Вы находитесь на странице: 1из 21

J Popul Econ (2004) 17:583–602 DOI 10.1007/s00148-004-0187-0

Popul Econ (2004) 17:583–602 DOI 10.1007/s00148-004-0187-0 Does time preference change with age? David M. Bishai

Does time preference change with age?

David M. Bishai

Department of Population and Family Health Sciences, Johns Hopkins University, Bloomberg School of Public Health, 615 N. Wolfe St., Baltimore, MD 21205, USA (Fax: +1-410-955-2303; e-mail: dbishai@jhsph.edu)

Received: 1 March 2002/Accepted: 6 October 2003

Abstract. This study looks at compensating differentials in the National Longitudinal Survey of Youth (NLSY) to derive estimates of the levels of time preference for labor force participants in each of 15 waves of data from 1979 to 1994. With these estimates the evolution of time preference over the life course is described. Future utility among labor force participants appears to be valued more highly by subjects who are older, more schooled, white, or male. Controlling for schooling level, a higher IQ is associated with a pref- erence for more immediate rewards. If social rates of time preference are correlated with individual rates of time preference then population aging could create intergenerational asym- metries in the social rate of time preference. This phenomenon could make the optimal investments of young popula- tions appear selfish to future generations that are older.

JEL classification: H43, J31, J28

Key words: Time preference, compensating differentials, aging

1. Introduction

The work of Thaler and Rosen (Thaler and Rosen 1976) and later Moore and Viscusi (Viscusi and Moore 1989; Moore and Viscusi 1990) established that

Support from Hopkins Population Center (R24) is gratefully acknowledged. I am thankful for helpful comments received from Jim Burgess, Mike Grossman, Bruce Hamilton, Fritz Laux, Athanasios Orphanides, Katie Roche, and Eric Slade. All errors are my own. Responsible editor:

Alessandro Cigno.

584

D. M. Bishai

observing labor markets where subjects trade higher wages in exchange for higher occupational fatality risk could be a useful means of measuring rates of time preference. An implicit assumption in this work is the accurate perception of fatality risk by the marginal worker. The term ‘‘time preference’’ refers to a subject’s capacity to value future events. The rate of time preference is defined as –log(b), where b is the familiar discount rate ranging from 0 to 1. The formation of consistent plans for inter-temporal consumption places restrictions on the relationship between future satisfaction and current sat- isfaction. In theory, there is little reason for an optimal consumption plan to change when nothing has changed but calendar time. Strotz (1956) described the alternative conditions that were necessary for time consistency, 1) The discount function is an exponential function of the time which would elapse between the present and future consumption; or 2) The discount function is purely a function of calendar time; or 3) The discount function is a function of other values whose evolution is completely determined by calendar time (Strotz 1956). Yet, if an individual’s discount rate expresses features of an evolving personality there may be lifecycle effects that violate Strotz’s con- ditions. The assumption of dynamic consistency and constant exponential dis- counting is a staple of models of inter-temporal choice. The assumption is defended by suggesting that agents can have perfect foresight into the future dynamic evolution of their discount rates and employ pre-commitment strategies to enforce adherence to an optimal plan. In reality the question of whether human subjects have discount functions that support dynamically consistent planning is an empirical one. There has been little evidence to support constant exponential discounting (Thaler 1981; Ainslie and Haslam 1992; Loewenstein and Prelec 1992; Cairns and van der Pol 1997). This has led to greater consideration of alternative discounting models and a concern that dynamic inconsistency, if true, would invalidate basic assumptions required to apply rational choice models to environmental, health, and safety regulations (Loewenstein 1996). One important form of dynamic inconsistency is the possibility that the cognitive limitations of young persons might render them incapable of foreseeing changes in their time preference rates that will render them more concerned with their future welfare when they are older (Laux 2000). If human psychological development during the life course imposes dynamic inconsistency, it would offer an alternative explanation for the wave of regret that accompanies retrospection on the indiscretions of youth. Accounts of ‘‘regret’’ require an unforeseeable shift in the recognition of potential harm from an activity. Orphanides and Zervos explain the regret from addictive behavior as rooted in unforeseeable shifts in the potential of young people to become addicted (Orphanides 1994). The model of Orphanides would not explain why other non-addictive behaviors of youth lead to regret. However, unforeseeable shifts in time preference would account for regret of non-addictive youthful behaviors such as unsafe sex, petty crime, and delinquency. The optimal social rate of time preference governs a society’s rate of investment and under a golden rule strategy would be set equal to the sum of productivity growth and population growth (Marini and Scaramozzino 2000). In market economies society’s actual rate of investment results from an

Does time preference change with age?

585

aggregate of individual investment decisions governed by individual rates of

time preference. Thus, the actual social discount rate is a complex aggregate

of individual rates of time preference. If aging affects individual rates of time

preference it will alter the de facto social rate of time preference. Shifts in the

age structure of a population could break intergenerational symmetry and lead to levels of investment that are individually optimal in one generation,

but too low or too high according to the age structure of a future population.

A preliminary step to this line of inquiry is the estimation of valid measures of

time preference during youth and across the lifecycle. This paper extends a compensating differentials model of Moore and Viscusi by 1) Relaxing the assumption that workers are perfectly informed about occupational fatality risk; 2) Implementing a model of endogenous time preference. The revised model highlights the way in which prior esti- mates required an assumption of perfect risk perception in order to identify both the discount rate and the value of a statistical life. The paper demon- strates that even though absolute estimates of discount rates may not be identifiable from current data, there is much to be learned through exploring the causes and consequences of interpersonal variation in the acceptance of compensating differentials both in cross section and over time. As an initial demonstration of this approach, estimates of individual discount rates based on data from the National Longitudinal Survey of Youth (NLSY) are used here to test Becker and Mulligan’s theory of the endogenous determination of time preference. In Sect. 2, the paper sets out a compensating differentials model with imperfect risk perception. Section 3 discusses the National Longitudinal Survey of Youth (NLSY) data and the estimation method. Section 4 first shows the results of estimating the individual discount rates under the assumption of perfect risk perception and how these would change if alter- native assumptions were employed. The validity of these measures is dis- cussed briefly, followed by a discussion of Becker and Mulligan’s (1997) suggestion that subjects adapt their rates of time preference as their future prospects and mental capacities develop. Section 5 discusses how to reconcile the implied dynamic inconsistency of human choice with policy suggestions to regard perfect information on risk magnitude as a sufficient condition for rational choice.

2. Imperfect risk perception in a compensating differentials model

Deferring for now a discussion of the market forces which establish an equilibrium gradient between wages (w) and the true risk (p j ) of dying on job ‘‘j ’’, we simply assume that each separate labor market is characterized by a unique gradient denoted w(p j ) with w’(p j )>0. Labor market partic- ipants confronting this gradient must choose among the various occupations for which they are qualified to maximize a lifetime utility function that trades off the competing goals of earning more money and having higher survival probability. The Viscusi and Moore model (Viscusi and Moore 1989) focuses on a single once in a lifetime choice of occupation. Occupations are assumed to be completely characterized by their wage and occupational fatality rate. The ‘‘one-shot’’ problem of occupational choice can be written as:

586

D. M. Bishai

Max V ¼ U

p j

w

1

p

j

X b

j

t ¼ 1

1

p j p j

p

x

ð

p

1 p j p j

p x ð p x

x

Þ

Þ

t 1 :

ð 1 Þ

where:

U ( ) is the utility function w(p j ) is the payoff of the jth agent as a function of the occupation (and risk) selected

p j (p j )

p j is true occupational risk p x (p x ) is the perceived life table fatality risk determined by p x . p x . is true life table risk (1-p j (p j ) -p x (p x )) t is the perceived probability of surviving at least t periods given exposure to the perceived job fatality risk p j, and life table-driven

perceived overall fatality risk p x b j is the discount rate of the jth agent

I assume that p j (p j ) is invertible and will later denote its inverse as p j (p j ). The relationship between p j and p j in Eq. (1) modifies the (1989) Viscusi and Moore model by breaking the identification of perceived risk with actual risk. Imperfect knowledge of risk is modeled by incorporating the process whereby subjects perceive the fatality risks as ‘‘prospects’’ (p) (Kahneman and Tversky 1979) which have a functional association with the true risks now symbolized with Greek characters ‘‘p j ’’ and ‘‘p x .’’ In this model, the subject selects once and for all their occupation which is parameterized only by its fatality risk. That choice determines both the subject’s wages according to w(p j ) and the subject’s occupational survival rate (1 ) p j ) for the subsequent t periods.

ð 1 p Þ , and noting that p j has a

functional relationship with perceived risk, p j , Eq. [1] can be expressed as:

is the perceived occupational fatality risk determined by p j

Noting that P ¼ 1 ½ b ð1 p Þ t 1 ¼

1

t

1

1

b

Max V

p j

¼ U ðw ð p j ð p j ÞÞÞ ½1 p j ð p j Þ p x ðp x Þ

1 b ½1 p j ðp j Þ p x ð p x Þ

:

ð 2 Þ

Writing p(p) as simply p, the job-seeker’s first order condition with respect to p j , is:

dU

dw

ð 1 p j p x Þ ð p j Þ ð @ p j Þ U ð wÞ

@

w

@

@

p j

1 b ð 1 p j p x Þ

Solving for

@

p j yields

w

@

@ w ¼ U ð wÞ

@

p j

dU

dw

1

b

1 p þ 1 b ð 1 p Þ

b U ð wÞ ð 1 p j p x Þ

½ 1 b ð 1 p j p x Þ 2

1

@

p j

@

p j

¼ 0

ð

ð

3 Þ

4 Þ

To estimate the model one must apply functional forms to U( ) and w(p j ). Viscusi and Moore (1989) explore the general implications of Constant Relative Risk Aversion (CRRA) utility functional form which can be

Does time preference change with age?

587

parameterized as U=(w c ) 1)/c. The limiting case of the CRRA utility function that turns out to be of interest for both exposition and estimation is c ! 0 in which case the CRRA utility function approaches the familiar logarithmic utility function. Later the empirical analysis will vary the parameter, c, to determine restriction the most appropriate value of c. For now, the loga- rithmic form simplifies exposition.

ð 5 Þ

With this functional form

U ¼ ½ log ðw Þ

U

dU

dw

¼ w log ð wÞ

and (4) can be rewritten as:

ð 6 Þ

@

w

@

p j

¼ w log ð wÞ

Noting that @ log ð w Þ ¼

@ p

@ w

@

p

w

1

1 p j p x

þ

b

1 b ð 1 p j p x Þ

and rearranging yields:

1

@

@ p j

p j

1

1 p j p x

þ

b

1 b ð 1 p j p x Þ

1 @ log ð wÞ @ p j

¼ log ð wÞ

which can be simplified to yield

½ð 1 p j p x Þ b ð 1 p j p x Þ 2 @ p j

p j

@ log ð wÞ

@ p j

@

¼ log ð w Þ

1

@

p j

@ p j

ð

ð

7 Þ

8 Þ

ð 9 Þ

Since mortality risk, p, is less than 1/1000 a good approximation of (1 )p) 2 is (1 ) 2p), for small p. However, a simplification that preserves relative relationships between risk and wage – though confounding them with a factor of 2 is (1 ) p) 2 » (1 ) p). With this approximation (9) permits factoring the (1 )p) terms and becomes:

log ð wÞ¼ð 1 bÞ @ p j

p j

@

ð1 p j p x Þ @ log @ p ð j wÞ

ð

10 Þ

Augmenting (10) with a vector of observed variables x j and a vector of unobservable variables l j that may account for individual differences in tastes yields:

log ð wÞ¼ð 1 b Þ @ p j

p j

@

ð1 p j p x Þ @ log @ p ð j wÞ þ X a k x jk þ l j

k

ð

11 Þ

2.1. The role of risk perception

There is extensive evidence that subjects systematically overestimate small risks and underestimate larger ones (Viscusi and O’Connor 1984; Viscusi 1992). Figure 1 displays Viscusi’s linear approximation of the relationship between perceived risk (p) and true risk (p). This suggests using an approximation for p( p) of the form: p j = A+ p j

k ,

where A and k are constants. With this approximation

and p x

= A+ p x

k

588

Perceived

Risk, p

Intercept A

D. M. Bishai

45° 1 Slope k
45°
1
Slope k
Perceived Risk, p Intercept A D. M. Bishai 45° 1 Slope k p( π ) True

p(π)

True Risk, π

Fig. 1. Approximate relationship between perceived risk and true risk. Source: Viscusi 1992

inserted in Eq. (11) the relationship between wages and the market can be expressed as:

log w j ¼ C þ ð 1 b j Þ ð 1

k þ 2 A p j p x Þ d log d p ð j wÞ þ a j X j þ l j

ð

12 Þ

Equation (12) requires a prior assumption on the values of A and k in order to identify (1 ) b). Many prior compensating differentials studies have used epidemiological measures of occupational fatality risk as a direct proxy for perceived risk measures. This assumes p ¼ p which implies that A ¼ 0 and k ¼ 1. This achieves identification at the expense of ignoring what is known

from studies of biases in risk perception. If it is correct to assume that A ¼ 0

is indeed a measure of (1 ) b j ).

and k ¼ 1 the coefficient of (1 ) p j ) p x ) ·

However given the psychological evidence that 1>A>0 and k > 1 the term 1 k +2A in (Eq. (12)) will be less than 1 leading to a bias in the coefficient on the

danger premium and estimates of 1 )b that are too large. If both risk perception and time preference are heterogeneously distributed in the population, they may be systematically correlated and estimates of the

coefficient on the danger premium will confound both processes. This will be discussed later. The intuition for Eq. (12) is straightforward. A positive coefficient on

in Eq. (12) would indicate that higher subject wages are better

@ log ðw Þ @ p j

explained by increases in ‘‘attractive local wage offers for dangerous jobs’’ when one or more of the following is true:

The subject faces a high probability of surviving (p j and p x approaching zero) The subject under perceives increases in job fatality risk (k approaching zero) The subject has a low discount rate (b approaching zero)

@ log ðw Þ

@

p j

Does time preference change with age?

589

To interpret (1) b j ) purely as a measure of time preference requires that statistical controls have been used for other mediators of the relationship between earnings and the danger premium. Three important confounders might be 1) economic desperation and 2) risk aversion and 3) risk per- ception. The analysis here can achieve some success towards control for desperation by using measures of family size, race, schooling, and gender as X j variables. Furthermore the analysis can parameterize risk attitude within specialized functional forms like the constant relative risk aversion

utility function to further control for this confounding. However, unless one is willing to assume that prospective employees have perfect infor- mation about occupational fatality risks as past authors have done, esti- mates of time preference will be biased by systematic errors in risk perception. Equation (12) imposes a restriction stating that other determinants of wages (X j ) such as schooling, age, gender etc. will independently shift the relationship between wages and wage-risk offers up and down (changes in intercept) but they will not alter its slope. This restriction will be loosened below. The danger premium, @ log ðw Þ , must be computed. To compute it one can essentially use NLSY respondents as informants. Each NLSY subject reports their region, their wage, and their occupation. NIOSH’s National Traumatic Occupational Fatality (NTOF) data is used to convert reported occupation into the fatality rate faced by that NLSY respondent in exchange for the wage they were given (National Institute for Occupational Safety and Health 1993). The use of occupation-based fatality rates avoids the bias induced by errors in variables inherent in the cruder industry-based fatality rates (Dillingham 1985). The NTOF data offer annualized estimates based on roughly a decade of reporting. The long time horizon represented by these risk estimates accords with the models depiction of a lifetime occupational choice. Furthermore the 1980-1989 period covered by NTOF is well aligned with the 1979–1994 period covered by NLSY. The statistical model imple- mented is a regression of the form:

@

p j

log w j ¼ C þ X

k

b

1 k

D k p j

þ b 2 k

D k p þ b X X þ e

2

j

ð 13 Þ

where D k are dummy variables for region, and period variables that separate different labor markets in space and time. Equation (13) is a representation of the hedonic market equilibrium for job safety (Kahn and Lang 1988) that was simply assumed to exist at the start of this discussion. Controlling for other determinants of wages, the coefficient on occupational fatality will represent an equilibrium. Competition for workers between firms in a local labor market (determined by space and time) establishes the local balance between firms’ willingness to invest in safer working conditions and their ability to simply raise wages to attract workers who are willing to jeopardize their survival for cash. Competition between workers for jobs establishes a similar balance between local workers’ hedonic demand for better occupational survival poised against decrements in wages. As pointed out by Kahn and Lang (1988) were one to simply regress log wage on the X variables and on occupational risk without the interactions, then the measures of @ logð w Þ would not be identified in Eq. (12). The X’s are other determinants of wage. After running this regression one can use the results to compute as follows

@

p j

590

D. M. Bishai

danger

premium ¼ @ log @ p ð wÞ ¼ X b 1 K D K þ 2 b 2k D k p j

k

ð

14 Þ

Viscusi and Moore (1989) treat the imputed value of @ logð w Þ from Eq. (14)

@ p

as endogenous to an individual’s wage equation and instrument it using age,

education gender, and race prior to inserting it in Eq. (12). In this analysis, each individual’s imputed @ logð w Þ from Eq. (14) is considered to be exogenous

@ p

during subsequent analyses because it is a linear combination of exogenous variables—the regional dummy, D k , and the risk variable, p j . This identi- fying assumption could be challenged (Brown and Rosen 1982). The assumption is necessary to permit the ensuing random coefficient models that look at age, education, gender, and race as factors that modify the

with measures of

( k þ 2 A ) p x ) p j ) in which it is explicitly assumed that k=1 and A= 0 forms a

new independent variable, Z jt

hand version of Eq. (12) describing wages and fatality risk for the j-th subject in the t-th period as follows:

ð 15 Þ

To estimate 1 ) b j , note that it seems plausible that in (12) and (15) not just the intercept but the slope of the tradeoff between the danger premium and wages is a random coefficient that differs idiosyncratically in general patterns suggested by Becker and Mulligan’s (1997) theory of endogenous time preference. Becker and Mulligan’s theory offers an account of the determination of time preference by suggesting that individuals may alter their time preference in part by spending time and effort in forming mental pictures of future pleasures. They cite as examples of such time and effort activities like acquiring information through schooling, access to print media, and time spent with older persons, particularly parents. Subjects endogenously invest effort in these activities in proportion to the agreeable- ness of their future life prospects. For instance, Becker and Mulligan speculate that because church attendance offers repeated promise of future heavenly splendor it may lead attendants to endogenously develop ways to discount future splendor less. I thus introduce a time preference determination equation inspired by Becker and Mulligan:

that permits a short-

response to the danger premium. Interacting @ log ðw Þ

@ p

1

(1) p xt ) p jt ) · @ logð w Þ

@ p

=

log

w

jt

¼ C þ

1 b j Z jt þ c j X jt þ l jt

ð 1 b j Þ ¼ C þ c a age þ c s schooling þ c g gender

þ c r race þ c f father þ c c church þ e j

ð 16 Þ

which forms a system together with Eq. (15) In this system, person-rounds define the level 1 process of Eq. (15) and persons define the level 2 process of Eq. (16). Random coefficients estimation methods will be used to estimate all of the above parameters including the individual specific error term, l bj . An alternative approach is to interact the observables from (15) with Z jt I in (16) and use the OLS coefficients on the interaction terms to learn how the background variables affect time prefer- ence. These OLS coefficients could be used to estimate the average discount rate for a population based on population averages of the background variables. The goal of obtaining individual specific error terms cannot be realized using OLS.

Does time preference change with age?

591

Note that all right hand side variables with the exception of ‘‘Z’’ are commonly collected in nationally representative datasets. Because Z is an interaction of fatality risk with the respondent’s local danger premium, an important preliminary step will be to derive measures of the danger premia facing each of the respondents using Eqs. 13 and 14.

3. Data and estimation method

The National Traumatic Occupational Fatality data from NIOSH contains occupation specific fatality rates for US workers. These data have been merged with National Longitudinal Survey of Youth (NLSY) data in order to perform the occupational risk study described in Eq. 14. Each employed NLSY respondent reports their occupation and wage. The NIOSH data permit the attribution of the actual fatality risk that each respondents were facing in exchange for the wage they report. The data do not indicate whether the subjects perfectly perceived their fatality risk. The impact of assuming perfect perception (k ¼ 1) will be explored in detail. The NLSY offers panel data on a cohort of 12,686 young persons aged 14 to 21 in 1979 followed up until when they are aged 29 to 37 in 1994. Table 2 describes the essential variables from NLSY to be included in the analysis of occupational choice. From Table 2 one can get a sense of the magnitude of selection bias. Attrition shrinks the sample size from 12,686 to 8,889, indi- viduals who can be followed over 16 years. Attrition means that observations on the effect of higher age and later period are conditional on sample non- attrition

3.1. Multilevel model estimation

Equation system (15) and (16) is the multilevel system that will be estimated using the 15 waves of observations on however many of the 12,686 subjects in the NLSY report wages and occupations. Equation (16) describes phenom- ena at the level of the individual across successive rounds, while Eq. (15) describes phenomena at the level of the individual within each round. Our approach is similar to that taken by Cairns and Van Der Pol (Cairns and van der Pol 1997) whose analysis of 2 rounds of survey data on intertemporal preferences using multilevel models decisively favored the multilevel approach over OLS.

3.2. Computing individual time preference measures

In addition to examining the coefficients for Eq. (16) to assess the effects of age, education, etc. on time preference, I will compute individual measures of b j that incorporate estimates of the individual specific fixed effects e j . Estimates of b j can be derived which maximize the joint likelihood function for the system set up by (16). The error terms generated by the same subject in different years are likely to be correlated and will result in heteroskedasticity.

592

D. M. Bishai

Table 1. Cohort structure of NLSY. Ages of subjects vs. Years of successive waves. The
Table 1. Cohort structure of NLSY. Ages of subjects vs. Years of successive waves. The final row indicates the number of years in which the
corresponding age group is observed.

Does time preference change with age?

593

Table 2. Variable definitions and means for wage-risk estimates. The full sample of labor market participants (column I ) is used to estimate the wage-risk locus (Eq. 13). The instruments and their interactions with job fatality rates are used to identify the locus. The sample over age 25 (column II) is subsequently used to estimate the discount rate

Variable

Column I Full sample of labor market participants (N = 116,405 )

Column II Labor market participants >age 25 (N = 44,827)

Source

Mean

SD

Mean

SD

Endogenous Annual real after tax wage Danger premium (dlog w/dp j ) % D per unit D in Deaths/100K (Calculated after running a regression of Eq. 13)

Exogenous included Job fatality rate (p j ) (deaths/100k) All cause mortality (deaths/100k) Years of schooling Race (% African American) Gender (% female) Age AFQT score Religious attendance frequency 1979 Lived with father in 1979 Number of dependents (Incl Self) Tenure at current job (months) Years in labor force Union status (% members)

$12,337

(7803)

$15,535

(8716)

NLSY

0.00518

(0.0052)

0.00515

(0.0052)

NTOF +

 

NLSY

4.9

(9.7)

4.8

(9.2)

NTOF

142.9

(103.2)

163.8

(113.5)

NCHS

12.6

(2.2)

13.0

(2.36)

NLSY

23.4%

25%

NLSY

48.5%

48.6%

NLSY

25.6

(4.7)

29.2

(2.8)

NLSY

41.8

41.9

NLSY

3.25

(1.7)

3.22

(1.68)

NLSY

60%

59%

NLSY

3.25

(1.8)

3.01

(1.6)

NLSY

121

(144)

167

(176)

NLSY

5.8

(3.75)

8.4

(3.05)

NLSY

16%

16.3%

NLSY

Exogenous (instruments for Eq. 13) Period (4 quadrennial dummies) Pd1 (1979-1982) Pd2 (1983-1986) Pd3 (1987-1990) Pd4 (1991-1994) Region (4 separate dummies) Northeast South North Central West

NA

NLSY

20%

NLSY

28%

NLSY

29%

NLSY

22%

NLSY

NA

NLSY

18%

NLSY

38%

NLSY

24%

NLSY

19%

NLSY

The multilevel model takes into account precisely this form of heteroskedas- ticity.

3.3. Self-selection into the workforce

The strategy for estimating rates of time preference is only successful if a respondent is participating in the workforce. College attendance is an important reason that younger NLSY respondents do not participate in the workforce. Because the sub sample that self-selects out of the work

594

D. M. Bishai

force and into college is likely to have different determinants of time preference the compensating differentials analysis is confined to observa- tions generated by NLSY respondents after they have attained age 25. Even after 25 work force participation is endogenous and could be strongly related to time preference. I implement a Heckman procedure where a probit equation predicts workforce participation as a function of the state unemployment rate, spouse’s income, savings, squares of spousal income and savings, number of children, and the interaction of a gender dummy and number of children. The resulting inverse Mills ratio is inserted in Eq. 15.

4. Results

4.1. Determining the US wage risk locus for 1979–1994

Table 3 presents estimates of Eq. (13) which is a first stage equation describing how the equilibrium wage risk locus shifts across local markets defined by quadrennial periods and 4 regions composed of roughly 15 states each. For the task of estimating the wage risk locus the full sample of NLSY labor market participants was used regardless of age and occupational category. Although this is hardly the proper specification to examine the effects of the productivity controls, it builds confidence to review the ways in which the long established facts from labor economics are reaffirmed in Table 3.

Marriage is associated with an increase in annualized wages of about 5% similar to previous findings (Daniel 1995). Each year of schooling is associated with an increase in annualized wages of about 5% similar to previous findings (Schultz 1988; Behrman, Pollak et al. 1995). Workforce experience more than tenure in a given job is associated with increased wages similar to previous findings (Altonji 1987). Black race and female gender reduce annualized wages significantly see also (England, Reid et al. 1996). Union membership has a profound effect on wages (Leigh and Gill 1991) Lower job satisfaction (eg., Job Stress) is associated with higher wages (French 1999). Occupational category controls appear to have reasonable effects.

All of the interaction variables are significant, and their coefficients are used together with Eq. (14) to obtain a total of 16 (4 regions · 4 periods) separate estimates of the local danger premia corresponding to the 16 separate labor markets confronting NLSY respondents. Job fatality interactions with the occupational dummies were not used because of the theoretical requirement that the dummies should mark off separate markets in which subjects could trade safety for money. To accomplish these ‘‘trades’’ subjects can choose alternative occupations, but not alternative regions or periods. By this criterion, substantial migration on the basis of danger premia, would invalidate the use of the regional dummies to identify the occupational risk vs. wage market locus. There is no evidence to date

Does time preference change with age?

595

Table 3. Identification of the regional and period specific danger premia. Dependent variable is log of annualized after tax earnings*

 

Coefficient

t statistic

N

116,405

R 2 Job fatality * North East

0.4282

0.0416

(16.24)

Job fatality 2 * North East Job fatality * North Central Job fatality 2 * North Central Job fatality * South Job fatality 2 * South Job fatality * West Job fatality 2 * West Job fatality * Period 1 Job fatality 2 * Period 1 Job fatality * Period 2 Job fatality 2 * Period 2 Job fatality * Period 3 Job fatality 2 * Period 3 Job fatality * Period 4 Job fatality 2 * Period 4 Highest grade completed Highest grade completed 2

)0.0003

)(9.17)

0.0289

(11.41)

)0.0002

)(6.23)

0.0306

(12.18)

)0.0002

)(6.5)

0.0416

(16.36)

)0.0003

)(9.27)

)0.0311

)(12.24)

0.0002

(7.05)

)0.0316

)(12.54)

0.0002

(7.12)

)0.0275

)(10.87)

0.0002

(5.84)

)0.0254

)(10.)

0.0002

(5.36)

)0.0160

)(4.)

0.0025

(16.04)

Black

)0.0498

)(14.34)

Female

)0.1519

)(47.34)

Age

0.0725

(18.39)

Age 2

)0.0010

)(13.69)

Job tenure

0.0010

(35.18)

Job tenure 2 Experience in job market Experience 2 Union member Job satisfaction Family size Married Managerial** Technical sales Operators & fabricators Precision production

0.0000

)(19.86)

0.0474

(25.41)

)0.0005

)(3.85)

0.1725

(44.47)

)0.0318

)(16.64)

)0.0144

)(17.8)

0.0490

(15.58)

0.2723

(51.58)

0.1831

(43.88)

0.1258

(25.65)

0.2327

(40.68)

Farm/forestry/fishing

)0.0513

) (6.39)

Constant

7.5184

(139.17)

* Hourly rate of pay * 2000 Hours * (1-tax rate). ** Excluded occupational dummy: Service industries. Period 1 = 79–82, Period 2 = 83–86, Period 3 = 87–90, Period 4 = 91–93.

for migration on the basis of geographical differences in compensating differentials.

4.2. Determining the US wage determinants for 1979–1994

Table 4 presents parameters from the wage Eq. (15) which is estimated jointly with the time preference determination Eq. (16) . The sample is now constrained to subjects over age 25. The OLS estimates in Column I are

596

D. M. Bishai

Table 4. Estimating the wage equation in the NLSY. Dependent variable is log of annualized after tax earnings

 

Column I

Column II

Column III

OLS

OLS

Unrestricted maximum

 

likelihood

 

Coefficient t-statistic

Coefficient t statistic

Coefficient t statistic

N

44827

44,827

44,827

Likelihood

2.86E+04

R 2

0.2897

0.4758

Constant

8.3002

(199.81)

8.3927

(196.23)

8.2727

(262.19)

Survival adjusted local danger premium* Union member

9.5739

(15.07)

Random Variable

Random Variable

0.1640

(20.49)

0.1665

(20.86)

0.1554

(24.98)

Married

0.0745

(10.05)

0.0684

(9.31)

0.0653

(12.64)

Family size Highest grade completed

)0.0170

)(7.08) )0.0159

)(5.53) )0.0132

)(7.64)

0.0659

(37.78)

0.0634

(28.9)

0.0629

(48.67)

Experience in job market

0.0361

(21.88)

0.0352

(21.26)

0.0341

(30.36)

Age

0.0052

(3.25)

0.0040

(2.28)

0.0082

(7.18)

Job tenure

0.0005

(21.71)

0.0005

(21.49)

0.0005

(32.74)

Job satisfaction

)0.0446

)(11.39) )0.0457

)(11.72) )0.0455

)(15.01)

Black

)0.0779

)(8.98) )0.0845

)(7.86) )0.0825

)(12.22)

Female

)0.1797

)(20.87) )0.2197

)(19.97) )0.2154

)(32.74)

Mill’s ratio

)0.1832

)(6.02) )0.1831

)(6.07) )0.1813

)(8.94)

* Computed value of dlog(wage)/d job fatality from Table 3, interacted with survival probability from lifetable.

Table 5. Estimating time preference in the level 2 equation. Dependent variable is coefficient on (1-p) dlogw/dp, the survival adjusted danger premium in Table 4

 

OLS*

Maximum likelihood

Coefficient

t statistic**

Coefficient

t statistic

Tau test for null of no Random coefficients N Constant Lived with father in 1979 Religious attendance Age Highest grade completed Black Female AFQT Family Size

 

989.2405

44827

44827

43.4167

(4.44)

1.3185

(1.33)

1.5351

(1.38)

0.1771

(.61)

)0.2839

)(.89)

0.2088

(1.86)

)1.5893

)

(4.65)

)0.8197

)(3.12)

)0.6069

) (1.88)

5.7377

(3.71)

5.5782

(3.45)

7.4678

(5.97)

7.6847

(6.15)

0.1977

(8.92)

0.2807

(11.16)

0.0864

(.24)

)0.2539

)(.75)

* Model from column II of Table 4.

** OLS standard errors adjusted for heteroskedasticity.

derived by simply including the uninteracted survival adjusted danger premium, Z j . The OLS estimates in Column II are derived by interacting each of the time preference determinants in Table 5 with survival adjusted local danger premium. OLS standard errors are adjusted for heteroskedas- ticity using White’s correction.

Does time preference change with age?

597

The Maximum Likelihood (multilevel) estimates in Column III are random coefficients estimates which impose Eq. (15) as written, including the person specific error term. A statistical package known as Hierarchical Linear Models (HLM) was used for the analysis. The (1–b) coefficients produced by the maximum likelihood technique were calculated for all subjects in the NLSY who ever participated in the labor market according to Eq. (15) which says

ð 1 b j Þ ¼ A þ c a age þ c s schooling þ c g gender

þ c r race þ c f father þ c c church þ e j

Note that the actual estimation of the random effect term, e j , makes the discount rate unique for each respondent. For any given subject, the last 5 terms of Eq. [15] are fixed, but age and (for younger subjects) schooling may vary over time. The inclusion of the c a · age term induces systematic variation over time for any given subject. The distribution is approximately normal and centered around a mean of 5.3. A Tau test (Bryk and Raudenbush 1992) rejected the null of a non-ran- dom coefficient on the danger premium, thus the multilevel model offers the preferred set of estimates as well as the preferred specification.

4.3. Can we infer the appropriate coefficient of risk aversion?

A grid search was conducted in which Eq. (15) was altered by substituting

(wage) c for log(wage). This substitution implemented a constant relative risk aversion utility of wages where ‘‘c’’ represents the coefficient of relative risk aversion. The resulting values of the likelihood function were tabulated and

asymptotically approached a maximum as c approached 0. This evidence was interpreted as suggesting that the coefficient of relative risk aversion for the full sample was very small and could be approximated as zero. The logarithmic utility function that was used for most of the analysis cannot

be rejected as the best fit to the data. Unfortunately the grid search method

can only identify risk aversion parameters for the entire population and not

for individual subjects.

4.4. Can we infer the proper discount rate for society?

As can be seen from Eq. (12) the interpretation of the danger premium coefficients as indicators of the discount rate require the assumption of perfect perception of changes in occupational fatality rates (e.g., A ¼ 0 and k ¼ 1). An empirical rejection of perfect risk perception in the NLSY can be based on the finding that the coefficient on the danger premium at 5.3 is roughly 10 fold higher than it would need to be to place the discount rate in its theoretical range of 0–1. If subjects underestimated the magnitude of changes in their job fatality risk by at least a factor of 10 (e.g., 1 k ¼ 10 and A » 0), it would bring the mean estimate of b into the theoretical range. Unfortunately in the absence of information about the quality of risk perception, the empirical methods based on compensating differentials cannot identify absolute estimates of the population discount rate (b) , nor of the value of a statistical life (VSL) . Those interested in studying compensating

1

598 D. M. Bishai 0.06 0.05 0.04 0.03 0.02 0.01 0 25 26 27 28
598
D. M. Bishai
0.06
0.05
0.04
0.03
0.02
0.01
0
25
26
27
28
29
30
31
32
33
34
35
-0.01
-0.02
Z-Score of Beta

Age

Fig. 2. Age related changes in time preference. A plot of the sample mean discount rate Vs. age. Discount rates have been normalized to facilliate graphing. Lower discount rates indicate that future consumption is more heavily discounted

differentials to obtain estimates of these parameters for public policy must rely on assumptions about the quality of risk perception to identify b and to form estimates of VSL. But this does not suggest that the study of compen- sating differentials is fruitless. The results presented in the following section suggest that the study of compensating differentials continues to have promise in explaining relative variation in discount rates and assessing their impact on health behavior.

4.5. The determination of time preference

Table 5 indicates that controlling for covariates young age, less schooling, white race, and male sex are associated with an immediate time preference and or a less acute perception of risk. When the level of attained schooling is controlled, intelligence as measured by the Armed Forces Qualification Test (AFQT) is associated with a more immediate time preference. This implies that the component of intelligence that is orthogonal to the qualities required to persist in school is associated with immediate time preference. Figure 2 shows the life course evolution of time preference based on the sample means from this dataset. The greatest increases in time preference between ages 25 and 35 occur after age 29. The hypothesis that life events such as family formation played a role was tested by including family size in

Does time preference change with age?

599

the model of time preference determinants. Table 5 indicates that family size tends to make time preference less immediate (b ¼ ) 0.2539), but the effect was not statistically significant.

5. Discussion

The results that age and schooling are associated with the development of greater patience is consistent with the theory of endogenous time preference (Becker and Mulligan 1997) With life experience and schooling subjects may obtain a more certain indication of their future utility and thus be inclined to invest in personal activities that will render them more patient. There is weak evidence from these results to support the conjecture of Becker and Mulligan that religious participation would encourage a less immediate time preference. The effect of religious attendance trends towards a preference for less immediate rewards but it is not statistically significant. There is no evidence to support the hypothesis that the presence of a father early in life (at age 14) fosters a less immediate time preference. The inferences drawn regarding the effect of father’s presence and the religiosity indicator suffer because these variables are assessed at age 14 and all markers of time preference are assessed after age 25. The finding of aging and schooling effects on time preference leads to a discount function that fails to satisfy any of Strotz’s 3 possibilities for time consistent discount functions. If young people fail to account for the likeli- hood that they will become more patient with age, they could make decisions that would have been rational had their rates of time preference remained constant, but which lead to regret in the face of life cycle changes in time preference. One could press the argument that the empirical estimates of time preference presented here admittedly could be confounded with risk attitude and risk perception. Thus one might stubbornly cling to a belief in time consistency by insisting that the aging and schooling effects apply only to these other risk perception parameters and not to the time preference parameter. The implication remains that improving individuals’ foresight into how their preferences will change through time can improve the quality of their decisions.

5.1. Validating the time preference measures: Do they predict risky health behavior?

To assess the validity of time preference measures derived from studying compensating differentials prior work has tested their performance in explaining differences in alcohol use. Measures of time preference have a prominent but infrequently tested association with health behavior in virtually all economic models of health behavior (Becker et al. 1990; Becker and Murphy 1988; Fuchs 1982; Grossman 1972; Grossman et al. 1994; Grossman et al. 1996; Ribar 1994). Thus, empirical measures of time preference can also be used to assess the strength of this association. Health- promoting behaviors that pay off in the future are more enjoyable to individuals who have lower time preference. The theory of rational addiction (Becker et al. 1991; Becker and Murphy 1988) highlights the way in which

600

D. M. Bishai

past consumption of an addictive good increases the marginal utility of current consumption. The theory predicts that individuals who greatly discount the future are more likely to become addicted and will be more sensitive to shifts in the price of the commodity. In work applying the compensating differentials approach outlined above to a larger sample of NLSY workers with ages as young as 14, I found that more patience (High b) is significantly negatively related to both any beer drinking and among drinkers to the measure of the number of days of drunkenness in the past 30 days (Bishai 2001). The personal background variables controlled for age, schooling, race, gender, marital status, and religiosity such that the contribution of the discount rates to explaining variation in drinking behavior was independent of family background. This finding suggests that the time preference measures obtained from an analysis of compensating differentials have predictive validity for other health behaviors.

6. Conclusion

This paper expands and refines a method pioneered by Viscusi and Moore (1989) to examine the response of subjects to various offers to trade their survival for money in the labor market. The paper demonstrates that estimates of time preference obtained from studying compensating differen- tials depend critically upon the assumption of perfect risk perception. Empirical evidence that risk perception is imperfect suggests that the study of compensating wage differentials yields biased estimates of the absolute rate of time preference in a population and biased estimates of the value of statistical life. The paper shows that parameterizing compensating differentials is a promising strategy for the study of relative differences in risky decision- making between subjects in a population. The measures obtained in this manner admittedly confound time preference with the ability to accurately perceive fatality risks and with risk aversion. This confounding is neither conceptually nor empirically unappealing. Each of these factors are cognitive responses to contingent states of the world that dampen the full consideration and valuation of possibilities other than the present moment. Each may shift endogenously in response to whether such a shift would complement the utility associated with the contemplation of alternative states of the world. The effect of schooling and aging on time preference rates were two of the chief predictions of Becker and Mulligan’s theory of endogenous time pref- erence. Subject to the limitations of the methods used in this study these theoretical predictions accord with the empirical findings. Note that the identification of the heterogeneous time preference rates in NLSY respon- dents relied on assuming that both risk perception and risk aversion rates were constant across the population. In actual fact both risk perception and risk aversion are heterogeneous too and may very well correlate with time preference. Indeed, schooling and aging may be just as effective at improving risk perception and risk aversion as they are at making time preference less immediate. Prior work (Bishai 2001) has demonstrated that time preference is inde- pendently correlated with drinking and the frequency of drunkenness among drinkers. This investigation found that subjects become more patient with

Does time preference change with age?

601

aging and schooling. The effects of aging and schooling lead to the possibility of dynamic inconsistency in planning intertemporal consumption paths. The fact that a ‘‘self’’ that starts down a consumption path could differ in an unforseeable way from one’s future ‘‘self’’ can lead to tragedy when the goods consumed have addictive or irreversibly harmful properties. Youth may not foresee that as they mature and gather lessons from life they might value the future more highly or see risks in a different way. Experience and reason can teach children that features of the personality that they consider immutable shift throughout the life course. Some youth can accept this truth simply by being told it. For others only time can properly instill this lesson. If future studies confirm the current finding that individual aging raises an individual’s discount rate this would have important implications for social policy. The actual social rate of time preference can be considered to be a complex politicized aggregate of individual rates of time-preference. As the median age of a population changes, one would predict that social time preference would also change. The rate of social time preference is a critical determinant of intertemporal transfers to future generations (Marini and Scaramozzino 2000). These transfers themselves may suffer from dynamic inconsistency due to population aging. It is possible that a more aged pop- ulation of the future with a low rate of time preference, may inherit invest- ments from the present era that are optimal under the current age structure and current social time preference, but which would be considered selfish by the older and hence more patient population of the future. If aging affects time preference, it could thwart the possibility of equitable intergenerational transfers during an era of shifting age structures.

References

Ainslie G, Haslam N (1992) Hyperbolic Discounting. Choice over Time. G. Loewenstein. Russell Sage Foundation, New York Altonji JG (1987) Do Wages Rise with Job Seniority. Review of Economic Studies 54:437–459 Becker GS, Mulligan CB (1997) The Endogenous Determination of Time Preference. Quarterly Journal of Economics: 729–758 Behrman J, Pollak RA et al. (1995) From Parent to Child. University of Chicago Press, Chicago Bishai D (2001) Lifecycle Changes in the Rate of Time Preference: Testing the Theory of Endogenous Preferences and its Relevance to Adolescent Substance use. In: Grossman M, Hsieh C-R, Economic Analysis of Substance Use and Abuse: The Experience of Developed Countries and Lessons for Developing Countries. Edward Elgar Press, London, 155–177 Brown JN, Rosen HS (1982) On the Estimation of Structural Hedonic Price Models. Econometrica 50(3):765–768 Bryk A, Raudenbush SW (1992). Hierarchical Linear Models. Sage Publications, London Cairns JA, van der Pol MM (1997) Saving Future Lives: A Comparison of Three Discounting Models. Health Economics 6(4):341–350 Daniel K (1995) The Marriage Premium. In: Tommasi M, Ierulli K, The New Economics of Human Behavior. Cambridge University Press, New York, 92–112 Dillingham AE (1985) The Influence of Risk Variable Definition on Value-of-Life Estimates. Economic Inquiry 24:227–294 England PL, Reid L et al. (1996) The Effect of the Sex Composition of Jobs on Starting Wages in an Organization: Findings from the NLSY. Demography 33(4):511–521 Kahn S, Lang K (1988) Efficient Estimation of Structural Hedonic Systems. International Economic Review 29(1):157–166

602

D. M. Bishai

Kahneman D, Tversky A (1979) Prospect theory: An Analysis of Decision Under Risk. Econometrica 47:263–291 Laux FL (2000) Addiction as a Market Failure: Using Rational Addiction Results to Justify Tobacco Regulation. Journal of Health Economics 19(4):421–438 Leigh JP, Gill AM (1991) Do Women Receive Compensating Wages for Risks of Dying on the Job? Social Science Quarterly 72(4):727–737 Loewenstein G (1996) Out of Control: Visceral Influences on Behavior. Organizational Behavior and Human Decision Processes 65(3):272–292 Loewenstein G, Prelec D (1992) Anomalies in Intertemporal Choice: Evidence and An Interpretation. Quarterly Journal of Economics: 573–597 Marini G, Scaramozzino P (2000) Social Time Preference. Journal of Population Economics

13(4):639–45

Moore MJ, Viscusi WK (1990) Models for Estimating Discount Rates for Long-Term Health Risks Using Labor Market Data. Journal of Risk and Uncertainty 3(4):381–401

National Institute for Occupational Safety and Health (1993). Fatal Injuries to Workers in the United States, 1980–1989: A Decade of Surveillance. U.S. Department of Health and Human Services, Washington, DC Orphanides A (1994) Rational Addiction with Learning and Regret. Journal of Political Economy

103(4):739–758

Schultz TP (1988) Education Investments and Returns. In: Chenery H, Srinivasan TN, Handbook of Development Economics, Vol 1. North-Holland, New York, 543–630 Strotz RH (1956) Myopia and Inconsistency in Dynamic Utility Maximization. Review of Economic Studies: 165–180 Thaler R (1981) Some Empirical Evidence on Dynamic Inconsistency. Economics Letters 8:201–

207

Thaler R, Rosen S (1976) The Value of Saving a Life: Evidence from the Labor Market. In:

Terleckyj NE, Household Production and Consumption. National Bureau of Economic Research, New York, 265–298 Viscusi WK (1992) Rational Models of Biases in Risk Perception. In: Viscusi WK, Fatal Tradeoffs: Public and Private Responsibilities for Risk. Oxford University Press, New York,

101–109

Viscusi WK, Moore MJ (1989) Rates of Time Preference and Valuations of the Duration of Life. Journal of Public Economics 38(3):297–317 Viscusi WK, O’Connor CJ (1984) Adaptive Responses to Chemical Labeling: Are Workers Bayesian Decision Makers? American Economic Review 74(5):942–956

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.