education and health: what is the role of lifestyles?
DESCRIPTION
Education and Health: What is the Role of Lifestyles?. Giorgio Brunello (University of Padova) Margherita Fort (University of Bologna) Nicole Schneeweis (University of Linz) Rudolf Winter Ebmer (University of Linz). Regensburg May 2011. Motivation. Research questions. - PowerPoint PPT PresentationTRANSCRIPT
EDUCATION AND HEALTH: WHAT IS THE ROLE OF
LIFESTYLES?
Giorgio Brunello (University of Padova) Margherita Fort (University of Bologna) Nicole Schneeweis (University of Linz) Rudolf Winter Ebmer (University of Linz)
Regensburg May 2011
1
MOTIVATION 2
RESEARCH QUESTIONS
Does education cause health outcomes?
Are lifestyles an important channel through which education improves health?
Problem: confounding factors affecting both education and health
2 strategies: IV strategy to identify total causal effect Aggregation and differencing to decompose total
effect into effect due to lifestyles3
Education
Health
Lifestyles
4
CHANNELS FROM EDUCATION TO HEALTH (LOCHNER, 2011)
Stress reduction Better decision making Health Insurance Better information gathering Better jobs (and higher income) Healthier peers and neighbourhoods Lifestyles (eating, drinking, smoking,
exercising...)5
LITERATURE6
PREVIOUS RESEARCH ON CAUSAL EFFECTS
Recent literature uses changes in mandatory schooling laws to identify causal effects Simple OLS models likely to be biased due to
confounders
Mixed results so far (see review by Lochner, 2011)
Important differences by gender7
PREVIOUS RESEARCH ON THE ROLE OF LIFESTYLES
Cutler – Lleras Muney, 2006: in the US the overall effect of education on mortality is reduced by 30% when controlling for lifestyles However, they ignore endogeneity issues and only
look at the effect of current lifestyles Contoyannis and Jones, 2004, estimate a
structural health equation and lifestyle equations by FIML, but treat education as exogenous. Using Canadian data, they find that treating
lifestyles as endogenous could change radically estimated mediating effects
8
THEORY9
STANDARD THEORETICAL APPROACH (ROSENZWEIG – SCHULTZ)
Individuals care about health H (health in the utility function)
They choose optimal lifestyles L by maximizing an inter-temporal utility function subject to a budget constraint
Education affects optimal lifestyles because it affects the valuation of health the discount factor preferences the health production function 10
Dynamic Health Production
Instantaneous utility
Health production function
1,1, tiititiit HYELH Budget constraint
ittitiit LpCEY )(
Inter-temporal utility
rtrT
r
0
itititit HEhLGCU )()()(
11
OPTIMAL LIFESTYLES
)|( itiit XELL
Using this in the health production function and substituting sequentially lagged health yields
)|( itiit XEHH
Where X includes prices and other exogenous factors. We call the latter
“reduced form” health function
Education can influence life style because it increases evaluation of health and increases the discount factor.
12
MEDIATING EFFECT OF LIFESTYLES
Mediating role of lifestyles: effect of education on health going through lifestyles
Current health status most likely depends not only on lifestyles in the previous year, but also on the entire history of lifestyles
Ex: smoking last year matters, but also smoking in the previous years (albeit with lower weights?) 13
SHORT AND LONG RUN MEDIATING EFFECTS
Short run mediating effect: the effect of education on health going through lifestyles lagged once
Long run mediating effect: the effect of education on health going through lifestyles lagged from 1 to T
14
DATA15
DATA: SHARE AND ELSA
Survey of Health, Ageing and Retirement Waves 1 and 2 We also use SHARELIFE
English Longitudinal Survey of Ageing Waves 2 and 3
Sample Females and males (separately), aged 50+ In IV estimates birth cohorts max 10 years
around pivotal age16
INTERESTING FEATURES OF THE DATA By using data on 50+ males and females, we
focus on the effects of education acquired close to or more than 30 years earlier
Not clear whether effects of education on health increase with age
We have information on self reported health self reported limited activity due to poor health long term illness 14 health conditions (heart-related, respiratory,
bones-related, cancer, diabetes, ...)17
MEANS OF HEALTH MEASURES (PERCENTAGE WITH CONDITION)
18
Country
self reported
poor health
Has chronic
diseasesLong term
illnessHeart
diseasesHigh blood pressure Diabetes
Bone related
diseasesRespiratory
diseases Cancer Years of education Age Observations
Austria 23.27 64.06 38.49 22.76 28.13 8.05 16.19 4.84 3.70 11.37 59.01 782Czech Republic 41.76 77.65 51.71 29.64 41.68 13.58 24.53 9.37 4.04 12.02 63.33 2452Denmark 20.81 69.98 46.78 23.01 26.54 6.10 25.48 10.16 5.43 11.80 59.29 1899England 37.30 82.88 62.00 41.24 38.64 3.62 44.88 15.56 4.54 10.70 72.27 4779France 33.06 70.94 45.74 31.30 26.13 8.32 29.17 7.31 5.35 11.31 63.41 2223Italy 33.73 71.12 35.81 26.00 33.89 8.17 31.20 7.19 3.29 8.82 59.83 2092Netherlands 33.80 73.09 44.97 28.47 31.63 11.19 19.88 10.50 5.92 10.60 70.10 1840
Total 33.51 72.97 48.12 32.27 32.96 8.67 22.27 10.43 4.43 11.19 65.54 7415Total males 35.04 77.36 51.28 31.02 34.91 6.98 39.08 10.86 4.80 10.64 65.71 8652Total females 33.87 75.34 49.87 31.59 34.01 7.76 31.33 10.66 4.63 10.89 65.63 16067
IV ESTIMATES19
CAUSAL EFFECT OF EDUCATION ON HEALTH
Use multi-country data (see Brunello, Fort and Weber, 2009; Brunello, Fabbri and Fort, 2010)
Identification
Compulsory schooling reforms in Europe as natural experiment
Reforms in the 1930s-60s in 7 European countries Country fixed effects Cohort fixed effects Country specific trends in cohorts
20
Note: clustered standard errors in parentheses
FIRST STAGE ESTIMATES BY GENDER
21
VARIABLES females males
years of compulsory education 0.268 0.325(0.057)*** (0.079)***
Observations 8,652 7,415F test 21.82 17.10
Note: “rejects” means that the null hypothesis of poolable education coefficients is rejected
POOLING TESTS, FIRST STAGE AND REDUCED FORM
22
Health variable
Pooling test - firstage males
Pooling test - reduced form
males
Pooling test - firstage females
Pooling test - reduced form
females
self reported bad healthdoes not
reject rejectsdoes not
rejectdoes not
reject
has chronic diseasesdoes not
reject rejectsdoes not
reject rejects
long term healthdoes not
reject rejectsdoes not
rejectdoes not
rejectlimited activities (b/c of poor health)
does not reject
does not reject
does not reject rejects
heart problemsdoes not
rejectdoes not
rejectdoes not
rejectdoes not
reject
high blood pressuredoes not
rejectdoes not
rejectdoes not
rejectdoes not
reject
diabetesdoes not
reject rejectsdoes not
rejectdoes not
reject
bone related problemsdoes not
rejectdoes not
rejectdoes not
rejectdoes not
reject
respiratory problemsdoes not
reject rejectsdoes not
reject rejects
cancerdoes not
reject rejectsdoes not
reject rejects
indicator of diseases linear does not
rejectdoes not
rejectdoes not
rejectdoes not
reject
EFFECTS OF EDUCATION ON HEALTH OUTCOMES. FEMALES (SEMI-ELASTICITIES)
23
Health variable
Probit estimate Reduced form 2SLS IV Observations
self reported bad health -0,079 -0,05 -0,189 -0,197 8,602 *** ** ** **
has chronic diseases -0,008 -0,038 -0,142 -0,129 8,652*** *** ** **
long term illness -0,015 -0,036 -0,135 -0,137 8,651 *** ** **
limited activities (b/c of poor health) -0,088 -0,01 -0,04 -0,193 8,652
***heart problems -0,036 0,0005 0,002 -0,008 8,652
***high blood pressure -0,041 -0,065 -0,245 -0,241 8,652
*** ** ** ***diabetes -0,071 -0,206 -0,776 -0,462 8,652
*** *** ** ***bone related problems -0,018 -0,029 -0,111 -0,128 8,652
***respiratory problems -0,042 0,05 0,189 0,235 8,652
*** (0.060)cancer 0,036 -0,134 -0,502 -0,347 8,652
** *indicator of diseases linear -0,017 -0,022 -0,084 -0,084 8,652
EFFECTS OF EDUCATION ON HEALTH OUTCOMES. MALES (SEMI-ELASTICITIES)
24
Health variable
Probit estimate Reduced form estimate 2SLS IV Probit Observa
tions
self reported bad health -0,063 -0,055 -0,171 -0,183 7358*** ** * **
has chronic diseases -0,007 0,009 0,029 0,023 7415*** **
long term illness -0,017 0,037 0,116 0,111 7415*** ** ** **
limited activities (b/c poor health) -0,07 0,009 0,029 0,062 7415
***heart problems -0,012 0,093 0,286 0,22 7415
*** *** ** ***high blood pressure -0,017 0,044 0,137 0,135 7415
*** *diabetes -0,038 -0,057 -0,175 -0,238 7415
***bone related problems -0,043 -0,05 -0,155 -0,181 7415
***respiratory problems -0,052 0,098 0,301 0,275 7415
*** ** * ***cancer 0,006 -0,13 -0,399 -0,326 7415
** **indicator of diseases (linear) -0,012 0,019 0,06 0,06 7415 *** ** * *
IV RESULTS (PERCENTAGES EVALUATED AT SAMPLE MEANS) Females: one additional year of education
reduces Self reported bad health (-19.7%) Presence of chronic diseases (-12.9%) High blood pressure (-24.1%) Diabetes (-46.2%)
Males: one additional year of education reduces Self reported bad health (-18.3%) INCREASES
Long term illness (11.1%) Hearth problems (22%) Respiratory problems (27.5%) Objective measure of conditions (6%) 25
IV RESULTS
We confirm important gender differences
Positive effect of education on health conditions is puzzling. Possible explanations include
Education moves males away from less sedentary occupations
Education moves males to more stressful occupations (or males are less able to cope with stress...) 26
HEALTH CONDITIONS AND SCREENING Conditions are reported by individual but
must have been detected by a doctor „Did your doctor tell you …?“
Marginal effect of education:
If more education induces e.g. males to go to the doctor more often, more diseases would be detected
Preliminary results: no effects of screening!!
27
( ) ( ) ( | )S...Screening D ... DiseaseP D P S P D S
( ) ( ) ( | )( | ) ( )P D P S P D SP D S P SE E E
POTENTIAL BIASES
Older cohorts (pre-treatment) are less healthy: we capture this with cohort dummies
Members of older cohorts who are still alive – positive selection and downward bias – we try to control for this by adding life expectancy at birth Using sampling weights that are inversely proportional to
the difference between age and life expectancy
Placebo treatment as in Black, Devereux and Salvanes (2008) ---- Placebo reforms should have no effect
28
REDUCED FORM ESTIMATES: WITH YEARS OF COMPULSORY EDUCATION 5 YEARS AHEAD
Health variable
Marginal effect of
YCOMP Males
Marginal effect of YCOMP -
placebo test – Males
Marginal effect of YCOMP 5 years
ahead - placebo test -Males
Marginal effect of YCOMP
- Females
Marginal effect of YCOMP
- placebo test - Females
Marginal effect of YCOMP 5 years
ahead - placebo test -
Females
self reported bad health -0,055 -0.038 0.052 -0,05 -0.039 0.043
** **
has chronic diseases 0,009 0.011 0.004 -0,038 -0.033 0.003*** **
long term illness 0,037 0.043 0.016 -0,036 -0.030 0.037** * **
limited activities due to poor health 0,009 0.023 0.040 -0,01 -0.040 0.008
heart problems 0,093 0.108 0.044 0,0005 -0.014 -0.044*** ***
high blood pressure 0,044 0.058 0.039 -0,065 -0.061 0.050* ** *
diabetes -0,057 -0.068 -0.033 -0,206 -0.252 0.026*** ***
bone related problems -0,05 -0.013 0.111 -0,029 -0.030 0.011**
respiratory problems 0,098 0.089 -0.029 0,05 0.115 0.112** *
cancer -0,13 -0.138 -0.022 -0,134 -0.178 -0.159
indicator of diseases linear 0,019 0,026 0.0207 -0,022 -0,017 0.0176 ** *** * *
THE MEDIATING EFFECTS OF LIFESTYLES
30
THE CARD ROTHSTEIN APPROACH We do not have credible instruments for lifestyles
We combine gender differencing (fixed effects) to remove common un-observables with selection on observables, using SHARELIFE info.
SHARELIFE variables control for early health conditions and parental background.
Fixed effects remove nature and nurture effects that are common between genders.
31
icgbticgbticgticgbicgbt HELH 132)1(1
icgbtcbtcgbticgbt eu
0],,,|[ tcgbeE icgbt
cbtFcbt
Fcbtcbt vZZu 21
cbtFcbt
Fcbt
Fc
Fc
Ftcb
Ftcbcbt uHHEELLH 13130202)1(1)1(1
where i=individual; c: cohort; g: gender; t=time. We assume
We take gender differences to remove cbt
We model the residual error as function of Z (parental background and early health from SHARELIFE)
32
ESTIMATES OF “REDUCED FORM” AND DYNAMIC HEALTH EQUATIONS
We add to the sample Germany and Sweden (in future work we plan to extend this approach to other countries included in SHARE)
We estimate these equations both on micro data using selection on observables only
and on cell data using gender differences plus selection
on observables (Card-Rothstein)
33
ESTIMATED EFFECTS OF EDUCATION ON SELF REPORTED BAD HEALTH, WITH AND WITHOUT HEALTH LIFESTYLES. LINEAR PROBABILITY MODELS MICRO DATA
34
Reduced form health equation -
Females
Dynamic health equation - Females
Reduced form health equation
- Males
Dynamic health
equation - Males
years of schooling -0.017 -0.005 -0.014 -0.006
*** *** *** ***lagged dependent variable 0.503 0.491
*** ***drunk alcohol every day in year t-1 -0.029 -0.032
** **
was smoking in year t-1 0.064 0.047*** ***
did vigorous activity in year t-1 -0.016 -0.023
*** ***BMI in year t-1 0.007 0.005
*** ***father drunk or had mental troubles (age 10) 0.033 0.021 0.021 0.004
** *
presence of parents in the house at 10 0.017 0.011 -0.001 -0.007
ESTIMATED EFFECTS OF EDUCATION ON SELF REPORTED BAD HEALTH, WITH AND WITHOUT HEALTH LIFESTYLES. GENDER DIFFERENCES. CELL DATA. WEIGHTED REGRESSIONS.
35
Reduced form health equation -
Females
Dynamic health equation with
income - Females
Reduced form health equation -
Males
Dynamic health equation with
income - Males
years of schooling -0,023 -0,018 0.003 0.013*** *
lagged dependent variable 0,170 0.299* ***
drunk alcohol every day in year t-1 0,082 0.079
was smoking in year t-1 0,028 -0.103
did vigorous activity in year t-1 -0,005 -0.035
BMI in year t-1 0,001 -0.001
real income -0,003 -0.004**
Observations 232 230 232 230
ESTIMATED EFFECTS OF EDUCATION ON LIMITED ACTIVITY DUE TO POOR HEALTH, WITH AND WITHOUT HEALTH LIFESTYLES. LINEAR PROBABILITY MODELS MICRO DATA
36
Reduced form health equation
-Females
Dynamic health equation with income
- Females
Reduced form health
equation - Males
Dynamic health
equation with income - Males
years of schooling -0.0065 -0.0026 -0.0034 -0.0008*** ** ***
lagged dependent variable 0.4320 0.4490*** ***
drunk alcohol every day in year t-1 -0.0112 0.0038
was smoking in year t-1 0.0186 -0.0029*
did vigorous activity in year t-1 -0.0141 -0.0116*** ***
BMI in year t-1 0.0052 0.0035*** ***
father drunk or had mental troubles (age 10) 0.0259 0.0125 -0.0005 -0.0054
**
presence of parents in the house at 10 0.0188 0.0150 0.0106 0.0018
ESTIMATED EFFECTS OF EDUCATION ON LIMITED ACTIVITY BECAUSE OF POOR HEALTH, WITH AND WITHOUT HEALTH LIFESTYLES. GENDER DIFFERENCES. CELL DATA. WEIGHTED REGRESSIONS.
37
Reduced form health
equation -Females
Dynamic health
equation with income - Females
Reduced form health
equation - Males
Dynamic health
equation with income -
Males
years of schooling -0,020 -0,024 -0.018 -0.017*** *** *** **
lagged dependent variable 0,146 0.156*
drunk alcohol every day in year t-1 0,018 0.029
was smoking in year t-1 -0,022 -0.076
did vigorous activity in year t-1 -0,001 -0.034*
BMI in year t-1 0,008 0.002*
real income 0,000 -0.001
Observations 232 230 232 230
ESTIMATION OF MEDIATING EFFECT We estimate both the dynamic health
equation and the „reduced form“ health equation
Total effect of education on health
Use also reduced form H=f(E)
Effect NOT going through lifestyle (could be pos. or neg.)
38
1 1
,
H L E Y H
L E Y YE
1 ( )1
EG Y
1 ( )1
DEG Y
EDUCATION GRADIENT MEDIATION BY LIFESTYLE - FEMALES
39
Education gradient
Gradient not
mediated by
lifestyles
Gradient mediated
by lifestyles
Gradient mediated by lagged lifestyles
self reported health -0.023 -0.024 0.001 0.001has chronic diseases 0.002 -0.003 0.006 0.004long term illness -0.051 -0.073 0.022 0.012limited activities due to poor health -0.020 -0.028 0.008 0.007heart problems -0.002 -0.012 0.010 0.006high blood pressure -0.004 0.001 -0.005 -0.002diabetes 0.001 0.009 -0.008 -0.003bone related conditions -0.005 0.000 -0.005 -0.002respiratory conditions 0.009 0.000 0.009 0.005cancer -0.009 0.001 -0.010 -0.008linear indicator of conditions -0.017 -0.026 0.008 0.005
EDUCATION GRADIENT MEDIATION BY LIFESTYLE - MALES
40
Education gradient
Gradient not
mediated by
lifestyles
Gradient mediated
by lifestyles
Gradient mediated by lagged lifestyles
self reported health 0.003 0.016 -0.013 -0.009has chronic diseases 0.009 0.004 0.005 0.004long term illness 0.008 -0.013 0.021 0.019limited activities due to poor health -0.018 -0.021 0.003 0.002heart problems -0.006 -0.013 0.007 0.004high blood pressure -0.000 0.006 -0.006 -0.004diabetes 0.005 0.011 -0.006 -0.003bone related conditions -0.008 -0.002 -0.006 -0.002respiratory conditions -0.001 -0.01 0.009 0.005cancer -0.009 -0.006 -0.003 -0.003linear indicator of conditions 0.008 -0.009 0.017 0.012
LONG AND SHORT TERM EFFECTS
In most cases short and long effects are not very different, which suggests that the first lag of lifestyles captures most of the mediating effect Impact of Ht-1 small (around 0.2)
Males generally small education gradient For Females negative effect for:
Self-reported health Long-term illness Limited activities Linear indicator of diseases
41
LONG TERM MEDIATING EFFECTS OF LIFESTYLES
Generally small Effects for females:
Blood pressure Diabetes Bones Cancer
Effects for males: Self-reported health Blood pressure Diabetes Bones Cancer
42
IMPORTANT QUALIFICATION
Finding that the mediating effect of lifestyles is small does not exclude that omitted lifestyles (unprotected sex or drug abuse) are important
vehicles of the education gradients
The effect of unobserved lifestyles is incorporated in the direct effect of education on health
43
CONCLUSIONS44
CONCLUSIONS
Education has important protective effects on the health of females
The evidence for males is less compelling: in some cases education increases bad health
The mediating effect of measured lifestyles (drink, smoke, exercise and calorie balance) is close to zero for several health outcomes
Measured lifestyles really matter for high blood pressure, cancer and respiratory diseases for females, and for bone related conditions for males
45
PROBLEMS AND THINGS TO DO
We omit several important lifestyles (for instance unprotected sex, drugs)
We need to produce standard errors for our measures of mediating effects
More data (countries)
Include “screening” among chosen “lifestyles” 46