Job Loss, Retirement and
the Mental Health of
Older Americans
Bidisha Mandal
Brian Roe
The Ohio State University
Outline







Motivation
Literature
Data
Model
Results
Conclusion
Future Research
Motivation

Increasing percentage of older individuals in the population.

General decline in job security in U.S. labor market.

Physical limitations, cognitive changes, bereavement are
commonly associated with aging.

Does work displacement cause additional distress? Are there any
long-term effects?
 Job loss – skills may not be transferable, loss of income
 Retirement – lifestyle changes.

Policy implication – increased private medical expenditure,
increased public spending for government medical programs.
Relevance

Mental health affects social behavior, morale, as well as work
productivity.

Deteriorating mental health can manifest in weakened physical
health and increase likelihood of suicide.

Declines in the mental health may negatively influence the wellbeing of other household members.

Older Americans may be less inclined to seek help for
psychological problems (as compared to physical decrements).

Job loss affects the quality of life
Literature

Retirement




Kim and Moen (2002): 458 New York employees; 1994, 1996, 1998
waves of Cornell Retirement and Well-being study.
Results: short-term boost in morale, and long-term increase in distress
levels for men.
Drentea (2002): 2 different cross-sectional national surveys
Mixed results: lower sense of control, but lower anxiety levels among
retirees.
Midanik et al. (2005): 595 members of a health maintenance
organization; short-term effect.
Result: lower stress levels among retirees.
No clear trend; No long-run national panel have been studied yet.

Related Literature on Retirement



Kerkhofs et al. (1999): health and retirement are endogenously
related.
Dwyer and Mitchell (1999), Disney et al. (2006): health problems
influence retirement plans more strongly than economic variables.
Involuntary Job Loss (business shut-down or lay-off)



Gallo et al. (2000): 1992 and 1994 waves of Health and Retirement
Study.
Different methodology to handle endogeneity
 Reverse causality and unobserved heterogeneity
 OLS vs. HT/IV and 2SLS
Alternative coding
Framework
Work
displacement
Involuntary
job loss
Retirement
Easily adjusts to
new lifestyle
Reentry
Unable to adjust to
new lifestyle
Retirement
plans affected
Reemployment
Long spell or
forced to retire
CESD Score

Mental health measure
 Developed by Radloff (1977) – short, self-reporting scale (20
items) for general population.
 HRS only includes 8 items – 6 negative and 2 positive binary
indicators


Negative items – felt depressed, everything an effort, sleep was restless,
felt lonely, felt sad, could not get going
Positive items – was happy, enjoyed life

CESD = sum (negative items) – sum (positive items)
Thus, higher score (0 to 8) means worse mental health.

Both versions commonly used in other studies to measure
distress and psychological well-being.
Summary – CESD Score

Reliability
 Cronbach’s alpha coefficient for 20 items = 0.85
 Cronbach’s alpha coefficient for 8 items = 0.71

Mean change in CESD score among those who suffered
involuntary job loss is 0.19

Mean change in CESD score among retirees is 0.17

Maximum increase in CESD score is reported between the first
two waves (1992 to 1994), when job loss rates were high.

CESD scores improve during latter waves for all.
Coding Involuntary Job Loss and Retirement


Unbalanced panel data
 6 waves – 1992, 1994, 1996, 1998, 2000 and 2002
 N=7,780 (all those employed in 1992, 51-61 years old)
Coding
 Survey does not ask R if suffered involuntary job-loss, but reason
for unemployment.
 Involuntary job loss


Retirement (voluntary)



If R reports business closure or layoff, and started looking for job
immediately.
If R accepts early retirement incentives, and does not look for job
immediately. These individuals also call themselves – ‘self-retired’.
Plus, those who report retirement as labor market status.
Data limitations
Data
Mean change in CESD score between 1992 and 1994
Labor market status Employed
Invol. exit
Vol. exit
(coded)
ΔCESD
0.44
0.95
Self-retired
(survey instrument)
0.61
‘Lost job’
0.63
Collapsed as ‘Retirees’
Distribution of HRS respondents in different labor market situations
Survey year
Retired
Lost job
Retired
Employed
1994
97
792
6153
1996
87
1451
5116
1998
67
1907
4373
2000
32
2365
3605
2002
47
2883
2932
0.65
Summary Statistics (selected variables)
Variables (Change)
1992-1994 1994-1996 1996-1998 1998-2000 2000-2002
Job loss (%)
3.32
6.25
4.54
3.17
4.13
Retirement (%)
11.25
25.23
31.80
43.31
53.75
Separated/divorced (%)
1.65
1.08
0.98
0.68
0.56
Married/re-married (%)
1.07
1.19
0.84
0.91
0.81
Widowed (%)
1.05
1.08
1.31
1.42
1.52
Worse physical health (%)
14.4
15.9
16.9
17.6
20.8
Δ ADLA
0.03
0.07
0.02
0.03
0.02
Δ Wealth ($10,000)
3.31
3.01
5.71
4.66
- 1.03
Response rate
91.04
86.58
83.06
78.77
76.62
Unobserved Heterogeneity

Compare fixed effects, random effects and Hausman-Taylor IV
random effects model using Hausman specification test
FE:
Yit  Yi.  ( X it  X i. )    it   i.
where, X it are time-varying independent variables
RE:
Yit  X it   Zi   i   it
where, Z i are time-invariant independent variables
and,  i denotes individual-specific effects
HT-IV: Yit  X 1it 1  X 2it  2  Z1i  1  Z 2i  2   i   it
where, the subscripts distinguish between exogenous and
endogenous variables

First difference model: Yit  X it   it
Comparing Model Properties

FE
Subtracts off group means
 Along with time-invariant regressors, latent effects are left out
FD
 Similar, but subtracts off last period’s observations
 Again, gets rid of both time-invariant factors and latent effects
 Unbiased, consistent estimates from both FE and FD
RE
 Can use time-invariant variables, as long as independent of latent effects
 Efficiency gain
HT-IV RE
 Allows time-invariant variables under lesser constraints
 Correct specification produces consistent, unbiased and efficient estimates
 Limitation – single-equation model; model misspecification




FE, RE and HT-IV RE Models
Estimates (SE) from different models for selected variables
Dependent variable: CESD score
Variables
FE
RE
Job loss
0.191* (0.042)
0.204* (0.040)
0.189*
Retired
0.051** (0.023)
0.036
0.051** (0.021)
Time
0.424* (0.019)
0.413* (0.019)
- 0.051* (0.003)
- 0.052* (0.003)
Time (squared)
(0.021)
HT-IV RE
0.425
(0.038)
(0.018)
- 0.052* (0.002)
Separated
0.037
(0.127)
0.168** (0.069)
0.034
Widowed
0.405* (0.133)
0.409* (0.073)
0.402* (0.121)
- 0.276** (0.128)
- 0.210* (0.066)
- 0.282** (0.117)
0.190* (0.016)
0.248* (0.009)
0.199* (0.014)
Married
Physical health
* p < 0.01; ** p < 0.05
(0.116)
Model Choice

Latent effects – motivation, productivity





Time-varying endogenous variables – involuntary and voluntary
exits, marriage/remarriage, separation/divorce, ADLA index,
physical health condition
Time-invariant endogenous variables – age, education, white/bluecollar job
Choice of model (FE vs. RE) depends on cost of efficiency gain
Specification test
RE
HT-IV RE
χ2 (df)
374.01 (21)
8.14 (7)
p-value
0.00
0.32
Only one time-invariant variable significant - gender
First difference model is adequate in controlling for latent effects
and is able to capture the change in mental health due to a shock
Compare with Previous Study

Gallo et al. (2000) use data from 1992 and 1994 HRS.

Sample selection is sufficient to take care of latent effects – exclude
retirees, self-employed individuals, disabled, and those who left their
jobs for reasons other than plant closure and lay-off.

Involuntary job loss – plant closure and lay-off

Method – OLS regression.

Replicate their coding and methodology, and obtain estimate of
involuntary exit similar to theirs.

Problem – unobserved heterogeneity still exists
Specification test
RE
HT-IV RE
χ2 (df)
102.32 (18)
16.62 (5)
p-value
0.00
0.01
Reverse Causality

Suspect endogenous variables





Involuntary exit
Voluntary exit
Separation/Divorce
Marriage/Re-marriage
Instruments (excluded exogenous variables)






Unemployment rate
Age at the beginning of each survey
Parents’ level of education
R’s level of education
If R’s parents are/were married to each other or to step-parents
Number of divorces and widowhoods reported in 1992
Validity

Three basic tests to
 Check if endogeneity actually exists
 Ho: suspect endogenous variables are exogenous
Compare 2 regressions – one where suspect regressors are treated as
endogenous, and the other where they are exogenous. Test statistic is
distributed χ2 with df = number of endogenous regressors

Check for weak instruments
 LR test - Ho: equation is underidentified
To check if the instruments are poor proxies for the endogenous
variables. Test statistic is distributed χ2 with df = total number of
exogenous regressors - endogenous regressors + 1

Check the validity of the instruments
 J statistic - Ho: instruments are uncorrelated with error
Test statistic is distributed χ2 with df = number of instruments - 1
Results from Labor Market Exit
2SLS regression: Dependent variable – ΔCESD score
Variables (Change)
All (N = 7780)
Estimate
SE
0.244*
0.042
Retirement E
- 0.261*
0.034
Sep./divorce E
- 0.121*
0.038
Married E
0.037
0.035
Widowed
0.897*
0.092
Death of child
0.220*
0.059
ΔADLA
0.426*
0.027
Worse physical health
0.331*
0.032
Involuntary exit E
Endogeneity χ2 (4)
87.98*
LR statistic χ2 (6)
24.79*
J statistics χ2 (5)
4.44
E - endogenous
* p < 0.01
Results from Re-entry after Job Loss
2SLS regression: Dependent variable – ΔCESD score
Variables (Change)
Only those who lost job (N = 418)
Estimate
Reemployment E
SE
- 0.426*
0.137
Sep./divorce
0.862**
0.440
Married
0.341
0.504
Widowed
1.999*
0.425
ΔADLA
0.723*
0.133
Worse physical health
0.451*
0.126
Endogeneity χ2 (1)
8.57*
LR statistic χ2 (3)
112.42*
J statistics χ2 (2)
4.12
E – endogenous; * p < 0.01; ** p < 0.05
Results from Re-entry after Retirement
2SLS regression: Dependent variable – ΔCESD score
Variables (Change)
Only retirees (N = 3578)
Estimate
Reemployment E
Sep./divorce E
Married E
SE
- 0.216*
0.044
0.170
0.041
- 0.065
0.042
Widowed
0.973*
0.125
ΔADLA
0.334*
0.037
Worse physical health
0.175*
0.040
Endogeneity χ2 (3)
23.37*
LR statistic χ2 (4)
16.67**
J statistics χ2 (3)
4.28
E – endogenous; * p < 0.01; ** p < 0.05
Involuntary Exit vs. Reemployment
2SLS regression: Dependent variable – ΔCESD score
Variables (Change)
Only those who lost job (N = 418)
Estimate
SE
0.172**
0.075
Involuntary exit E
Reemployment E
- 0.156*
0.050
Sep./divorce
0.579
0.446
Married
0.185
0.545
Widowed
1.073**
0.424
ΔADLA
0.517*
0.118
Worse physical health
0.208
0.135
Ho: Effect of exit = - Effect of re-entry
F-value (p-value)
E – endogenous; * p < 0.01; ** p < 0.05
0.36 (0.55)
Summary of Steps

First Difference Model
 Accounts for Unobserved Heterogeneity
 To capture the effect of change in labor market status on change
in mental health

Reverse Causality
 Mental health may decide labor market status
 Two Stage Least Squares (2SLS)
 Endogenous Regressors
 Labor market status
 Marital status (except widowhood)

Effect of Reemployment
Conclusion

Endogenous regressors are labor market exit and re-entry,
separation or divorce, and marriage or remarriage

Involuntary job loss negatively impacts mental health

Similar in magnitude and direction to effect of death of child

Highest negative effect due to death of spouse

Retirement has a positive effect on mental health (short-term)

Re-entering labor market has a positive effect on the mental health
for all

Re-entry recaptures the previous mental health status of those who
lost job involuntarily