Earnings responses to payroll and income taxes Exploiting variation in Dutch pension rates (CPB Netherlands Bureau for Economic Policy Analysis) Casper van Ewijk (University of Amsterdam, CPB, Netspar) Maja Micevska Scharf (University of Amsterdam) February 2016
INTRODUCTION Payroll taxes amount to 23% and income taxes to 14% of labour costs in the OECD and the Netherlands in 2014 (OECD, 2014). On top of this, the Dutch pension contributions are among the highest worldwide. Welfare effects of taxation are driven by both behavioural effects as well as wage rate effects (incidence). Empirical literature on payroll taxes is scarce (Saez et al., 2012). Current paper uses panel approach exploiting exogenous tax variation over time largely resulting from pension contributions. Goals Separately estimate responses to payroll taxes and to income taxes Shed additional light on the issue of incidence of payroll taxes Empirical analysis closely related to Lehmann, Marical & Rioux (JpuB, 2013).
Dutch pension system The Dutch pension system consists of three pillars. Old age pensions, occupational pensions and individual pension savings. Even without statutory obligation, employers voluntarily offer occupational pensions. Once such an agreement exists at an industry level, it is compulsory. Pension funds are supervised by the Netherlands Bank and are restricted to regulation (funding ratio, interest rate) but have large discretionary power. Funds invest abroad (not own sector). There is a clear contribution-benefit linkage, but because of the collective nature, the system is not actuarially fair. Most are defined benefit schemes, some are defined contribution. Contributions apply to earnings between a lower limit and - often- an upper limit. Large up-and-down movements in average and marginal rates (see next Figures).
Marginal pension rates in 2012
Variation in marginal pension rates over time
Variation in lower limits over time
Variation in lower limits over time, including extreme cases E /E' W, d ^
Are changes in pension rates exogenous? log w i,t = α + β P log τ P i,t + γx i,t 1 + u i,t (1) where w i,t indicates labour costs for person i in year t and τ P is the net-of-payroll tax rate. Reversed causality: change in w i,t affects change in τ due to nonlinear tax Solutions Omitted vars: unobs. (pension fund) chars may impact both changes in τ and w Instrument tax rates (Gruber and Saez, 2002; Weber, 2014) Include pensionfund dummies (e.g. demographics), yeardummies (financial crisis)
Are changes in pension rates exogenous after including pensionfund and timedummies? Source of variation in rates due to legal rules (funding ratios, life expectancy) that do not affect wages directly Variation in international interest rate impact asset portfolio and may lead to different rates, but do not have a direct impact on wages. Pensionfunds have large discretionary power in setting pension rates (lower limit, division between employer and employee, indexation retirees). Sectoral downturns (construction sector)? Lower wages and higher rates go together? Impact on rates depends foremost on financial position fund. Funds have large discretionary power, and can set different rates even in same financial position.
What is incidence? Statutory incidence: who pays the tax Economic incidence: who bears the burden (employment and wages) Invariance of Incidence Proposition: statutory incidence irrelevant Labour demand and labour supply elasticities determine incidence In equilibrium, incidence on employee/ full shifting to employee (Gruber, 1997) if: Elastic demand; Inelastic supply or Full tax-benefit linkage
Complications/ disequilibrium Institutions (minimum wage, sticky wages, labour unions) Competitiveness market (small open economy, monopoly, prices) In the end, it is an empirical question whether costs are shifted to employees. Earlier micro-econometric studies find varying effects: No shifting to employee because of sticky wages and significant labour supply effects (Lehmann et al., 2013) Incidence of higher employee (employer) contributions paid by employee (employer). Different costs and net wages for new employees (Saez et al. 2012) Economic incidence equals statutory incidence (Neumann and Mueller, 2014)
Infer incidence from panel estimates Let s start from changes in labour costs w = h W (Lehmann et al., 2013). log w i,t = α+β P τ log τ P i,t +βi τ log τ I i,t +βp ρ log ρp i,t +βi ρ log ρi i,t +γx i,t 1+u i,t (2) marginal net-of-payroll tax τ P and net-of-income tax τ I (substitution effects) average net-of-payroll tax ρ P and net-of-income tax ρ I (income effects/incidence) (assuming small income effects) Full incidence EE ER on ER EE on EE Full incidence ER Benchmark (el.ld) e.g. Sticky wages Labour cost βρ P = βi ρ = 0 βp ρr = 1; βp ρe = βi ρ = 0 βp ρr = βp ρe = βi ρ = 1 βτ I > βp τ ; βi τ > 0 βi τ > βp τe ; βi τ > 0 βi τ = βp τ = 0
Hourly wage costs log( w i,t h i,t ) = α+β P τ log τ P i,t +βi τ log τ I i,t +βp ρ log ρp i,t +βi ρ log ρi i,t +γx i,t 1+u i,t marginal net-of-payroll tax τ P and net-of-income tax τ I (marginal incidence) average net-of-payroll tax ρ P and net-of-income tax ρ I (incidence) (assuming hours is only behavioural effect) (3) Full incidence EE ER on ER,EE on EE Full incidence ER Benchmark (el.ld) e.g. Sticky wages Labour cost/h βρ P = βρ I = 0 βρr P = 1; βρe P = βρ I = 0 βρr P = βρe P = βρ I = 1 βτ I = βτ P = 0 βτ I = βτ P = 0 βτ I = βτ P = 0
Empirical model First, we estimate changes in labour costs w = h W (Lehmann et al., 2013). log w i,t = α+β P τ log τ P i,t +βi τ log τ I i,t +βp ρ log ρ P i,t +βi ρ log ρ I i,t +γx i,t 1+u i,t (4) The vector X i,t 1 includes demographics, pensionfund dummies, yeardummies and a function of base-year income. We assume log SSCs = 0. The error term u i,t captures time-varying and unobserved heterogeneity. β P ρ = 0 and βp τ > 0 (incidence on EE, lower posted wages and/or lower hours) βρ P = 1 and βτ P = 0 (incidence on ER) Income versus payroll taxes: βρ P > βi ρ and βp τ > βi τ > 0 (more salient payroll tax) β I ρ > βp ρ and βi τ > βp τ > 0 (tax-benefit linkage)
Hourly labour cost Second, we estimate the following equation where we control for hours: log( w i,t h i,t ) = α+β P τ log τ P i,t +βi τ log τ I i,t +βp ρ log ρ P i,t +βi ρ log ρ I i,t +γx i,t 1+u i,t Assuming hours is only behavioural response (neglect effort, avoidance), we can interpret coefficient as incidence. (5) β P ρ = 0, β I ρ = 0 (full incidence on EE) β P ρ = 1 (full incidence on ER)
Separating employee and employer taxes Our third estimation separates employer rate ( ρr P ) from employee pension rate ( ρe P ). log w i,t = α+β P τ log τ P i,t +βi τ log τ I i,t +βp ρe log ρe P i,t +βp ρr log ρr P i,t +βi ρ log ρ I i,t +γx i (6) (assuming small income effects) βρr P = 0 (incidence of ER on EE, full shifting) βρr P = 1 (incidence of ER on ER, e.g. posted wage stickiness) βρe P = 0, βi ρ = 0 (incidence of EE on EE) βρe P = 1, βi ρ = 1 (incidence of EE on ER) βρr P > βp ρe (less salient employer payroll tax, Iturbe-Ormaetxe, 2015)
Analogue to ETI literature, estimating previous equations leads to two problems: Endogenous (marginal) tax rate > instrument with synthetic tax rate based on (lags of) base-year income (Gruber and Saez, 2002; Weber, 2014) Mean reversion and diverging income trends > add functions of base-year income Include income controls can mob up changes in tax rates. Especially if income controls are close to large changes in tax rates (thresholds). We include synthetic changes in average tax rate and not actual change in average tax rates, because they do not depend on labour cost and pension system is not proportional.
Data Dutch Social Statistical Database on Jobs 2006-2012 is a longitudinal dataset for entire working population. Containing (uncapped) monthly wage earnings, hours, CLA code, firm code, and employer and employee SSCs. Earnings and SSCs are reported by employers to tax administration and are reliable. Combined with municipality data and a self-constructed database with pension rates for each collective labour agreement (CLA) Income rates are simulated by using CPB s detailed gross-net calculator Sample selection: age 20-55, balanced panel, same household type, same CLA This leaves us with 3 mln person-year observations (approx. 700K each year).
Estimation results: Labour cost log w i,t = α+β P τ log τ P i,t +βi τ log τ I i,t +βp ρ log ρp i,t +βi ρ log ρi i,t +γx i,t 1+u i,t (1) (2) (3) (EE) (ER) βτ P 0.0281 0.0227 0.0193 >=0 0 (0.0002) (0.0018) (0.0018) βτ I 0.0035-0.0019-0.0021 >=0 0 (0.0007) (0.0007) (0.0007) βρ P -0.1770-0.1130-0.0900 0-1 (0.0023) (0.0023) (0.0023) βρ I -0.0299-0.0230-0.0149 0-1 (0.0057) (0.0056) (0.0056) Income control none 10-piece log income Notes: controls for gender, age and marital status, year and pensionfund dummies, labour costs > 10K. Two-year changes. N=2.93 mln First-stage regressions: strong instruments
Estimation results: hourly labour costs log( w i,t h i,t ) = α+β P τ log τ P i,t +βi τ log τ I i,t +βp ρ log ρp i,t +βi ρ log ρi i,t +γx i,t 1+u i,t (1) (2) (3) (EE) (ER) βτ P 0.1410 0.1280 0.1250 >=0 0 (0.0020) (0.0016) (0.0016) βτ I 0.0017-0.0033-0.0038 >=0 0 (0.0008) (0.0006) (0.0006) βρ P -0.2400-0.1780-0.1680 0-1 (0.0025) (0.0020) (0.0020) βρ I 0.0414 0.0488 0.0632 0-1 (0.0114) (0.0048) (0.0049) Income control none 10-piece log income Notes: controls for gender, age and marital status, year and pensionfund dummies, labour costs > 10K. Two-year changes. N=2.93 mln First-stage regressions: strong instruments
Estimation results Labour costs: separating employer and employee LC Penmr Penpmr Taxmr Taxpmr PenEpar PenRpar Taxp LC 1 Penmr 0.055 1 Penpmr 0.026 0.788 1 Taxmr -0.145-0.009-0.004 1 Taxpmr 0.012 0.014 0.015 0.437 1 PenEpar -0.200 0.029 0.043 0.005 0.001 1 PenRpar -0.052 0.137 0.166 0.01-0.003 0.758 1 Taxpar -0.033-0.061-0.064-0.052-0.043 0.022 0.048 1
Estimation results Labour costs: separating employer and employee log w i,t = α+β P τ log τ P i,t +βi τ log τ I i,t +βp ρe log ρe P i,t +βp ρr log ρr P i,t +βi ρ log ρ I i,t +..+u (1) (2) (3) (EE) (ER) (EE,ER) βτ P 0.0226 0.0177 0.0137 >=0 0 >=0 (0.0018) (0.0018) (0.0018) βτ I 0.0034-0.0020-0.0022 >=0 0 >=0 (0.0007) (0.0007) (0.0007) βρe P 0.1760 0.1010 0.0622 0-1 0 (0.0046) (0.0046) (0.0046) βρr P -0.3730-0.2130-0.1390 0-1 -1 (0.0063) (0.0063) (0.0063) βρ I -0.0328-0.0246-0.0165 0-1 0 (0.0057) (0.0056) (0.0056) Income control none 10-piece log income Notes: controls for gender, age and marital status, year and pensionfund dummies, labour costs > 10K. Two-year changes. N=2.93 mln First-stage regressions: strong instruments Correlation ρe P i,t, ρrp i,t = 0.75
Sensitivity: labour costs (1) (2) (3) (4) βτ P -0.0590 0.0144 0.2160 0.159 (0.0014) (0.0018) (0.0027) βτ I -0.0002-0.0021-0.0018-0.0020 (0.0058) (0.0009) (0.0037) βρ P -0.0420-0.1190 0.0399-0.527 (0.0018) (0.0023) (0.0026) βρ I 0.0253-0.1040-0.0446-0.0148 (0.0043) (0.0070) (0.0056) Notes: (1) splines and difference log(w i,t 1 ) log(w i,t 2 ), (2) excluding public sector, (3) fully interacted pensionfund and timedummies, (4) actual changes in tax rates (instrumented)
Findings Small and significant coefficients for marginal tax rates. Zero effect income tax. Small and negative coefficients for (net-of) average income tax. Small but sizeable negative coefficients for (net-of) average payroll tax (range of -0.1 to -0.2). Somewhat sensitive to inclusion of income controls. Employer pension contributions raise labour costs (10% increase leads to 2% higher costs) Employee contributions lower labour cost. Contradicting result, implying negative income effects or more than offsetting incidence.
Preliminary conclusions Pension schedule provides ample opportunity to analyse incidence. Small variation in income tax rates limits comparison income and payroll tax. Rich and reliable administrative data. Quality hours needs further exploration. We find less negative effects of payroll taxes than earlier studies. This suggests that Dutch wages are flexible or that employees fully value benefits. We conclude that incidence is on the employee, although not fully. Tentative results show that employers are less able to shift their contributions to employees which could be explained by the fact that these are less salient. Overall, more to be done.
Future plans Estimate hours equation and explore reliability hours Technical issues like rescaling and multiplicative structure Focus on separate pensionfunds Alternative instruments: grouping variable (pensionfund*year) Sample selection (more heterogeneous sample) Focus on subgroup with income tax reform