Labor Supply and the Tax Reform of the Century¤

Labor Supply and the Tax Reform of the Century¤ Martin Ljungeyz University of Chicago October 27, 2004

Abstract We estimate the elasticity of earned income to the net of tax rate following the very large 1991 tax reform in Sweden. Earned income measures both quantitative and qualitative margins of labor supply. We estimate an explicit behavioral model in static and dynamic settings. We ¯nd large responses to the tax reform, with elasticities in response to an unanticipated tax reform ranging from 0.37 to 0.56 on the intensive margin.

JEL codes: J22, H31 Keywords: Labor supply, tax reform, policy evaluation

1

Introduction

The magnitude of the Swedish tax reform of 1991 (TR91) is large by any metric. One measure of the size of the reform is the change in the average net of tax rate (1 ¡ ¿ ), the share of the last krona earned that you keep. Table 1.1 provides a comparison of the Swedish tax reform to the largest tax reforms in the US. The change in the average net of tax rate due to TR91 is 24.6%. This is ¯ve times larger than the 1986 tax reform in the US. With such a large change in ¤ This paper is based on the paper "Who Responded to the Tax Reform of the Century?", which is joint with Kelly Ragan (University of Chicago). y I would like to thank Austan Goolsbee, James Heckman, Casey Mulligan, and Kelly Ragan for valuable discussions and comments on the paper. I am also grateful for comments from seminar participants at the University of Chicago, IIES, SOFI, IFAU, SSE, and the EEA 2004 meeting in Madrid. z Contact info: [email protected], Department of Economics, University of Chicago, 1126 E. 59th Street, Chicago, IL 60637.

1

the marginal incentives to earn it is natural to ask if and by how much Swedes responded to these changes in incentives.

Table 1.1. The Swedish tax reform in comparison to the largest US tax reforms. % change 1-τ Country Year Sweden 90-91 24.6 US 41-42 -8.8 Sweden 89-90 6.9 US 40-41 -6.2 US 47-48 6.1 US 43-44 -5.4 US 86-87 4.8 US 50-51 -4.8 US 81-82 4.5 Note: US numbers are from Barro & Sahasakul (1983) and Marion & Mulligan (2004). τ is the marginal tax rate. Tax rates are weighted by assessed gross income.

In this paper we take a broader view of the nature of behavioral responses of earnings e®ort to marginal tax rates than most labor supply studies. We estimate the elasticity of earned income to the net of tax rate, an alternative measure of labor supply. The motivation for this alternative approach is simple. While hours worked may be a fairly ¯xed component of the labor contract, in particular in Sweden, other dimensions of labor supply such as e®ort may be more °exible. Consider the simple example of an individual who responds to the tax reform by working the night shift. The total hours worked by this individual may be unchanged, and she may appear to be inelastic along the hours worked margin, but she has responded to the tax reform. Her earned income has 2

likely increased as she receives a premium for exerting extra e®ort and working inconvenient hours. Similar examples where individuals exert additional e®ort on the job or acquire unmeasured skills that result in increased earnings are generally not captured by analyses that focus solely on the response of hours worked. For this reason we believe that our study is an addition to the extensive literature on the response of hours worked in Sweden to the tax reform of 1991, as well as the labor supply literature in general. Our behavioral model builds on a long tradition of work in labor supply, along the lines of MaCurdy (1981) and Blundell and MaCurdy (1998). We write down an explicit intertemporal model of individual behavior and the economic environment.

We consider speci¯cations under di®erent sets of assumptions

regarding market completeness. We estimate the model based on assumptions that the tax reform is anticipated as well as with an unanticipated reform. The analysis of an unanticipated reform requires a careful account of how new information a®ects the individual's decision problem. The large and very detailed data we use allow the analysis of these questions and it is, we believe, a signi¯cant contribution to the existing literature. Estimating the behavioral responses to the changes in marginal tax rates brought about by TR91 is no small task. Several micro simulation studies, most recently Blomquist, Eklof and Newey (2001) and several previous studies by Blomquist ¯nd positive labor supply responses to changes in the after tax wage rate, though relatively small. Direct estimation using instrumental variable methods has frequently found negative uncompensated labor supply elasticities,

3

and occasionally negative compensated elasticities. Compensated labor supply elasticity estimates have generally been small, around 0.10 or less. Klevmarken (2000) ¯nds positive responses for women but not for men. His study is notable since it uses panel data and looks at the 1985 to 1992 time di®erence.

In

an earlier study Flood and MaCurdy (1993) use cross-sectional data but they don't ¯nd much evidence of elastic labor supply. This literature, like much of the labor supply elasticity literature (Heckman (1993), Blundell, Duncan, and Meghir (1998)), has generally found inelastic labor supply, in terms of hours worked on the intensive margin. This paper estimates a more complete measure of the real response to changes in the net of tax rate, the elasticity of earned income. These elasticity estimates provide a point of comparison to previous measures of the elasticity of hours worked. This paper is interesting not only as an analysis of an important policy reform in Sweden, but also as a contribution and extension of several existing literatures. We use very detailed panel data over 6 years to examine the questions. The data is a large random sample of the population, and it allows for careful controls of individual and business cycle factors. We ¯nd positive and signi¯cant responses. The magnitude of the response is substantially larger than what is found in the labor supply literature.

4

2

Model

Consider a model where a ¯nite number of consumers maximize life time utility. Agents are active for a ¯nite number of periods T. Uncertainty in this economy is treated as the probability space (-; F; P ), where ! 2 - represents a particular realization of all random variables in the economy through all time periods. Let the sequence of ¾-algebras F 1 µ F 2 µ ::: µ F characterize how information accumulates over time. E t denotes the expectations operator associated with the information set F t. All information is assumed to be public. Households are assumed to obey the expected utility hypothesis and have rational expectations. Preferences take a time additive form E0

T X t= 1

¯ t U (Cit ; 1 ¡ Hit )

(1)

C it denotes consumption and Hit denotes labor supply by individual i at time t, respectively. Assume that market are not necessarily complete but that there exist assets r 2 R which households may trade each period for current consumption and leisure. Let srit denote the quantity of asset r held by the nth household in period t, qrt its price and drt the dividend. The household face budget constraints of the form X

r2R

[qrt (srit ¡ sri;t+1 ) + sritdr t] + W (Hit) ¡ T (W (Hit )) ¸ C it

(2)

for every time period t. The individual maximizes life time utility with respect to the budget constraints, as well as given initial and terminal conditions. We

5

can formulate the following Lagrangian function 2 L0 = E 0

T X t=1

6 ¯t 6 4

+¸ nt

¡P

3

U (C it ; 1 ¡ Hit) +

r2R

[qrt (srit ¡ sri;t+ 1) + srit drt ] + W (Hit ) ¡ T (W (Hit)) ¡ Cit (3)

The ¯rst order conditions with respect to consumption, labor supply and savings at date t are

U1

=

¸it

U2

¸

¸ nt

=

¸it W 0 (1 ¡ T 0 ) · ¸ qrt+1 + dr t ¯E t ¸i;t+1 = ¯E t [¸ i;t+1 ¼ rt+ 1] qrt

where the last equality de¯nes the return on asset r. These equations are quite hard to use empirically since they depend on the variable ¸ it that is speci¯c to both the household and the time period. However, if we assume that markets are complete this will impose further structure on the multiplier (following Altug and Miller 1990). ¸it = ´ i¸ t

(4)

The multiplier can be decomposed into a household speci¯c term and a common time speci¯c term. The ´ i term captures the individual's potential life time wealth.

The ¸ t term describes aggregate factors. We can then rewrite the

intertemporal FOC as 1 = ¯E t

·

¸ t+1 ¼ rt+1 ¸t 6

¸

(5)

7 7 ¢ 5

We may note that the intertemporal allocation rule does not depend on anything speci¯c to the nth household. Denote by "t+1 the error from forecasting ¯ ¸t+1 ¼ rt+1 at time t. We can then write (assuming "t+ 1 > ¡1) ¸t ln ¸ t+1 ¡ ln ¸t = ln (1 + "t+1 ) ¡ ln (¯¼ r t+1 )

(6)

The second term on the right hand side captures realized returns in the asset markets. The term ln (1 + "t+ 1 ) captures other factors that in°uence lambda. One such shock may be changes in the economic environment like shifts in the tax function due to tax reform. Tax reform a®ects the whole economy since the same tax schedule applies to all individuals. Since certain parts of the tax schedule only are relevant for some sub groups of the population we may want to recognize some heterogeneity of these shocks. The realized shocks to lambda can be captured by appropriately de¯ned time e®ects, as discussed by Altug and Miller (1990). In our case we may want time e®ects that capture aggregate shocks to asset markets, as well as shocks to the tax function.

The timing of the forecasting error is an important consideration as well. The di®erence between the expectation and the outcome depends on the evolution of the information sets. In the case of aggregate business cycle shocks these will be realized on an annual basis. When considering changes in the tax function the situation is somewhat di®erent. Changes in the tax function are arguably permanent changes in the environment. The second issue is when agents incorporate the tax reform in their information sets. There are several 7

possible cases and I outline a few below. 1. The tax reform is expected. Then there is no impact on lambda since it has already been incorporated in the agent's optimization. Lambda has adjusted before our sample period. 2. The tax reform is a surprise some time before the implementation. People know of the reform some time before. Lambda adjusts when the surprise is revealed, for example when the reform is passed into law in 1989. Need to account for changed lambda in time di®erences that include the year 1989. 3. The tax reform is a surprise at the time of implementation. A case of 'seeing is believing', the reform is only credible when implemented. Lambda changes when the tax reform is implemented. All time di®erences that include the tax reform should account for the changed lambda. 4. "Ricardian equivalence." The changed paths of tax liabilities and prices cancel out and there is no change in lambda. Equivalent to case 1 in terms of constant lambda.

2.1

Parametric Utility

Consider a utility function where consumption and leisure are separable. U (C nt ; 1 ¡ Hnt ) =

1+° ° C 1¡ ® H t ° + ®0 t 1¡° 1¡ ®

From the FOC for labor supply we get

8

(7)

1=°

=

¸ nt W 0 (1 ¡ T 0)

Ht

=

[¸ ntW 0 (1 ¡ T 0 )]

Ht

°

(8)

Under the assumption of complete markets we have ¸ nt = ¸ t .At this point we may apply the earnings function to labor supply. £ °¤ W [Ht ] = W [¸ t W 0 (1 ¡ T 0)]

(9)

In order to proceed from here with a parametric solution we need to make some assumption regarding the earnings function. Consider the following speci¯cation. W [Ht ] = Ht ¤ Sj;t

(10)

The subscript j de¯nes a group which has a common earnings function. The term Sj; t shifts earnings functions across groups. The shifter term could take many forms. Consider the following form. ¡ ¢ Sj;t = exp ± 1t Sj1 + ±2 Sj2

(11)

In this speci¯cation there are two types of determinants of earnings functions (or wage functions), variables with time varying returns and variables without time varying returns. The main example of a time varying return is working in a sector where a union negotiates for a general pay raise for the group. Another time varying factor may be business cycle shocks. A characteristic of the wage contract that does not vary over time may be the return to gender. Given 9

the speci¯cation of the earnings function we can express labor supply with the following expression.

ln W (Ht ) = ° ln (1 ¡ T 0 ) + ° ln ¸ t + ° ln W 0 + ±1t Sj1 + ±2 S 2j

(12)

By time di®erencing we get (¢ denotes the time di®erence operator) ¢ ln W (Ht )

= °¢ ln (1 ¡ T 0 ) + °¢ ln ¸ t + °¢ ln W 0 + ¢±1t Sj1 = °¢ ln (1 ¡ T 0 ) + °¢ ln ¸ t + ¢± 1t (1 + °) Sj1

(13)

where the last line follows from the fact that the wage rate equals the derivative of the earnings function with respect to hours worked. The change in the earnings function is now equal to the last term. This term captures changes in earnings that are not related to hours worked or e®ort, only aggregate changes like uniform income growth. These terms could be uniform for the whole economy in which case year e®ects would capture them. However, if these e®ects di®er at a lower level of aggregation we need time e®ects at this level. It may be reasonable to think that general income growth di®ers across sectors and we would need sector speci¯c time e®ects to account for that.

2.1.1

Changes in the Marginal Utility of Wealth

There are two reasons why we need to control for changes in lambda. The ¯rst is aggregate shocks to asset returns and the second is changes in the tax function. Shocks to asset returns can be captured by year ¯xed e®ects.

10

In some cases there is no change in the marginal utility of wealth in the sample due to a tax reform. One example of that is if the tax reform is anticipated before the observed periods. If there are unanticipated changes to the tax code during the observed periods, however, we need to account for the arrival of new information and its impact on the marginal utility of wealth. In the rest of this section we consider what happens when new information arrives, that is, if the tax reform is not anticipated. The e®ect of a tax reform on the relative change of the marginal utility of wealth can be divided into three components 1 . The components depend on how the reform changes tax liabilities, discount factors, and marginal income tax rates. The complete future time paths of these changes will a®ect the marginal utility of wealth, ¸. Let d describe the di®erence operator between the old and the new tax schedule. Then we can write the e®ect on ¸ as T T T X X X d¸ dT 0 ¡1 = B0 R¡1 dT (W (H )) ¡ B [W (H ) ¡ T (W (H ))] dR + B ¯t t 0 t t 1 1;t 1;t ¸ 1 ¡ T0 t=1 t=1 t=1

(14) The ¯rst term on the right hand side is the discounted sum of changed tax liabilities. The second term is the change in the present value of after tax income. The term arises from the e®ect of the tax reform on assets' after tax returns that a®ect individuals' discount factors. The last term captures the new sequence of marginal tax rates. 1 This

is based on a model with Cobb-Douglas utility under certainty where we can compute

the precentage change in the marginal utility of wealth in response to changes in the tax schedule.

11

2.1.2

Measuring Changes in the Marginal Utility of Wealth

The ¯rst component of the change in the marginal utility is the discounted sum of changed tax liabilities. For a measure of this consider the following approach. The tax reform induces a change in the tax bill an individual faces. Assume that the change can be described by an individual ¯xed e®ect and a polynomial in age.. dT (W ) = f i + ¯ 1 ¤ Age + ¯ 2 ¤ (Age)

2

By estimating this equation on time di®erences that include the tax reform we can obtain an estimate of the ¯xed e®ect, f^i . The discounted sum of changed tax bills as of time period t through the agent's last period T¹i is ¹

Ti X

¹

1 R¡ 1; s dT

s=t

(W ) = fi

Ti X s=t

¹

1 R¡ 1;s

+ ¯1 ¤

Ti X s=t

¹

1 R¡ 1; s Ages

+ ¯2 ¤

Ti X

2 R¡1 1;s (Ages )

s=t

The discounted sum of changed tax bills is proportional to the percentage change of the marginal utility of wealth, conditional on age. As long as everyone of the same age faces the same discount factors the two last sums will not vary within age groups. The impact of changed life time tax liabilities can then be captured by ¡ ¢ ¡ ¢ f^i ¤ g1 T¹i ¡ t + g 2 T¹i ¡ t that is feasible to compute and include in the main earnings growth regression. If discount rates vary across individuals of the same age, we can estimate discount factors for the population in pre reform years. With such an estimate we can 12

¡ ¢ PT¹i ^ ¡ 1 ^ ¡1 represents the estimated discount replace g1 T¹i ¡ t by s=t R1;s where R 1;s

factor for that period. One may be concerned about the exogeneity of f^i since it is based in part on post reform behavior. Under the null hypothesis of no behavioral response to the tax reform the f^i will be exogenous. If the null is rejected, however, there will be an endogenous component of f^i but we will argue that it will work to reduce the income e®ect and hence reduce our estimate of the substitution e®ect if leisure is a normal goo d. If there is a behavioral response to the tax reform we would expect individuals to increase their labor supply and earnings. Since the marginal tax rate is always positive, higher earnings translate into a higher tax bill. The observed change in the tax bill will hence be smaller than the 'constant behavior' change. By using the observed changes in tax bills we will potentially downward bias our measure of the income e®ect. The second component is the change in the discounted value of after tax income. To obtain an estimate of the life time path of after tax income we can regress after tax income on individual characteristics using pre reform years. Based on the ¯tted values and estimates on age we can trace out individual income paths. The change in the discount factor over the life time can be obtained in a similar manner. We regress changes in the discount factor (based on individual capital income tax rates) on individual characteristics. We can then construct a series of changed discount factors, which we combine with the projected series of after tax incomes in order to obtain a measure of the second component of (14).

The same approach can be used to get measures of the

third component.

13

2.2

The Concept of Labor Supply

At this point we should elaborate on what we mean by labor supply.

In the

model above labor supply is represented by H , a one dimensional measure. Many empirical studies interpret labor supply quite literally to mean the number of hours worked. Here we take a broader view of labor supply, incorporating an e±ciency units approach where the quantity and quality of labor supply is allowed to vary. We argue that labor supply can incorporate many dimensions in addition to the number hours worked. Many of these aspects of labor supply may not be as easily observed as hours, yet they should not be ignored when estimating the behavioral response to tax. One important dimension of labor supply is e®ort. E®ort may be observable in some cases, such working the day shift versus the night shift. In many cases e®ort may not be readily observable to the econometrician.

An example of

unobserved e®ort may be the acquisition of skills on the job. To incorporate these qualitative dimensions of labor supply we elaborate on the simple model above to incorporate a type of e±ciency units measure of labor supply.

We

allow labor supply to denote some e®ort adjusted quantity of hours provided by the individual. We represent labor supply as the product of e®ort, e, and hours worked, l, H = e ¢ h; where e and l can take on some range of values, the product of which cannot exceed H. Labor supply is still a one dimensional measure of productive e®ort,

14

but it is a function of both the quantity and quality of e®ort exerted by the individual. This elaboration on the simple model does not change the qualitative results of the equivalence of the earned income elasticity to changes in the net of tax rate and the labor supply elasticity to the wage rate. What this elaboration makes clear is the advantage of estimating the earned income response to changes in the tax rate when the total labor supply response is unobservable. In the context of this simple model we have established the equivalence of the elasticity of earned income to the net of tax rate and the labor supply elasticity to the wage rate.

The elasticity of earned income to the net of tax

rate is equivalent to the familiar labor supply elasticity to the wage in this simple model. The theoretical motivation is the same. What is di®erent about this approach to previous studies is that it allows for labor supply to vary not only in the number of hours worked, but along other less easily measured margins such as e®ort. In practice, by estimating the response of earned income to changes in the net of tax rate we hope to capture not just quantitative changes in the number of hours spent at work, but also qualitative changes in labor e®ort. It is reasonable to believe that for many full time workers in Sweden qualitative labor supply responses to changes in the net of tax rate are the most elastic. To ignore these responses to changes in the net of tax rate may underestimate the true labor supply elasticity, in particular for those who already participate in the labor force. By using earned income as a measure of productive e®ort we hope to provide a means of comparing how signi¯cant qualitative labor responses are relative to hours of work alone, as documented in an extensive literature on

15

hours worked. There is second and more practical motivation for looking at incomes directly. In most studies of hours worked the wage is computed as earned income divided by hours worked, so the wage is a transformation of earned income. Since hours worked exhibit bunching at certain levels it is usually assumed that there is measurement error in hours worked. Introducing measurement error in hours mechanically translates in to measurement error in wages. By examining earnings directly we bypass two confounding steps, that is, we don't infer wages from incomes and we don't introduce measurement error as a way to align theory with observed hours choices.

2.3

Income E®ects

TR91 was not revenue neutral.

As illustrated in an earlier table, the income

e®ects to the household sector in terms of both the reduction in tax paid, as well as the steady increase in transfers received may have been signi¯cant. In this respect, it is not surprising that uncompensated elasticity estimates to changes in the tax price may be small, as the income e®ects of the tax reform are potentially large.

We are interested in estimating both the compensated

and uncompensated elasticity of earned income to tax. To compute the e®ect of taxes on e±ciency we would like to know the compensated elasticities.

To

control for income e®ects from the tax reform we not only need to consider the change in the total tax bill to the individual, but also changes in other incomes associated with the reform.

The redistributive aspects of tax systems should

16

not be ignored. The income e®ects of the tax reform may vary not just in the change in taxes paid, but also in the amounts of transfers and bene¯ts received from the government. To illustrate the idea behind controlling for income e®ects in a static setting consider a simple example.

Imagine an economy where the proportional tax

is such that the maximum attainable consumption is given by point A in the graph below, and the agent chooses leisure L0 .

Suppose now that there is a

reduction in the tax rate such that the maximum feasible consumption increase from point A to point B. The tax reform induces two e®ects, the relative price of leisure has increased and the agent is richer since less tax is paid. To obtain a compensated elasticity we want to examine the e®ect of the change in prices holding income constant. Graphically, we shift the new budget line left such that the old consumption leisure bundle is just a®ordable, which is the dotted line parallel with the new budget line. The amount by which we shift the budget line left is the change in the tax bill.

This exactly accounts for the income

e®ect if there is no behavioral response to the tax reform, that is, if the agent chooses L0 in both states. Given that individuals respond to the tax reform, the observed change in the tax bill is arguably smaller than if behavior was constant. Using the observed change of the tax bill will hence understate the true income e®ect of the tax reform and potentially downward bias estimates of the substitution e®ect. We will address the endogeneity of the tax bill by instrument for it.

When controlling for how much tax individuals pay the

estimated tax elasticities will be compensated in terms of holding real income

17

Leisure

L0

∆ Taxbill

A

B

Consumption

Figure 1: Simple illustration of the substitution and income e®ects of a tax reform in a static setting. constant.

2.4

Price Fluctuations

A key assumption imposed when estimating the labor supply elasticity to changes in the tax price is that other prices are constant. Even when labor supply is assumed to be observable, regressions of change in labor supply on change in tax price will be biased if wage changes are not accounted for. When earned income is the measure of interest the need to control for price changes is of great importance, as prices enter directly into the dependent variable. In the results reported here we control for wage growth di®erentials based on detailed individual level observables, a variety of business cycle controls, and time varying

18

region and sector interaction terms.

In this respect we can control for varia-

tions in individual income due to observed characteristics such as the level of education obtained, or changes in the number of household members, in addition to asymmetric aggregate shocks to wages at the region-industry level. The controls we have included are much more extensive than those previously used in the literature.

3

Empirical Model

We specify an empirical model based on the theoretical model discussed in the previous section. The log earned income for individual i, at time t, is a function of the log of the net of tax rate, ln(1 ¡ ¿ ), and the log of the marginal utility of wealth, ln ¸ t.

We control for variation in wages and skill prices

by incorporating individual demographic controls as well as region/sector/year speci¯c wage trends contained in Xi . To control for individual level variation in preferences we also include an individual ¯xed e®ect. The following is a version of this model: ln Yit = i i + atXi + ° ln(1 ¡ ¿ t ) + Á ln ¸t + " it :

(15)

We estimate this model in di®erences, instrumenting for the terminal period marginal and average tax rates. By di®erencing the above process we get ¢ ln Y it+ 1 = (at+ 1 ¡ at)Xi + °[ln(1 ¡ ¿ t+1 ) ¡ ln(1 ¡ ¿ t )] + Á [ln ¸ t+1 ¡ ln ¸ t ] + "i;t+1 ¡ "i;t (16)

19

The coe±cient ° on the net of tax rate is the parameter of interest, the elasticity of income to the net of tax rate. Depending on the assumptions and the variables included to control for changes in the marginal utility of wealth the interpretation of ° will change.

In the case of a static model (or no capital

markets) the change in the marginal utility of wealth can be captured by ¢ ln ¸ t+1 = ¢ ln ¿¹ i;t+ 1 + ¢Pi;t+ 1

(17)

where ¹¿ i;t is the average tax rate and Pi;t is non labor income. In this speci¯cation we interpret ° as the (income constant) compensated elasticity of labor supply with respect to the net of tax rate. Next consider a dynamic model where the tax reform is anticipated and markets are complete.

The change in the marginal utility of wealth is then

only time e®ects to capture common shocks to asset markets.

Under these

assumptions we estimate a 'lambda constant' elasticity of labor supply, usually referred to as the intertemporal elasticity of substitution of labor supply, ° IE S . The third main speci¯cation is a dynamic model where the tax reform is unanticipated.

Under these assumptions the change in the marginal utility

of wealth is more involved.

As discussed in the previous section the change

in the marginal utility of wealth can be decomposed in to three parts.

Mea-

sures of these parts can be obtained, although it requires strong assumptions. We need to assume that we know individuals' planning horizons and that the currently middle aged can tell us something about the life time paths of the currently young. Given these assumptions we can for estimates of the components

20

a®ecting the marginal utility of wealth. We get ¢ ln ¸ t+1 = b0

T X t=1

\ ^ ¡1 dT (W R (Ht )) ¡ b 1 1;t

T h X t=1

T i X \ ¡1 dT 0 \ c 1;t + b2 W (Ht) ¡ T (W (Ht )) dR ¯t 1 ¡T0 t=1

(18)

where hats indicate estimated values. In this speci¯cation the ° is the intertemporal elasticity of substitution °I ES , given by anticipated movements along price paths, plus the e®ect of the unanticipated change in price paths by the tax reform.

This estimate is the best suited to evaluate the e®ects of tax reforms

since it incorporates the unanticipated shift in future price paths. After constructing our instruments we estimate the elasticities directly by instrumental variable methods. In order to control for business cycle e®ects and changes in skill prices or wages related to observed characteristics we incorporate a rich set of demographic, income, sector, and business cycle controls. In our most °exible speci¯cations we control for asymmetric trends in income growth and business cycle shocks by yearly ¯xed e®ects for sectors within regions. We also allow for di®erent income groups to respond di®erentially to business cycle shocks and control for widening income distribution.

3.1

Identi¯cation

We identify the parameter ° by comparing the growth rates of income to the growth rates of the net of tax rate across individuals. The model we use is linear and we capture the e®ect on income growth that is proportional to the change in the net of tax rate. The estimate of ¯ will be positive if individuals with relatively larger increases of the net of tax rate have higher income growth. 21

We exploit the fact that di®erent individuals get di®erent treatment, that is, the reduction of the net of tax rate varies across individuals. Since there is an increase in the net of tax rate on average it may seem that we can not estimate the e®ect of the average change. However, the linearity assumption is important in this instance. We identify the e®ect of taxes based on di®erences between individuals, which is only a part of the treatment the individuals receive. Linearity tells us that the e®ect of the treatment is proportional to the treatment. We can apply the treatment parameter ° to the whole change in net of tax rates to evaluate the e®ect of tax rate changes on income. The linearity assumption is important in our analysis and we can test it2 .

3.2

Instrumenting

Marginal tax rates are typically a function of income. If the schedule of marginal tax rates were °at and everyone faced the same marginal tax rate regardless of their income it would not be necessary to instrument for marginal tax rates in period t + 1. Under a °at tax system, the marginal tax rate you face would be independent of your terminal period income. In this case OLS would provide an unbiased estimate of the elasticity of earned income to the exogenous change in tax rates. This can be illustrated by decomposing the change in the net of tax rate into two components. The change in tax rate associated with the tax reform (from schedule ¿ t to ¿ t+ 1 ) is an exogenous source of variation in the tax 2 We

test linearity by adding the square of the treatment term, (¢ ln(1 ¡ ¿ t))2 , in the

regression. The squared term is not signi¯cantly di®erent from zero in the regression and we do not reject linearity.

22

rate. The second component of observed tax rate variation comes from changes in terminal period income.

That is the di®erence between observed income,

Y t+1; and trend income, (1 + g)Y t ; or what your income would be if it grew at some average rate of income growth, g.

The di®erence between the tax rates

faced on these two terminal period incomes is a function of Y t+1 , part of the left hand side variable, and therefore endogenous. Consider expanding the change in the log of the net of tax rate to show these two components of the observed change in marginal tax rates: ¢ ln (1 ¡ ¿ t+1 ) = ln(1 ¡ ¿ t+1 ((1 + g)Y t)) ¡ ln(1 ¡ ¿ t (Y t )) + ln(1 ¡ ¿ t+1 (Y t+1 )) ¡ ln(1 ¡ ¿ t+1 ((1 + g)Y t)): | {z } | {z } exo genous chang e

endog enou s change

(19)

In the case of a °at tax schedule, ¿ t+1 would be the same for all Y . The last two terms would cancel, and there would be no endogenous component of the change in tax rates. Consider a tax system that is progressive in the sense that those with higher incomes face higher marginal tax rates, and lower net of tax rates. A tax rate reduction, ¿ t > ¿ t+ 1 , would increase the incentive to work and earn income, and for those whose tax rates were reduced relatively more we would expect their income to rise relative to trend. Under a progressive tax system this increase in income could in turn increase observed marginal tax rates. This increase in marginal tax rates would translate into a reduction in the observed net of tax rate.

As a result the endogenous component of the change in the net of tax

rate would be negative. Estimating the empirical model directly with OLS on

23

observed changes in marginal tax rates would therefore introduce upward bias in our estimates of ° since the observed total change in the net of tax rate would be less than the exogenous change. Likewise, in a regressive tax system where those who earn higher incomes face lower marginal tax rates the use of OLS would result in a downward bias in estimates of °; since the endogenous change in the net of tax rate would be positive. By instrumenting for the change in the net of tax rate we isolate the response to the exogenous change in taxes. This estimate may be greater than or less than the OLS estimate of °, depending on the amount of progressivity or regressivity in the tax system. The relation between OLS and IV estimates of the tax elasticity term ° is the following: ° IV =

¢ ln Y ¢ ln Y ? = ° OLS : ln(1 ¡ ¿ t+ 1((1 + g)Yt )) ¡ ln(1 ¡ ¿ t (Y t)) ln(1 ¡ ¿ t+1 (Y t+1 )) ¡ ln(1 ¡ ¿ t (Y t)) (20)

We forecast terminal period income, Y t+1 , as initial income growing at the region-sector observed income growth rate, (1 +g)Y t , where g is the growth rate. We then use this forecasted income to compute the constant behavior tax rate under the post-reform tax scheme, ¿ t+1 ((1 + g)Y t ). This is the instrument for the observed terminal period marginal tax rate, ¿ t+1 (Yt+1 ). We then construct the instrument for the change in the net of tax rate as the di®erence between the log of this constant behavior net of tax rate and the log of the initial period net of tax rate, [ln(1 ¡ ¿ t+ 1((1 + g)Yt )) ¡ ln(1 ¡ ¿ t(Y t ))]. This instrument is similar to what is used in for example Klevmarken (2000) and Feldstein (1995), although we apply the income pro jections at a lower level of aggregation. 24

The instrument is still a function of initial period income and therefore correlated with initial period income shocks. This could introduce bias. Mean reversion, in particular, is one type of bias that may plague this estimator, since in some speci¯cations our sample is truncated from below. To address this we add various initial income controls in the speci¯cations we estimate.

It may

also be noted that the tax rates are based on taxable income, which includes more components of income than earned income. Since we examine three year di®erences we can allow for error terms to be autocorrelated up to the second order without a®ecting our estimates. The mean reversion controls are used to capture potential higher orders of autocorrelation. Autocorrelation of income that in°uence the tax rate instrument is still a concern. Temporary shocks to initial income will in°uence pro jected income and the tax rate instrument. We have experimented with using additional instruments to control for this mean reversion, but without substantial impact on the results. In the case of Sweden after TR91 the actual schedule of marginal tax rates is somewhat regressive.

Even excluding the rate reductions observed among

the highest income groups, sixty percent of the tax payers in our sample in any given year are located on regressive portions of the tax schedule, that is, portions of the tax schedule where if their income increased to the next bracket they would face lower marginal tax rates. The average tax rates are also functions of income, and are endogenous to the choice of ¯nal period income. To address this we instrument for the ¯nal period average tax rate. The approach is similar to the marginal tax rate instrument.

25

The instrument for the average tax rate is the individuals initial average tax rate plus the change for the sector and region group that the individual is in, denoted by ¹g . Consider a decomposition of the average tax rate.

ln ¹¿ t+1 ¡ ln ¹¿ t = ln (¹¿ t + ¹g ) ¡ ln ¿¹ t(Yt ) + ln ¹¿ t+1 (Y t+1 ) ¡ ln (¹¿ t + g¹) | {z } | {z } exog eno us change

(21)

e ndogenou s chang e

The exogenous change is the percentage change of the average tax rate when there is no individual response to the change in ¿¹ , which we call the 'behavior constant' change. The exogenous change is used as an instrument for the observed change in the marginal tax rate. The endogenous change is the di®erence between the observed individual average tax rate and the behavior constant average tax rate. The sign of the di®erence is unknown in general and depends on the relation between the average and marginal tax rates. Consequently, the relation between the IV and the OLS estimator of the average tax rate is not known in general. The computations of the tax rate instruments are discussed in more detail in a data appendix.

4

The Tax Reform of the Century

The background of the Swedish tax reform of 1991 (TR91) was an increasing awareness of the deadweight losses from the highly progressive tax code. One objective of the tax reform was to correct many of the perverse incentives induced by the old tax code. These incentives were particularly pronounced at the higher end of the income distribution. One argument for TR91 was that it would make people work and earn more, and that would actually increase tax 26

revenues. It is natural to ask if earned income increased as marginal income tax rates decreased, and if so to what extent? The magnitude of the tax rate change due to TR91 in Sweden is large. In the U.S., TRA86 increased the marginal net-of-tax income per dollar earned by 44% for the wealthiest. In Sweden the wealthiest had a much more dramatic gain in net-of-tax income per dollar earned. Before TR91 the highest income Swedes faced statutory marginal tax rates of 85%, or a net-of-tax rate of 15%. After the reform the maximum marginal tax rate declined to 50%, and the netof-tax rate increased to 50% from 15%. This implies a possible increase in the net-of-tax rate of 230%.

The scale of the TR91 tax reforms are much larger

than any seen in the history of U.S. income taxation, where the largest changes in net of tax rates for the wealthiest were less than 150%, observed after the Revenue Act of 1964 (Goolsbee, 1999). The tax reductions in TR91 were broad and not limited to the highest income earners. As mentioned above, the average net of tax rate was increased by 24.6%. The 1986 tax reform in the US is by many accounts a large reform with an increase in the net of tax rate of 4.8%, but it dwarfs in comparison to TR91. TR91 entailed a broad revision of personal income tax rates, as well as consumption taxes, corporate taxes and the transfer system. The multiple reforms included under the umbrella of TR91 imply many changes to the de¯nitions of income and deductions in addition to dramatic tax rate reductions. The table below summarizes some of the key tax rates in Sweden over the years we consider:

27

Tax Rates 1989 1990 1991 Earned Income Taxes: Kommun Rate minimum rate 26.7 26.7 26.7 maximum rate 33.25 33.49 33.49 State Rate minimum rate 5.0 3.0 0.0 maximum rate 42.0 35.0 20.0 Other Taxes: Corporate Tax Rate 52.0 52.0 30.0 Capital Income Tax Rate EI tax EI tax 30.0 VAT Rate 25.0 25.0 25.0 Public Pension 0.0 0.0 0.0 Payroll Taxes: a. Social Security Tax 28.0 28.0 28.0 b. Salary Tax 18.0 18.0 18.0 c. Unincorp. Bus. Tax 25.0 25.0 25.0 A key similarity between TR91 and TRA86 in the US

1992

1993

1994

26.5 33.48

26.47 33.47

26.47 33.47

0.0 20.0

0.0 20.0

0.0 20.0

30.0 30.0 25.0 0.0

30.0 30.0 25.0 0.95

30.0 30.0 25.0 1.95

28.0 28.0 28.0 18.0 18.0 18.0 25.0 25.0 25.0 is the di®erential tax

treatment of capital gains and other sources of non-labor income post-reform. After TR91 capital income was taxed at a °at rate of 30%, regardless of earned income. Pre-reform, capital income had been taxed at the same rate as earned income. Because of this fundamental shift in the tax treatment of capital income we turn our attention to narrower income measures which exclude capital gains. The elasticity of earned income is the focus of our analysis. While an analysis of how capital income responded to the tax reform would be interesting, we will not address this here. In addition to the changing tax treatment of various forms of income, the de¯nition of taxable income also changed.

Pre-reform, thirty percent of cap-

ital losses were deductible from reported labor and business income. Many deductions claimed against labor income before the tax reform, such as interest deductions, were no longer directly deductible from income after TR91. Instead, interest deductions were reported as deductions against capital income. 28

Net

Table 2.1: Aggregate Income and Deductions Data (in 100,000 SEK) Income/Expenditure 1989 1990 1991 1992 1993 1994 Total Income 782,600 795,325 737,234 741,439 731,459 744,801 Taxable Income 619,685 638,086 635,899 641,215 624,387 640,787 Deductions/Exclusions (inc. basic deduction) 162,915 157,239 101,335 104,271 107,072 104,014 Share of Taxable Income 0.26 0.25 0.16 0.16 0.17 0.16 Deductions/Exclusions (exc. basic deduction) 95,203 95,417 8,508 7,697 12,460 19,760 Share of Taxable Income 0.15 0.15 0.01 0.01 0.02 0.03 *Note 1991-1994 values exclude capital income (average capital income is negative). Source Statistisk Årsbok 1991-1997.

Figure 2: capital losses were then partially deductible from labor income. These changes in tax income de¯nitions had a sizable e®ect on the absolute and relative importance of deductions and exclusions as a share of total income to the household sector. In the case of Sweden before TR91, and to a lesser degree after TR91, the magnitude of tax favored forms of income and consumption are large relative to total reported income. According to aggregate income accounts, the tax code reforms and base broadening measures of TR91 resulted in a dramatic drop in deductions and exclusions as a share of taxable income.

These mea-

sures understate the true magnitude of tax avoidance behavior since they cannot capture the unreported or non-taxable income/consumption (certain perquisites for example) some may enjoy in lieu of taxable income and consumption. Tax avoidance through use of deductions and exclusions has been a subject of much interest.

Table 1 suggests that avoidance behavior in Sweden was signi¯cant

before TR91. An empirical evaluation of the response of taxable income and quanti¯cation of the importance of avoidance will be addressed in a future paper. The base broadening aspects of TR91did not fully o®set the tax rate reductions, at least for the household sector. It should be noted that while the tax 29

Fiscal Policy and the Household Sector Overview (1994 SEK) Income/Expenditure 1989 1990 Taxes on Income and Wealth 360,856 339,394 Transfers (Local and State) 323,600 323,178 Net Transfers (Transfers-Taxes) -37,257 -16,216

1991 335,849 372,506 36,657

1992 275,234 363,740 88,506

*Values are 100,000 1994 SEK. www.scb.se.

Figure 3: Overview of taxes and transfer. rate reductions associated with TR91 were phased in the household sector realized a real reduction in the amount of taxes paid of approximately 20 percent. This reduction in the total tax bill was accompanied by an increase in transfers to the household sector of almost 15 percent.

In the aggregate this suggests

that the income e®ects of TR91 and the accompanying increase in real transfers to the household sector may be large. The ¯nal distinguishing feature of the tax reform in Sweden is the macroeconomic environment of the time. Unlike TRA86 which occurred during a time of relative economic expansion, TR91 in Sweden predated a severe recession in Sweden that began in late 1990 and continued through 1993. The late 1980's were a time of economic expansion in Sweden. In 1989 GDP growth was 2.7%, while private consumption grew at 1.8%. The boom soured by the early 1990's when real growth declined 1.8%, and private consumption declined by 2.0% in 1993. [Figure 1. Aggregate hours graph] The evolution of hours worked by the employed shows a steady increase during the 1990's, as reported by the OECD. This suggests a behavioral response 30

1993 277,932 376,040 98,108

to the tax reform among the employed. When we consider the fraction of the population that is employed the picture is di®erent. There large decline in the employment share starts in the early 1990's and can in a large part be associated with the increase in unemployment during the recession. [Figure2. Aggregate employment graph] The next graph shows three time series (source: Statistics Sweden). solid line is the real GDP growth that captures the business cycle.

The

We can

note the negative growth in the early 1990's, which rebounds later. The long dashed line shows annual changes in the consumer price index. The development is quite dramatic with double digit in°ation around 1990 and a sharp decline thereafter. The high in°ation rate made it hard to maintain the ¯xed exchange rate, the Swedish krona was pegged against a basket of European currencies. On November 19, 1992, the exchange rate peg was dropped and the krona saw a substantial depreciation.

This produced a burst of imported in°ation after

which prices have been stable. The third time series is the change in salaries, the short dashed line.

The series is for white collar workers in the private

sector. Salaries are what ¯rms have contracted to pay their employees in full time equivalents. The salaries do es not include bonuses etc. It can be seen as a measure of the wage structure. We note that real salary growth is negative in both 1990 and 1991. This could be a response to the tax reform.

If the tax

reform increases labor supply one would expect wages to decrease in response, and what we observe could be an immediate response in prices. In an estimation strategy with aggregate time e®ects, the general equilibrium response would be

31

picked up by the time e®ects and the labor supply response as measured by earnings would be underestimated. [Figure 3. GDP in°ation salary graph] This macroeconomic environment is signi¯cantly di®erent than that of the US in the late 1980's. To what extent these macroeconomic factors in°uence income elasticity estimates will be addressed later.

5

The Data

The data used in this study is the Longitudinal Individual Dataset from Statistiska Central Byran (SCB) Sweden. The data contains approximately 3% of the Swedish population that are followed over time, and the sample is crosssectionally representative in any given year. We examine six years, from 1989 through 1994. This sample is a random sample of the Swedish population, and does not over-represent any income or demographic group. The 1989-1994 data sets consists of approximately 270,000 sampled individuals per year. Of those sampled, 109,000 individuals are between the ages of 25 and 55 in 1989.

We

follow these prime aged individuals from the 1989 cohort for six years. No new individuals enter the sub-sample we construct. The data is a registry based database, and reports information from not only the Swedish taxing authority, the RSV, but also from other local and national agencies that administer ¯nancial support programs.

With this additional

data on transfers and subsidies we are able to select a sample to control for

32

various circumstances that may in°uence individuals' ability to respond to tax rate changes. The summary statistics for our sample describe a much di®erent economic environment than that which existed in the U.S. in the late eighties. The macroeconomic downturn is working against the income gains we would anticipate in response to the tax reform. Tax rates fall, while real labor income remains constant, and earned income falls.

Transfers, both taxed and

untaxed are growing over this period, as well as unemployment.

The table

below summarizes these averages (amounts are in 1994 SEK). 1989 1990 1991 1992 1993 Age 39 40 41 42 43 % Male 51 51 51 51 51 % Public Sector 38 38 39 40 40 % Private (Non-Fin) 56 56 55 54 53 % Financial Sector 2.3 2.2 2.2 2.3 2.5 % Unemployed 4.7 4.4 6.0 10.4 14.3 Total Labor Income 174,000 178,000 174,000 175,000 171,000 Earned Income 154,000 157,000 152,000 152,000 145,000 Taxable Transfers 15,900 16,100 16,400 17,200 18,600 Non-taxed Transfers 7,000 6,800 8,000 8,100 8,200 Marginal Tax Rate .498 .471 .368 .354 .355 Tax on Labor Income 63,800 62,400 53,100 52,500 51,800 Total Taxes 66,700 65,200 55,000 53,100 54,300 House Price Index 182 204 218 197 174 We focus on earned income as our measure of labor supply. This measure is constructed by subtracting taxable transfers, as well as annuity and pension payments from the total labor income measure reported by the taxing authority. Earned income declines in real terms over the sample period, while total labor income remains fairly constant. This decline in earned income is coupled with growing transfers. As incomes stagnate and tax rates fall the tax bill on labor income declines, as does the total tax bill which includes additional taxes such as wealth and capital income taxes.

33

1994 44 51 40 53 2.7 15.3 174,000 147,000 18,300 8,200 .362 53,700 59,400 183

We compute the marginal tax rate that each taxpayer faces as the sum of both the Kommun and State tax rates they face for each year using the Swedish tax schedule and the taxable income reported.

For 1989 and 1990 we make

the additional assumption that the last dollar of income is subject to all state taxes that the agent is liable for (both the basic and additional state tax if applicable). Prior to TR91 the state tax (national tax) was a two tiered system that was a function of two di®erent income computations (basic amount and additional amount) and two tax schedules for these income computations (basic tax and additional tax). This method of double income measurement, and double taxation of income, was a means to combat tax evasion. In practice this method meant agents faced a marginal tax rate that was the sum of two state tax rates and the local kommun tax rate. Kommuns set the local tax rate. The rates are quite dispersed with di®erences up to 9 percentage points across about 280 Kommuns. Taxes are assessed individually in Sweden. There is no joint taxation of income for married couples. With the net of tax rate we can de¯ne a tax price variable as the di®erence in logs of the net of tax rate in each year. A simple IV procedure was describe before. We construct an instrument by computing the terminal period tax price using initial period income pro jected ahead to terminal period levels according to the growth rate of income over the same period by sector and region. The instrument is the di®erence in logs of this synthetic tax price (at the terminal period pro jected value) and it instruments for the actual tax price computed from the data.

34

The three overlapping di®erences considered here include two large increases in the net of tax rate (tax rate reduction) and one small decline in the net of tax rate (tax rate increase) in the ¯nal three year comparison due to an increase in the social security tax which went into e®ect in 1994.

This shift

in marginal tax rates is illustrated by the empirical cumulative distribution functions plotted below. We see how the curves shift to the left, which means that more individuals face lower marginal tax rates. [Figure 4. CDF plot of marginal tax rates] In addition to detailed income source data it also contains detailed data on non-tax factors such as education (level attained, area of study, and year completed), employment sector by SIC code, occupation, and household size. Of equal importance, the data also contains detailed information on numerous forms of wealth as well as non-taxed forms of income for all household members. Such detailed and reliable data yields interesting insights into behavioral responses beyond the scope of previous studies. The results are reported in the following section for the elasticity of earned income to the net of tax rate. The regressions reported here incorporate control variables and selection criterion to control for the e®ects of the business cycle. We consider two samples in the analysis. In the unrestricted sample we include everyone who is between 25 and 55 years old in 1989, and includes about 100,000 individuals. The second sample is restricted in two ways. We require that in the initial year of each time di®erence the individual earns at least SEK 60,000 3 from 3 For

comparison, 1 USD is roughly equal to 8 SEK.

35

employment and receives less than SEK 50,000 in taxable transfers. With the earned income exclusion we focus on the intensive margin of labor supply. The transfer exclusion drops individuals who may face incentives which we don't measure properly, although it would only be a problem if changes in those incentives are correlated with changes in the tax code. The restricted sample contains about 90,000 individuals.

When looking at the unrestricted sample

we allow for adjustments on both the intensive and extensive margin of labor supply.

However, the inclusion of the extensive margin is only partial since

individuals with zero earned income will turn into missing values when we look at the log of earned income. We do not at this point model the participation decision nor do we control for potential selection. The time periods we look at are three 3 year di®erences from 1989 to 1994. Two of the time di®erences starting in 1989 and 1990 include the large tax reform.

6

Results

In this section we report the response of earned income to changes in the net of tax rate under di®erent empirical speci¯cations. We control for demographic, income and business cycle factors that may e®ect the growth of earned income. We have found compensated elasticity estimates of earned income with respect to the net of tax rate ranging from 0.34 to 0.56 depending on the estimated model. These estimates are larger than previous labor supply estimates in Swe-

36

den which have been approximately zero or less when using the hours worked measure. The income e®ects of the tax reform are signi¯cant in the unrestricted sample. The compensated elasticity estimates we report are fairly close to the uncompensated elasticity estimates. Though these earned income responses may seem small relative to the taxable income elasticity estimates of Feldstein (1995) which ranged from one to greater than three, the results are in the same range as the more recent estimates of Gruber and Saez (2002) which were about 0.40.

Our estimates are

substantially larger than the estimated labor supply elasticities found in Swedish studies, which have found uncompensated elasticities that are frequently negative or in the neighborhood of zero, and compensated elasticities of about 0.10. We will ¯rst present our results from the static model. In this model income e®ects are captured by the change in average tax rates and changes in non labor income. The results are presented in Table 6.1. The estimated responses to the net of tax rate are positive and highly signi¯cant. In the ¯rst column the elasticity of earned income with respect to the net of tax rate is 0.348.

The

estimated income e®ect of the tax reform as captured by the average tax rate is positive and signi¯cant, though the magnitude is much smaller than for the net of tax rate.

Note also that the positive estimate has the expected sign when

leisure is a normal good. The typical change in the average tax rate is negative so individuals get more after tax income after the reform, which would tend to reduce labor supply. The demographic variables are signi¯cant and have the expected signs. In

37

the Married category are both legally married couples and cohabitating couples with children. Singles are not yet married and the excluded category is divorced and widows/widowers. Earned Income Lag is a three year lag, that is the initials income in each time di®erence. It controls for mean reversion in income. second set of controls for mean reversion is income deciles.

A

The deciles are

based on 1989 total income (not only earned income) and allow for asymmetric mean reversion across income groups. The non labor incomes we control for are capital and business income, transfers, and labor income of other members of the household. Most point estimates have the expected negative sign although several estimates are not signi¯cantly di®erent from zero. We control for the business cycle in several ways. We control for the changes in the employment rates for sectors of the economy. The interaction of income deciles with GDP growth allows di®erent income groups to respond di®erentially to the business cycle. The last set of controls are sector year e®ects, which are time e®ects for the three main sectors in the economy (public, private non¯nancial, and private ¯nancial sectors). These sector time e®ects control for both shocks to the business cycle as well as shifts in the wage structure.

The

standard errors are based on robust (Huber/White/sandwich) standard errors, and we allow for autocorrelated individual error terms. In the second column we expand the number of time e®ects as indicated on the last line. Region/Sector/Year is a set of time e®ects for each sector within each of 8 regions.

This is an expansive set of controls, but the e®ect on the

estimates are limited. The elasticity of the net of tax rate is slightly reduced

38

and the e®ect of the average tax rate turns insigni¯cant. In the third column we look at the unrestricted sample, which adds about 7,000 individuals with low initial earned income and/or high transfers. The estimated responses to both the net of tax rate and the average tax rate increase substantially to 0.43 and 0.14, respectively. In Table 6.2 we turn to a dynamic model with complete markets. We assume that individuals anticipate the tax reform and that agent's have incorporated it in their information set prior to out sample period. The only shocks to the marginal utility of wealth that individuals face is the di®erences of expected and actual returns in asset market, which are common to everyone and can be captured by time e®ects. We use the same demographic controls as in the static model, including the age polynomial. The estimates on the net of tax rate is somewhat larger in the dynamic model in the restricted sample compared to the static model, and still very signi¯cant. The estimates can be interpreted as the intertemporal elasticity of substitution of labor supply. The di®erence between the restricted and unrestricted samples is small in this model. The next step we take is to change our assumption regarding the individuals' expectations. We will now assume that the tax reform is unanticipated. When the reform is not anticipated there will be an e®ect on individuals' marginal utilities of wealth, which we need to control for. As described in equation (18) this e®ect can be quite elaborate. Before we can estimate income growth equation we need to obtain measures of the components in (18). The levels denote

39

quantities pre reform. The di®erential terms describe changes from pre to post reform where we will use time di®erences covering the reform.

For the level

quantities we will use the pre reform years 1989 and 1990, and for the di®erences we look at the 3 year time di®erences starting in those years. Throughout these computations we assume that the individuals planning horizon is to age 65, the age by almost everyone has retired. For the change in the tax bill we run an individual ¯xed e®ects regression including age and age squared and it produces the observed U-shaped pattern of reduced tax bill over the life cycle, and using the ¯tted values and the age coe±cients we get individual time paths. For all other variables we run random e®ects regressions4 , where we included age, agesquared, gender, and education e®ects. We generate life cycle paths by using ¯tted values and the age coe±cients. The discount factors are based on a pre tax real return of 5%. The rate the individual faces is the after tax rate and we will use the marginal tax rate on capital income. Pre reform the capital income tax rate equals the marginal tax rate on labor income. Post reform the tax rate on capital income drops to a °at 30% rate. Discount factor are based on after tax returns, which we regress on age terms and time invariant characteristics. The after tax income in the second term is de¯ned as total reported income minus all direct taxes the individual is liable for. The last term captures the changed time path for marginal tax rates. We may be concerned there is a 4 We

tried ¯xed e®ects regressions as well but they had a very high correlation (usually

over -0.9) between the ¯xed e®ect and the age terms. This resulted in that the estimates on the age terms were counter to what we observe. The change in tax bill was di®erent with a much smaller such correlation of -0.05.

40

behavioral component in this measure which is correlated with age To address this we will instrument for this term using gender and education interacted with age and age squared. It should also be noted that the terms to control for the change in lambda are only included in time di®erences that span the reform year 1991. The time di®erence starting in 1991 does not include these terms since the tax reform has been implemented and we assume it is in the individuals' information set. In the presented result we will use the following terminology ¢ ln ¸ t+1 = b0 ¤

T X

t= 1

|

\ ^ ¡1 dT (W R (Ht )) i ¡ b1 1;t {z

}

Change in T ax B ill F acto r

T h i ¡1 X c 1;t + b2 ¤ W (Ht) \ ¡ T (W (Ht)) dR t=1

|

{z

Chang e in Discount F actor

}

(22)

The results are presented in Table 6.3. In the ¯rst column we present, for comparison, the dynamic model under the assumption of an anticipated tax reform. time.

In the subsequent columns we introduce the factors in (22) one at a In the second column we introduce the tax bill factor.

The elasticity

of the net of tax rate increases slightly. The tax bill factor has the expected positive sign (lower tax liabilities would tend to reduce labor supply if leisure is a normal good) but has a low level of signi¯cance. In the third column we add the discount factor. The net of tax rate elasticity is reduced. The negative sign on the discount factor is what we would expect, a reduced discount factor (dR ¡1 < 0) from the reduction in capital tax rates reduces the discounted value of future earnings which would tend to increase labor supply. signi¯cant.

The factor is

In the last column we add the term capturing changes in future

tax rates. It has a large impact on the net of tax rate elasticity, which jumps to 41

T X t=1

|

\ dT 0 1 ¡ T0 {z }

¯t

Change in F uture T ax

0.567. The e®ect of the new future tax rates have a positive impact on earned income as expected.

The typical change in marginal tax rates is negative so

work in the future is more rewarding. Individuals are hence richer and this tends to decrease labor supply.

6.1

Alternative Explanations

Our main hypothesis is that the net of tax rate a®ects labor supply as measured by earned income. There are a number of alternative explanations that at least partially may explain our results. One such explanation is that earned income is determined by changes in the wage structure and that our estimates doesn't pick up a behavioral e®ect. Wage formation is quite centralized in Sweden, with national unions and employer organizations negotiating labor contracts.

The

contract are usually for rather broad groups on the labor market. These general shifts in the earnings function should be captured by our sector time e®ects or sector/region time e®ects. During the period following the tax reform there was also a government appointed group of mediators, the Rehnberg commission, who helped to mediate contracts with low nominal wage increases. This environment contrasts the years preceeding the reform when both price and wage in°ation was high. Another concern may be shifting forms of compensation.

Since the re-

form made it more rewarding to compensate employees with earned income as compared to non-pecuniary forms of compensation one might expect a shift into earned income, which could in°uence our estimates. One form of non-pecuniary

42

compensation we have data on before the reform is the value of a company car that the individual has for private use. We run the regressions including the value of car bene¯ts in the pre reform years.

The point estimate is positive

so individuals with car bene¯ts have higher income growth than others. However, including this variable has no e®ect on the estimate of the net of tax rate. This suggests that there is some income shifting but that it doesn't a®ect our estimates of the elasticity of labor supply to the net of tax rate. A third concern may be de¯nitional changes in how income is recorded. For example, pre reform there are lunch coupons that could be used to pay for restaurant meals. They were not taxed but are de¯ned as labor income after the reform. To the extent that these forms of compensation are uniform within certain groups our time e®ects at the sector or sector/region level should control for these changes.

When a form of compensation becomes taxable its use is

usually reduced substantially. This is also the case with the lunch coupons, they virtually disappeared over night.

Another de¯nitional change relates to sick

pay. In 1992 the ¯rst two weeks of sick pay becomes part of earned income. To evaluate the e®ect of this change we rede¯ne earned income to include all sick pay in all years.

The results are not signi¯cantly di®erent with this income

measure and we conclude that it does not a®ect our analysis.

We may also

note that sick leave is substantially reduced in the years following the reform, which may be in response to the reform5 . 5 To

further investigate this we put the relative change in sick pay on the left hand side

in our regression.

We ¯nd that large increases in the net of tax rate signi¯cantly decrease

sick pay, providing evidence of another margin of response to the tax reform. Sick pay is not

43

7

Conclusion

We ¯nd that the tax reform had a large impact on labor supply, as measured by earned income. very large.

The response is not only large but the tax reform was

This result contrasts the usually modest response found when

examining hours worked. Our ¯ndings suggest that the qualitative margin of labor supply is of large importance.

It is not surprising that the response is

higher when individuals can respond on more margins, but the magnitude can not be determined without estimation. We have quanti¯ed the importance of including these additional margins in response to changes in incentives. The estimates most relevant for the evaluation of tax reforms are given in Table 6.3 where we control for the e®ect of the tax reform on the marginal utility of wealth, lambda. When we include all three factors capturing the changes in lambda we obtain an elasticity estimate of 0.56. This is much larger than the estimates of labor supply in previous studies of the tax reform, granted they didn't control for changes in the marginal utility of wealth. The analysis of the unanticipated reform requires strong assumptions but produce estimates with an behavioral interpretation that is relevant for evaluating tax reforms. measured perfectly but if the included ¯xed e®ects capture the de¯nitional change the result would hold.

44

8

References

References [1] Aarbu, Karl O., and Thoresen, T. O., 2001. Income Responses to Tax Changes- Evidence from the Norwegian Tax Reform. National Tax Journal. 54(2), 319-35. [2] Ackum Agell, Susanne, and Costas Meghir, 1995. Male Labour Supply In Sweden: Are Incentives Important?, Swedish Economic Policy Review. 2, 391-418. [3] Agell, Jonas, Peter Englund and Jan SÄo dersten, 1998. Incentives and Redistribution in the Welfare State: The Swedish Tax Reform. Macmillan, London. [4] Agell, Jonas, Peter Englund and Jan SÄo dersten, 1998. Tax Reform of the Century - The Swedish Experiment. National Tax Journal, 18(4), 643-664. [5] Agell, Jonas, Mats Persson and Jan Sacklen, 2003. The E®ects of Tax Reform on Labor Supply, Tax Revenue and Welfare When Tax Avoidance Matters. Stockholm University Working Paper. [6] Altug, and Miller, 1990. Household Choices in Equilibrium. Econometrica vol. 58 No.3, 543-570.

45

[7] Aronsson and M. Palme, 1998. A Decade of Tax and Bene¯t Reforms in Sweden - E®ects on Labour Supply, Welfare and Inequality. Economica, 65, 39-67. [8] Auten, G., Carroll, R., 1999. The E®ect of Income Taxes on Household Income. Review of Economics and Statistics. 81(4) 681-693. [9] Blomquist, S., EklÄo f, M., and Newey, W., 2001. Tax reform evaluation using non-parametric methods: Sweden 1980-1991. Jounal of Public Economics. 79, 543-568. [10] Barro, R. and Chaipat Sahasakul, 1983. Measuring the Average Marginal Tax Rate from the Individual Income Tax. Journal of Business. 56(4), 419452. [11] Blundell, R, A Duncan and C Meghir, 1998. "Estimating Labor Supply Responses Using Tax Reforms" Econometrica. [12] Blundell, R, and T MaCurdy, 1998. Handbook of Labor Economics. [13] Burman, L. E., and Randolph, W. C., 1994. Measuring Permanent Responses to Capital-Gains Tax Changes in Panel Data. American Economic Review, 84, 794-809. [14] Chiappori, P-A, B. Fortin, and G. Lacroix, 2002. Marriage Market, Divorce Legislation, and Household Labor Supply. Journal of Political Economy, 110(1), 37-72.

46

[15] Feldstein, M., 1995. The E®ect of Marginal Tax Rates on Taxable Income: A Panel Study of the 1986 Tax Reform Act. Journal of Political Economy, vol 103, 551-572. [16] Feldstein, M, 1999. Tax Avoidance And The Deadweight Loss Of The Income Tax. The Review of Economics and Statistics, 81(4), 674-680. [17] Flood, L. and T. MaCurdy, 1992. Work disincentive e®ects of taxes: an empirical study of Swedish men. Carnegie-Rochester Conference Series on Public Policy, 37, 239-278. [18] Goolsbee, A., 2000. What Happens When You Tax the Rich? Evidence from Executive Compensation. Journal of Political Economics, vol 108 no. 2, 352-378. [19] Goolsbee, A., 1999. Evidence on the High-Income La®er Curve from Six Decades of Tax Reform. Brookings Papers Economic Activity, no. 2. [20] Gruber, J., Saez, E, 2002. The elasticity of taxable income: evidence and implications. Journal of Public Economics 84,1-32. [21] Gruber, J., Saez, E, 2000. The elasticity of taxable income: evidence and implications. NBER Working Paper 7512. [22] Heckman, James, 1993. What Have We Learned About Labor Supply In the Past Twenty Years? Anaheim, 1993 Meetings.

47

[23] Holmlund, B, and A Kolm, 1995. Progressive Taxation, Wage Setting, and Unemployment: Theory and Swedish Evidence. Swedish Economic Policy Review, 423-460. [24] Klevmarken, N.A., 2000. Did the Tax Cuts Increase Hours of Work? A Statistical Analysis of a Natural Experiment. KYKLOS, 53, 337-362. [25] MaCurdy, 1981. An Empirical Model of Labor Supply in a Life Cycle Setting. Journal of Political Economy, Vol. 89 No.6, 1059-1085. [26] Marion and Mulligan, 2004. Working Paper, University of Chicago. [27] Ministry of Finance, 1991. The Swedish Tax Reform of 1991. [28] Mo±tt, R., and Mark Wilhelm, 1998. Taxation and the Labor Supply Decisions of the A²uent. NBER Workng Paper 6621. [29] Saez, Emmanuel, 2000. Using Elasticities to Derive Optimal Income Tax Rates. NBER Working Paper 7628. [30] Selen, Jan, 2002. Taxable Income Responses to Tax Changes - a Panel Analysis of the 1990/91 Swedish Reform. FIEF Working Paper No.177. [31] Showalter, M. H., and Thurston, K. T., 1997. Taxes and labor supply of high-income physicians. Journal of Public Economics, 66, 73-97. [32] Sillamaa, Mary-Anne, 2001. The E®ect of Marginal Tax Rates on Taxable Income: A Panel Study of the 1988 Tax Flattening in Canada. Journal of Public Economics, 80, 341-356.

48

[33] Slemrod, Joel, 2001. A General Model of the Behavioral Response to Taxation. International Tax and Public Finance, 8, 119-128. [34] Slemrod, Joel, 1998. Methodological Issues in Measuring and Interpreting Taxable Income Elasticities. National Tax Journal, 51(4), 773-788. [35] Slemrod, Joel, and Wojciech Kopczuk, 2002. The Optimal Elasticity of Taxable Income. Journal of Public Economics, 84, 91-112. [36] Triest, Robert K., 1998. Econometric Issues in Estimating the Behavioral Response to Taxation: A Nontechnical Introduction, 51(4), 761-772.

9

Appendix A. Parameterizing Changes in the Marginal Utility of Wealth

In this section we consider a speci¯c model in order to parameterize the e®ect of a tax reform on the marginal utility of wealth. The parameterization will guide the empirical speci¯cation of how the marginal utility of wealth in response to an unanticipated tax reform. Consider a model with separable intratemporal utility and non-linearities in the budget constraint. Individuals live in a world with certainty.

T X t= 1

¯

t¡1

"

1¡ °

(1 ¡ Ht ) 1¡°

49

C 1¡® + ®0 t 1¡®

#

(23)

Budget constraint allows for non-linear gross earned income W(.) and nonlinear taxes T(.).

T X

R¡1 1;t C t =

t= 1

where R1;s ´

Qs

u= 1

X

R¡1 1;t (W (Ht ) ¡ T (W (Ht )))

(24)

(1 + ru ).

FOCs (lambda is the multiplier on the budget constraint)

®0 = ¸t Ct®

, Ct =

1 = W 0 (1 ¡ T 0 ) ¸ t (1 ¡ Ht )°

,

µ

µ

®0 ¸t

¶1=®

1 W 0 (1 ¡ T 0 ) ¸ t

(25)

¶ 1=°

= 1 ¡ Ht

(26)

W' and T' are the derivatives of the earnings function and the tax function with respect to labor supply and earnings, respectively.

1 , ¸t = ¯ ¡t R¡ 1; t ¸ 0

¸ t = ¯ (1 + rt+ 1 ) ¸ t+1

(27)

Use the optimal choices in the intertemporal budget constraint.

T X t=1

R¡1 1;t

µ

®0 ¸t

¶ 1=®

=

X

1 R¡ 1; t

Ã

W

Ã

1¡

1 1=°

(W 0 (1 ¡ T 0 ) ¸ t )

!

¡T

Ã

W

Ã

1¡

1

(28)

50

1=°

(W 0 (1 ¡ T 0) ¸ t)

!!!

¡1=® ¸0

T X

1 R¡ 1;t

t=1

=

X

0

Ã

0

®0 1 ¯ ¡ tR¡ 1;t

@ W @1 ¡ ¡ R¡1 1;t

! 1=® 1

W0

(1 ¡

T 0)

¯

¡t

¢1=° 1 R¡ 1; t ¸ 0

1

0

0

A ¡ T @W @1 ¡ ¡

1 0

W (1 ¡

Now di®erentiate this expression with respect to ¸0 , T , T 0 , and R ¡1 1;t to get

T 0)

¯

¡t

¢1=° R¡1 1;t ¸ 0

expressions for how marginal utility of wealth responds to changes in marginal tax rates and asset return rates. Assume ® = ° = 1 for simplicity . When evaluating the change in lambda in response to changes in the environment there are several approaches. One is to allow for behavioral responses and the other is to restrict behavioral responses. Consider the case with no behavioral response ("constant behavior") to changes in taxes and interest rates. We consider changes in lambda in response to changes in the tax function and discount factors.

T X

1 R¡ 1; t

t =1

=

¡

T X

t =1

¸ ¡2 0

T X t= 1

®0

2 ¡ ¡t ¡1 ¸¡ 0 ¯ R1;t

R¡1 1;t dT

t=1

T X

"

1 R¡ 1; t

"

¡

¯ t [¡® 0 ¡ 1] d¸ 0

¡

(W (Ht)) +

W 0 (1 ¡ T 0 )

2 W 0 (1 ¡ T 0 ) ¯ ¡t R¡1 1;t ¸ 0

T X t=1

W 0 (1 ¡ T 0 )2 ¯ ¡t R¡1 1;t ¸ 0

¡

T X

#

dT 0

R¡1 1;t dT (W (Ht )) +

t=1

¡¸ ¡1 0

t

¯ dT 0 0 1 ¡ T t= 1 51

(30)

(31)

T X t=1

T X

d¸ 0

[W (Ht ) ¡ T (W (Ht))] dR¡1 1;t +

W 0 (1 ¡ T 0 )

=

#

[W (Ht ) ¡ T (W (Ht ))] dR¡1 (32) + 1;t (33)

1 11

A(29) AA

The change in lambda is proportional to the discounted change in the tax bill, the change in the discounted value of net income, and a term that depends on the relative change of the net of marginal tax rate. We can also rewrite the expression in terms on the percentage change in lambda in response to the reform T

T

T

t=1

t=1

t=1

X ¡1 X X t dT 0 d¸ 0 ¡1 = B0 R1;t dT (W (Ht)) ¡ B0 [W (Ht ) ¡ T (W (Ht ))] dR¡1 + B ¸ ¯ 0 0 1;t ¸0 1¡T0 (34) 1 where B0¡1 ´ ¸ ¡ 0

PT

t=1

¯ t [®0 + 1].

Given the number of time periods in the summation, B0¡1 is a constant.

10

Appendix B. Data

In this appendix we'll discuss sample selection and the construction of some variables. There are basically two samples we examine in the paper. The largest sample includes everyone who is between 25 and 55 years old in 1989. There are about 100,000 individuals in the sample.

We study three 3 year di®erences in the

regressions and there are potentially 300,000 observations. Due to some missing observations the typical regression has about 280,000 observations. The more restricted sample imposes two additional restrictions beyond the age criteria. For each of the initial years (1989-1991) in the time di®erences we require that the individual had at least real SEK 60,000 in earned income from employment and less than real SEK 50,000 in taxable transfers. The reason for 52

the earned income criterium is to restrict the study to labor market participants. The income cut o® roughly corresponds to half time employment at a low paying job. The taxable transfer exclusion is set because individuals with large transfer income may face marginal incentives that we don't measure properly. As long as the marginal incentives of the transfers change such that they are uncorrelated with marginal tax changes it should not a®ect our estimate of the substitution e®ect.

In case the changes in marginal incentives are correlated, we exclude

those individuals.

About 90,000 individuals enter the sample for at least on

time di®erence. A typical regression has about 250,000 observations. The instrument for the marginal tax rate is constructed in two steps. First, by pro jecting base year income to grow at the rate speci¯c to the region and sector that the individual is in we obtain a 'constant behavior' income for the ¯nal year in each time di®erence. We then apply the ¯nal period tax code to this projected income, and this is the marginal tax rate instrument. The marginal tax rate instrument is then used to compute the log of the instrumented marginal tax rate minus the log of the initial period net of tax rate. This log di®erence instruments for the observed log di®erence in the net of tax rates. The average tax rate is computed as the tax bill on earnings divided by assessed labor income plus business income declared as labor income.

Labor

income is a wider concept than the dependent variable earned income in the regressions. Labor income includes various kinds of transfer that are taxable. Since the log of the average tax rate enters the main speci¯cation it can not be pro jected by using growth rates as with the marginal tax rates. The log

53

function would separate the variation from the multiplication procedure, and the instrument would fail to provide any variation.

To break the log linear

relationship we use an additive projection. We compute the average change in the average tax rate by sector-region groups.

This is the ¹g in the discussion

in the main text. The instrument for the ¯nal period average tax rate ¹¿ t+1 is then constructed as ¿¹ t + ¹g.

The log of the projected average tax rate varies

independently of the other controls in the model.

The projected average tax

rate predicts actual ¯nal period average tax rates correctly on average. In the second model speci¯cation in table 6.3 the average tax rate enters linearly. It would be problematic to use the additive pro jection as an instrument in this setting since the additive group constant would not exhibit much independent variation when other controls are accounted for. In this speci¯cation we project the average tax rate using sector-region growth rates of average tax rates, call it ^g.

The instrument for the ¯nal period average tax rate ¹¿ t+1 is

then constructed as ¹¿ t (1 + ^g ).

54

Annual hours worked for employed, ages 15-64.

1660

1620 1600 1580 1560 1540 1520 1500 1480

Year

Figure 4: Figure 1. Hours worked of employed in Sweden.

55

1999

1998

1997

1996

1995

1994

1993

1992

1991

1990

1989

1988

1987

1986

1460 1985

Average hours worked

1640

Employment ratio for population 15-64 in Sweden. 0.86 0.84 Fraction employed.

0.82 0.80 0.78 0.76 0.74 0.72 0.70 0.68 0.66 0.64 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 Year

Figure 5: Figure 2. Fraction employed.

56

Aggregate time series for Sweden. Real GDP growth

Inflation

Real Salary growth

12 10 8

%

6 4 2 0 -2 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 -4 -6 Year

Figure 6: Figure 3. Three time series in Sweden

57

1 density .2 cumulative .4 .6 .8 0 0

.2 1989

.4 Marginal Tax Rate 1990

1991

1992

.6

.8 1993

Figure 7: Figure 4. CDF of marginal tax rates, by year.

58

1994

Dependent variable: dLog(Earned Income). IV/2SLS Regression. Economic Model

Sample dLog (1- τ) dLog (avg τ) Demographic Controls: Age Age Sqaured/100 ∆ # Children Married Single Male High School