Pension Wealth and Household Saving in Europe: Evidence from SHARELIFE

Pension Wealth and Household Saving in Europe: Evidence from SHARELIFE Rob Alessie, Viola Angelini and Peter van Santen University of Groningen and Netspar PHF Conference 2012 12 July 2012

Motivation The life cycle/permanent income hypothesis suggests people save for retirement Social security and occupational pensions should therefore affect household wealth accumulation We try to estimate the Displacement effect : How much do households save more (or less) when pension wealth is reduced? Crucial parameter to evaluate the effects of pension reforms in Europe, for households and economy-wide SHARE and SHARELIFE in particular provide the necessary information to estimate the displacement effect

Literature and our contribution Feldstein (1974, JPE): US aggregate time-series data Gale (1998, JPE): US cross-sectional data Attanasio and Rohwedder (2003, AER) and Attanasio and Brugiavini (2003, QJE): time series of cross sectional data for the UK and Italy Kapteyn, Alessie, Lusardi (2005, EER): Dutch panel data Engelhardt and Kumar (2011, JHR): US administrative data with information on past labour earnings Estimates of displacement effect vary not only in magnitude, but also in sign! Contributions 1 New estimates for Europe, using retrospective survey data 2 Wrong sign likely to occur if measurement error not taken seriously

A stylized model max c τ s.t. = L (1 + ρ) τ=1 1 τ c1 γ τ 1 γ L (1 + r) 1 τ c τ = τ=1 R (1 + r) 1 τ E τ + τ=1 (1a) L (1 + r) 1 τ y τ (1b) τ=1 L τ=r+1 (1 + r) 1 τ B τ (1c) c τ denotes consumption at age τ, y τ income, E τ pre retirement earnings (labour income), B τ pension benefits, R exogenous retirement date, L maximum length of life, ρ the discount rate, γ coefficient of relative risk aversion, r constant real interest rate.

Optimal consumption ( ( ) ) 1 + r 1/γ τ 1 c τ = c 1 τ = 2,..., L (2a) 1 + ρ ( L ) c 1 = (1 + r) 1 τ y τ (2b) λ = τ=1 λ τ 1) 1 ( L τ=1 ((1 + r)/(1 + ρ))1/γ 1 + r Standard result: consumption path is flat when r = ρ (2c)

Optimal wealth By definition, wealth at age t equals A t = Plug in optimal consumption to obtain ( t A t = (1 + r) t τ E τ Q(λ, t) τ=1 t τ=1 (1 + r) t τ (y τ c τ ). ) R (1 + r) t τ E τ τ=1 } {{ } z1t L Q(λ, t) (1 + r) t τ B τ τ=r+1 } {{ } z2t where the so-called Gale (1998) s Q is defined as ( L ) Q(λ, t) = λ τ 1 τ=1 λ τ 1) 1 ( t τ=1 (3) (4)

Equation to be estimated A t = β 0 + β 1 z 1t + β 2 z 2t + x tγ + ε t (5) The life cycle model predicts full displacement (β 2 = 1) and β 1 = 1 SHARE(LIFE) contains sufficient retrospective and prospective information to proxy z1t (present value of lifetime earnings) and z2t (pension wealth) However, severe measurement error problems bias the results, leading to spurious rejection of the life cycle model

Measurement error I We observe z 1 = z 1 + η 1 and z 2 = z 2 + η 2 A = β 0 + β 1 z 1 + β 2 z 2 + x γ + ε β 1 η 1 β 2 η 2 (6) Most likely, Cov(η 1, η 2 ) = σ η1 η 2 0, not a classical measurement error problem! We show in the paper that, under the validity of the life cycle model, the OLS estimator ˆβ 1 OLS may easily become negative, and ˆβ 2 OLS may become positive Main reason: correlated measurement errors amplify impact of measurement error in any single regressor

Measurement error II To limit the impact of measurement error, we impose the restriction β 1 = 1 and estimate A z 1 = β 0 + β 2 z 2 + x γ + ε η 1 β 2 η 2 (7) The bias in the OLS estimator now equals (+) ( ) ( ) plim ˆβ OLS 2 β 2 = (σ η1 η 2 + β 2 σ 2 η 2 ) c Presumably, we still face attenuation bias towards zero. We perform the cleanest test of the life cycle model using a subsample of retirees for which σ η1 η 2 0. A significantly negative estimate of β 2 for this group strongly suggests (at least partial) displacement

SHARELIFE Retrospective survey conducted in 13 European countries on 50+ individuals as part of the SHARE project: review social, economic and health events occurred during life Interviewers use life history calendar to guide respondents to answer questions as accurately as possible E.g. first questions: date of birth of children We use information on job history: start date of each job in career, first wage earned on the job, end date, main job in career, last wage main job, current wage if employed Combined with waves 1 and 2 ( standard socio-economic survey) we obtain a potentially very rich dataset

Sample selection Our sample: 3,590 males participating in SHARELIFE Those with no wage or just one wage reported are dropped Those who have worked for less than 20 years are dropped Age restriction: 55-75 years old in wave 2 Household non pension wealth (A t ): Net worth, net financial wealth All 5 imputations are used, estimates and standard errors corrected

Data I Compounded labour income: t (1 + r) t τ E τ τ=1 From SHARELIFE, we construct a panel with one observation per year per individual, from birth to 2009 First monthly wage on each job Last monthly wage on main job / current wage Wage from wave 2 (if still employed) All amounts PPP-adjusted German Euros of 2006 Interpolation between all wages to obtain complete wage path Period 1: first year in first job Retirement: first year receiving pension benefits (if retired) Labour income compounded using r = 3%

Data II Future labour income (from now to retirement): R (1 + r) t τ E τ τ=t+1 Only needed for those still working in wave 2 Retirement (R): expected retirement age from wave 2 or statutory retirement age by country if missing Simplifying assumption: future real wage path is flat All future amounts weighted by survival probabilities from country-specific HMD life tables

Data III Pension wealth: z 2t = L τ=r+1 For retired: pension benefits (1 + r) t τ B τ For employed: use expected replacement rate from wave 2, multiplied by current income Simplifying assumption: constant real pension benefits Future amounts weighted by survival probabilities Maximum age (L): 110

Some descriptives Figure : Wages and pensions in Europe 0 10,000 20,000 30,000 40,000 Median income by country AT DE SE NL ES IT FR DK GR CH BE CZ PL Annual pension income Annual labour income

Main results Table : Estimates of the displacement effect (1) (2) Variables Robust regression Median regression Pension wealth -0.471*** -0.609*** (0.0878) (0.151) Age 55-60 -1.406*** -1.409*** (0.155) (0.173) Age 70-75 1.797*** 1.614*** (0.172) (0.177) Gaps in career -0.824*** -0.756*** (0.146) (0.164) Constant -4.786*** -4.385*** (0.329) (0.374) Observations 3590 3590 p-value β 2 = 1 0.000 0.011 p-value Country effects 0.000 0.000 Standard errors in parentheses; *** p <0.01, ** p <0.05, * p <0.1 Bootstrapped standard errors for median regression, 1000 replications. Additional controls: Second earner, Married, No children, Education, Bad Health, Country dummies

Robustness checks Table : Robustness checks displacement effect (1) (2) (3) (4) Variables Retired Old Low High sample sample educated educated Robust regression -0.205** -0.173* -0.215* -0.833*** (0.0936) (0.0965) (0.122) (0.153) p-value β 2 = 1 0.000 0.000 0.000 0.277 Median regression -0.296-0.306* -0.275-1.099*** (0.180) (0.177) (0.192) (0.286) p-value β 2 = 1 0.000 0.000 0.000 0.729 Observations 2487 2415 3590 3590 Standard errors in parentheses; *** p <0.01, ** p <0.05, * p <0.1 Bootstrapped standard errors for median regression, 1000 replications.

Results by country group Figure : Displacement effect across country groups Estimated displacement effect by country group Estimate, 90% CI 1.5 1.5 0.5 North Mid West South East Displacement effect Pooled displacement effect Lower bound/upper bound

Results by country group II Highest displacement in North, limited displacement in South and East Generous welfare systems less precautionary savings More access to capital markets relax liquidity constraints Jappelli (2010, Economic Journal): significant positive relationship between GDP/capita and economic literacy; highest literacy scores in Denmark, Switzerland and the Netherlands, lowest scores in Poland, Italy and Spain Table : Financial literacy: household surveys NL SE DE IT Russia 46.2% 26.7% 56.8% 28.3% 3.4%

An IV approach Pension wealth may be endogenous due to unobserved heterogeneity (taste for saving, patience) or endogenous retirement We compute median pension income by country and employment status (employee, civil servant, self employed) Transformed to pension wealth, using country- and time-specific statutory retirement age; no individual characteristics used in computation of the instrument Assumption: conditional on demographic characteristics, education, wealth and the country of residence, workers do not sort across employment sectors based on the taste for saving Chernozhukov-Hansen s median regression IV estimator

Variation in the instrument Figure : Instrument 0 500 1,000 1,500 2,000 Median pension income by country and work status AT DE SE NL ES IT FR DK GR CH BE CZ PL Employees Self employed Civil servants

IV results Table : IV Median regression estimates (1) (2) (3) Variable Full Retired Old sample sample sample Pension Wealth -1.232-0.622-0.955 (0.876) (0.863) (0.813) p-value β 2 = 1 0.804 0.684 0.960 F -statistic first stage 41.895 31.232 38.567 Observations 3590 2487 2414 Bootstrapped standard errors in parentheses; 1000 replications

Concluding remarks We provide new estimates for the displacement effect using retrospective life history data for Europe We show that addressing measurement error problems is crucial to estimate the displacement effect when using survey data We find substantial displacement: each euro of pension wealth is associated with a 47 (61) cent decline in non pension wealth using robust (median) regression, with a lower bound of 17(30)%. Full displacement using IV median regression but not precisely estimated However, we cannot reject zero displacement for the low educated and for Southern and Eastern European countries Research agenda: eliminate remaining biases, due to both measurement error and endogeneity of retirement decision