who are under the age of 35 when answering the risk question
in 2004, 2006 and 2008. From this sample, I keep everybody who entered the labor
market in the year or after the risk question was answered the first time. I define
entering the labor market by starting to work full-time, defined as declaring to work
full-time in 2004 or later but not the year before. For all variables, I use the first date
of availability starting from 2004.
Sample of Experienced Workers
The sample is the same used for the estimation of income processes across occupations
and time. For all answers to the risk question I use the average over the years 2004,
2006 and 2008 and the round to an integer and for wealth averages over the years
2002-2007. Age is used from the year 2004.
19