Evidence from matched data for the US, England and European countries - - PowerPoint PPT Presentation

National Institute of Economic and Social Research

Health impacts of retiring: Evidence from matched data for the US, England and European countries

Ser Sergio

• Sali

Salis (N (NIESR) Lon London, 26 26th

th Sep

September 2018 2018 This research has been funded under the ESRC/MRC grant ES/L002884/1 FACTAGE Workshop on Gender Inequalities in Extending Working Lives

National Institute of Economic and Social Research

Outline

• Background
• Aims
• Data
• Empirical methodology
• Findings
• Conclusions

National Institute of Economic and Social Research

• An ‘extended working lives’ policy agenda promotes working in later life on the basis of

pension sustainability and health benefits (OECD, 2015; DWP, 2014; and WHO, 2002)

• However, empirical evidence on the health impact of retirement is inconclusive:
• Evidence of beneficial effects (Bound and Waidman, 2007; Neuman, 2008; Coe and

Lindeboom, 2008; Coe and Zamarro, 2011; De Grip et al., 2012; Eibich, 2015; Insler, 2014)

• Retirement as detrimental for health (Behncke, 2012; Siegrist, et al., 2004; Hartlapp and

Schmid, 2008; Bonsang et al., 2007; Wu et al. 2016)

• No evidence (Hernaes et al., 2013; Coe and Lindeboom, 2008)
• These studies are difficult to compare as thy differ in many respects (e.g., econometric

strategy, data, country, definition of retirement and health outcomes)

Background

National Institute of Economic and Social Research

• Investigating the causal effect (impact) of retirement on older people’s subjective and objective

health in a multi-country setting (US, England and 11 European countries)

• The endogeneity of the retirement decision potentially biasing impact estimates (health is both a

determinant and a consequence of retirement) is addressed using two alternative non-parametric estimation techniques:

• Propensity Score Matching (PSM) in tandem with Difference-in-Differences (DID)
• Instrumental Variable (IV)
• Exploring the heterogeneity of impacts
• By retiree’s gender: Women experience a smoother transition to retirement do to established

roles and routines in the home (Price and Nesteruk, 2010); loss of an important social role for the men (Coppola and Spizzichino, 2014)

• By nature of the job they retire from: Those in physically-demanding jobs typically have less

leisure physical activity so less experience from which to build (Berger et al., 2010)

Aims

National Institute of Economic and Social Research

• Individual-level data come from three different surveys:
• RAND HRS (US)
• Harmonised ELSA (England)
• Harmonised SHARE (Italy, Germany, Austria, Sweden, Netherlands, Spain,

France, Denmark, Switzerland, Belgium and Greece)

• Information about health and socio-economic characteristics and

circumstances for people aged 50 or over (over 50 for the RAND HRS) and their spouses or partners

• Comparability across these three sources

Data - So Sources

National Institute of Economic and Social Research

• Different waves of the three surveys are pooled together to form the waves
• f a multi-country dataset

Data - Ti Time str tructure

Wave 1 Wave 2 Wave 3 Wave 4 Wave 5 Wave 6 ELSA

✓ ✓ ✓ ✓ ✓ ✓

HRS

✓ ✓ ✓ ✓ ✓ ✓

SHARE

NA ✓ ✓ NA ✓ ✓

• Attrition: Sample size progressively reduces for later waves

National Institute of Economic and Social Research

• Two types of individuals are selected:
• Retirees (employed at Wave 2 interview and retired at Wave 3 interview)
• Non-retirees (employed at both Wave 2 and Wave 3 interviews)
• We observe almost 1,600 retirees and more than 8,000 non-retirees

Data - Defi finition of f retirees

Wave 2 interview ELSA Wave 2 (Jun 2004-Jul 2005) HRS Wave 7 (Feb 2004-Jan 2005) SHARE Wave 1 (May 2004-Jul 2005) Wave 3 interview Wave 3 (May 2006-Aug 2007) Wave 8 (Mar 2006-Feb 2007) Wave 2 (Jan 2006-Dec 2007)

Retirement window

National Institute of Economic and Social Research

• Impact of retirement on two health outcomes:
• Self-assessed health (whether the individual reported being in excellent, very

good or good health)

• Physical health (whether a doctor had ever diagnosed individuals with one or

more conditions among heart problem, stroke, cancer, lung problem, arthritis, high blood pressure or diabetes)

• Health outcomes are observed immediately after retiring (at Wave 3 interview) and

two later time points (Wave 5 and Wave 6 interviews)

Data - Health outcomes

National Institute of Economic and Social Research

Empirical methodology - Health ti time tr trends

 Pre-retirement health gap between retirees and non-retirees  Risk to overstate the (negative) impact of retirement  Need to control for baseline health differences between retirees and non-retirees

Based on 1,588 retirees and 8,219 non-retirees (sample sizes get lower in later waves due to attrition)

75.0 77.0 79.0 81.0 83.0 85.0 87.0 89.0 Wave 2 Wave 3 Wave 4 Wave 5 Wave 6 % with good or better health

Time trends in self-reported health

Retirees Non-retirees

National Institute of Economic and Social Research

Empirical methodology - Se Selection bia ias

Table 1 Baseline compositional differences between retirees and non-retirees Retirees Non-retirees Age (mean) 64.2*** 57.8 Female (%) 49.0*** 53.1 Reaches State Pension age by Wave 3 (%) 50.5*** 12.1 University degree (%) 24.6*** 30.3 Married or living with partner (%) 76.8** 79.0 Has children (%) 91.8** 90.2 Family assets (mean of standardised log) 0.116***

• 0.001

Earnings (mean of standardised log)

• 0.228***

0.100 Job tenure (mean) 17.8*** 15.3 Self-reported health is excellent, very good or good (%) 83.5*** 87.6 Has/has had a chronic condition (%) 65.6*** 51.3 Health limits work (%) 19.3*** 14.4 Sample size varies from 1,590 retirees and 8,227 non-retirees (for variables Age, Female and Reaches State Pension age by Wave 3) to 1,496 retirees and 8,094 non-retirees (for Health limits work); *** and **: difference significant at the 1 and 5% significance level, respectively.

 At baseline, retirees and non-retirees differ in several other respects*  Need to control for all confounders * Having reached State Pension age by Wave 3 is highly predictive of retirement for samples defined by Wave 3 outcomes (so a good instrument)

National Institute of Economic and Social Research

• Average Treatment Effect on the Treated (ATT)
• Conditional Independence Assumption (CIA) or selection on observables

Empirical methodology - PSM SM & DID ID

• PSM is used in tandem with DID to improve on the estimates (Blundell and

Costa Dias, 2000)

National Institute of Economic and Social Research

• What if immediately before retiring (and after our baseline time point) people

experienced a health shock which cannot be observed?

• Local Average Treatment Effect (LATE)
• C=1 denotes Compliers, i.e. those who respond to the instrument (State

Pension age)  Retire if SPA=1 and stay employed if SPA=0

• LATE conditional on confounders (Frölich, 2007). In our case, the confounders

are age, gender and country of residence

Empirical methodology - IV IV

National Institute of Economic and Social Research

Fin indings - PSM SM & & DID ID esti timates

Table 2 Impact of retirement on retirees' self-assessed and objective health (ATT estimates) Self-reported health Objective health Self-reported health Objective health Self-reported health Objective health Impact on the Impact on the Impact on the Impact on the Impact on the Impact on the proportion of proportion of proportion of proportion of proportion of proportion of retirees reporting retirees having had retirees reporting retirees having had retirees reporting retirees having had good, very good a health problem good, very good a health problem good, very good a health problem

• r excellent health

diagnosed

• r excellent health

diagnosed

• r excellent health

diagnosed 1-to-1 matching (with replacement)

• 0.057***

0.028**

• 0.062

0.024

• 0.115***

0.039 (0.018) (0.012) (0.035) (0.028) (0.042) (0.037) LLR matching

• 0.051***

0.020

• 0.069**

0.039

• 0.091***

0.065** (0.015) (0.010) (0.027) (0.021) (0.034) (0.027) Number of treated individuals 1,291 [1,279] 1,292 [1,280] 856 [838] 862 [843] 739 [722] 743 [726] Number of untreated individuals 891 [6,946] 899 [6,949] 450 [3,284] 439 [3,288] 318 [2,246] 327 [2,252] Impacts in percentage points are obrained multiplying the estimates in the table by 100; standard errors are reported in round brackets; *** and **: statistically significant at the 1 and 5% significance level, respectively; The numbers of individuals in square brackets refer to LLR matching; Standard errors for impact estimates obtained by means of 1-to-1 matching are based on Abadie and Imbens (2012); Standard errors for impact estimates obtained by means of LLR matching are bootstrapped (1,000 replications). Wave 3 Wave 5 Wave 6

• Waves 5 and 6 results not reliable due to

covariate balancing issues (sample sizes)

• Retirement was found to have reduced the

proportion of retirees who experienced good/better health by over 5ppts in Wave 3

• Some evidence of a ‘negative’ impact also for
• bjective health (3ppts increase)

The matching algorithm is chosen based on the observed distribution of the propensity score

National Institute of Economic and Social Research

Fin indings - IV IV estimates

Table 3 Impact of retirement on compliers' self-assessed and objective health (LATE estimates) Self-reported health Objective health Self-reported health Objective health Self-reported health Objective health Impact on the Impact on the Impact on the Impact on the Impact on the Impact on the proportion of proportion of proportion of proportion of proportion of proportion of retirees reporting retirees having had retirees reporting retirees having had retirees reporting retirees having had good, very good a health problem good, very good a health problem good, very good a health problem

• r excellent health

diagnosed

• r excellent health

diagnosed

• r excellent health

diagnosed LATE (controlling for age, gender and country) 0.196 0.320** 0.009 0.339*** 0.049 0.265** (0.105) (0.132) (0.121) (0.111) (0.134) (0.105) LATE (controlling for gender and country)

• 0.035

0.321***

• 0.106**

0.309***

• 0.121**

0.264*** (0.040) (0.048) (0.053) (0.041) (0.056) (0.040) Number of observations 9,798 9,813 7,687 8,894 7,361 8,878 Proportion of compliers 0.202 [0.414] 0.203 [0.414] 0.180 [0.390] 0.188 [0.408] 0.177 [0.386] 0.194 [0.411] Wave 3 Wave 5 Wave 6 The nonparametric instrumental variable estimation of LATE is based on Frölich (2007); Standard errors are reported in round brackets; *** and **: statistically significant at the 1 and 5% significance level, respectively; The proportion of compliers in square brackets refers to the LATE estimate controlling for age, gender and country.

• Retirement found to have increased the

proportion of compliers who were diagnosed with a physical condition by 32ppts

• Size of LATE is tenfold that of the ATT
• Controlling for age of individual does not

change this result

National Institute of Economic and Social Research

• Health impacts were estimated separately by:
• Gender: Do women experience a smoother retirement transition than men?
• Nature of the job (physical vs non-physical): do those who retire from physically

demanding jobs gain more from retiring?

• PSM-DID impacts were not found to vary across these subgroups (so no evidence of a

gender gap)

Heterogeneity of f im impact?

National Institute of Economic and Social Research

Health gender gap: PSM SM-DID estimates

Table 4 Impact of retirement on retirees' self-reported and objective health, by gender Self-reported health Objective health Self-reported health Objective health Self-reported health Objective health Impact on the Impact on the Impact on the Impact on the Impact on the Impact on the proportion of proportion of proportion of proportion of proportion of proportion of retirees reporting retirees having had retirees reporting retirees having had retirees reporting retirees having had good, very good a health problem good, very good a health problem good, very good a health problem

• r excellent health

diagnosed

• r excellent health

diagnosed

• r excellent health

diagnosed Female 1-to-1 matching (with replacement)

• 0.080***

0.027

• 0.129***

0.086***

• 0.017

0.097*** (0.024) (0.018) (0.042) (0.027) (0.041) (0.028) LLR matching

• 0.078***

0.032**

• 0.116***

0.056**

• 0.141***

0.097*** (0.021) (0.015) (0.038) (0.027) (0.051) (0.030)

• N. of treated individuals

623 [611] 625 [615] 417 [412] 421 [416] 366 [361] 370 [365]

• N. of untreated individuals

456 [3,669] 439 [3,671] 224 [1,791] 226 [1,792] 171 [1,256] 174 [1,258] Male 1-to-1 matching (with replacement)

• 0.018

0.031

• 0.005

0.016

• 0.046

0.016 (0.028) (0.017) (0.055) (0.041) (0.052) (0.070) LLR matching

• 0.028

0.017

• 0.024

0.031

• 0.064

0.041 (0.020) (0.013) (0.040) (0.032) (0.039) (0.041)

• N. of treated individuals

668 [663] 667 [662 ] 439 [431] 441 [433] 373 [351] 373 [351]

• N. of untreated individuals

448 [3,277] 439 [3,278] 199 [1,493] 206 [1,496] 143 [990] 141 [994] Wave 3 Wave 5 Wave 6 Impacts in percentage points (ppts) are obrained multiplying the estimates in the table by 100; standard errors are reported in round brackets; *** and **: statistically significant at the 1 and 5% significance level, respectively; Standard errors for impact estimates obtained by means of one-to-one matching are based on Abadie and Imbens (2012); Standard errors for impact estimates obtained by means of local linear regression matching are bootstrapped (1,000 replications); The sizes of the treated and untreated samples used to obtain impact estimates by means of local linear regression matching are reported in square brackets.

• Waves 5 and 6 results not reliable due to

covariate balancing issues

• Retirement was found to have reduced

the proportion of female retirees who experienced good/better health by 8ppts in Wave 3

• However, the difference between the

Female and Male subgroup impacts is not statistically significantly different from zero

National Institute of Economic and Social Research

• Health impacts were estimated separately by:
• Gender: Do women experience a smoother retirement transition than men?
• Nature of the job (physical vs non-physical): do those who retire from physically

demanding jobs gain more from retiring?

• PSM-DiD impacts were not found to vary across these subgroups (so no evidence of a

gender gap)

• However, LATE estimates suggested that retirement was more deleterious for men than

women (objective health)

Heterogeneity of f im impact?

National Institute of Economic and Social Research

Health gender gap: IV IV estimates

Table 5 Impact of retirement on compliers' self-assessed and objective health, by gender Self-reported health Objective health Self-reported health Objective health Self-reported health Objective health Impact on the Impact on the Impact on the Impact on the Impact on the Impact on the proportion of proportion of proportion of proportion of proportion of proportion of retirees reporting retirees having had retirees reporting retirees having had retirees reporting retirees having had good, very good a health problem good, very good a health problem good, very good a health problem

• r excellent health

diagnosed

• r excellent health

diagnosed

• r excellent health

diagnosed Female LATE (controlling for age and country) 0.214*** 0.199***

• 0.051

0.359*** 0.192** 0.319*** (0.057) (0.065) (0.078) (0.073) (0.085) (0.066) LATE (controlling for country)

• 0.063

0.309***

• 0.155**

0.303***

• 0.057

0.283*** (0.060) (0.068) (0.066) (0.057) (0.069) (0.053) Number of observations 5,137 5,143 4,118 4,699 3,986 4,700 Proportion of compliers 0.391 [0.178] 0.390 [0.178] 0.368 [0.152] 0.384 [0.158] 0.360 [0.146] 0.391 [0.170] Male LATE (controlling for age and country) 0.228*** 0.546***

• 0.014

0.485***

• 0.018

0.396*** (0.060) (0.069) (0.082) (0.070) (0.080) (0.067) LATE (controlling for country)

• 0.019

0.374***

• 0.150*

0.359***

• 0.194**

0.286*** (0.061) (0.074) (0.085) (0.065) (0.088) (0.064) Number of observations 4,661 4,670 3,569 4,195 3,375 4,178 Proportion of compliers 0.408 [0.186] 0.411 [0.190] 0.385 [0.164] 0.410 [0.185] 0.392 [0.174] 0.407 [0.184] Wave 3 Wave 5 Wave 6 The nonparametric instrumental variable estimation of LATE is based on Frölich (2007); Standard errors are reported in round brackets; *** and **: statistically significant at the 1 and 5% significance level, respectively; The proportion of compliers in square brackets refers to the LATE estimate controlling for age and country.

The instrument does not predict retirement for Wave 5 and Wave 6 male samples

• The difference between the impacts for

Female and Male compliers (34.7ppts) is significantly different from zero (objective health)

• No such evidence for self-reported health (but

the impact difference between Female and Male is only -1.4ppts)

National Institute of Economic and Social Research

• Overall, evidence of a detrimental impact of retirement on older people’s

health in the immediate post-retirement period. However, the mechanism is not known (e.g., intellectual stimulation, sense of purpose or physical activity?)

• Impact magnitude and statistical significance vary depending on the estimation

method used. This suggest that different subpopulations of older people are likely to experience different impacts

• Some evidence of impact heterogeneity (the effect of retirement varies by

gender, with men being affected more than women)

• Policy recommendations: implementation of preventative measures (e.g.,

workplace support, local communities and financial support/advice)

Conclusions

National Institute of Economic and Social Research

Any question? Or please email me: s.salis@niesr.ac.uk Working paper by Salis, Smeaton and Icardi (2017) available at http://workandretirement.uk/working-papers And keep an eye on NIESR website for an updated version!

Th Thanks!

National Institute of Economic and Social Research

Choice of f th the PSM SM alg lgorithm

• Not all retirees have a non-

retiree counterpart with a similar propensity score (PS>0.2)

• Week common support region

for PS>0.6  1-to-1 matching with replacement (Dehejia and Wahba, 1999)  LLR matching (see Caliendo & Kopeinig, 2008)

2 4 6 8 10 .2 .4 .6 .8 1 Propensity score Retirees Non-retirees

National Institute of Economic and Social Research

Estimatin ing propensity sc scores (probit it results for r Wave 3 sa sample)

Variable Coefficient

• Std. err.

z P>|z| Age 0.541 0.048 11.37 0.000 Age square

• 0.004

0.000

• 9.57

0.000 Male (dummy) 0.044 0.042 1.06 0.288 Reaches State Pension age by Wave 3 (dummy) 0.346 0.065 5.32 0.000 University degree (dummy)

• 0.008

0.046

• 0.16

0.869 Married or has partner (dummy) 0.022 0.052 0.41 0.679 Has children (dummy)

• 0.027

0.070

• 0.38

0.702 Assets 0.109 0.023 4.80 0.000 Earnings

• 0.114

0.022

• 5.19

0.000 Job tenure 0.012 0.002 7.28 0.000 Excellent health (dummy)

• 0.297

0.075

• 3.96

0.000 Very good health (dummy)

• 0.177

0.065

• 2.72

0.007 Good health (dummy)

• 0.082

0.062

• 1.33

0.184 Health limits work (dummy) 0.077 0.057 1.35 0.176 Variable Coefficient

• Std. err.

z P>|z| England (dummy)

• 0.005

0.062

• 0.08

0.935 Italy (dummy) 0.739 0.115 6.41 0.000 Germany (dummy) 0.316 0.101 3.12 0.002 Austria (dummy) 0.757 0.135 5.60 0.000 Sweden (dummy) 0.041 0.081 0.51 0.611 Netherlands (dummy) 0.343 0.096 3.57 0.000 Spain (dummy) 0.221 0.151 1.46 0.144 France (dummy) 0.356 0.097 3.68 0.000 Denmark (dummy) 0.341 0.099 3.44 0.001 Switzerland (dummy)

• 0.044

0.144

• 0.30

0.761 Belgium (dummy) 0.337 0.093 3.62 0.000 Greece (dummy)

• 0.351

0.132

• 2.66

0.008 Constant

• 20.994

1.533

• 13.70

0.000

Number of observations = 8,237 LR Chi square (25) = 1624.44

• Prob. > Chi square = 0.000
• Having reached State Pension age is a really good predictor of retirement