Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
Three essays on health & aging
(USC Thesis Other)
Three essays on health & aging
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
THREE ESSAYS ON HEALTH & AGING
by
Urvashi Jain
A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulllment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(Economics)
August 2018
Copyright 2018 Urvashi Jain
Acknowledgements
I am deeply indebted to my advisor, John Strauss, for his diligent guidance, endless patience,
unconditional support and the seasonal kumquats! I am grateful to Jinkook Lee for her continuing
support and the exciting opportunities that she has always included me in. Her enthusiasm and
energy are infectious, she leads and motivates by example. From John, I learned the value of
academic rigor and hard work. From Jinkook, I learned persistence, and how interdisciplinary
work can help us do better economics research.
I would also like to thank Jerey Nugent and Marco Angrisani for serving on my committee. I
have not not met a kinder soul than Professor Nugent, or somebody as intuitive with econometrics as
Marco. I have no words to express my gratitude for my friend, colleague, and co-author, Mingming.
I would've been adrift for far too long had it not been for her friendship, intelligence, and most
importantly, the countless times we racked our brains to nd food around USC. I owe special thanks
to Young Miller, Sandy Chien, Morgan Ponder and Sara Burnett- they have always been ready to
help and made my life pleasant and easier.
Last but not the least, I thank my parents, Devender Kumar Jain and Kamlesh Jain. They
ensured I had a comfortable upbringing despite the many constraints they faced. I enjoyed a level
of freedom unheard of in my extended family. They have trusted me to make the right decisions,
and supported my pursuit of a Ph.D. so far away from home, even when they knew they would
miss me immensely.
ii
Table of Contents
Acknowledgements ii
List Of Tables v
List Of Figures viii
Abstract ix
Chapter 1: Introduction 1
Chapter 2: Long-Term Health Eects of Malaria Eradication: Evidence from India 5
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Previous Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Basic Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4.1 Study on global AGEing and adult health (SAGE) . . . . . . . . . . . . . . . 9
2.4.2 Pre-eradication malarial intensity . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4.3 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5.1 Dierence-in-dierences Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5.2 Cohort Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.5.3 Heterogeneity by Gender . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.5.4 Robustness Check . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
Chapter 3: Height Shrinkage, Cognition and Health among the Elderly: Compar-
isons across England and Indonesia 24
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2 Existing Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.3.1 England . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.3.2 Indonesia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.4 Shrinkage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.4.1 Extent of Shrinkage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.4.2 (Erroneous) Height Gain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.4.3 Covariates of Shrinkage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.5 Shrinkage and Health . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
iii
3.5.1 Shrinkage and Cognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.5.2 Shrinkage and Physical Health . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.5.3 Shrinkage and Mortality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.6 Strengths and Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
Chapter 4: Together in Sickness and in Health: Spousal In
uence in Health and
Health Behaviors of Elderly in England 64
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.3 Theoretical Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.3.1 Assortative Mating . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.3.2 Shared Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.3.3 Spillover Eect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.4 Emprical Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.4.1 Empirical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.4.2 Assumptions and Empirical Methods . . . . . . . . . . . . . . . . . . . . . . . 71
4.5 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.5.1 Outcome Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.5.2 Control Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.5.3 Sample and Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.6 Estimation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.6.1 Health Status . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.6.2 Health-related Behaviors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.7 Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.7.1 Dierence GMM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.7.2 Path Dependence and Spousal In
uence . . . . . . . . . . . . . . . . . . . . . 100
4.7.3 Sample Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
Chapter 5: Conclusion 104
Bibliography 106
Appendix A
Chapter 2 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
Appendix B
Chapter 3 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
Appendix C
Chapter 4 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
iv
List Of Tables
2.1 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Childhood malaria exposure and Education . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Childhood malaria exposure and Height . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4 Childhood malaria exposure and Episodic Memory . . . . . . . . . . . . . . . . . . . 16
2.5 Childhood malaria exposure and Attention . . . . . . . . . . . . . . . . . . . . . . . 17
2.6 Childhood malaria exposure and Verbal
uency . . . . . . . . . . . . . . . . . . . . . 18
2.7 Heterogeneity by Gender . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.8 Urban sample . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.1 Shrinkage (in cm) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Shrinkage (in cm) Covariates in ELSA . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.3 Shrinkage (in cm) Covariates in IFLS . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.4 Descriptive Statistics: ELSA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.5 Descriptive Statistics: IFLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.6 Height Shrinkage and Cognition (Word Recall) in ELSA . . . . . . . . . . . . . . . . 47
3.7 Height Shrinkage and Fluency (Animal Naming) in ELSA . . . . . . . . . . . . . . . 48
3.8 Height Shrinkage and Cognition (Word Recall) in IFLS . . . . . . . . . . . . . . . . 49
3.9 Height Shrinkage and Fluency (Animal Naming) in IFLS . . . . . . . . . . . . . . . . 50
3.10 Height Shrinkage and Lung Function (Forced Expiratory Volume) in ELSA . . . . . 54
v
3.11 Height Shrinkage and Grip Strength in ELSA . . . . . . . . . . . . . . . . . . . . . . 55
3.12 Height Shrinkage and Lung Function (Peak Expiratory Flow) in IFLS . . . . . . . . 56
3.13 Height Shrinkage and Grip Strength in IFLS . . . . . . . . . . . . . . . . . . . . . . 57
3.14 Height Shrinkage and Disability in ELSA . . . . . . . . . . . . . . . . . . . . . . . . 58
3.15 Height Shrinkage and Diculty in Kneeling, Stooping or Crouching in ELSA . . . . 58
3.16 Height Shrinkage and Disability in IFLS . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.17 Height Shrinkage and Diculty in Kneeling, Bowing or Squatting in IFLS . . . . . . 60
3.18 Height Shrinkage and Mortality in IFLS . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.1 Descriptive Statistics of Main Variables for Men . . . . . . . . . . . . . . . . . . . . . 78
4.2 Descriptive Statistics of Main Variables for Women . . . . . . . . . . . . . . . . . . . 79
4.3 Correlation Coecients of Health Status and Health Behaviors Within Couples . . . 80
4.4 Results on Good Self-reported Health (SRH) . . . . . . . . . . . . . . . . . . . . . . 83
4.5 Results on Depressive Symptoms (CES-D 8 Score) . . . . . . . . . . . . . . . . . . . 84
4.6 Results on Smoking Status (= 1 if smokes now) . . . . . . . . . . . . . . . . . . . . . 86
4.7 Results on Smoking Intensity (# cigarettes/day) . . . . . . . . . . . . . . . . . . . . 87
4.8 Results on Drinking Frequency (# days/week drinks) . . . . . . . . . . . . . . . . . 88
4.9 Results on Frequency of Exercise (# days/week participating in vigorous physical
activities, VPA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.10 Results on Social Contact (= 1 if any weekly contact with relatives in person) . . . . 90
4.11 Dierence GMM Results on Good Self-reported Health (SRH) . . . . . . . . . . . . . 93
4.12 Dierence GMM Results on Depressive Symptoms (CES-D 8 Score) . . . . . . . . . 94
4.13 Dierence GMM Results on Smoking Status (= 1 if smokes now) . . . . . . . . . . . 95
4.14 Dierence GMM Results on Smoking Intensity (# cigarettes/day) . . . . . . . . . . 96
4.15 Dierence GMM Results on Drinking Frequency (# days/week drinks) . . . . . . . . 97
vi
4.16 Dierence GMM Results on Frequency of Exercise (# days/week participating in
vigorous physical activities, VPA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.17 Dierence GMM Results on Social Contact (= 1 if any weekly contact with relatives
in person) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
A1 Additional birth-year weather controls . . . . . . . . . . . . . . . . . . . . . . . . . . 114
B1 Limb Length and Shrinkage in IFLS . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
B2 Height Shrinkage and ADL in ELSA . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
B3 Height Shrinkage and IADL in ELSA . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
B4 Height Shrinkage and ADL in IFLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
B5 Height Shrinkage and IADL in IFLS . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
B6 Height Shrinkage and Diculty in Lifting in ELSA . . . . . . . . . . . . . . . . . . . 121
B7 Height Shrinkage and Diculty in Getting up from Chair in ELSA . . . . . . . . . . 122
B8 Height Shrinkage and Diculty in Carrying a Heavy Load in IFLS . . . . . . . . . . 123
B9 Height Shrinkage and Diculty in Getting up from Chair in IFLS . . . . . . . . . . 124
C1 Results on Self-reported Health (1{3 Scale) . . . . . . . . . . . . . . . . . . . . . . . 126
C2 Results on Frequency of Exercise (=1 if any weekly vigorous physical activities) . . . 127
C3 Results on Chronic Conditions (# of severe diseases) . . . . . . . . . . . . . . . . . . 128
vii
List Of Figures
2.1 Malaria Endemicity Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Cohort-specic coecient plots for memory. Points are birth cohort specic co-
ecients,
c
, from Equation 2.2, 95% condence intervals displayed around point
estimates. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3 Cohort-specic coecient plots for attention. Points are birth cohort specic co-
ecients,
c
, from Equation 2.2, 95% condence intervals displayed around point
estimates. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.1 Shrinkage over 4 years in ELSA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.2 Shrinkage over 8 years in ELSA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.3 Shrinkage over 7 years in IFLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.4 Shrinkage over 17 years in IFLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.5 Shrinkage over 4 and 8 years by gender and age in ELSA . . . . . . . . . . . . . . . 37
3.6 Shrinkage over 14 and 17 years by gender and age in IFLS . . . . . . . . . . . . . . . 37
viii
Abstract
This dissertation is an endeavor to better understand the health of older adults, using datasets
and measures comparable across countries. I study early-life origins of later-life health, useful
population level biomarkers of aging, and peer eects in health and health behaviors among older
couples.
Chapter 2 examines the eects of early life malaria exposure on later-life cognition in a sample of
Indian adults, using a nationwide eradication program and geographic variation in pre-eradication
malaria endemicity. Dierence-in-dierence estimates suggest that the program was associated
with increases in the measures of episodic and working memory. I nd no eects on the cognitive
domain of retrieval/
uency, as well as no eects on height and education.
Chapter 3 documents the extent of height loss and explores the associations of early life health
measured by height, depletion of health endowments measured by height shrinkage and later life
health among the English and Indonesian elderly populations. Older women and Indonesian elderly
are found to shrink more. Height shrinkage is signicantly associated with lower cognition and
reduced physical function for elderly men and women in both countries, even when baseline height
is controlled for. Extreme height loss is also found to be a major predictor of mortality among the
Indonesian elderly.
Chapter 4 analyzes the widely documented spousal concordance in health status and health be-
haviors, utilizing data from the English Longitudinal Study of Ageing (ELSA). GMM estimations
ix
of a dynamic panel model with individual xed eects imply signicant and positive contempora-
neous spillover eects in health between elderly couples, with women's mental health being aected
more by their spouse. Strong causal eects are also found for spousal smoking, drinking and phys-
ical activity, as well as social contact. While wives are a bigger in
uence when it comes to their
husband's smoking likelihoods, men aect more than women as to how much they smoke.
x
Chapter 1
Introduction
Beneting from the advancements in medicine and public health, the 20th century witnessed un-
precedented declines in mortality rates and increases in life expectancy. Combined with lower
fertility rates, these changes have resulted in demographic transition throughout the world, albeit
at diering times and to varying extents. Hence, in the 21st century, the world has begun to
experience population aging, which is already advancing in Northern America and Europe, and is
projected to increase rapidly in the developing, yet densely populated, regions of Latin America,
Africa and Asia. Population aging will aect the economy in many ways. First, the burden of
disease is shifting from communicable diseases to chronic and non-communicable diseases. Aic-
tions traditionally associated with old age like hypertension, diabetes, and cognitive decline are
gradually becoming public health concerns. Second, population aging will increase the need for
both social support (e.g. health benets and pensions) as well as familial support (e.g. economic,
care-giving). However, all of these vital questions regarding the pace of aging, disease burden, and
the need for social and familial support can only be answered accurately with high quality data,
ideally also representative at the population level. Panel surveys are especially useful for studying
aging and the accompanying changes in health. In this dissertation, I focus on better understand-
ing the epidemiological determinants of later-life cognition, measures of declining health, and how
1
peer eects contribute to the dynamics of health in old age. I utilize both cross section and panel
datasets, employing both causal and descriptive empirical approaches.
Chapter 2 contributes towards better understanding why some individuals have better health
in older ages simply due to better early-life environments. While throughout the globe, the battle
against infectious diseases is an ongoing one, immense gains have been made over the course of last
century. It raises the question as to whether the adverse eects of these diseases are only temporary,
or whether depending upon the timing of morbidity, especially if it coincides with a critical devel-
opment stage, can have long lasting health impacts (Barker (1992), Currie and Almond (2011)).
One disease in particular, malaria, has been contained to a large extent due to WHO sponsored
eradication campaigns in the 1940's and 1950's, which relied upon the then newly available insec-
ticide, DDT. Malaria is known to potentially cause neuropsychological decits as well as general
growth impairment among infants and children (up to the age of 5), and these are likely to manifest
as lower cognitive capacity in adulthood. Hence early life exposure to malaria could have economic
implications for productivity via aecting human capital. Using a nationwide malaria eradication
program in India, I investigate whether reduced exposure to malaria in early life was associated
with gains in cognitive capacity. Geographic variation in malaria endemicity, combined with the
plausibly exogenous timing of the eradication program (driven by the advent of DDT), provide a
quasi-experimental set up ideal for a dierence-in-dierences identication strategy. I use individ-
ual level data for India from the rst wave in 2008 of WHO-Study on Global Ageing and Adult
Health (SAGE), Cutler et al. (2010) provide the data on pre-eradication malaria endemicity. I nd
that cohorts born after the eradication program, in areas with pre-eradication malaria endemicity
greater by one unit, experienced gains of 0.2 standard deviations on an average in the cognitive
domain of episodic as well as working memory. There is no evidence of positive treatment eects
on the cognitive domain of retrieval/
uency, nor on height or education. These results suggest that
2
gains in cognition are still feasible even in the absence of improved schooling, by virtue of a better
epidemiological environment.
Chapter 3 explores the interaction of early life health measured by height, deterioration of
health endowment measured by height shrinkage and later life health. This research is only feasible
due to high quality, long-term, panel datasets from England and Indonesia, which also carefully
collected biomarker data. When studying early life origins of later life cognition, height is often
used as a proxy measure of early life health (Case and Paxson, 2008a). In this paper, we argue
that a current measure of height among the elderly can be separated into pre-shrinkage height
and height loss. While pre-shrinkage height is still a summary measure of early life health, height
loss (shrinkage) is also an informative variable regarding age related health deterioration. By
using only the measure of height in estimations, previous literature missed out upon potential
associations of health variables with height shrinkage (Huang et al., 2013). We rst document
the extent of height loss based on repeated measurements of height over time among the English
and Indonesian elderly populations, using panel data from English Longitudinal Study of Ageing
(ELSA) and the Indonesian Family Life Survey (IFLS). We nd that older women lose more height
than men, and Indonesian elderly women in particular experience more height shrinkage. We then
explore associations between height, height loss and cognitive and physical functions. We nd that
height shrinkage is signicantly associated with lower cognition and reduced physical function for
elderly men and women in both countries, even when baseline height is controlled for. Extreme
height loss is also found to be a major predictor of mortality among the Indonesian elderly. Our
results, although purely correlational, indicate that height shrinkage in old age might serve as an
informative yet underutilized biomarker for later life health.
Chapter 4 examines whether there are causal spillovers of health status and behaviors among
married and cohabiting English older adults. Although spousal concordance in health is widely
3
documented, direct evidence on the existence of spousal spillovers is relatively scarce due to con-
foundedness from other non-causal factors: assortative mating and shared environment. Disentan-
gling the causal spillover from the other factors is important, because of its relevance for ecient
policy designs and comprehensive program evaluations in health domain. We utilize the data from
the English Longitudinal Study of Ageing (2000-2014) which is representative of the population
aged over 50 in UK and estimate the contemporaneous eect of spousal health status and behaviors
in a dynamic panel model which allows for path dependence of one?s own health. Using rst dif-
ferencing and System GMM to address endogeneity problems, we nd there is strong and positive
spousal concordance in health status and health behaviors. GMM results conrm that there exist
spillover eects between spouses for English couples in health status measured by self reported
health and depressive symptoms, as well as health-related behaviors including smoking, drinking,
exercising and social contacting. There exists certain gender dierences but not much. While wives
have larger impact on their husband's smoking likelihoods, men aect more than women as to their
wives' mental health and smoking intensity.
4
Chapter 2
Long-Term Health Eects of Malaria Eradication: Evidence from
India
2.1 Introduction
Malaria is a global disease, and it has particularly taken a toll in the developing countries. While
immense gains have been made over the last decade, in particular by the WHO, the battle is not
over. It has been especially dicult to contain malaria in endemic areas, and half of the world's
population is still at risk
1
. Children and pregnant women are especially vulnerable to the adverse
eects of malaria. Among children, it could have lifetime eects on human capital via direct adverse
eects on health, as well as indirect eects on educational attainment- for example children having to
miss school due to poor health, or not able to learn as much due to neuropsychological decits. The
severity of health eects of malaria also varies depending upon the prevalent parasite. Plasmodium
falciparum, predominant in Africa, is the most dangerous and potentially life-threatening. Others,
like P. vivax, while causing morbidity, generally do not lead to mortality.
With the advent of DDT in the mid 1940's, WHO coordinated malaria eradication campaigns
were launched throughout the world. Since DDT was cheap to produce, the developing word
1
http://www.who.int/gho/malaria/epidemic/deaths/en/
5
benetted in particular. India was one of those countries. Combined with the geographic variation
in malaria endemicity, the national malaria eradication program provides a quasi-experimental,
dierence-in-dierences, setup to study the causal eects of early life exposure to malaria on health
and economic outcomes. In this paper, I study the eects on cognition. I nd that cohorts born
after the eradication program, in areas with pre-eradication malaria endemicity greater by one unit,
experienced gains of 0.27 standard deviations on an average in the cognitive domain of episodic
memory, and gains of 0.18 standard deviations in working memory.
The rest of this paper is organized as follows. I begin with a review of the existing literature
on early life malaria exposure and later life outcomes in Section 2.2. Empirical strategy is outlined
in Section 2.3. Data sources and variables used are described in Section 2.4. Results are presented
and discussed in Section 2.5, Section 2.6 concludes.
2.2 Previous Literature
It is by now well documented and widely acknowledged that shocks to early life health can have
eects that last a lifetime (Currie and Almond (2011), Currie (2009)). There exist many kinds of
shocks, be it economic or epidemiological, that could aect various aspects of health, and hence
aect human capital accumulation and economic outcomes (see Currie and Vogl (2013) for a de-
tailed review focusing on developing countries). In this paper, I specically look at the eects of
early life exposure to a particular disease, malaria, on cognition. There are many biological path-
ways through which malaria aiction could impede child development, causing neuropsychological
as well as growth impairment (Holding and Kitsao-Wekulo, 2004). Hence, a growing literature in
economics is dedicated to studying the eects of exposure to malaria in early life and later life
6
outcomes. Since many malaria eradication programs around the world, around the 1950s, were af-
ter the (arguably exogenous) advent of DDT, these programs provide a useful setting for studying
causal eects of malaria, across many countries.
Lucas (2010) studied the eect of malaria eradication on female educational attainment in
Paraguay and Sri Lanka, and documents positive program eects in both countries. Bleakley
(2010) is also a multi-country analysis, covering the United States, Brazil, Colombia and Mexico.
He nds that cohorts who were exposed to the program, especially in erstwhile malarious areas, have
signicantly higher incomes as adults. Cutler et al. (2010) study the eects of malaria eradication
in India. They found no evidence for eects on educational attainment, but positive eects on
household consumption for men. Rawlings (2016) also nds some evidence of positive treatment
eects on schooling in Brazil, but points out that there are possibly heterogeneous eects by race.
Venkataramani (2012) nds positive treatment eects on
uid intelligence (Raven's Progressive
Matrices) in Mexico. He also nds that men in the post-eradication cohorts entered and exited
school at younger ages, even though there are no eects on total years of schooling. Chang et al.
(2014) studying the long term eects of a malaria eradication program in Taiwan, nd that exposure
to malaria around birth is associated with higher likelihood of cardiovascular disease and lower
cognition in old age.
To summarize, this literature has looked at the relationship between malaria eradication and
economic outcomes like productivity (wages) and consumption. The underlying mechanisms for
these associations are through higher human capital- which could be improved health, and more
education. Building upon Cutler et al. (2010) ndings of positive treatment eects on consump-
tion in India, this paper further explores as to the channels of these eects. Since the eradication
program was not associated with increases in educational attainment, the other possible channel
is improved health, which are also the rst order eects of a health intervention. To that end, I
7
nd some evidence, albeit in a limited sample, that malaria eradication in India was associated
with improvements in cognition, specically in the domain of memory. These ndings also con-
tribute to the broader literature on the causal eects of childhood conditions on cognitive skills
(Cunha and Heckman, 2008).
2.3 Basic Empirical Strategy
This paper follows the empirical approach used in Bleakley (2010), Cutler et al. (2010), Lucas
(2010) and Venkataramani (2012). These studies use a dierence-in-dierences (DID) setup, uti-
lizing geographic variation in malaria intensity prior to eradication campaigns. Humans are more
vulnerable to malaria and the ensuing morbidity during early life, that is, during fetal, infancy and
early childhood phases. Hence, following most of this literature, I also focus on malaria exposure
during early life. Pre-eradication malarial endemicity is interacted with exposure to the eradication
program. Individuals born in erstwhile highly malarious areas are expected to experience greater
improvements in health outcomes compared to the previous cohorts than those for whom the erad-
ication program did not change much due to low levels of malaria to begin with. To identify the
causal impact of early life malaria exposure on later life cognitive health, I estimate the following
regression, which is very similar to Cutler et al. (2010):
Y
icd
=(Post
c
) (Malaria
d
) +
d
+
c
+
pre
Z
icd
+
post
Z
icd
+
icd
; (2.1)
where Y
icd
are outcome variables for individual i, belonging to birth cohort c, in geographical
district d. Post is equal to 1 if the individual was born in or after 1960, equal to 0 if born in or
before 1952. Malaria
d
is the district level measure of pre-eradication malarial intensity.
d
are the
district xed eects,
c
are birth-cohort xed eects. Vector Z includes basic covariates like gender,
8
and those relevant in the Indian context- dummies for scheduled caste, scheduled tribe and religion,
as well as whether the individual's mother went to school (Thomas, Strauss and Henriques, 1991).
is the DID estimator of treatment eect. Standard errors are clustered at the district level.
Additional state xed eects and district-time trends are added as robustness checks.
While Cutler et al. (2010) estimated equation 2.1 separately for men and women, I run single
regressions for men and women combined since I have a much smaller sample size to work with. I
discuss eects by gender in section 2.5.3.
Next, I estimate the eects by cohort, because it helps understand the heterogeneity of treatment
eects over time.
Y
icd
=
X
c
c
Malaria
d
+
d
+
c
+
Z
icd
+
icd
(2.2)
where
c
is the DID estimators for cohort c, using which I explore visual patterns in trends over
time.
2.4 Data
2.4.1 Study on global AGEing and adult health (SAGE)
The individual-level data in this chapter come from the 2007/08 (Wave 1) Study on Global Ageing
and Adult Health (SAGE), India. SAGE is the second wave of the multi-country World Health Sur-
vey (2003), run by the World Health Organization. SAGE Wave 1 was conducted in six countries:
China, Ghana, India, Mexico, Russian Federation and South Africa, in nationally representative
samples. Data on health and well-being, including biomarkers, were collected for older adults aged
50 years and above, plus a smaller comparison sample of adults aged 18-49 years. SAGE wave 1
9
India was implemented in six states: Assam, Karnataka, Maharashtra, Rajasthan, Uttar Pradesh
and West Bengal, with a sample size of 11,230 (Arokiasamy et al., 2013).
I use data on health of individuals born up to fteen years before and fteen years after the
eradication eorts (1953-1959)
2
. I chose to include fteen years in the post eradication group in
order for the birth cohorts to be comparable to the program exposed cohorts in Cutler et al. (2010).
Fifteen years before the eradication campaign are chosen with the aim of achieving symmetry.
However, since SAGE is an ageing study, older cohorts are proportionally more- individuals born
before the program comprise 61.39% of the analysis sample.
Self reported birth year is missing for 61% of the respondents. For these, I generate birth year
as per their reported age. Since the data exhibits age heaping, I accordingly use ve-year birth
cohorts. Hence I generate three birth cohorts each for the pre and post-eradication groups. Using
ve-year birth cohorts also helps in achieving district-cohort cell sizes that are not too small. I
also keep respondents who have been living in the same region in which they were born, which is
97% of the sample. As per Cutler et al. (2010), I further restrict the analysis sample to individuals
in rural areas. Finally, I only keep respondents who attempted all the cognitive tests I use as
outcome variables. The resulting sample size is 3960. Education outcome variables used are a
dummy for whether attended any school, and years of schooling. I use logged values of height, as
an indicator of early-life health. I study three domains of cognition, constructing z-scores for each
test. The cognitive domain, memory, is measured by immediate word recall (10 words) and delayed
word recall. SAGE administers immediate word recall thrice, I use the score on the rst attempt,
since the next two attempts capture not only memory but also learning. Domain of retrieval is
measured by verbal
uency, where the respondent is asked to name as many animals as possible
2
(Cutler et al., 2010) dene eradication era from 1953-1961. As per the maps they provide, the entire country had
been brought under the program by 1960. Hence I include individuals born in 1960 or after since they were exposed
to, and hence are expected to benet from, the program. For earlier cohorts from 1953 to 1959, it is not clear as to
when the program began in their district.
10
within a minute. Finally, digit span, consisting of digits forward and digits backward is a measure
of attention and working memory. Respondent's score is the longest number of sequential digits
that they could accurately remember.
2.4.2 Pre-eradication malarial intensity
Similar to many other countries battling malaria as a public health problem, especially in the
developing world, major eradication eorts were launched in India after the advent of DDT. India's
National Malaria Control Program (NMCP) was launched in April 1953, and had been extended to
the entire country by 1960-61, albeit in rural areas only (see Cutler et al. (2010) for more details).
The geographic expansion of the eradication eorts was rapid, though the selection of areas early
on was not exogenous but based on malaria endemicity and agricultural capacity (Cutler et al.,
2010). I merge SAGE wave 1 India with district level measures of malaria endemicity before the
eradication campaign, as provided by (Cutler et al., 2010). They digitized a 1948 map (Figure 2.1)
from the Ministry of Health and Family Welfare, Government of India. The map classies areas
into six endemicity areas, as in 1948: (1) areas above 5000 feet; (2) non-malarious; (3) known
healthy plain areas, spleen rate under 10 percent; (4) variable endemicity associated with dry
tracts, potential epidemic areas; (5) known areas liable to fulminant epidemic diluvial malaria; (6)
moderate to high endemicity, fulminant epidemics unknown; and hyperendemicity of jungly hill
tracts and terai land. I follow their approach in constructing two measures of malaria intensity.
First is a continuous score which is the average of all values within a district, which they call the
malaria index. Second is a categorical classication. They classify areas 1 and 2 as non-malarious,
areas 3 and 4 as potential epidemic, areas 5 and 6 as malarious. Districts are then assigned a
classication as per its model category. Cutler et al. (2010) dropped bimodal districts in their
analysis when using these classications. In the SAGE sample, 23 districts are bimodal, hence
11
VOL. 2 NO. 2 79 CUTLER ET AL.: EARLY-LIFE MALARIA EXPOSURE AND ADULT OUTCOMES
The digitization procedure subdivided districts into polygons of roughly equal
size, so that some districts have more than one possible classifi cation. To aggregate
the polygons at the district level, we take two approaches. In the fi rst approach,
we average all polygon values (ranging from 1 to 6, as described above) within a
district to generate a continuous measure of endemicity, which we call the malaria
index. However, the effects of malaria eradication may be nonlinear, so our second
approach uses a categorical classifi cation of pre-eradication endemicity. To generate
this classifi cation, we fi rst map the original six-category endemicity measure into
a new three category variable. Categories 1 and 2 (as described above) are classi-
fi ed as non-malarious, categories 3 and 4 are classifi ed as potential epidemic, and
categories 5 and 6 are classifi ed as malarious.
12
We then categorize each district
by its modal polygon malaria category. Seventy-seven districts do not have unique
modes. For example, some mountainous districts in northern India have equal num-
bers of non-malarious, high-altitude polygons and malarious, low-altitude polygons.
12
In areas where malaria is endemic, individuals can acquire limited immunity over time through years of
continued exposure and multiple infections. The effects of malaria are therefore most pronounced in childhood
and youth, when individuals have not acquired immunity. Immunity may also be reduced during pregnancy. In
areas where malaria is epidemic, individuals may have little or no acquired immunity. In these areas, malaria can
affect children and adults, and can result in severe adverse health consequences.
Figure 2. Malaria Endemicity Map
Areas above 5,000 feet—Non-malarious
Known healthy plain areas—Spleen rates under 10 percent
Variable endemicity association with dry tracts—Potential epidemic areas
Known areas liable to fulminant epidemics
Moderate to high endemic rate—Fulminant epidemics unknown
Hyperendemicity—Jungly hill tracts and terai land
Malaria endemicity
Source: Cutler et al. (2010)
Figure 2.1: Malaria Endemicity Map
12
Table 2.1: Descriptive Statistics
Variable N mean sd min max
Age 3960 52.27 11.26 28 79
Female 3960 0.57 0.50 0 1
Whether attended school 3960 0.45 0.50 0 1
Years of schooling 3960 3.24 4.44 0 23
Height (cm) 3960 156.51 9.11 112 198.5
Immediate recall (1st attempt) 3960 4.51 1.52 0 10
Delayed recall 3960 4.75 1.93 0 10
Verbal
uency 3960 10.57 3.38 0 51
Digit forward 3960 4.39 1.02 3 9
Digit backward 3960 2.07 1.40 0 8
Treatment
Malaria index 3960 4.18 1.00 2.67 5.67
Non-malarious 3396 0.02 0.14 0 1
Potential epidemic 3396 0.45 0.50 0 1
Endemic malarious 3396 0.53 0.50 0 1
Bimodal 3960 0.14 0.35 0 1
Covariates
Whether mother attended school 3960 0.08 0.27 0 1
Hindu 3960 0.84 0.37 0 1
Muslim 3960 0.13 0.34 0 1
Scheduled caste 3960 0.21 0.40 0 1
Scheduled tribe 3960 0.08 0.27 0 1
Notes: Table re
ects sample characteristics for cohorts born between 1938-1952, and 1960-1974. Sample excludes
urban residents, and those born between 1953-1959.
these are dropped when using the categorical classication. Indicators of pre-eradication malaria
endemicity can be matched for all 133 districts of SAGE Wave 1, leaving only 0.04% observations
unmatched.
2.4.3 Descriptive Statistics
Table 2.1 provides summary statistics for the analysis sample. As discussed before, the sample
excludes urban residents, and those born between 1953-1959. The sample consists of middle-
aged and older adults. In the estimations, I use logged values of height. Regarding the cognition
variables, delayed recall scores are higher than immediate recall since SAGE administers immediate
13
word recall thrice. Digit span backward is a more challenging task, hence the scores are lower than
digit span forward. Since non-malarious areas constitute only 2% of the sample, I combine non-
malarious areas and potential epidemic as the low malaria group when estimating Equation 2.1.
2.5 Results
2.5.1 Dierence-in-dierences Analysis
Table 2.2 shows the results for specication 2.1 for any schooling and years of schooling. Columns
(1) and (4) show results for baseline specication, which are followed by robustness checks added
in further columns as state-xed eects, and district specic linear trends. Following Cutler et al.
(2010), I report treatment eects using a continuous malaria endemicity index, as well as categorical
classication, where sample size is smaller since bimodal districts are dropped. While the direction
of program treatment eects is positive for educational attainment, the results using continuous
index are statistically insignicant. The coecient on Post X endemic malarious is also not sta-
tistically signicant, and changes sign across specications for the any schooling variable. Overall,
these results suggest a similar pattern as in Cutler et al. (2010), who found no robust evidence
of positive treatment eects of malaria eradication on education, even in a much bigger sample
covering a majority of Indian states.
Next, I look at eects on health. I begin with height since it has long been used as a measure
of childhood health (Strauss and Thomas, 2008). As per the estimates reported in Table 2.3, there
are no treatment eects of malaria eradication on height. One plausible explanation is that since
among the malarial parasites, the less virulent Plasmodium vivax was more prevalent than the
severe P. falciparum, the gains might not be apparent for a stock measure like height.
14
Table 2.2: Childhood malaria exposure and Education
Dependent variable: Any school Years of schooling
(1) (2) (3) (4) (5) (6)
Continuous
Post x malaria index 0.00880 0.0295 0.0323 0.104 0.451*** 0.639
(0.0171) (0.0180) (0.0376) (0.140) (0.160) (0.417)
Observations 3,960 3,960 3,960 3,960 3,960 3,960
Categories
Post x endemic malarious 0.00691 0.0187 -0.00277 -0.0963 0.0847 0.284
(0.0344) (0.0438) (0.0885) (0.272) (0.337) (1.001)
Observations 3,396 3,396 3,396 3,396 3,396 3,396
State x post xed eects x x x x
District-specic linear trends x x
Notes: Standard error clustered at the district level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1.
Omitted group in categories is non-malarious or potential epidemic.
Table 2.3: Childhood malaria exposure and Height
Dependent variable: Log(height)
(1) (2) (3)
Continuous
Post x malaria index 0.000399 0.000799 -0.00282
(0.00121) (0.00161) (0.00315)
Observations 3,960 3,960 3,960
Categories
Post x endemic malarious 0.00341 0.00457 -0.00822
(0.00265) (0.00411) (0.00697)
Observations 3,396 3,396 3,396
State x post xed eects x x
District-specic linear trends x
Notes: Standard error clustered at the district level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1.
Omitted group in categories is non-malarious or potential epidemic.
15
Table 2.4: Childhood malaria exposure and Episodic Memory
Dependent variable: Immediate word recall Delayed word recall
(1) (2) (3) (4) (5) (6)
Continuous
Post x malaria index 0.0275 0.0994** 0.146 0.0391 0.110** 0.267***
(0.0410) (0.0422) (0.105) (0.0334) (0.0505) (0.0964)
Observations 3,960 3,960 3,960 3,960 3,960 3,960
Categories
Post x endemic malarious 0.0435 0.0653 -0.00791 0.00851 -0.0598 0.0190
(0.0824) (0.106) (0.249) (0.0727) (0.109) (0.239)
Observations 3,396 3,396 3,396 3,396 3,396 3,396
State x post xed eects x x x x
District-specic linear trends x x
Notes: Standard error clustered at the district level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1.
Omitted group in categories is non-malarious or potential epidemic.
Dependent variables are standardized test scores.
I study three domains of cognition- episodic memory, working memory or attention, and re-
trieval. Results for episodic memory are presented in Table 2.4. Looking at the continuous measure,
birth cohort exposure to malaria eradication was associated with increases in both immediate and
delayed recall, with the results stronger for delayed word recall and statistically signicant at the
5% level for model with additional state xed eects and at 1% level with district specic time
trends. The estimates increase in magnitude with additional state-level xed eects and even more
so with district-level time trends. This could be due to state level resources, like health infras-
tructure, being accounted for, which were omitted in the baseline model. With district-level time
trends, secular gains in health as well as resources are accounted for. In the most stringent model,
column (6), a one unit increase in the pre-eradication malaria index is associated with delayed recall
scores higher by 0.27 standard deviations. Coecient on Post X endemic malarious is insignicant
and also changes signs across specications. One potential explanation for this could be that while
eradications eorts did help in reducing prevalence, the disease was not eradicated, and this might
be driven by the endemic areas where it has been a more dicult battle.
16
Table 2.5: Childhood malaria exposure and Attention
Dependent variable: Digit span: digit forward Digit span: digit backward
(1) (2) (3) (4) (5) (6)
Continuous
Post x malaria index -0.0260 0.0492 0.0866 0.00311 0.101** 0.183**
(0.0272) (0.0344) (0.0656) (0.0317) (0.0488) (0.0771)
Observations 3,960 3,960 3,960 3,960 3,960 3,960
Categories
Post x endemic malarious -0.0821 -0.00920 -0.171 -0.0442 0.00630 0.0709
(0.0520) (0.0530) (0.148) (0.0674) (0.0904) (0.183)
Observations 3,396 3,396 3,396 3,396 3,396 3,396
State x post xed eects x x x x
District-specic linear trends x x
Notes: Standard error clustered at the district level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1.
Omitted group in categories is non-malarious or potential epidemic.
Dependent variables are standardized test scores.
Results for working memory, or attention, are presented in Table 2.5. The dependent variables
are z scores for performance on the digit span tests. While there are no signicant eects for
digit span forward, column (6) indicates that a one unit increase in malarial index is associated
with a gain of almost 0.18 standard deviations in digit span backward scores after the eradication
program. As was the case with episodic memory, we don't see any signicant associations with the
categorical classication of endemic-malarious. Lastly, results for retrieval are presented in Table
2.6. The dependent variable is standardized values for animal naming task. When looking at the
continuous index, there were gains of 0.17 standard deviations, but the coecient is only weakly
signicant. With the categorical classication, we do not see any signicant eects of malaria
eradication on this cognitive domain.
To summarize, while there is some evidence that malaria eradication led to gains in cognitive
domains of episodic and working memory, no strong eects are found for retrieval. A plausible
reason for this could be that the domain of retrieval, which relies on an individual's stock of
knowledge, is more liable to be built by education, as compared to memory. It has been shown by
17
Table 2.6: Childhood malaria exposure and Verbal
uency
Dependent variable: Animal naming
(1) (2) (3)
Continuous
Post x malaria index 0.0325 0.0811 0.174*
(0.0389) (0.0766) (0.0965)
Observations 3,960 3,960 3,960
Categories
Post x endemic malarious 0.0557 -0.0229 0.118
(0.0759) (0.109) (0.175)
Observations 3,396 3,396 3,396
State x post xed eects x x
District-specic linear trends x
Notes: Standard error clustered at the district level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1.
Omitted group in categories is non-malarious or potential epidemic.
Dependent variable is standardized test score.
both Cutler et al. (2010), and Table 2.2 in this paper, that malaria eradication did not lead to gains
in educational attainment in the Indian context. However, it is possible that eradication eorts did
lead to some improvements in the epidemiological environment, which shows as gains in memory
for individuals born later. These results are also consistent with Venkataramani (2012), who found
positive treatment eects of a similar malaria eradication program in Mexico on a measure of
general intelligence (12-item Raven Progressive Matrices test).
2.5.2 Cohort Analysis
In this section, I examine cohort-specic eects of malaria eradication. Going by the results from
Section 2.5.1, I restrict attention to the two measures of memory. Figure 2.2 presents estimates
of coecients
c
, from Equation 2.2, that is, the interaction of ve-year birth cohorts and pre-
eradication malaria index, for episodic memory. For immediate word recall, we do not see a trend
break in the cohort right after the eradication eorts, which is consistent with results from Table
2.4. For delayed word recall, there is a jump in the coecient estimate for the rst cohort from
18
Figure 2.2: Cohort-specic coecient plots for memory. Points are birth cohort specic coecients,
c
, from Equation 2.2, 95% condence intervals displayed around point estimates. 19
Figure 2.3: Cohort-specic coecient plots for attention. Points are birth cohort specic coe-
cients,
c
, from Equation 2.2, 95% condence intervals displayed around point estimates. 20
Table 2.7: Heterogeneity by Gender
Dependent variable: Immediate Delayed Digit span Digit span Verbal
recall recall forward backward
uency
(1) (2) (3) (4) (5)
Post x malaria index 0.128 0.248** 0.0703 0.155** 0.173*
(0.106) (0.0964) (0.0653) (0.0762) (0.0950)
Post x malaria index x female 0.0295* 0.0268* 0.0259* 0.0347** 0.00761
(0.0158) (0.0161) (0.0134) (0.0137) (0.0134)
Observations 3,960 3,960 3,960 3,960 3,960
Notes: Standard error clustered at the district level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1.
Dependent variable are standardized test scores.
Specications include state-post xed eects and district-specic time trends.
post-eradication group. Moreover, while the coecients for the subsequent two post-eradication
cohorts are lower, they are still slightly above the pre-eradication cohort estimates.
Figure 2.3 presents cohort eects for working memory. While coecient estimates for digit
span forward were declining with subsequent cohorts in the pre-eradication group, we see a jump
for the rst cohort in the post-eradication group. For digit span backwards, the jump is even
bigger. However, for both these tests, the subsequent cohorts again have lower magnitudes, and
wider condence intervals. These also raise a question as to how long-lasting the eects of such
eradication programs are.
2.5.3 Heterogeneity by Gender
Table 2.7 presents the results for interaction between the dierence-in-dierence estimator and
gender. For word recall and digit span tests, in areas with higher malaria intensity, gains are
higher for women by about 0.03 standard deviations. These gains are only weakly statistically
signicant, except for digit span backward where women experience higher gains by an average
of 22.4% compared to men. There are no dierential treatment eects among women for verbal
uency.
21
Table 2.8: Urban sample
Immediate Delayed Digit span Digit span
recall recall forward backward
Post x malaria index -0.177 -0.257 -0.182 0.0414
(0.227) (0.221) (0.221) (0.129)
Observations 1,347 1,333 1,361 1,336
Notes: Standard error clustered at the district level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1.
Dependent variable are standardized scores.
Sample consists of individuals from urban areas, born between 1938-1952, and 1960-1974.
Estimations include state x post xed eects and district-specic linear trends.
2.5.4 Robustness Check
The graphs in section 2.5.2 indicate clearly that improvements in the cognitive domain of memory
were following the malaria eradication eorts. Since these eorts were only in rural areas in the
1950's, the analysis sample was restricted accordingly. As a robustness check, I estimate equation
2.1 for the same birth cohorts, but in urban areas. The results are presented in Table 2.8. None
of the memory tests are signicantly associated with the interaction of pre-eradication malaria
endemicity and being born after the eradication program. These results provide some evidence
that the eects were indeed only in rural areas. It is to be noted though that the urban sample
sizes here are about a third of the rural.
I also check for sensitivity of the estimates by controlling for birth-year average rainfall and
temperature. These are reported in Appendix Table A1. The dierence-in-dierence estimates
from Tables 2.4 to 2.6 are robust to the inclusion of these additional weather covariates.
2.6 Conclusion
This paper contributes to the literature on the eects of early life health shocks on human capital
formation. Specically, it contributes to the literature on the early-life origins of later-life cognitive
function, as well as the long-lasting eects of childhood exposure to malaria. Utilizing a national
22
level, and plausibly exogenous, malaria eradication campaign in India during the 1950's, along with
estimates of pre-eradication malaria endemicity, I estimate the causal eects of malaria exposure
in early-life on domains of cognition.
I nd that cohorts born after the eradication program, in previously more malarious areas,
experience gains in memory tests scores. I nd no evidence of similar gains in the cognitive domain
of retrieval, or in height. Conrming results from some previous studies, I also nd no evidence for
eects on schooling. These suggest that since reduced exposure to malaria did not increase educa-
tional attainment, and hence perhaps the cognitive domain of retrieval/
uency which is shaped to
a greater extent by education. However, a cognitive domain which is relatively more independent
of education, memory, was improved, due to a safer disease environment post eradication.
There are several limitations to this study, mostly driven by the data. Given the small sample
size, I am unable to analyze the gender dierences, as well as birth-year specic eects, in greater
detail. While the eradication program was at the national level, the results are for a sample from
6 states in India (out of a total of 29 states and 7 union territories). Since improvements in health
and schooling are the mechanisms by which reduced malaria exposure is hypothesized to increase
productivity (and hence income and/or consumption), it would be interesting to explore the eects
on individual wages (and not household income), along with health.
23
Chapter 3
Height Shrinkage, Cognition and Health among the Elderly:
Comparisons across England and Indonesia
1
3.1 Introduction
There is a rich literature at the intersection of biodemography and economics, devoted to the study
of human height both as an outcome as well as a predictor of economic factors. Childhood and
adult height have been found to be aected by socioeconomic micro and macro level factors, like
parental investment and disease environment respectively. e.g. see Thomas, Strauss and Henriques
(1990), Thomas and Strauss (1992) and Deaton (2007). Adult height, considered to be a good
proxy for childhood height and hence childhood health, has been established as an important co-
variate for adult labor market outcomes and health, e.g. see Case and Paxson (2008b). However,
as Huang et al. (2013) explain, while height is a marker of early life health, when studying health
of the elderly, height shrinkage is a potential marker for later life health. They show that height-
health associations for the elderly ought to take into account age-related height loss since many of
the old-age health outcomes they look into are signicantly (and negatively) associated with height
1
This is a joint work with Urvashi Jain. This analysis uses data or information from the Harmonized ELSA dataset
and Codebook, Version E as of April 2017 developed by the Gateway to Global Aging Data. The development of the
Harmonized ELSA was funded by the National Institute on Aging (R01 AG030153, RC2 AG036619, 1R03AG043052).
For more information, please refer to www.g2aging.org.
24
shrinkage, in some cases the coecients on height loss being larger and/or signicant as compared
to height. This is especially important for studies using self reported height since there exists some
literature documenting that respondents overestimate current and recalled heights, and this over-
estimation increases with age (Cline et al., 1989; Briot et al., 2010; Cawley, Maclean and Kessler,
2017), leading to a systematic bias in the height-health coecients. There have been attempts
to estimate pre-shrinkage height for the elderly populations by using various limb lengths, since
limbs do not shrink with age the way height does
2
. In case of cross sectional data, the association
between limb length and height among a younger sample (which has not yet begun to experience
shrinkage) is estimated, and then this association is applied to measured limb length of the elderly
to predict their pre-shrinkage heights, as explained by Huang et al. (2013). However, this approach
needs to be careful about cohort eects. Later born cohorts are taller on average due to improv-
ing living conditions (Fogel, 1986; Steckel, 1995), hence it is to be assumed that the relationship
between height and limb lengths does not change at higher levels of height. Limb length can also
be used more directly, as instruments or proxies for pre-shrinkage height or early life health. e.g.
see Huang et al. (2008) and Sohn (2015). A simpler, and more direct, way to measure shrinkage
would be using long term panel data which measures pre-shrinkage height, i.e. when respondents
are younger, and continues to measure their height at regular intervals as they age. We follow the
latter approach, taking advantage of aging and household panel surveys in England and Indonesia
respectively. While the English population belongs to the taller end of human height spectrum, In-
donesians are among the shorter populations of the world
3
. We begin by documenting the extent of
height shrinkage among the elderly in these two very dierent populations. England and Indonesia
2
Using lower limb length, i.e. up to the ankle, might be marginally better than knee height since
attening of feet
could also decrease measured height.
3
https://www.disabled-world.com/artman/publish/height-chart.shtml
25
are at vastly dierent levels of economic development and social welfare support, and have racially
dierent populations.
We see that women across both England and Indonesia exhibit more height shrinkage than
men. This nding is consistent with the medical literature which has established that women
experience greater height loss (Sorkin, Muller and Andres, 1999), with one of the reasons being
that are more prone to osteoporosis and broken bones in old age than men (Cawthon, 2011).
Indonesian elderly women exhibit the most height shrinkage despite being the shortest, ranging
from average of 1.1 cm over seven years to 2.2 cm over seventeen years. We then look into age,
current SES and childhood SES gradients in height shrinkage. Consistent with the existing medical
literature, we nd that height loss increases with age in both England and Indonesia. However, we
do not nd any signicant associations with current or childhood SES variables. Since height loss is
associated with spinal compression and/or curving (see Krege et al. (2015)) as well as muscle loss,
we estimate the extent to which height loss aects cognition and physical functioning (measured
functional capacity, reported diculties). Finally, we also explore whether height loss before death
is a signicant predictor of mortality. There are important non-linearities in the association between
height-shrinkage and health. Among the English elderly, while we do not nd a continuous measure
of shrinkage to be signicantly associated with lower cognition (memory), extreme height loss, which
is dened as height loss above 3 cm, is negatively associated with word recall scores for men and
women. Extreme height loss is also associated with lower
uency scores for both men and women.
Among the Indonesian elderly, shrinkage is related to lower memory among women but not men, and
the association becomes stronger as we look into longer time spans. Interestingly, this association
is
ipped for genders when we look at measures of verbal
uency, with stronger results for men.
Since height-shrinkage is a marker for worsening health in old age, it is negatively associated with
physical function and increased disability in both English and Indonesian elderly. Finally, we nd
26
that height loss in the years preceding death is a signicant predictor of subsequent mortality
among the Indonesian elderly. To summarize, in most of the height-shrinkage-health associations
we study, shrinkage is found to be an important component of height-health associations that is
signicantly correlated with worse old-age health.
Given rapid population aging around the world, there is an increasing interest in later life health
among both academia and policy makers. The growing number of aging surveys over the past and
present decade, in both developing and developed countries, is testimony to this. We contribute to
this growing literature by decomposing the usual height-health associations into height-health and
shrinkage-health associations.
The rest of this paper is organized as follows. We begin with a review of the existing literature
on height, height shrinkage and later life health in Section 3.2. We describe the datasets used, in
Section 3.3. We then go on to document the extent of height shrinkage in England and Indonesia in
Section 3.4. Section 3.5 explores height-health and shrinkage-health relationships. Section 3.6 dis-
cusses strengths and limitations, along with suggestions for future research. We provide concluding
remarks in Section 3.7.
3.2 Existing Literature
Height has long been established as a key measure of the health component of human capital,
a component especially important for labor market outcomes in developing countries. Some of
the earliest population level evidence of the association between height and wages is from Brazil
(Thomas and Strauss, 1997). See Thomas and Strauss (1998) and Schultz and Strauss (2008) for
detailed reviews on the importance and measurement of health as a key component of human
capital. The next set of key contributions, linking adult height with cognition, health and labor
27
market outcomes are by Case and Paxson (2008a), who study these among the elderly in United
States, utilizing the Health and Retirement Study. Steckel (2009) provides a literature review
of these associations. Height in childhood and as adults has been found to be associated with
health and economic outcomes across many countries, using multitudes of population surveys.
E.g., height and later life cognition in Latin America (Maurer, 2010), England (Guven and Lee,
2013), and specically among the Irish elderly (Mosca and Wright, 2016). Relationship between
adult height and employment in Mexico has been recently explored in LaFave and Thomas (2016).
Given the various populations in which height and later life cognition has been studied, there
have also been comparisons across various countries, e.g. Guven and Lee (2015) across Europe,
Weir, Lay and Langa (2014) across United States, China and India, and McGovern (2014) who
compares China, Ghana, India, Mexico, Russia and South Africa.
Studying the relationship between height and later life health outcomes for the elderly needs
a more careful approach though since human beings lose height as they age, particularly after age
40 (Cline et al., 1989), with women being more prone to height loss (Sorkin, Muller and Andres,
1999). The main reason behind age related height loss is vertebral disc degeneration due to disc
dehydration and compression. Second important reason for height loss is muscle degeneration in
old age, or sarcopenia, which aects spinal curvature. Other reasons are the
attening of feet
arcs, and in extreme cases, vertebral compression fractures. One of the mechanisms/factors due
to which we expect a relation between shrinkage and health is osteoporosis- if someone has severe
osteoporosis, they experience more loss in height (Old and Calvert, 2004), osteoporosis also aects
functional status and/or ability. Vitamin D deciency increases the risk for osteoporosis and has
also been shown in a few studies to increase the risk for cognitive decline, e.g. see Llewellyn et al.
(2010). Height shrinkage is a marker for overall aging and frailty (Wannamethee et al., 2006). The
28
associations between shrinkage and cognition/physical-function are mostly driven by behaviors
aecting both- e.g., poor diet/nutrition, low levels of physical activity, smoking, alcohol intake.
Age related height loss is a commonly known biological fact, and recent height-cognition later life
research acknowledges and addresses this. E.g., Maurer (2010) instruments height with knee height,
Guven and Lee (2013) use the earliest height measurements in their analysis, Mosca and Wright
(2016) restrict their sample to age 51-70 to avoid the eects of shrinkage biasing the height co-
ecients. However, very little research has been done in the health-height economics literature
regarding how not only height but also height loss could be associated with health in old age. Two
notable and recent contributions to the study of height shrinkage in this literature are Huang et al.
(2013) and Fernihough and McGovern (2015).
Huang et al. (2013) employ techniques from the human biology literature by using the associ-
ation between limb lengths and height of a younger cohort (age 45-49) to estimate pre-shrinkage
height, and shrinkage among the Chinese elderly cohort (age 60 and above) in the CHARLS Base-
line cross section sample. They also elucidate shrinkage-health associations and clearly show how
shrinkage, along with pre-shrinkage height, is in fact a signicant predictor of cognition and phys-
ical function in old age. Fernihough and McGovern (2015) are able to calculate shrinkage using
repeated height measurements for the English elderly, utilizing Waves 0, 2 and 4 of ELSA. They
nd a strong age gradient, but no SES gradient in height shrinkage. They also argue that change in
physical function (peak
ow and grip strength) over time is a predictor of height loss. Finally, they
argue that using current height instead of pre-shrinkage height does not bias the health production
function estimations for the elderly.
There are two main contributions of this paper. First, we extend Fernihough and McGovern
(2015)'s work for the English elderly, by looking into how later life cognition is related to height-
loss. Second and more importantly, we document height loss and explore the shrinkage-health
29
relationships for another population- the Indonesian elderly. Indonesia is an interesting and valuable
comparison for many reasons. First, Indonesians are among the shorter populations in the world,
which indicates worse early life health (e.g. see Currie and Vogl (2013) for a detailed review about
early life environment and height). Second, since it is still a developing economy with constrained
resources, later life health for older adults is also expected to be worse than their elderly counterparts
in a developed country. Third, estimates from Indonesia are also more useful for comparisons with
China, shrinkage estimates for which have been obtained by Huang et al. (2013).
3.3 Data
In order to study height shrinkage among the elderly, longitudinal data with measured height is
preferred over panel data that rely on self-reported height since it has been found that reporting
error in height increases with age. For example, see Cawley, Maclean and Kessler (2017) who study
reporting errors in height and weight among the elderly in the widely used HRS sample. Hence,
objective height measurements over a relatively long time period, along with cognition and health
variables were the criteria for selecting the datasets for our study. Similarity in cognition and health
measures is also desirable to compare eects across countries.
3.3.1 England
We use the nationally representative English Longitudinal Study of Aging (ELSA)
4
, which began in
2002
5
. It is one of the international sister studies
6
of the Health and Retirement Study in the United
States. ELSA panel respondents are aged 50 and above, and their spouses are also interviewed,
4
https://discover.ukdataservice.ac.uk/series/?sn=200011
5
Respondents for the baseline wave were recruited from the Health Survey of England (HSE) conducted in 1998,
2000 and 2001. This is also referred to as ELSA Wave 0 by Fernihough and McGovern (2015).
6
https://hrs.isr.umich.edu/about/international-sister-studies
30
regardless of age. We use Harmonized versions
7
of Wave 2 (2004-05), Wave 4 (2008-09), Wave 6
(2012-13) and Wave 7 (2014-15). Anthropometric measures (including height, grip strength and
lung function) were collected by nurses every alternate wave. Cognition and self-reported health
measures are available for every ELSA wave. In addition, we also use Wave 3 (2006-07) which
asked retrospective life history questions. As per Cline et al. (1989), height loss becomes apparent
only after 60 years of age, a nding also conrmed by Fernihough and McGovern (2015)
8
. Hence,
we only keep respondents who are aged 60 and above as in Wave 6
9
. Attrition for men is higher
than for women across waves. 72.3% (7,664 observations) of the sample is of age 60 and above, as
in Wave 6. In our estimations, we only keep respondents for whom height measures are available
across all waves, which is for 47.77% of the 60-plus sample, giving us 3,661 observations, from which
we drop a few more outliers which show implausible height gain or loss across waves.
3.3.2 Indonesia
We use the Indonesian Family Life Survey (IFLS), which began in 1993 and is representative of 83%
of the Indonesian population
10
. We utilize four waves: IFLS-2 (1997) (Frankenberg and Thomas,
2000), IFLS-3 (2000-01) (Strauss et al., 2004), IFLS-4 (2007-08) (Strauss et al., 2009) and IFLS-5
(2014-15) (Strauss, Witoelar and Sikoki, 2016). For each of these waves, anthropometric variables
were measured by trained nurses. Cognition measures for the elderly were introduced in IFLS-4,
as was grip strength measurement. 13.04 % (9,343 observations) of the IFLS-5 sample is aged 60
and above. As we do for the ELSA sample, we keep respondents who were aged 60 and above
in Wave 5. We checked for mean shrinkage levels by age groups, and it is only after age 60 that
7
Gateway to Global Aging https://g2aging.org/
8
We checked for mean shrinkage levels by age groups in ELSA and conrmed this pattern.
9
This allows us to get pre-shrinkage heights for some respondents (those aged 52 and less in Wave 2), and for
others who had already begun to shrink, that is respondents who were above 60 in Wave 2, we can still compute the
extent of further shrinkage as they age.
10
http://www.rand.org/labor/FLS/IFLS.html
31
14-year and 17-year mean shrinkage exceeds 1 cm for women. Within the 60-plus sample, height
measurement is available across all waves for 29.02%, giving us a sample size of 2,711, from which
we drop outliers with extreme values of height change.
3.4 Shrinkage
Height shrinkage is dened simply as the dierence in height (in centimeters) from an older wave
and the most recent wave.
shrinkage
t
=height
t
height
c
(3.1)
shrinkage
t
is height loss over the course of t years. height
t
is height (cm) measured t years ago.
height
c
is current height, that is height measured in the most recent available wave. These are
Wave 6 in ELSA and Wave 5 in IFLS. Since height was measured every alternate wave in ELSA,
t2f4,8g. That is, shrinkage over four years is the change in height from Wave 4 (2008-09) to Wave 6
(2012-13), and shrinkage over eight years- from Wave 2 (2004-05) to Wave 6 (2012-13)
11
. Similarly,
for IFLS, t2f7,14,17g. That is, shrinkage over seven years is the change in height from Wave 4
(2007-08) to Wave 5 (2014-15), shrinkage over fourteen years is the height dierence from Wave
3 (2000-01) to Wave 5 (2014-15), and shrinkage over seventeen years is the height dierence from
Wave 2 (1996) to Wave 5 (2014-15). Note that positive (increasing) shrinkage
t
indicates height
loss, and negative shrinkage
t
indicates height gain.
11
Due to data access restrictions, we could not link ELSA respondents from Wave 1 onwards to ELSA Wave 0,
which would have given us even longer term measurements of shrinkage.
32
3.4.1 Extent of Shrinkage
We drop a few respondents for whom either measured height is less than 100cm, or for whom height
gain or loss over subsequent waves exceeds 10cm
12
. See Appendix A for details on data cleaning.
Table 3.1: Shrinkage (in cm)
VARIABLES N mean sd min max N mean sd min max
ELSA Male ELSA Female
Current Height (cm) 1608 172.42 6.96 152.1 202 2035 158.76 6.55 134.5 179.8
Height 4 years ago 1608 173.09 6.93 151.8 201.6 2035 159.47 6.44 136.5 181
Height 8 years ago 1608 173.63 6.79 153 201.4 2035 160.1 6.33 137.6 181.4
Shrinkage over 4 years 1608 0.67 1.68 -7.2 8.10 2035 0.7 1.65 -7.2 9.4
Shrinkage over 8 years 1608 1.2 1.73 -5.9 9.5 2035 1.33 1.79 -9.8 10
IFLS Male IFLS Female
Current Height (cm) 1162 158.3 6.28 126.4 179 1382 146.42 5.73 129.2 166.8
Height 7 years ago 1162 158.9 6.15 126.6 178.9 1382 147.56 5.50 130.9 168.6
Height 14 years ago 1162 159.77 5.95 127.2 178.1 1382 148.62 5.37 132.1 169.5
Height 17 years ago 1162 159.78 5.88 126.5 178.6 1382 148.64 5.31 132 168.7
Shrinkage over 7 years 1162 0.6 1.81 -8.6 10 1382 1.14 1.99 -9.7 10
Shrinkage over 14 years 1162 1.47 1.85 -9 9.3 1382 2.2 2.05 -8.6 9.8
Shrinkage over 17 years 1162 1.48 2.02 -9.2 8.8 1382 2.22 2.27 -8.9 9.9
Note: ELSA Waves 1 to 6, and IFLS Waves 2 to 5.
Table 3.1 presents the extent of shrinkage in ELSA and IFLS. Height shrinkage is evident from
the decrease in mean height over time in both the samples, with current height being the lowest.
On average, men lost 0.67 cm height over the course of four years (baseline average height was
173.1 cm), and 1.2 cm over eight years (baseline average height was 173.63 cm) in ELSA which
is a mean height loss of 0.7%. These are consistent with what Fernihough and McGovern (2015)
found for previous waves of ELSA. Women lost 0.7 cm height over the course of four years (baseline
average height was 159.47 cm) and 1.33 cm over eight years (baseline average height was 160.1
cm) which is a mean height loss of 0.85%. Even though Indonesians are considerably shorter,
12
Wannamethee et al. (2006) look at how height loss over 20 years predicts cardiovascular disease and mortality
in a sample of English men. They drop respondents for whom height dierence over time indicates an increase in
height of more than 5 cm. Huang et al. (2013) drop any height observations less than 120 cm.
33
they experience greater shrinkage than their English elderly counterparts. Height loss over seven
years among Indonesian elderly men was 0.6 cm (baseline height of 158.9 cm). Height loss over
fourteen and seventeen years for men was close, 1.47 and 1.48 cm respectively (baseline height of
159.77 cm). Indonesian elderly women experience the greatest shrinkage- 1.14 cm over seven years
(baseline height of 147.56 cm), approximately 2.20 and 2.22 cm over fourteen and seventeen years
respectively. That is, elderly women in Indonesia lose 1.6% of their height over the course of 17
years. Note that elderly men in Indonesia lose slightly more height over 7 years than the men in
England lose over 8 years. In comparison, mean height shrinkage estimates for the Chinese elderly
computed by Huang et al. (2013) were 3.3 cm for men and 3.8 cm for women, over a ten-year and
higher time period
13
.
Gender dierences in height shrinkage are also evident from Figures 3.1 and 3.2 for ELSA,
Figures 3.3 and 3.4 for IFLS, where we plot the kernel densities of height shrinkage over time.
All of these distributions have a positive mean, which indicates height loss in the population, and
becomes more prominent with greater time lags. That is, the distribution shifts towards the right
for values of change in height measured over a longer time period. In all four graphs, the density
plot for women's height loss is further right than the plot for men. In Figures 3.5 and 3.6, we plot
mean height shrinkage by gender and ve-year age categories. We clearly see that gender dierences
are more stark in Indonesia than in England.
13
We looked at change in height from CHARLS Wave 1 (2011) to Wave 4 (2015), which gave us mean height loss
over 4 years at 0.49 cm for men and 0.84 cm for women, 60 years of age and above as in Wave 4.
34
Figure 3.1: Shrinkage over 4 years in ELSA
Figure 3.2: Shrinkage over 8 years in ELSA
35
Figure 3.3: Shrinkage over 7 years in IFLS
Figure 3.4: Shrinkage over 17 years in IFLS
36
Figure 3.5: Shrinkage over 4 and 8 years by gender and age in ELSA
Figure 3.6: Shrinkage over 14 and 17 years by gender and age in IFLS
37
3.4.2 (Erroneous) Height Gain
Kernel density plots of shrinkage (Figures 3.1, 3.2, 3.3 & 3.4 ) indicate that for some respondents,
shrinkage is less than zero, that is, they appear to have been gaining height. In the ELSA male
sample, 29.73% indicate height gain over four years, and 20.71% over eight years. Corresponding
percentages in ELSA female sample are 29.58% and 18.03%. 28.17% of Indonesian male sample
show increase in height over seven years, 12.33% over fourteen years, and 16.01% over seventeen
years. In the female sample, 19.68%, 6.20% and 9.23% show height increase over seven, fourteen
and seventeen years respectively. Our percentages of height gain in ELSA are well within what
(Fernihough and McGovern, 2015) found, and IFLS percentages of height gain are even lower.
Height gain is biologically implausible among the age groups we study. Fernihough and McGovern
(2015) discuss this issue and oer two explanations. First attributes to the nature of population
surveys- measurement error by the interviewer. Second attributes to human biology- individual
height
uctuations within a day, which could be up to 1 cm. We are all taller in the morning
than in the evening. Hence, if the nurses measured respondents' heights earlier in the day in the
most recent wave, and towards the latter part of the day in the previous wave, this could lead to a
reported increase in measured height over time. We check for correlation between the time of height
measurement and measured height in the IFLS sample. A simple regression of measured height in
Wave 5 on the time of biomarker interview (continuous) as well as on various dummy measures
indicating time of biomarker interview to be in the latter part of the day does not indicate any
systematic pattern. Therefore, the main reason for supposed gain in height for some respondents
is measurement error which can be due to many reasons- the respondents not standing as erect
as they could, tilting their head, hair-dos for women and caps for men etc. It is not intuitively
obvious whether the expected measurement error is classical. Older respondents with a stooping
38
posture would have greater diculty in standing erect. It might be more dicult for interviewers
to measure extremely tall respondents. While this measurement error shows up as height gain
over time, it could also lead to misclassication regarding height loss. Hence, we trim the data
allowing for equal extent of height gain and loss (10 cm), even though the medical literature
would allow us to cap height gain more strictly than height loss in these age groups, for example
Wannamethee et al. (2006) drop respondents with height gain of more than 5 cm.
We now summarize height loss across these four demographic groups. First, Indonesian elderly
women experience greater height shrinkage than English women over a comparable period. Indone-
sian elderly men on the other hand experience lesser height loss on an average than English men.
Second, women in both England and Indonesia exhibit higher shrinkage than the men in their re-
spective countries, a nding consistent with the medical literature which has established that post
menopausal women are more prone to osteoporosis and fractures as they experience declining bone
density due to lower estrogen levels. Third, Indonesian men are on an average, shorter than, and
experience more height shrinkage, than English women. Finally, Indonesian elderly women seem to
bear a double burden- they have the lowest heights and experience highest shrinkage. This points
to lower early life nutrition and health, as well as worse later life health, in a developing country
compared to a developed one.
3.4.3 Covariates of Shrinkage
Next, we examine SES covariates for height shrinkage, using the specication below, separately for
men and women.
shrinkage
i;t;tw
= +height
i;tw
+
Age
i;t
+SES
i;t
+
i;t
(3.2)
39
Shrinkage
i;t;tw
is change in height as measured between a previous wave which was w years ago,
and the most recent wave,t. Age
i;t
are ve year dummies on the basis of current age. w2f4,8g for
ELSA respondents, w2f7,14,17g for IFLS respondents. SES
i;t
are current, i.e. in the most recent
available wave, socio-economic variables which include respondent's education, marital status, race
(white or non-white in England) or ethnicity in Indonesia
14
, self reported health in childhood,
household characteristics in childhood, and mother's current or age at death. Most of the SES
controls are time-invariant, given the age group, except for marital status. We also include lagged
per capita monthly household consumption corresponding to the time gap w.
Tables 3.2 and 3.3 present results for ELSA and IFLS respectively. As expected, age is the most
signicant predictor of height shrinkage, especially after 70 years among both English and Indone-
sian elderly; shrinkage is highest among the oldest group, aged 80 and above. Height shrinkage over
the course of w years is signicantly associated with height w years ago. That is, taller respondents
exhibit more shrinkage. Since even lagged measures of height could already be subject to shrinkage,
especially for oldest respondents, and hence not capturing their maximum, pre-shrinkage, height,
we also estimate equation (2) replacing height
i;tw
with limb length (which do not shrink with
age) for the IFLS sample. However, we do not nd either upper arm length or knee length to be
signicant covariates of height shrinkage (Appendix Table B1).
We do not nd any education gradients for shrinkage in neither the ELSA nor the IFLS, which
is in contrast to Huang et al. (2013) who nd signicant correlations of shrinkage with education
levels in the Chinese sample. We also do not see any race gradients for shrinkage in the ELSA
sample. We do see that Tionghoa (ethnic Chinese migrants) women experience lower shrinkage in
Indonesia. We don't nd any of the childhood SES variables to be signicantly associated with
height shrinkage in either ELSA or IFLS sample.
14
Dummy variables for Java, Sunda, Bali, Batak, Bugis, Tionghoa, Madura, Sasak, Minang
40
Table 3.2: Shrinkage (in cm) Covariates in ELSA
(1) (2) (3) (4)
Male sample Female sample
Shrinkage Shrinkage Shrinkage Shrinkage
VARIABLES over 4 yrs over 8 yrs over 4 yrs over 8 yrs
Height 4/8 years ago 0.0372*** 0.0229*** 0.0260*** 0.0198**
(0.00783) (0.00778) (0.00764) (0.00788)
Age: \65 to 69" 0.0844 0.320** 0.222 0.340**
(0.150) (0.147) (0.135) (0.139)
Age: \70 to 74" 0.565*** 0.731*** 0.240 0.406***
(0.163) (0.159) (0.149) (0.153)
Age: \75 to 79" 0.298* 0.696*** 0.463*** 1.018***
(0.171) (0.167) (0.156) (0.159)
Age: \80+" 0.786*** 1.467*** 0.873*** 1.458***
(0.182) (0.176) (0.176) (0.179)
High-school graduate 0.00708 0.0137 -0.199 -0.274**
(0.152) (0.149) (0.122) (0.125)
Some college -0.193 -0.0406 -0.0277 -0.0819
(0.140) (0.137) (0.130) (0.133)
College and above -0.0546 0.00141 -0.0642 -0.0378
(0.156) (0.152) (0.165) (0.170)
Partnered -0.179 -0.0476 -0.459 -0.296
(0.335) (0.328) (0.288) (0.294)
Separated -0.480 -0.792 0.736 0.101
(0.842) (0.821) (0.572) (0.585)
Divorced -0.0198 -0.218 0.208 0.103
(0.242) (0.234) (0.158) (0.161)
Widowed -0.0803 0.0178 0.0277 -0.0497
(0.203) (0.197) (0.128) (0.130)
Never married 0.494* 0.315 -0.121 -0.0103
(0.287) (0.280) (0.274) (0.280)
Non-white -0.316 -0.262 0.344 0.639
(0.428) (0.418) (0.542) (0.554)
Reported Poor Childhood health -0.282* -0.186 -0.196 -0.0908
(0.170) (0.166) (0.134) (0.137)
Number of bedrooms of residence -0.107* -0.0473 -0.0626 0.00107
lived in when aged 10 (0.0640) (0.0626) (0.0547) (0.0559)
Mother's age: \60 to 64" 0.353 0.452 0.679** 0.640**
(0.330) (0.323) (0.299) (0.306)
Mother's age: \65 to 69" 0.269 0.474 0.165 0.391
(0.308) (0.300) (0.266) (0.272)
Mother's age: \70 to 74" 0.358 0.367 0.330 0.470*
(0.289) (0.283) (0.252) (0.258)
Mother's age: \75 to 74" 0.402 0.385 0.359 0.513**
(0.279) (0.273) (0.242) (0.248)
Per capita Monthly Household -0.0896 -0.0395
Consumption 4 years ago (0.0920) (0.0573)
Per capita Monthly Household -0.116 -0.0301
Consumption 8 years ago (0.0964) (0.0686)
Constant -5.460*** -2.871** -3.588*** -2.638**
(1.441) (1.425) (1.303) (1.337)
Observations 1,042 1,042 1,166 1,166
R-squared 0.055 0.087 0.049 0.090
Note: Omitted category in age: \60 to 64", in education: \No School", in marital
status: \Unmarried". Robust standard errrors in parenthesis. *** p < 0:01, **
p< 0:05, * p< 0:1.
41
Table 3.3: Shrinkage (in cm) Covariates in IFLS
(1) (2) (3) (4) (5) (6)
Male sample Female sample
Shrinkage Shrinkage Shrinkage Shrinkage Shrinkage Shrinkage
VARIABLES over 7 yrs over 14 yrs over 17 yrs over 7 yrs over 14 yrs over 17 yrs
Height 7/14/17 years ago 0.0433*** 0.0170* 0.0204 0.0320** 0.0259** 0.0326**
(0.0106) (0.00981) (0.0121) (0.0134) (0.0112) (0.0132)
Age: \65 to 69" -0.0489 -0.0971 0.132 0.114 0.492*** 0.509***
(0.186) (0.188) (0.177) (0.0983) (0.118) (0.153)
Age: \70 to 74" 0.0931 0.630*** 0.578*** 0.543*** 0.917*** 1.104***
(0.140) (0.154) (0.145) (0.179) (0.167) (0.177)
Age: \75 to 79" 0.663** 1.140*** 1.492*** 0.291 1.250*** 1.380***
(0.247) (0.251) (0.276) (0.204) (0.180) (0.248)
Age: \80 to 100" 0.588* 1.181*** 1.784*** 0.596* 1.740*** 2.037***
(0.338) (0.350) (0.406) (0.329) (0.372) (0.440)
Urban -0.0996 -0.267 -0.339*** 0.349** 0.264 0.347
(0.0920) (0.163) (0.104) (0.153) (0.160) (0.218)
Education: Primary 0.0388 0.393 0.251 0.125 -0.0749 0.0738
(0.289) (0.287) (0.273) (0.139) (0.156) (0.165)
Education: Junior -0.380 0.195 -0.0104 0.0678 -0.218 -0.231
(0.368) (0.365) (0.362) (0.214) (0.222) (0.173)
Education: Senior & above -0.127 0.109 0.0515 -0.0500 -0.267 -0.0410
(0.339) (0.295) (0.300) (0.233) (0.223) (0.256)
Married 1.314*** 1.471*** 1.025 0.976 0.457 0.518
(0.298) (0.361) (0.828) (0.792) (0.566) (0.491)
Separated -0.811 -0.276 0.639 1.778** -0.719 -1.574***
(1.440) (1.493) (0.955) (0.839) (0.590) (0.569)
Divorced 2.054*** 1.241 1.430 1.029 0.751 0.447
(0.626) (0.891) (0.937) (0.684) (0.608) (0.586)
Widow 1.771*** 1.693*** 1.349* 1.208 0.687 0.744
(0.367) (0.301) (0.774) (0.774) (0.567) (0.510)
Java -0.0500 -0.0195 0.0784 -0.129 0.308 -0.00486
(0.161) (0.180) (0.208) (0.170) (0.231) (0.224)
Sunda 0.290 0.120 0.0822 0.272 0.516 0.325
(0.233) (0.145) (0.339) (0.217) (0.340) (0.319)
Bali -0.443 -0.276 -0.00160 -0.705*** -0.00184 -0.238
(0.265) (0.259) (0.375) (0.196) (0.227) (0.231)
Batak 0.0103 -0.103 -0.0368 -0.0690 0.227 -0.309
(0.459) (0.652) (0.709) (0.333) (0.395) (0.452)
Bugis -0.0826 0.138 -0.350 -0.366 0.643** -0.316
(0.193) (0.197) (0.420) (0.340) (0.313) (0.314)
Tionghoa -2.583** -0.266 -0.920 1.164 -0.522** -0.901**
(1.233) (1.050) (0.901) (1.078) (0.226) (0.419)
Madura -0.0952 -0.241 0.0471 0.730 0.101 -0.482
(0.343) (0.278) (0.419) (0.571) (0.345) (0.423)
Sasak 0.342 0.263 0.456 -0.104 -0.0363 -0.454
(0.294) (0.212) (0.381) (0.259) (0.293) (0.281)
Minang -0.379 -0.439 -0.537* 0.0854 0.313 -0.255
(0.248) (0.285) (0.314) (0.180) (0.345) (0.237)
Reported poor childhood health 0.0532 -0.107 0.0693 -0.170* -0.130 -0.211*
(0.136) (0.0848) (0.107) (0.0940) (0.147) (0.112)
Number of rooms of residence 0.0350 0.0121 0.0293 -0.00120 -0.00672 0.000833
lived in when aged 12 (0.0299) (0.0458) (0.0388) (0.00432) (0.00542) (0.00371)
Whether electricity in house 0.0839 0.217 0.497** -0.123 -0.0302 -0.0280
lived in age 12 (0.148) (0.328) (0.186) (0.205) (0.276) (0.261)
Whether lived with biological 0.0599 0.123 0.282 0.142 0.127 0.184
mother at age 12 (0.204) (0.212) (0.252) (0.164) (0.138) (0.221)
Per capita monthly household 0.0119 -0.0674 0.00827 -0.0244 0.0800 0.201**
consumption 7/14/17 years ago (0.0325) (0.0729) (0.0753) (0.101) (0.104) (0.0804)
Constant -7.902*** -2.388 -3.556 -4.772 -4.042* -6.296***
(1.772) (2.028) (2.563) (3.057) (2.136) (1.998)
Observations 970 970 970 1,089 1,089 1,089
R-squared 0.071 0.082 0.102 0.052 0.100 0.112
Omitted category in age:\60 to 64", in education:\No School", in marital status:\Unmarried", in Whether
lived with biological mother at age 12:\Not applicable". Estimations include district (Kabupaten) xed
eects, standard errors clustered at district level in parenthesis. *** p< 0:01, ** p< 0:05, * p< 0:1.
42
3.5 Shrinkage and Health
We now explore whether shrinkage is signicantly associated with later life health, even after con-
trolling for baseline height (proxy for childhood health). Among the health variables, we look at
current measures of cognition, measured physical strength and reported functional diculties. Ta-
bles 3.4 and 3.5 present the summary statistics for health measures in ELSA and IFLS respectively.
Mean cognition levels are similar for men and women in England. In Indonesia, women have lower
cognition scores. What is striking is the dierence in cognition across the two countries. Both word
recall and verbal
uency in England are only slightly less than twice than in Indonesia. One of
the major potential explanations for these dierences is education levels- the causal eect of which
has been showed in many contexts, e.g. see Glymour et al. (2008), Banks and Mazzonna (2012)
and Huang and Zhou (2013). A similar pattern exists for other measures of physical health- no
remarkable gender dierences in ELSA, but measures for women are worse than men in IFLS; and
the English elderly are overall healthier than the Indonesian elderly.
Table 3.4: Descriptive Statistics: ELSA
VARIABLES N mean sd min max N mean sd min max
Male Female
Current Age 1608 71.28 7.63 60 90 2035 71.43 7.68 60 90
Word Recall 1608 5.07 1.73 0 10 2035 5.43 1.82 0 10
Fluency (wave 7) 1442 20.95 7.19 0 62 1822 20.41 7.04 0 57
Any ADL/IADL diculty 1608 0.2 0.40 0 1 2035 0.26 0.44 0 1
Some Diculty-Lift/carry 10lbs 1608 0.13 0.34 0 1 2034 0.31 0.46 0 1
Diculty walking 100 yards 1608 0.11 0.31 0 1 2034 0.13 0.34 0 1
Diculty stooping, kneeling or crouching 1608 0.33 0.47 0 1 2034 0.46 0.50 0 1
Diculty getting up from chair 1608 0.21 0.40 0 1 2034 0.29 0.45 0 1
ADL Diculties 1608 0.26 0.72 0 5 2035 0.29 0.75 0 5
IADL Diculties 1608 0.13 0.54 0 5 2035 0.16 0.52 0 5
Lung function (FEV) 1419 2.67 0.79 0 4.9 1752 1.86 0.57 0 3.5
Lung function (HTPEF) 1419 7.61 2.37 0 17.7 1752 5.18 1.63 0 10.6
Grip Strength 1582 35.25 9.18 0 63.7 1969 20.86 6.22 0 40.7
Self Reported Health as Poor 1606 0.27 0.45 0 1 2035 0.27 0.44 0 1
Note: ELSA Waves 1 to 7.
43
Table 3.5: Descriptive Statistics: IFLS
VARIABLES N mean sd min max N mean sd min max
Male Female
Current Age 1162 68.62 7.35 60 95 1382 68.48 6.97 60 94
Word Recall 980 2.99 1.44 1 8 1106 2.76 1.50 1 8.5
Fluency 979 13.95 5.16 0 42 1097 11.99 4.81 0 31
Any ADL/IADL diculty 1155 0.38 0.49 0 1 1373 0.4 0.49 0 1
Some di- carrying heavy load 1155 0.31 0.46 0 1 1373 0.51 0.50 0 1
Some di- walking 1 km 1155 0.3 0.46 0 1 1373 0.48 0.50 0 1
Some di- bowing, squatting or kneeling 1155 0.17 0.37 0 1 1373 0.27 0.44 0 1
Some di- getting up from chair 1155 0.06 0.24 0 1 1373 0.1 0.30 0 1
ADL Diculties 777 0.19 0.66 0 5 1162 0.25 0.76 0 5
IADL Diculties 1155 1.03 1.74 0 6 1373 0.99 1.75 0 6
Lung Function (PEF) 1126 306.37 114.03 0 733.3 1293 201.76 70.40 0 396.7
Grip Strength 1097 24.82 7.32 0 48 1276 15.8 5.35 0 47
Self Reported Health as Poor 1155 0.37 0.48 0 1 1373 0.41 0.49 0 1
Dead in Wave 5 1643 0.23 0.42 0 1 1918 0.16 0.37 0 1
Note: IFLS Waves 2 to 5.
Following Huang et al. (2013), we rst regress each of these variables on current height in
equation (3), which has been standard in the literature. We then regress current health on height
w years ago and height loss over time t-w and t, as in equation (4).
health
i;t
= +height
i;t
+
Z
i
+
i;t
(3.3)
health
i;t
= +
0
height
i;tw
+
1
shrinkage
i;t;tw
+
Z
i
+
i;t
(3.4)
We also replace the continuous measure of height-shrinkage into four groups, as classied by
Wannamethee et al. (2006), dummies for whether shrinkage
i;t;tw
falls under categories < 1cm,
1-1.9cm, 2-2.9cm, or 3 cm, to explore non-linearities
15
.
health
i;t
= +
0
height
i;tw
+
3
X
c=0
!
c
(shrinkage
c
)
i;t;tw
+
Z
i
+
i;t
(3.5)
As much as each panel data allows us to, we can look at short term and long term shrinkage
eects on health. That is, ELSA allows us to explore height shrinkage over four and eight years,
15
Using dummy variables also reduces bias due to measurement error in the continuous shrinkage variable.
44
IFLS allows to look at height shrinkage over seven, fourteen and seventeen years. Age categories
and SES variables are all used as controls (Z
i
) as was done in estimating equation (2). Coecients
in equations (4) and (5) do not have a causal interpretation since there is no exogenous variation
in shrinkage, because of omitted variables (like health behaviors), and potential reverse causality.
We expect an association since the mechanisms related to height shrinkage, like osteoporosis and
specic health behaviors, also aect the health outcomes we look at. Note that we limit our sample
to respondents for whom both 4-year and 8-year shrinkage can be calculated in ELSA, and to those
for whom 7-year, 14-year and 17-year shrinkage can be calculated in IFLS. Since respondents are
more likely to drop out of the the sample over subsequent panel rounds due to declining health
and/or mortality given the age groups we are looking at, the coecients would be underestimated
since we would expect the healthier of respondents to survive over the years. We conrm this
in a later section where we do in fact nd that extreme height shrinkage in the years preceding
death is signicantly associated with mortality. The strength of these associations might also be
underestimated in our sample since we allow equal maximum values for height gain and loss.
3.5.1 Shrinkage and Cognition
In order to compare estimates across England and Indonesia, we rst look at word recall which is
a measure of short term memory (McArdle, Smith and Willis, 2011). This is a mean of immediate
and delayed word recall, hence ranges from 0 to 10
16
. Results for word recall from ELSA are
presented in Table 3.6. First, coecients on current height and lagged height (by 8 years) are of
similar magnitudes. These have been interpreted in the literature as the eects of early life health
on later life cognition. Hence, both current and lagged height capture early life health to a similar
extent. Next, we look at the coecients on shrinkage. While shrinkage over both 4 and 8 years
16
Other variables measuring cognition in ELSA- verbal
uency and prospective memory task, while asked from
Wave 1 to 5, are not available for Wave 6. We use verbal
uency from Wave 7.
45
is associated with reduced memory scores, the relationship is not signicant. However, extreme
height-shrinkage over eight years, i.e. over 3 cm, is signicantly associated with lower immediate
word recall for both men and women. For men, extreme height shrinkage over 8 years is associated
with almost one less word recalled
17
.
Next, we look at verbal
uency (measured by number of animals named within a given time
period) in Table 3.7. For English elderly men, both current and lagged heights are signicantly
related with higher scores, and the strength of association is stronger for lagged height measures.
Only extreme height shrinkage over 8 years is signicantly associated with lower
uency. For
women, verbal
uency is not related with measures of height to begin with. We do see a negative
signicant association extreme 4 year shrinkage. While we would expect the associations to be
stronger for a longer time measure of shrinkage since height loss occurs over a longer term, losing
height in a short time span could indicate more rapidly worsening health.
As in ELSA, we look at word recall in the Indonesian sample. IFLS-5 introduced more cognition
variables, allowing us to also look at verbal
uency, measured similarly as in ELSA. In the IFLS
estimations, we also include Kabupaten (district) xed eects, and cluster standard errors at the
same level. Among the sample of Indonesian elderly men, we do not see word recall scores to be
related with either height (current or lagged) or height-shrinkage (Table 3.8). For women, however,
we do see a signicant negative association with height-shrinkage over 17 years. Verbal
uency is
positively related with current height for both genders (Table 3.9), but the association doesn't hold
for lagged height measures. Among the men that we also see a signicant and negative association
with 14-year and 17-year shrinkage, which in turn is seemingly driven by measures of extreme
height shrinkage. For the women, while shrinkage coecients are negative, there are no signicant
non-linearities.
17
0.5 less on memory score translates into one word less on either immediate or delayed recall
46
Table 3.6: Height Shrinkage and Cognition (Word Recall) in ELSA
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Panel A: Male sample Panel B: Female sample
Dependent Variable: Word Recall 4 years 8 years 4 years 8 years
Current Height 0.0125* 0.0135*
(0.00708) (0.00716)
Height 4/8 yrs ago 0.0129* 0.0120* 0.0123* 0.0134* 0.0124* 0.0119 0.0115 0.0121*
(0.00715) (0.00712) (0.00721) (0.00718) (0.00727) (0.00727) (0.00730) (0.00732)
Height loss over 4/8 yrs -0.0213 -0.0362 -0.0326 -0.0420
(0.0283) (0.0289) (0.0281) (0.0274)
Height loss: 1-1.9cm 0.284** 0.140 -0.0674 -0.0745
(0.121) (0.116) (0.109) (0.111)
Height loss: 2-2.9cm -0.256* -0.0714 -0.0997 -0.217*
(0.147) (0.138) (0.147) (0.130)
Height loss: 3cm + -0.234 -0.433*** -0.00195 -0.267*
(0.183) (0.149) (0.190) (0.150)
Observations 1,040 1,040 1,040 1,040 1,040 1,165 1,165 1,165 1,165 1,165
R-squared 0.238 0.234 0.243 0.234 0.243 0.229 0.226 0.225 0.228 0.229
Note: Robust standard errors in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. Age, Education, SES controls suppressed.
47
Table 3.7: Height Shrinkage and Fluency (Animal Naming) in ELSA
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Panel A: Male sample Panel B: Female sample
Dependent Variable: Fluency 4 years 8 years 4 years 8 years
Current Height 0.0637** 0.0313
(0.0323) (0.0308)
Height 4/8 yrs ago 0.0682** 0.0674** 0.0660** 0.0677** 0.0239 0.0237 0.0241 0.0223
(0.0325) (0.0325) (0.0328) (0.0327) (0.0311) (0.0310) (0.0314) (0.0315)
Height loss over 4/8 yrs -0.0575 -0.147 -0.245** -0.169
(0.133) (0.135) (0.122) (0.120)
Height loss: 1-1.9cm 0.812 0.925* -0.464 -0.497
(0.557) (0.535) (0.463) (0.475)
Height loss: 2-2.9cm -0.528 0.487 -0.531 0.111
(0.679) (0.630) (0.638) (0.562)
Height loss: 3cm + -1.164 -1.339* -2.335*** -1.058
(0.854) (0.693) (0.834) (0.649)
Observations 947 947 947 947 947 1,059 1,059 1,059 1,059 1,059
R-squared 0.170 0.166 0.171 0.166 0.174 0.186 0.189 0.192 0.186 0.188
Note: Robust standard errors in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. Age, Education, SES controls suppressed.
48
Table 3.8: Height Shrinkage and Cognition (Word Recall) in IFLS
(1) (2) (3) (4) (5) (6) (7)
Panel A: Male sample
Dependent Variable: Word Recall 7 years 14 years 17 years
Current Height 0.00474
(0.00808)
Height 7/14/17 yrs ago 0.00355 0.00286 0.00279 0.00229 0.00214 0.00147
(0.00752) (0.00731) (0.00862) (0.00858) (0.00868) (0.00888)
Height loss over 7/14/17 yrs -0.0149 -0.0209 -0.0237
(0.0258) (0.0264) (0.0217)
Height loss: 1-1.9cm 0.0254 -0.0907 0.0837
(0.171) (0.102) (0.159)
Height loss: 2-2.9cm 0.165 0.0763 -0.0132
(0.187) (0.115) (0.126)
Height loss: 3cm + -0.156 -0.160 0.0147
(0.178) (0.214) (0.182)
Observations 821 823 823 823 823 823 823
R-squared 0.144 0.145 0.146 0.145 0.147 0.145 0.145
(1) (2) (3) (4) (5) (6) (7)
Panel B: Female sample
Dependent Variable: Word Recall 7 years 14 years 17 years
Current Height 0.0112
(0.00898)
Height 7/14/17 yrs ago 0.0127 0.0130 0.0106 0.0110 0.00558 0.00672
(0.00947) (0.00945) (0.00903) (0.00886) (0.00973) (0.00951)
Height loss over 7/14/17 yrs -0.00486 -0.0309 -0.0513***
(0.0242) (0.0279) (0.0172)
Height loss: 1-1.9cm -0.114 -0.236* 0.0160
(0.124) (0.130) (0.130)
Height loss: 2-2.9cm -0.172 -0.236 -0.320***
(0.145) (0.180) (0.111)
Height loss: 3cm + -0.0158 -0.251* -0.236*
(0.184) (0.148) (0.124)
Observations 828 828 828 828 828 828 828
R-squared 0.211 0.210 0.212 0.207 0.210 0.213 0.217
Note: Robust standard errors in parentheses.*** p < 0:01, ** p < 0:05, * p < 0:1. Age, Education, SES controls
suppressed. IFLS results control for district xed eects, cluster errors at the same level.
49
Table 3.9: Height Shrinkage and Fluency (Animal Naming) in IFLS
(1) (2) (3) (4) (5) (6) (7)
Panel A: Male sample
Dependent Variable: Fluency 7 years 14 years 17 years
Current Height 0.0467**
(0.0202)
Height 7/14/17 yrs ago 0.0477** 0.0494** 0.0332 0.0320 0.0297 0.0303
(0.0214) (0.0213) (0.0215) (0.0215) (0.0242) (0.0248)
Height loss over 7/14/17 yrs -0.0175 -0.152** -0.239***
(0.0811) (0.0586) (0.0573)
Height loss: 1-1.9cm -0.0124 -0.177 -0.325
(0.390) (0.394) (0.376)
Height loss: 2-2.9cm -0.0943 -0.498 -1.296***
(0.595) (0.489) (0.333)
Height loss: 3cm + -0.742 -1.092*** -1.051***
(0.584) (0.387) (0.347)
Observations 966 969 969 969 969 969 969
R-squared 0.162 0.164 0.165 0.167 0.169 0.169 0.171
(1) (2) (3) (4) (5) (6) (7)
Panel B: Female sample
Dependent Variable: Fluency 7 years 14 years 17 years
Current Height 0.0365*
(0.0201)
Height 7/14/17 yrs ago 0.0412* 0.0397 0.0324 0.0308 0.0305 0.0296
(0.0232) (0.0235) (0.0204) (0.0201) (0.0212) (0.0212)
Height loss over 7/14/17 yrs -0.0189 -0.0749 -0.0866**
(0.0586) (0.0531) (0.0345)
Height loss: 1-1.9cm 0.0446 0.0784 -0.0243
(0.205) (0.302) (0.383)
Height loss: 2-2.9cm 0.505 0.208 -0.187
(0.356) (0.376) (0.225)
Height loss: 3cm + -0.0175 -0.119 -0.263
(0.422) (0.340) (0.284)
Observations 1,082 1,083 1,083 1,083 1,083 1,083 1,083
R-squared 0.219 0.217 0.218 0.219 0.219 0.222 0.221
Note: Robust standard errors clustered at district level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. Age,
Education, SES controls suppressed. IFLS results control for district xed eects.
50
3.5.2 Shrinkage and Physical Health
We rst look at two measured (or objective) health variables: lung function, and grip strength of the
dominant hand which are predictors of old age physical health, including disability (Rantanen et al.,
1999). While ELSA provides multiple measures of lung function, we use Forced Expiratory Volume
since this was measured in all the relevant waves. In IFLS-5, lung capacity was measured using a
Vitalograph peak
ow meter for all respondents 9 years of age and above. The variable is Peak
Expiratory Flow (L/min), which measures how fast the respondent can exhale. We use an average
of the three measurements. For grip strength in both ELSA and IFLS, we use an average of the
second and third measurements.
Current height, as well as longer term height-shrinkage (that is eight years in ELSA sample) are
signicantly associated with lung function for both genders (Table 3.10). Note that the absolute
value of height-shrinkage coecients are of similar magnitudes as those of the height coecients.
Patterns for grip strength are very similar (Table 3.11). Hence, taller people have better lung
function and higher grip strength, but height-shrinkage is even more strongly related with reduced
physical functioning capacity. Patterns for lung function and grip strength are very similar in the
IFLS samples (Tables 3.12 & 3.13). The extent to which shrinkage is associated with reduced grip
strength is higher for women than men since they have lower grip strength to begin with.
Finally, we also look at manifested functional diculties, measured by reported disability, and
diculty in activities requiring muscle strength. We use a binary measure of disability- an indicator
variable which equals one if the respondent reports some or more diculty with any of the 5 ADL's
or 6 IADL's. In both the ELSA and IFLS sample (Tables 3.14 & 3.16), this measure is very
weakly (and negatively) related with height. Height-shrinkage is, however, signicantly associated
with higher disability among English men, and both Indonesian men and women. We also explore
51
these separately for ADL and IADL Indexes, which measure the number of activities with which
respondents report some or more diculty (Appendix Tables B2 to B5). The patterns are similar
to those for the binary disability variable, hence we do not report these in the main tables. We
also look at diculty in kneeling and related actions, which indicates decreased muscle strength.
We do not nd signicant relation with height-shrinkage in the ELSA sample (Table 3.15). For
Indonesian men, we see that long term shrinkage is associated with higher likelihood of reported
diculty with squatting or kneeling (Table 3.17). For Indonesian women, we see an even stronger
association, for all measures of shrinkage. The coecients are from 1.65 to 1.85 percentage points,
which is of non trivial magnitude given that 27% women report this diculty
18
.
Based on the above discussion, we can see that height-shrinkage is systematically associated
with old-age specic health measures, even when these measures are often not correlated with
height.
3.5.3 Shrinkage and Mortality
Since height-shrinkage is a marker for worsening later life health, an association conrmed by Sec-
tions 3.5.1 and 3.5.2, we explore whether it is also a predictor of mortality in these age groups. There
is some evidence about this association in the medical literature. For example, Wannamethee et al.
(2006) look at how height loss over 20 years predicts cardiovascular disease and mortality in a sam-
ple of English men. They nd total mortality risk to be higher for men who experience height loss
of 3 cm or more. Auyeung et al. (2010) nd that all cause mortality is higher for men experiencing
height loss exceeding 2 cm over 4 years, but not for women, in a sample of Chinese elderly aged 65
and above. Masunari et al. (2012) nd a similar association in a Japanese sample- while they also
18
We also checked for diculty with carrying a heavy load, or diculty with getting up from a chair after sitting-
also measures of old age muscle strength. We found very similar patterns, hence those results are reported in Appendix
Tables B6, B7, B8 and B9.
52
dene marked height loss as above 2 cm, their sample is younger at baseline, starting from age 47.
Hillier et al. (2012) nd that height loss exceeding 5 cm over 15 years is associated with increased
mortality among American women aged 65 years and above.
We are able to investigate this hypothesis in the IFLS sample, by using whether the respondent
was reported dead in wave 5 as the outcome variable. We keep respondents who were of age 53 or
higher in Wave 4, and either are or would have been of age 60 or higher by Wave 5. We use change
in height from wave 4 (which would be the last available measurement of height for respondents
dead in wave 5) and wave 3 or wave 2, giving measures of height-shrinkage over 7 or 10 years,
going back in time 7 years or less depending on when the Respondent died after 2008 and before
2014-15. Note that this sample size would be bigger than the sample we used for studying health
associations
19
.
While a continuous measure of height-shrinkage is not signicantly correlated with mortality,
height loss exceeding 3 cm over 7 as well as 10 years is a positive and signicant predictor for
both men and women (Table 3.18). Coecient of 7-year pre death shrinkage is higher than that
on 10-year shrinkage for men, suggesting that extreme height loss over a shorter time period is
an indicator of more rapidly worsening health. The magnitude of the coecients is large, at 5
percentage points or higher, when looking at mean mortality- 23% for men and 16% for women. In
these estimations, we also control for whether the respondent self reported his/her health as poor
in the previous waves, which has been found to be an important predictor of mortality among the
elderly (Benjamins et al., 2004).
19
since health measurements are available for respondents alive in Wave 5, a subset of the mortality sample.
53
Table 3.10: Height Shrinkage and Lung Function (Forced Expiratory Volume) in ELSA
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Panel A: Male sample Panel B: Female sample
Dependent Variable: FEV 4 years 8 years 4 years 8 years
Current Height 0.0338*** 0.0230***
(0.00332) (0.00256)
Height 4/8 yrs ago 0.0344*** 0.0338*** 0.0339*** 0.0341*** 0.0234*** 0.0235*** 0.0229*** 0.0228***
(0.00334) (0.00334) (0.00337) (0.00337) (0.00259) (0.00258) (0.00261) (0.00261)
Height loss over 4/8 yrs -0.00771 -0.0342** -0.0131 -0.0222**
(0.0138) (0.0139) (0.0104) (0.0102)
Height loss: 1-1.9cm 0.102* -0.0289 0.0263 0.0137
(0.0572) (0.0552) (0.0386) (0.0394)
Height loss: 2-2.9cm -0.0696 -0.0469 -0.0942* -0.0320
(0.0704) (0.0660) (0.0528) (0.0465)
Height loss: 3cm + -0.0561 -0.235*** -0.0692 -0.126**
(0.0865) (0.0714) (0.0723) (0.0552)
Observations 850 850 850 850 850 893 893 893 893 893
R-squared 0.340 0.343 0.347 0.340 0.344 0.295 0.293 0.296 0.293 0.294
Note: Robust standard errors in parentheses. *** p< 0:01, **p< 0:05, *p< 0:1. Age, Education, SES controls suppressed. FEV indicates the volume
of air exhaled under forced conditions.
54
Table 3.11: Height Shrinkage and Grip Strength in ELSA
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Panel A: Male sample Panel B: Female sample
Dependent Variable: Grip Strength 4 years 8 years 4 years 8 years
Current Height 0.334*** 0.198***
(0.0366) (0.0261)
Height 4/8 yrs ago 0.339*** 0.346*** 0.335*** 0.338*** 0.198*** 0.198*** 0.198*** 0.200***
(0.0367) (0.0367) (0.0371) (0.0370) (0.0264) (0.0263) (0.0265) (0.0265)
Height loss over 4/8 yrs -0.154 -0.405*** -0.177* -0.214**
(0.145) (0.147) (0.104) (0.102)
Height loss: 1-1.9cm -0.682 -0.364 -0.591 0.108
(0.620) (0.594) (0.392) (0.406)
Height loss: 2-2.9cm -1.558** -1.053 -0.995* -0.742
(0.744) (0.704) (0.537) (0.465)
Height loss: 3cm + -1.584* -2.650*** -1.065 -1.530***
(0.939) (0.763) (0.695) (0.555)
Observations 1,007 1,007 1,007 1,007 1,007 1,097 1,097 1,097 1,097 1,097
R-squared 0.315 0.316 0.320 0.316 0.319 0.239 0.240 0.242 0.237 0.242
Note: Robust standard errors in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. Age, Education, SES controls suppressed.
55
Table 3.12: Height Shrinkage and Lung Function (Peak Expiratory Flow) in IFLS
(1) (2) (3) (4) (5) (6) (7)
Panel A: Male sample
Dependent Variable: PEF 7 years 14 years 17 years
Current Height 3.263***
(0.657)
Height 7/14/17 yrs ago 3.230*** 3.277*** 3.029*** 3.002*** 3.297*** 3.301***
(0.696) (0.709) (0.620) (0.615) (0.633) (0.614)
Height loss over 7/14/17 yrs: -2.882* -3.802 -3.204*
(1.567) (2.322) (1.862)
Height loss: 1-1.9cm -5.688 -1.538 9.796
(9.880) (8.172) (9.035)
Height loss: 2-2.9cm -28.54** -12.62 -13.57
(13.10) (8.503) (10.36)
Height loss: 3cm + -25.29** -18.25 -21.68**
(12.35) (12.21) (9.975)
Observations 947 949 949 949 949 949 949
R-squared 0.228 0.230 0.234 0.234 0.234 0.229 0.236
(1) (2) (3) (4) (5) (6) (7)
Panel B: Female sample
Dependent Variable: PEF 7 years 14 years 17 years
Current Height 1.924***
(0.458)
Height 7/14/17 yrs ago 1.855*** 1.854*** 1.735*** 1.724*** 1.729*** 1.755***
(0.500) (0.505) (0.423) (0.413) (0.485) (0.477)
Height loss over 7/14/17 yrs: -2.773** -4.048*** -3.612***
(1.041) (1.303) (0.967)
Height loss: 1-1.9cm -5.862 -5.053 -4.375
(4.590) (6.096) (4.935)
Height loss: 2-2.9cm -11.69** -15.07* -12.12*
(5.207) (7.892) (6.727)
Height loss: 3cm + -23.64*** -20.80*** -22.11***
(5.525) (5.696) (5.011)
Observations 1,014 1,015 1,015 1,015 1,015 1,015 1,015
R-squared 0.171 0.173 0.177 0.173 0.173 0.172 0.174
Note: Robust standard errors clustered at district level in parentheses. *** p < 0:01, ** p < 0:05, * p < 0:1.
Age, Education, SES controls suppressed. IFLS results control for district xed eects. PEF is the highest forced
expiratory
ow measured with a peak
ow meter.
56
Table 3.13: Height Shrinkage and Grip Strength in IFLS
(1) (2) (3) (4) (5) (6) (7)
Panel A: Male sample
Dependent Variable: Grip Strength 7 years 14 years 17 years
Current Height 0.433***
(0.0408)
Height 7/14/17 yrs ago 0.442*** 0.441*** 0.430*** 0.422*** 0.437*** 0.436***
(0.0417) (0.0435) (0.0407) (0.0397) (0.0408) (0.0402)
Height loss over 7/14/17 yrs: -0.194* -0.365** -0.370***
(0.0982) (0.145) (0.111)
Height loss: 1-1.9cm 0.212 0.662 -0.0272
(0.390) (0.481) (0.518)
Height loss: 2-2.9cm -0.373 0.338 -1.172**
(0.911) (0.617) (0.472)
Height loss: 3cm + -1.991** -2.518*** -2.150***
(0.978) (0.606) (0.716)
Observations 890 893 893 893 893 893 893
R-squared 0.296 0.300 0.302 0.297 0.308 0.296 0.299
(1) (2) (3) (4) (5) (6) (7)
Panel B: Female sample
Dependent Variable: Grip Strength 7 years 14 years 17 years
Current Height 0.241***
(0.0283)
Height 7/14/17 yrs ago 0.246*** 0.245*** 0.234*** 0.232*** 0.235*** 0.236***
(0.0279) (0.0273) (0.0286) (0.0286) (0.0304) (0.0296)
Height loss over 7/14/17 yrs: -0.226*** -0.329*** -0.325***
(0.0738) (0.0932) (0.0716)
Height loss: 1-1.9cm -0.284 -0.0156 0.463
(0.357) (0.493) (0.453)
Height loss: 2-2.9cm -1.253** -0.223 -0.685**
(0.461) (0.381) (0.315)
Height loss: 3cm + -1.629*** -1.720*** -1.479***
(0.551) (0.451) (0.380)
Observations 1,016 1,017 1,017 1,017 1,017 1,017 1,017
R-squared 0.214 0.209 0.214 0.211 0.215 0.210 0.213
Note: Robust standard errors clustered at district level in parentheses. *** p < 0:01, ** p < 0:05, * p < 0:1. Age,
Education, SES controls suppressed. IFLS results control for district xed eects.
57
Table 3.14: Height Shrinkage and Disability in ELSA
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Panel A: Male sample Panel B: Female sample
Dep Var: Any diculty in ADL/IADL 4 years 8 years 4 years 8 years
Current Height -0.00176 -0.00325
(0.00186) (0.00199)
Height 4/8 yrs ago -0.00142 -0.00117 -0.00153 -0.00170 -0.00234 -0.00227 -0.00233 -0.00220
(0.00187) (0.00187) (0.00188) (0.00188) (0.00201) (0.00201) (0.00202) (0.00202)
Height loss over 4/8 yrs 0.00512 0.00964 0.0234*** 0.0213***
(0.00739) (0.00755) (0.00775) (0.00758)
Height loss: 1-1.9cm -0.0333 -0.0416 0.0212 -0.0648**
(0.0319) (0.0304) (0.0301) (0.0307)
Height loss: 2-2.9cm -0.0221 -0.00505 0.104** 0.0239
(0.0385) (0.0362) (0.0407) (0.0357)
Height loss: 3cm + 0.0377 0.0916** 0.132** 0.137***
(0.0479) (0.0391) (0.0521) (0.0412)
Observations 1,042 1,042 1,042 1,042 1,042 1,166 1,166 1,166 1,166 1,166
R-squared 0.053 0.054 0.055 0.055 0.063 0.086 0.092 0.094 0.091 0.102
Note: Robust standard errors in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. Age, Education, SES controls suppressed.
Table 3.15: Height Shrinkage and Diculty in Kneeling, Stooping or Crouching in ELSA
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Panel A: Male sample Panel B: Female sample
Dep Var: Any diculty in kneeling etc. 4 years 8 years 4 years 8 years
Current Height 0.00165 0.00463**
(0.00215) (0.00230)
Height 4/8 yrs ago 0.00220 0.00201 0.00240 0.00230 0.00577** 0.00604*** 0.00543** 0.00534**
(0.00216) (0.00216) (0.00218) (0.00217) (0.00232) (0.00232) (0.00234) (0.00235)
Height loss over 4/8 yrs 0.00601 0.0118 0.0191** 0.0122
(0.00854) (0.00873) (0.00894) (0.00877)
Height loss: 1-1.9cm -0.00470 -0.0141 0.0667* -0.00311
(0.0369) (0.0352) (0.0347) (0.0357)
Height loss: 2-2.9cm 0.0186 -0.0433 0.0491 0.0541
(0.0445) (0.0420) (0.0470) (0.0416)
Height loss: 3cm + 0.107* 0.0981** 0.0441 0.0434
(0.0554) (0.0453) (0.0602) (0.0480)
Observations 1,042 1,042 1,042 1,042 1,042 1,166 1,166 1,166 1,166 1,166
R-squared 0.068 0.069 0.072 0.069 0.075 0.066 0.076 0.076 0.070 0.071
Note: Robust standard errors in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. Age, Education, SES controls suppressed.
58
Table 3.16: Height Shrinkage and Disability in IFLS
(1) (2) (3) (4) (5) (6) (7)
Panel A: Male sample
Dep Var: Any diculty in ADL/IADL 7 years 14 years 17 years
Current Height -0.00422
(0.00275)
Height 7/14/17 yrs ago -0.00337 -0.00314 -0.00295 -0.00259 -0.00244 -0.00250
(0.00260) (0.00267) (0.00272) (0.00268) (0.00255) (0.00256)
Height loss over 7/14/17 yrs: 0.0181** 0.0198** 0.0269***
(0.00845) (0.00900) (0.00842)
1-1.9cm -0.0204 -0.0519 0.0626
(0.0445) (0.0452) (0.0392)
2-2.9cm 0.0450 0.0333 0.0776*
(0.0782) (0.0418) (0.0416)
3cm + 0.179** 0.131** 0.158***
(0.0767) (0.0544) (0.0397)
Observations 967 970 970 970 970 970 970
R-squared 0.054 0.057 0.061 0.058 0.066 0.065 0.066
(1) (2) (3) (4) (5) (6) (7)
Panel B: Female sample
Dep Var: Any diculty in ADL/IADL 7 years 14 years 17 years
Current Height -0.00502**
(0.00238)
Height 7/14/17 yrs ago -0.00399 -0.00389 -0.00368 -0.00347 -0.00398 -0.00376
(0.00262) (0.00261) (0.00245) (0.00243) (0.00252) (0.00254)
Height loss over 7/14/17 yrs: 0.0117 0.0149** 0.0110*
(0.00964) (0.00646) (0.00587)
Height loss: 1-1.9cm 0.0139 -0.0156 0.0135
(0.0436) (0.0330) (0.0414)
Height loss: 2-2.9cm -0.0300 -0.0285 -0.0199
(0.0522) (0.0409) (0.0375)
Height loss: 3cm + 0.161** 0.0479 0.0666*
(0.0635) (0.0388) (0.0386)
Observations 1,088 1,089 1,089 1,089 1,089 1,089 1,089
R-squared 0.094 0.093 0.102 0.093 0.093 0.092 0.094
Note: Robust standard errors clustered at district level in parentheses. *** p < 0:01, ** p < 0:05, * p < 0:1. Age,
Education, SES controls suppressed. IFLS results control for district xed eects.
59
Table 3.17: Height Shrinkage and Diculty in Kneeling, Bowing or Squatting in IFLS
(1) (2) (3) (4) (5) (6) (7)
Panel A: Male sample
Dep Var: Any diculty in kneeling etc. 7 years 14 years 17 years
Current Height -0.00195*
(0.00105)
Height 7/14/17 yrs ago -0.00200* -0.00198* -0.00111 -0.000943 -0.00139 -0.00133
(0.00104) (0.00103) (0.00108) (0.00109) (0.00118) (0.00115)
Height loss over 7/14/17 yrs: -0.000330 0.00956 0.00909*
(0.00418) (0.00665) (0.00533)
Height loss: 1-1.9cm 0.0154 0.00259 0.00755
(0.0271) (0.0251) (0.0276)
Height loss: 2-2.9cm -0.0555* -0.00265 0.0378
(0.0315) (0.0300) (0.0363)
Height loss: 3cm + 0.0190 0.0579 0.0156
(0.0493) (0.0437) (0.0368)
Observations 967 970 970 970 970 970 970
R-squared 0.081 0.086 0.088 0.086 0.087 0.084 0.083
(1) (2) (3) (4) (5) (6) (7)
Panel B: Female sample
Dep Var: Any diculty in kneeling etc. 7 years 14 years 17 years
Current Height -0.000518
(0.00230)
Height 7/14/17 yrs ago 0.000992 0.00106 0.00143 0.00128 0.00176 0.00166
(0.00245) (0.00245) (0.00238) (0.00234) (0.00220) (0.00216)
Height loss over 7/14/17 yrs: 0.0135** 0.0185*** 0.0165**
(0.00547) (0.00528) (0.00757)
Height loss: 1-1.9cm 0.0713** 0.0476** 0.0251
(0.0305) (0.0209) (0.0269)
Height loss: 2-2.9cm 0.0741* 0.0609** 0.0354
(0.0403) (0.0278) (0.0271)
Height loss: 3cm + 0.0662 0.144*** 0.133***
(0.0394) (0.0309) (0.0400)
Observations 1,088 1,089 1,089 1,089 1,089 1,089 1,089
R-squared 0.093 0.095 0.099 0.099 0.106 0.098 0.105
Note: Robust standard errors clustered at district level in parentheses. *** p < 0:01, ** p < 0:05, * p < 0:1. Age,
Education, SES controls suppressed. IFLS results control for district xed eects.
60
Table 3.18: Height Shrinkage and Mortality in IFLS
(1) (2) (3) (4) (5)
Panel A: Male sample
Dependent Variable: Whether dead in wave 5 7 years 10 years
Last measured Height (Wave 4) -0.00113
(0.00167)
Height 7/10 yrs before death (Wave 3/2) -0.000952 -0.000915 -0.00127 -0.00133
(0.00173) (0.00173) (0.00175) (0.00174)
Height loss over 7/10 yrs before death 0.00747 0.00271
(Wave 3/2- Wave 4) (0.00545) (0.00475)
Height loss: 1-1.9cm -0.00175 -0.0385
(0.0229) (0.0236)
Height loss: 2-2.9cm -0.0150 -0.0353
(0.0300) (0.0300)
Height loss: 3cm + 0.118*** 0.0588*
(0.0386) (0.0342)
Observations 1,642 1,642 1,642 1,642 1,642
(1) (2) (3) (4) (5)
Panel B: Female sample
Dependent Variable: Whether dead in wave 5 7 years 10 years
Last measured Height (Wave 4) -0.00171
(0.00148)
Height 7/10 yrs before death (Wave 3/2) -0.00118 -0.00119 -0.00117 -0.00119
(0.00155) (0.00155) (0.00156) (0.00156)
Height loss over 7/10 yrs before death 0.00586 0.00504
(Wave 3/2- Wave 4) (0.00393) (0.00365)
Height loss: 1-1.9cm -0.00769 0.00934
(0.0198) (0.0202)
Height loss: 2-2.9cm 0.0274 0.0290
(0.0243) (0.0246)
Height loss: 3cm + 0.0537* 0.0552**
(0.0289) (0.0266)
Observations 1,918 1,918 1,918 1,918 1,918
Note: Robust standard errors clustered at district level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1.
Age, Education, SES controls suppressed. IFLS results control for district xed eects.
61
3.6 Strengths and Limitations
We get estimates of shrinkage based on repeated measurements of height for the same individuals,
these would be more accurate than using self-reported height. Most of the medical literature on
shrinkage has used clinical data or community level samples, whereas we use nationally represen-
tative samples. We are also able to demonstrate the negative association between shrinkage and
health using both measured and self reported variables.
Both the extent of height loss and the strength of shrinkage-health associations are underes-
timated in our study. We are also unable to provide causal estimates of the eect of shrinkage
on worsening old-age health. It is also unclear whether early life health environment would aect
shrinkage only through aecting height, or whether there exist other biological pathways. Twin
studies tracking respondents up until their old age would be very informative. It would be interest-
ing to explore whether identical twins experience dissimilar levels of shrinkage, and if so, whether
there are any SES or behavioral factors. Another potentially helpful approach would be to study
height-shrinkage in a scenario where health insurance coverage is introduced in a phased manner
while respondents are in their middle age so that insurance only aects later life health.
3.7 Conclusion
In this paper, we document and compare the magnitude of height-shrinkage over time and across
English and Indonesian elderly. We nd that women suer from greater height loss, which is
consistent with the medical literature. We also nd that Indonesian elderly, even though shorter
to begin with, lose more height than the elderly in England. This suggests worse early life as
well as later life health. While we nd strong age gradients in height-shrinkage, we do not nd
62
any relation with current or childhood SES variables. We then decompose the standard height-
health associations into height-health and shrinkage-health associations. We use both continuous
measures and categories of the extent loss, to address measurement error in height, as well as to
explore non-linearities in the relationship between shrinkage and later-life health. Extreme height
shrinkage is signicantly associated with lower cognitive scores measuring memory and
uency
in both Indonesian and English older adults. We also nd negative associations between height
shrinkage and physical function. Finally, extreme height loss is found to be a predictor of mortality
in the Indonesian elderly sample. The association between shrinkage and later life health is expected
to be underestimated in our study samples. Overall, shrinkage is found to be an equal or in some
cases, even more important, covariate than baseline height for the older-age health outcomes we
study. Our results suggest that shrinkage is an informative, and easy to measure, biomarker
for worsening old-age health at the population level where more detailed measures might not be
available.
63
Chapter 4
Together in Sickness and in Health: Spousal In
uence in Health
and Health Behaviors of Elderly in England
1
4.1 Introduction
A considerable number of studies in social science and medicine have associated health status and
health behaviors between spouses (Meyler, Stimpson and Peek, 2007). The general consensus of
ndings is that there exists marital concordance of health, both in terms of physical health (Wilson,
2002; Davillas and Pudney, 2017), mental health (Fletcher, 2009) as well as health related behaviors
(Fletcher and Marksteiner, 2017; Banks, Kelly and Smith, 2013; Brown, Hole and Roberts, 2014).
Theories of the concordance draw upon assortative mating, shared environment and resources,
spillover or social learning and aective contagion (e.g., Wilson, 2002; Meyler, Stimpson and Peek,
2007). Although some of the existing studies in the area have dedicated themselves to link one
or two of the theories to their correlation ndings, most of them fail to disentangle the eect of
dierent channels from each other, especially the causal channels from the non-causal ones. While
the rst two channels are of no less importance than the latter two, assortative mating and shared
1
This is a joint work with Urvashi Jain. This analysis uses data or information from the Harmonized ELSA dataset
and Codebook, Version E as of April 2017 developed by the Gateway to Global Aging Data. The development of the
Harmonized ELSA was funded by the National Institute on Aging (R01 AG030153, RC2 AG036619, 1R03AG043052).
For more information, please refer to www.g2aging.org.
64
environment and resources are more of a selection process and re
ection problem as in Manski
(1993). Understanding the causal pathways (such as spillover or social learning) between spouses
generates increasing interest both in observational and experimental studies, because of its strong
link with social network and peer eects theories, its relevance for policy designs and program
evaluations in health domain and help with benet calculations when there are spillover eects
along social ties (Fletcher and Marksteiner, 2017; Brown, Hole and Roberts, 2014).
This chapter examines whether there are causal spillover eects of spousal health status and
behaviors among married and cohabiting English older adults, by estimating dynamic panel mod-
els and controlling for the endogeneity problem due to marriage market matching and shared
environment using Generalized Method of Moments (GMM). We use the seven waves of the En-
glish Longitudinal Study of Ageing (2000{2014) which is representative of the population aged
over 50 in UK. We nd there is strong and positive spousal concordance in health status and
health behaviors. GMM results conrm that there exist spillover eects between spouses for En-
glish couples in health status measured by self reported health and depressive symptoms, as well
as health-related behaviors including smoking, drinking, exercising and social contacting. There
are some gender dierences for them but not much. While most studies on spousal health con-
cordance have been documenting correlations, this chapter is among the few to provide causal
evidence on the spillover eects among spouses, which is based upon similar methodologies ap-
plied in Michaud and van Soest (2008)'s study on US elderly couples. Our analysis focuses on the
inter-spousal relationships in health domains among a comparable yet dierent population. We
also extend the current literature by evaluating various health outcomes and health behaviors at
the same time.
This chapter is organized as follows. Section 4.2 reviews the related literature. Section 4.3
and 4.4 provides theoretical background, the dynamic model of health and health behaviors and
65
the empirical strategies. Section 4.5 describes the dataset we use. Section 4.6 shows the empirical
results. Section 4.7 discusses the robustness of the causal estimation results. Section 4.8 concludes.
4.2 Literature Review
While the correlations in health status and health behaviors within couples have been widely
recognized in the literature (Meyler, Stimpson and Peek, 2007), relatively little is known about the
mechanisms resulting in spousal concordance, and the importance of dierent mechanisms.
Wilson (2002) serves as an example by documenting the spousal concordance in physical health
and evaluating the causal and non-causal underlying channels. He uses data from Health and
Retirement Study (HRS) wave 1992 and nds that assortative mating in the observed characteristics
{ such as age, education, income, and other socioeconomic and demographic determinants of health
status as well as behavioral risk factors { helps explain about 33% to 57% of the overall correlation
in health between spouses. The remaining unexplained part of the correlations could either work
through non-causal pathways such as sharing environment and assortative mating in unobserved
characteristics, for example, health endowment, or causal pathways such as spillover. A recent paper
by Davillas and Pudney (2017) explores multi-dementional health data from the UK Understanding
Society Panel (origially the British Household Panel Survey) and by estimating a cummulative
health exposure model, they nd that health correlation within spouses is unrelated to the length
of marriage, which is interpreted as evidence of equal importance of assortative mating and shared
in
uences. However, these two papers still fail to distinguish the causal spillover eects of spousal
health from the shared environment and resources eects.
Dierent empirical strategies have been applied to distinguish the causal factors from non-
causal ones in the observed spousal concordance in health behaviors. Clark and Etil e (2006) use 9
66
waves of British Household Panel Survey data to examine interactions between spouses in smoking.
They show from their bivariate Probit panel estimations that the correlation of smoking status
between spouses are mostly due to assortative mating which is captured by the (correlated) indi-
vidual random eects instead of endogenous eects rising from bargaining within marriage that
is captured by partner's past smoking behaviors or social learning from partner's past health.
Brown, Hole and Roberts (2014) and Clark and Etil e (2011) look at spousal correlation in body
weight. Clark and Etil e (2011) use life satisfaction as proxy for individual utility and recover in
semi-parametric analysis that own Body Mass Index (BMI) has a negative well-being eect but the
eect is lower when the BMI of the individuals is larger, which is consistent with social spillover
eects in weight. Brown, Hole and Roberts (2014) in their Seemingly Unrelated Regressions with
random eects demonstrate the importance of matching in explaining correlated body weight. Their
results using single equation random eects model with spousal BMI as a right hand side variable
and a rich set of controls are also suggestive of social in
uence and spillover eect.
Although these ndings have provided insights on the mechanisms of spousal correlations in
health and health behaviors, it is still less clear whether there exists causal spillover from spouse in
the health domain. In order to have causal interpretations and provide policy implications, Fletcher
(2009) studies the spillover eect of mental health between working spouses using data from the Na-
tional Survey of Midlife Development in the United States (MIDUS). His results from xed eects es-
timation where spousal mental health is instrumented with spousal job problems show that spousal
mental health status has an economically signicant eect on individuals' mental health which is
even larger than the eect of individuals' own mental health endowment. Michaud and van Soest
(2008) which employs 6 waves (1992-2002) of HRS data in the US to study wealth-health gradient
in a dynamic panel framework nds in the\Granger" sense a causal impact of the health of elderly
men on their wives' mental health two years later by using system GMM. Fletcher and Marksteiner
67
(2017) on the other hand, utilize experimental data from the Lung Health Study and COMBINE
study and nd evidence for signicant causal spousal spillover eects from addiction interventions
in smoking and drinking.
This chapter builds upon Michaud and van Soest (2008) and Powdthavee (2009) in develop-
ing a dynamic panel model to further the study in spousal concordance in health, controlling
for assortative mating and confounding factors from shared environment and unobservables. Ap-
plying the econometric methods by Arellano and Bond (1991), Arellano and Bover (1995) and
Blundell and Bond (1998), we are able to test specically the theories on spillover eects.
4.3 Theoretical Background
There are three broadly dened channels through which spousal health and health behaviors are
correlated: assortative mating which happened when the knot is tied and hence time invariant,
shared environment and direct spousal spillover (social learning or aective contagion).
4.3.1 Assortative Mating
Becker's marriage market matching theory (Becker, 1973, 1974) predicts that people are married
based on their individual characteristics. When the utility of marriage is transferable and com-
plementary in individual characteristics, there will be positive assortative mating which implies
that people tend to marry people with similar backgrounds and personal traits. Homogamy has
been found in education (Mare, 1991), culture and occupation (Kalmijn, 1994), and anthropometric
traits (Chiappori, Orece and Quintana-Domeque, 2012). Dohmen et al. (2012) also provide evi-
dence for the positive assortative mating on risk and trust preferences. While it is likely that people
would select their partners based on their health status, sorting in socio-economic characteristics
68
and lifestyles (e.g., smoking) would also lead to the concordance of health and health behaviors
between spouses (Banks, Kelly and Smith, 2013).
4.3.2 Shared Environment
Couples who live in the same household, share resources and risk, and make joint decisions on
many issues. Thus health behaviors and health status of husbands and wives might be correlated
because they are aected by common factors simultaneously. For example, shared exposure to
environmental hazard, diet and stress could have similar eects on both partners' health at the
same time. This correlated eect might also interrelate with partner selection and be reinforced
by the assortative mating between spouses (Davillas and Pudney, 2017). Spouses that are more
homogeneous in personal traits and lifestyles are more likely to experience simultaneous shocks.
However, it is empirically challenging to estimate shared environment eects because they are
normally not observed. Nevertheless, we need to control for this channel in order to causally
estimate the spillover eect in health and health behaviors between spouses.
4.3.3 Spillover Eect
Social networks have been found to signicantly in
uence one's health (Fowler and Christakis,
2008; Raspe, Hueppe and Neuhauser, 2008) and health behaviors (Christakis and Fowler, 2007;
Eisenberg, Golberstein and Whitlock, 2014; Powell, Tauras and Ross, 2005; Clark and Loh eac, 2007;
Fletcher, 2010). Marriages or partnerships are arguably the most important social network that
individuals form in lifetime, which shape one's knowledge, decisions, behaviors and preferences.
Multiple theories indicate the existence of spillover eects (contagion). If spousal health and health
behaviors are complementary to each other and endogenous, positive spillover eects exist. Having
an obese partner decreases the cost for oneself to be obese and that is why we see the similar
69
body weight trends of couples. On the
ip side, smoking has negative externalities to the health
of non-smoking spouses, thus an altruistic individual might quit smoking for his or her partner.
It is also possible that spouses can learn from or control each other's behavior, and hence be-
come more alike. For health status, couples could aect each other directly through transmitting
of infectious diseases and caring for each other. There is evidence showing that informal care of
a sick family member leads to a reduction in both the physical and mental health of the care-
giver (Schulz and Sherwood, 2008). Couples might also directly benet from each other's mental
well-being, as shown by Powdthavee (2009) and Fletcher and Frisvold (2009).
4.4 Emprical Strategies
4.4.1 Empirical Model
The main goal of this chapter is to identify the causal spillover eects in health and health behaviors
within couples. We develop a linear dynamic model for health status and health behaviors, allowing
for their persistence and spousal in
uence over time.
Following Powdthavee (2009) and Michaud and van Soest (2008), let (y
it
;y
jt
) denote the health
status and behaviors of interest of individual i and spouse j as measured by self-reported health,
depressive symptoms, behaviors such as smoking, drinking, exercising and social contact respec-
tively, in one household at time t. The system of dynamic conditional health demand functions
with cross-spouse dependence for a couple can be written as follows:
y
it
=
0
+
1
y
it1
+::: +
m
y
itm
+
0
y
jt
+::: +
g
y
jtg
+
1
X
it
+
2
X
jt
+
i
+u
it
(4.1)
70
y
jt
=
0
+
1
y
jt1
+::: +
m
y
jtm
+
0
y
it
+::: +
g
y
itg
+
1
X
jt
+
2
X
it
+
j
+u
jt
(4.2)
where
i
and
j
are individual-specic and time-invariant characteristics which could be correlated,
such as family background, personality, genetic features and endowments. u
it
andu
jt
capture time
and individual-specic unobservables that have impact on either spouse's outcomes, such as time
varying tastes and health shocks which are often common to both spouses, but are not observed
by econometricians. X
it
and X
jt
are other health covariates (inputs). For simplicity, we are going
to omit X
it
and X
jt
to better illustrate our empirical strategies.
By construction,y
it
andy
jt
and all their past and future realizations are related to the individual
heterogeneity, which result in endogeneity problems. For instance, individuals born with poor health
endowments and inheritable conditions are less likely to be healthy as they grow older. What is
more,y
it
andy
jt
are correlated with each other, due to correlation in individual heterogeneity, i.e.,
i
and
j
, which capture the eect from assortative mating. y
it
and y
jt
might be correlated with
both u
jt
and u
it
as a result of shared environments and health shocks. 's and
's are parameters
of interest that identify the causal channel, among which
0
and
0
evaluate instantaneous spousal
spillover eects. First dierencing of the above system of equations could easily purge out all
time-invariant unobserved heterogeneity as well as the eect of assortative mating.
y
it
=
1
y
it1
+::: +
m
y
itm
+
0
y
jt
+::: +
g
y
jtg
+ u
it
(4.3)
y
jt
=
1
y
jt1
+::: +
m
y
jtm
+
0
y
it
+::: +
g
y
itg
+ u
jt
(4.4)
4.4.2 Assumptions and Empirical Methods
By construction, y
it1
and y
jt1
are correlated with u
it
and u
jt
and so are their current
values. For the model to be tractable, we make the following assumptions.
71
Assumption 1. (Dynamics of health and health related behaviors): A 1st order linear Markov
process is used to model the state dependence of health and health behaviors, i.e., m = 1. It also
implies a separability assumption: health in the previous period serving as a sucient statistic for
all health inputs and outcomes for all other prior periods (Strauss and Thomas, 2008).
We will relax the baseline assumption and compare the results when m = 1 with cases with
m = 2. The empirical result section shows that for certain health status and health behaviors, it is
more appropriate to control for further lags based on the specication tests. However, the estimated
spillover eects coecients for most outcomes are insensitive to the length of lagged values of own
health or behaviors.
Assumption 2. Spousal lagged outcomes are irrelevant in the structural model, i.e., g = 0
. This implies the absence of direct lagged spillover eects from spouses. In other words, lagged
spousal health or behaviors only aects individuals current health or behaviors through the current
level without having any independent eect. We argue that this is a reasonable assumption as the
lagged spousal outcomes are 2 years from the current time and we report results of estimations in
the reduced form where we only include spousal lags (g = 1) as a robustness check. We will show
that for most outcomes of interest, the lagged spillover coecients are insignicant, as evidence for
the lack of lagged spousal spillover eects.
Assumption 3. Serial uncorrelatedness inu
it
andu
jt
. This assumption is standard in time se-
ries and dynamic panel literature and can be tested (Arellano and Bond, 1991; Arellano and Bover,
1995).
Under Assumption 1 to Assumption 3, the instantaneous spousal spillover eects in the struc-
tural model are identiable and can be estimated by using instruments. Valid instruments are
available as below:
72
For the rst dierences of lagged outcomes: (y
it2
;y
it3
;:::y
i1
) and (y
jt2
;y
jt3
;:::y
j1
) respec-
tively. For the rst dierences of current outcomes: (y
it2
;y
it3
;:::y
i1
) and (y
jt2
;y
jt3
;:::y
j1
)
respectively.
Inclusion restriction is satised because y
it1
and y
jt1
are correlated with y
it2
and y
jt2
and further lags. Exclusion restriction is satised because (y
it2
;y
it3
;:::y
i1
) and (y
jt2
;y
jt3
;:::y
j1
)
are uncorrelated with u
it
and u
jt
.
Blundell and Bond (1998) proposed to use the rst dierences of the endogenous variables to
serve as instruments for the level equations, if the following assumption holds as well.
Assumption 4. (Stationarity) There is no correlation between the dierences of the explana-
tory variables and the unobserved individual xed eects, i.e., E(y
it1
i
) = 0 orE(y
it
i
) is time
invariant for alli. This assumption relies on the initial conditions of the data generating process as
formalized in Blundell and Bond (1998): the deviations ofy
i1
from the steady states are orthogonal
to the individual xed eects, i.e., no \regression to the mean".
Under the four assumptions, we have the following moment conditions:
For the rst dierence equations: s>= 2, t = 3;;T
E(y
its
u
it
) = 0;
E(y
jts
u
jt
) = 0;
E(y
jts
u
it
) = 0;
E(y
its
u
jt
) = 0:
73
For the level equations: s = 1, t = 3;;T
E(y
its
(
i
+u
it
)) = 0;
E(y
jts
(
i
+u
it
)) = 0;
E(y
jts
(
j
+u
jt
)) = 0;
E(y
its
(
j
+u
jt
)) = 0:
The previous moment conditions outline what are called Dierence and System GMM approach
developed by Arellano and Bond (1991), Arellano and Bover (1995) and Blundell and Bond (1998).
It can be applied to estimate the \small T , large N" linear dynamic panel model with not strictly
exogenous or endogenous independent variables and individual xed eects. Dierence GMM uti-
lize the moment conditions for rst dierence equations while System GMM approach increases
eciency by imposing the additional moment conditions. System GMM deal with the weak in-
strument problem in Dierence GMM if y is close to random walk, and hence the levels bear little
information about future changes (Roodman, 2009). System GMM is our preferred estimation
strategy because of its improved eciency, however, given that it requires further assumptions,
we report Dierence GMM results as a robustness check. Two-step Dierence and System GMM
estimates with Windmeijer (2005) corrected robust standard errors are reported. Single equations
for men and women are estimated separately.
Arellano-Bond test for autocorrelation is applied to test the assumption of the absence of au-
tocorrelation in the error term (Assumption 3) which is crucial for the validity of y
it2
and further
lags as instruments for the rst dierences. After eliminating individual xed eects, the test of
74
rst order autocorrelation in the error term is equivalent to test second order autocorrelation in
residuals in dierences since rst order autocorrelation is created by construction. The exogeneity
of the instruments is tested by Hansen test of overidentifying restrictions for the whole model and
Dierence in Hansen test is conducted for the subset of level equation instruments, given that the
additional assumptions in System GMM are non-trivial.
4.5 Data
England is one of most developed economies in the world with per capita GDP in PPP international
dollars reaching 41,458.7 by 2015. Population in England is older on average than in the U.S. with
ratio of elderly over 65 in the national population in 2015 around 18%
2
. We use the English
Longitudinal Study of Ageing (ELSA), one of the sister studies of the Health Retirement Study
(HRS) in the US, which is representative of the national population aged 50 and above. Respondents
with eligible age and their spouses irrespective of age in the survey year are interviewed. The
eldwork is conducted biennially since 2002, with refreshment samples added from the wave 1 age-
ineligible households in wave 3/4/6/7. We use the harmonized data of the waves 1{7 (2002{2015)
of ELSA, which following the RAND version of HRS provides information on the demographics,
health and health behaviors of all respondents as well as their spouses
3
.
4.5.1 Outcome Variables
We consider two groups of outcomes related to health: health status and health behaviors.
Health Status. Health status of interest include self-reported good health measured by a
binary variable which equals one if the respondent reports good, very good or excellent health,
2
https://www.census.gov/content/dam/Census/library/publications/2016/demo/p95-16-1.pdf. Accessed on Jan-
uary 15, 2017.
3
Gateway to Global Aging https://g2aging.org/ accessed on January 15, 2017.
75
zero if otherwise. In appendix, we also report results using a self-reported health measure on a
scale from 1 to 3 which are calculated based on the self-reported health question in wave 3 with
European scale
4
and the ones in all other waves with US scale
5
. We regroup the answers to a
Likert-scale measure, which equals 1 if the self-report is either excellent, very good or good; 2 if it
is fair health; 3 if the respondent reports bad, very bad or poor health. For mental health, we also
utilize a depressive symptom index which is based on an 8 item CES-D (total score from 0 to 8)
6
.
The health outcomes are chosen such that they cover both general and mental health domains and
have been measured relatively consistently and frequently.
Health Behaviors. The health related behaviors considered consist of smoking and drinking
behaviors for which we examine whether the respondents smoke currently, their intensity as mea-
sured by number of cigarettes per day
7
and number of days in a week they drink
8
. Frequency of
physical exercise measured in number of days in a week participating in vigorous physical activi-
ties
9
is examined as well. We also test in appendix on the extensive margin of exercise, which is
measured by a binary indicator for any weekly participation in vigorous physical activities. We also
investigate the spousal spillover eects in social activity which is measured by a binary variable
indicating whether the respondent has any weekly contact with relatives in person. Biomarkers are
not included as they are only available for every alternate wave. Behavior related outcomes are
selected as they are important predictors in either physical or mental health.
4
Very good, good, fair, bad and very bad.
5
Excellent, very good, good, fair and poor.
6
In ELSA, there are only 8 questions on CES-D and the harmonized value for each item is 1 if the respondent
answers that he or she has the following feeling much of the time last week and 0 otherwise: feeling depressed, feeling
that everything was an eort, sleep was restless, felt lonely, felt sad, felt he/she could not get going, felt happy
(reverse), felt he/she enjoyed life (reverse). The CES-D 8 score is the sum of the values of the rst 6 items and 1
minus the values of the last 2 items. Thus the higher the score, the more depressive the respondent is.
7
Number of cigarettes per day is recorded as 0 if the respondent does not smoke when interviewed.
8
Number of days in a week of drinking is recorded as 0 if the respondent does not drink when interviewed.
9
e.g., running/jogging, swimming, cycling, aerobics/gym workout, tennis, and digging with a spade. We impute
the number of days participating in vigorous physical activities based on respondents' choice from 4 categories. We
assign 4 if they chose \more than once per week"; 1 if \once per week"; 0.25 if \one to three times per month";
0 if \hardly ever". We have experimented with other values for the \more than once per week" category, and the
conclusions are not changed. Results are available upon request.
76
4.5.2 Control Variables
We control for a wide array of individual and household level characteristics, including: 1) respon-
dent and spouse age dummies (5 years interval)
10
, numbers of living daughters and sons, respondent
and spouse levels of education which are categorized as less than high school, high school, some
college or college and above, with less than high school being the omitted group, as well as length
of marriage; 2) logarithm of de
ated per capita household wealth
11
In OLS and FE specications,
all control variables are treated as exogenous. In GMM estimations, the rst set of covariates is
treated as exogenous variables or not estimable when taking rst dierences in Dierence GMM
and FE estimations. Household wealth is treated as endogenous and instrumented with valid lags
or dierence of lags in GMM estimations.
4.5.3 Sample and Descriptive Statistics
In our analysis, we keep couples that remained married or partnered and lived in the same household
(i.e., not separated), and both of whom responded with non-missing values in any of the outcome of
interest and all control variables, for at least 3 consecutive waves. For dierent outcomes of interest,
the number of couples and size of estimation sample varies. Table 4.1 and 4.2 summarize for men
and women respectively, the main variables under analysis. We report year-specic descriptive
statistics in wave 3 and 5 when the original ELSA cohorts and refreshment cohorts rst became
estimable by GMM, as well as wave 7, which is most recent wave with publicly available data.
10
<50, 50-54, 55-59, 60-64, 65-69, 70-74, 75-79, 80 and older.
11
Household wealth is constructed as the sum of the dierent wealth components: Net value of primary residence
+ Net value of business + Net value of non-housing nancial wealth + net value of the secondary home residence
and other property after paying all the debts + total value of other physical assets for those reporting having other
land, money owed by others, a trust, a covenant or inheritance, or other assets (Phillips et al., 2017). In practice,
the formula for transformation is: log(wealth) =log(wealth + 1) if wealth>= 0 or log(wealth) =log(1wealth)
if wealth< 0, following Michaud and van Soest (2008).
77
Across these waves, the levels of self-reported health status of both men and women are fairly
stable and similar to each other. The CES-D scores are higher on average for women, which is
consistent with the fact that depression in more common among women. There have been a steady
decline in the fraction of smokers and intensity of drinking and smoking for both genders. This
could be due to either the decline across cohorts or over the years as respondents age. Wives on
average are younger and less educated than their husbands. Women also drink and exercise less
frequently, although they smoke more intensively than men.
Table 4.1: Descriptive Statistics of Main Variables for Men
Wave 3 Wave 5 Wave 7
VARIABLES N mean N mean N mean
Health Outcomes
Have good self reported health 1,510 0.718 1,489 0.765 1,232 0.762
CESD Score 1,508 0.897 1,479 1.060 1,228 0.894
Health Behaviors
Smokes now 1,510 0.100 1,512 0.0847 1,269 0.0678
# cigarettes/day 1,456 0.947 1,457 0.681 1,229 0.429
# days/week drinks 1,358 3.140 1,396 3.003 1,150 2.891
Freq vigorous phys activ : # days/week 1,510 1.077 1,519 1.048 1,269 1.162
Any weekly contact w/ relative in person 1,271 0.297 1,303 0.303 1,073 0.285
Demographic & Socioeconomic Characteristics
Age (years) 1,510 66.12 1,519 66.81 1,270 69.85
Education 1,510 Fraction 1,519 Fraction 1,270 Fraction
Less than high-school 494 0.327 430 0.283 324 0.255
High-school graduate 261 0.173 279 0.184 241 0.190
Some college 402 0.266 409 0.269 351 0.276
College and above 353 0.234 399 0.263 354 0.279
Length of marriage 1,510 37.38 1,519 37.87 1,270 40.66
Log per capita de
ated household wealth 1,510 11.54 1,519 11.45 1,270 11.56
Family Composition
Number of living son 1,510 1.113 1,519 1.047 1,270 1.072
Number of living daughter 1,510 1.111 1,519 1.078 1,270 1.107
Number of living siblings 1,510 1.705 1,519 1.706 1,270 1.653
Note: Data are restricted to married couples living together, both of who responded with non-missing
values in any of the outcome of interest and all of the control variables, for at least 3 consecutive waves.
78
Table 4.2: Descriptive Statistics of Main Variables for Women
Wave 3 Wave 5 Wave 7
VARIABLES N mean N mean N mean
Health Outcomes
Have good self reported health 1,571 0.733 1,526 0.782 1,291 0.778
CESD Score 1,568 1.369 1,518 1.439 1,283 1.242
Health Behaviors
Smokes now 1,572 0.108 1,533 0.0978 1,316 0.0783
# cigarettes/day 1,549 1.301 1,505 1.118 1,290 0.791
# days/week drinks 1,418 2.315 1,430 2.235 1,205 2.135
Freq vigorous phys activ : # days/week 1,572 0.907 1,539 0.892 1,315 0.906
Any weekly contact w/ relative in person 1,371 0.373 1,351 0.376 1,144 0.354
Demographic & Socioeconomic Characteristics
Age (years) 1,572 63.30 1,540 64.06 1,316 67.17
Education 1,572 Fraction 1,540 Fraction 1,316 Fraction
Less than high-school 635.1 0.404 530 0.344 415 0.315
High-school graduate 389.9 0.248 400 0.260 351 0.267
Some college 320.7 0.204 320 0.208 287 0.218
College and above 226.4 0.144 290 0.188 263 0.200
Length of marriage 1,572 37.33 1,540 37.80 1,316 40.63
Log per capita de
ated household wealth 1,572 11.52 1,540 11.43 1,316 11.54
Family Composition
Number of living son 1,572 1.125 1,540 1.040 1,316 1.067
Number of living daughter 1,572 1.118 1,540 1.062 1,316 1.106
Number of living siblings 1,572 1.866 1,540 1.890 1,316 1.837
Note: Data are restricted to married couples living together, both of who responded with non-missing
values in any of the outcome of interest and all of the control variables, for at least 3 consecutive waves.
79
Table 4.3 describes the pairwise correlations between spouses in health status and health be-
haviors in wave 3, 5 and 7. Self-reported good health and CES-D scores are signicantly correlated
within couples, with Pearson correlation P value well below 1%. The same patterns are found
for smoking, drinking, exercising and social contacting behaviors. There is no clear time trend in
correlations in health and health behaviors among the ELSA elderly couples, which is also true if
we look at ELSA original cohorts and refreshment cohorts separately
12
. The signicant spousal
concordance in health and health behaviors can be the result of assortative mating, shared envi-
ronment and resources, or causal spillover, as we have discussed previously. The next section will
provide further evidence on the spillover channel specically.
Table 4.3: Correlation Coecients of Health Status and Health Be-
haviors Within Couples
VARIABLES Wave 3 Wave 5 Wave 7
Have good self reported health 0.26 0.24 0.27
CESD Score 0.23 0.26 0.30
Smokes now 0.29 0.31 0.27
Intensity of smoking: # cigarettes/day 0.19 0.27 0.18
Intensity of drinking: # days/week drinks 0.55 0.55 0.55
Freq vigorous phys activ : # days/week 0.33 0.29 0.32
Any weekly contact w/ relative in person 0.31 0.31 0.40
Note: All correlation coecients are signicant at 1% level.
4.6 Estimation Results
4.6.1 Health Status
Table 4.4 and 4.5 show results on the spousal spillover eects in health outcomes including self-
reported health and depressive symptoms (CES-D). OLS and FE results are reported as well.
12
Results are available upon request.
80
Self-reported good health. The self-reported health is strongly and positively correlated
between spouses: OLS and FE coecients before spousal self-reported good health are similar in
magnitude for both men and women. Column (3) and (8) report the system GMM results when we
only include one period lagged value of own health outcome, while in column (4) and (9) we allow
for state dependence up to a 2-period lag. Although all coecients for spousal self-reported good
health are signicantly positive for both husbands and wives, the specication tests column (3) and
(4) imply violations of Assumption 4 and the overall exclusion assumption for male sample. On
the other hand, the GMM results for English women indicate that if the husband has good health
as apposed to fair or poor health, there is a 27 percentage points increase in the probability for
the wife being in good health. A comparison between column (8) and (9), as well as between (3)
and (4) also indicate that the lag dependence of self-reported good health seem to dier between
men and women. While the estimated instantaneous spillover eects from spouses decline for both
genders when we include more lagged values of own health outcomes, the extent of the decline is
larger for men and the eect from lagged 2 period own self-reported health is only signicant for
husbands. This nding suggests it might be important to consider dierent lag structures for men
and women.
It is also noteworthy that the estimates of own lagged self-reported good health decline in GMM
estimations, compared with OLS results, to the extent that they are smaller than the instantaneous
spousal eects. This leads to concern about the credibility of our GMM results especially the
exclusion assumption such that the past spousal health only aects one's own health only through
the current health of spouses. The dierence between the OLS and GMM estimation results on
one's own lagged health is re-stated in reduced form estimation results where we include spousal
lagged health instead of their current health. From column (6) we see that even we don't include
spousal current health, the coecient before the own previous health barely changes. Comparisons
81
between column (3) and (5), as well between column (8) and (10) in addition imply that the
current health of one's spouse matters but spousal past health does not, which is consistent with
our Assumption 2. Clark and Etil e (2006) found that spousal past health status does not have
signicant impact on their health behaviors (smoking). If the spousal past health does not have
eect on ones behavior, it could lead to the case that one's current health status is not strongly
aected by their spousal health status. In appendix, we also show as a robustness check the results
for the self-reported health measured on a scale from 1 to 3, which are similar to what we nd for
a binary outcome for good self-reported health.
Depressive Symptoms. Powdthavee (2009) found that there is strong causal impact of
happiness of one's partner on that of one's own. We nd that such spillover eects also exist in
depressive symptoms. CES-D 8 score of one's spouse is signicantly and positively correlated with
one's own CES-D 8 score for both men and women. Furthermore, the causal estimates for the
instantaneous spillover eects are larger than the OLS or FE estimates, while the state dependence
in depression is much smaller in GMM specications than in OLS. Arrelano-Bond AR(2) tests,
Hansen J tests and Dierence in Hansen J tests prove the validity of the instruments used for
women irrespective of the lag structure. Reduced form estimation results reveal in column (5) and
(10) that spousal lagged depression does not signicantly aects one's current depression level,
which is true for both genders. Comparing results between column (3) and (4) points out that
mental health from 2 periods ago of men still have strong impact on their current mental health.
For women, however, it is only the most recent depressive episodes that matter. For both genders,
state dependence on one's own depression is smaller than the instantaneous spousal eect. The
magnitudes of system GMM estimates for the spillover eects are larger for women than those for
men, implying that women are aected more by their husband's depression than men are by their
wives.
82
Table 4.4: Results on Good Self-reported Health (SRH)
(1) (2) (3) (4) (5)
Men OLS FE System GMM
L.(own good SRH) 0.513*** -0.148*** 0.111*** 0.192*** 0.115***
(0.0174) (0.0203) (0.0277) (0.0353) (0.0286)
L2.(own good SRH) 0.0943***
(0.0297)
spouse's good SRH 0.140*** 0.145*** 0.287*** 0.169*
(0.0135) (0.0188) (0.0926) (0.0940)
L.(spouse's good SRH) -0.00803
(0.0222)
Observations 6718 6718 6718 4838 6759
Number of couples 1881 1881 1881 1606 1886
AR(1) test p val 1.71e-57 1.70e-43 2.01e-57
AR(2) test p val 0.351 0.246 0.279
Hansen J test p val 0.000205 0.0653 0.0000192
Di in Hansen p val for levels 0.0477 0.279 0.0360
(6) (7) (8) (9) (10)
Women OLS FE System GMM
L.(own good SRH) 0.519*** -0.143*** 0.127*** 0.144*** 0.128***
(0.0174) (0.0196) (0.0255) (0.0331) (0.0262)
L2.(own good SRH) 0.0301
(0.0281)
spouse's good SRH 0.138*** 0.136*** 0.297*** 0.265***
(0.0126) (0.0174) (0.0760) (0.0774)
L.(spouse's good SRH) -0.00753
(0.0189)
Observations 6850 6850 6850 4916 6916
Number of couples 1944 1944 1944 1636 1956
AR(1) test p val 1.79e-59 3.10e-41 4.08e-59
AR(2) test p val 0.959 0.832 0.972
Hansen J test p val 0.0898 0.127 0.0243
Di in Hansen p val for levels 0.191 0.220 0.332
Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05,
* p < 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living
daughters and sons, number of living siblings, respondent's and spouse's education dummies and
length of marriage; 2) log per capita de
ated household wealth.
83
Table 4.5: Results on Depressive Symptoms (CES-D 8 Score)
(1) (2) (3) (4) (5)
Men OLS FE System GMM
L.(own CESD score) 0.438*** -0.174*** 0.0438* 0.0851** 0.0498**
(0.0216) (0.0183) (0.0235) (0.0381) (0.0246)
L2.(own CESD score) 0.0556*
(0.0295)
spouse's CESD score 0.137*** 0.146*** 0.372*** 0.358***
(0.0122) (0.0161) (0.0927) (0.0941)
L.(spouse's CESD score) -0.00991
(0.0154)
Observations 6641 6641 6641 4771 6680
Number of couples 1872 1872 1872 1597 1877
AR(1) test p val 8.98e-42 4.78e-32 3.46e-46
AR(2) test p val 0.818 0.624 0.577
Hansen J test p val 0.0338 0.0354 0.00109
Di in Hansen p val for levels 0.333 0.566 0.00878
(6) (7) (8) (9) (10)
Women OLS FE System GMM
L.(own CESD score) 0.457*** -0.143*** 0.111*** 0.142*** 0.118***
(0.0179) (0.0169) (0.0221) (0.0308) (0.0225)
L2.(own CESD score) 0.0301
(0.0262)
spouse's CESD score 0.204*** 0.201*** 0.529*** 0.454***
(0.0174) (0.0228) (0.107) (0.110)
L.(spouse's CESD score) -0.00463
(0.0247)
Observations 6773 6773 6773 4857 6829
Number of couples 1931 1931 1931 1626 1940
AR(1) test p val 4.97e-64 3.14e-45 1.21e-63
AR(2) test p val 0.343 0.325 0.121
Hansen J test p val 0.476 0.153 0.00961
Di in Hansen p val for levels 0.793 0.262 0.179
Note: Two-step corrected robust standard errors in the parenthesis. *** p< 0:01, **p< 0:05,
* p< 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living
daughters and sons, number of living siblings, respondent's and spouse's education dummies
and length of marriage; 2) log per capita de
ated household wealth.
84
4.6.2 Health-related Behaviors
As health behaviors are closely related to health status, it is natural to ask whether there are also
spillover eects in health behaviors among couples. Although we do not attempt to test explicitly
whether the spousal spillovers in behaviors also contribute to the spousal concordance in health
status, some hypotheses accounting for spousal health concordance suggest that shared environment
within couples translates into shared health risks which benet or harm health, depending on the
health behaviors of the spouses (Meyler, Stimpson and Peek, 2007). Therefore, estimating the
spillover eects in behaviors within couples is not only interesting to its own end, but could also
help further the understanding of spousal concordance in health. In this section, we examined 4
important health related behaviors of adults: smoking, drinking, exercising and social contacting.
Smoking. Table 4.6 shows the results on smoking status, which is a binary outcome. Since
smoking is not just a conscious behavior but a compulsive addiction, state dependence are found to
relative high. Moreover, we nd evidence that this addiction can thrive further in case of spousal
support. OLS and FE results in Table 4.6 indicate of positive correlation in smoking status between
husband and wife. What is more, GMM estimations in Column (3) and (4) imply that there is
signicant spousal spillover eect in smoking behaviors for English elderly men. In contrast, the
in
uence of husband's smoking on their wife is smaller and not statistically signicant as shown in
column (8) and (9): having a smoking wife tends to increase the likelihood for the husband to be
a smoker by about 14% while it is only around 8% for wife if the husband smokes. A comparison
between these columns also suggest the smoking is persistent for both genders and allowing for
further path dependence might be important, given the signicance of 2-period lagged smoking
status and the rejection of specication tests in columns with fewer lags.
85
Table 4.6: Results on Smoking Status (= 1 if smokes now)
(1) (2) (3) (4) (5)
Men OLS FE System GMM
L.(own smoking status) 0.761*** 0.121*** 0.368*** 0.431*** 0.372***
(0.0183) (0.0394) (0.0635) (0.0596) (0.0604)
L2.(own smoking status) 0.118***
(0.0413)
spouse's smoking status 0.0675*** 0.102** 0.188*** 0.140*
(0.0122) (0.0440) (0.0715) (0.0757)
L.(spouse's smoking status) 0.0224
(0.0337)
Observations 6981 6981 6981 5048 6969
Number of couples 1926 1926 1926 1665 1918
AR(1) test p val 5.42e-11 1.02e-08 2.76e-11
AR(2) test p val 0.0779 0.110 0.101
Hansen J test p val 0.0900 0.733 0.0437
Di in Hansen p val for levels 0.659 0.989 0.613
(6) (7) (8) (9) (10)
Women OLS FE System GMM
L.(own smoking status) 0.826*** 0.0513 0.308*** 0.476*** 0.304***
(0.0159) (0.0452) (0.0640) (0.0844) (0.0659)
L2.(own smoking status) 0.136***
(0.0385)
spouse's smoking status 0.0460*** 0.0839*** 0.0340 0.0871
(0.0107) (0.0272) (0.0614) (0.0619)
L.(spouse's smoking status) 0.0386*
(0.0201)
Observations 7174 7174 7174 5147 7138
Number of couples 2019 2019 2019 1712 1997
AR(1) test p val 6.16e-09 0.0000105 2.86e-08
AR(2) test p val 0.396 0.786 0.472
Hansen J test p val 0.00648 0.163 0.0106
Di in Hansen p val for levels 0.600 0.976 0.624
Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05,
* p < 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living
daughters and sons, number of living siblings, respondent's and spouse's education dummies and
length of marriage; 2) log per capita de
ated household wealth.
86
Table 4.7: Results on Smoking Intensity (# cigarettes/day)
(1) (2) (3) (4) (5)
Men OLS FE System GMM
L.(# cigarettes/day smokes) 0.801*** 0.0812 0.246*** 0.297*** 0.226***
(0.0218) (0.0527) (0.0790) (0.107) (0.0846)
L2.(# cigarettes/day smokes) 0.0893*
(0.0458)
# cigarettes/day spouse smokes 0.0660*** 0.0792*** 0.0803 0.0461
(0.0187) (0.0275) (0.0607) (0.0478)
L.(# cigarettes/day spouse smokes) 0.0169
(0.0199)
Observations 6734 6734 6734 4839 6694
Number of couples 1919 1919 1919 1632 1900
AR(1) test p val 0.0000364 0.00353 0.000151
AR(2) test p val 0.310 0.280 0.527
Hansen J test p val 0.131 0.224 0.0670
Di in Hansen p val for levels 0.768 0.594 0.722
(6) (7) (8) (9) (10)
Women OLS FE System GMM
L.(# cigarettes/day smokes) 0.732*** 0.167** 0.320*** 0.391*** 0.309***
(0.0325) (0.0745) (0.0870) (0.0899) (0.0902)
L2.(# cigarettes/day smokes) 0.137***
(0.0404)
# cigarettes/day spouse smokes 0.0679*** 0.0863** 0.154* 0.188**
(0.0188) (0.0346) (0.0924) (0.0872)
L.(# cigarettes/day spouse smokes) -0.00783
(0.0227)
Observations 6528 6528 6528 4701 6524
Number of couples 1824 1824 1824 1569 1819
AR(1) test p val 0.000208 0.00361 0.000317
AR(2) test p val 0.0726 0.478 0.0650
Hansen J test p val 0.316 0.513 0.234
Di in Hansen p val for levels 0.400 0.770 0.384
Note: Two-step corrected robust standard errors in the parenthesis. ***p< 0:01, **p< 0:05, *p< 0:1.
All columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons,
number of living siblings, respondent's and spouse's education dummies and length of marriage; 2) log
per capita de
ated household wealth.
87
Table 4.8: Results on Drinking Frequency (# days/week drinks)
(1) (2) (3) (4) (5)
Men OLS FE System GMM
L.(# days/week drinks) 0.660*** -0.199*** 0.0943*** 0.171*** 0.135***
(0.0175) (0.0239) (0.0336) (0.0426) (0.0355)
L2.(# days/week drinks) 0.127***
(0.0398)
# days/week spouse drinks 0.217*** 0.203*** 0.484*** 0.468***
(0.0161) (0.0316) (0.103) (0.100)
L.(# days/week spouse drinks) 0.0567*
(0.0342)
Observations 4023 4023 4023 2678 4020
Number of couples 1448 1448 1448 1185 1457
AR(1) test p val 2.06e-36 8.08e-22 8.05e-33
AR(2) test p val 0.612 0.179 0.377
Hansen J test p val 0.136 0.621 0.00334
Di in Hansen p val for levels 0.398 0.786 0.176
(6) (7) (8) (9) (10)
Women OLS FE System GMM
L.(# days/week drinks) 0.671*** -0.142*** 0.153*** 0.242*** 0.206***
(0.0168) (0.0268) (0.0348) (0.0457) (0.0370)
L2.(# days/week drinks) 0.0937***
(0.0313)
# days/week spouse drinks 0.191*** 0.198*** 0.563*** 0.582***
(0.0144) (0.0280) (0.101) (0.0923)
L.(# days/week spouse drinks) 0.0549*
(0.0285)
Observations 4083 4083 4083 2697 4090
Number of couples 1483 1483 1483 1203 1496
AR(1) test p val 1.80e-29 8.24e-17 3.78e-34
AR(2) test p val 0.476 0.555 0.0970
Hansen J test p val 0.840 0.982 0.00508
Di in Hansen p val for levels 0.745 0.857 0.228
Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05,
* p < 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living
daughters and sons, number of living siblings, respondent's and spouse's education dummies and
length of marriage; 2) log per capita de
ated household wealth.
88
Table 4.9: Results on Frequency of Exercise (# days/week participating in vigorous
physical activities, VPA)
(1) (2) (3) (4) (5)
Men OLS FE System GMM
L.(# d/w w/ VPA) 0.441*** -0.105*** 0.126*** 0.217*** 0.130***
(0.0150) (0.0158) (0.0218) (0.0245) (0.0226)
L2.(# d/w w/ VPA) 0.120***
(0.0200)
# d/w spouse w/ VPA 0.222*** 0.271*** 0.244** 0.261***
(0.0146) (0.0186) (0.0959) (0.101)
L.(# d/w spouse w/ VPA) 0.0263
(0.0196)
Observations 7020 7020 7020 5077 7013
Number of couples 1934 1934 1934 1672 1931
AR(1) test p val 6.33e-86 7.02e-70 4.39e-87
AR(2) test p val 0.0197 0.184 0.0542
Hansen J test p val 0.000303 0.346 0.000636
Di in Hansen p val for levels 0.387 0.834 0.498
(6) (7) (8) (9) (10)
Women OLS FE System GMM
L.(# d/w w/ VPA) 0.436*** -0.119*** 0.150*** 0.150*** 0.162***
(0.0164) (0.0168) (0.0210) (0.0270) (0.0210)
L2.(# d/w w/ VPA) 0.0216
(0.0225)
# d/w spouse w/ VPA 0.181*** 0.234*** 0.131* 0.133*
(0.0131) (0.0160) (0.0722) (0.0788)
L.(# d/w spouse w/ VPA) 0.00335
(0.0159)
Observations 7212 7212 7212 5182 7193
Number of couples 2022 2022 2022 1717 2019
AR(1) test p val 1.55e-73 2.82e-51 2.61e-77
AR(2) test p val 0.183 0.600 0.234
Hansen J test p val 0.376 0.202 0.422
Di in Hansen p val for levels 0.212 0.311 0.259
Note: Two-step corrected robust standard errors in the parenthesis. *** p< 0:01, **p< 0:05,
* p< 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living
daughters and sons, number of living siblings, respondent's and spouse's education dummies
and length of marriage; 2) log per capita de
ated household wealth.
89
Table 4.10: Results on Social Contact (= 1 if any weekly contact with relatives in person)
(1) (2) (3) (4) (5)
Men OLS FE System GMM
L.(wkly contact w/ relativ) 0.411*** -0.137*** 0.147*** 0.175*** 0.154***
(0.0185) (0.0226) (0.0289) (0.0367) (0.0301)
L2.(wkly contact w/ relativ) 0.0427
(0.0360)
spouse's wkly contact w/ relativ 0.180*** 0.115*** 0.191** 0.205**
(0.0155) (0.0236) (0.0885) (0.0902)
L.(spouse's wkly contact w/ relativ) 0.0243
(0.0250)
Observations 4371 4371 4371 2991 4408
Number of couples 1456 1456 1456 1131 1472
AR(1) test p val 2.42e-50 1.15e-31 4.82e-51
AR(2) test p val 0.325 0.908 0.724
Hansen J test p val 0.216 0.250 0.0282
Di in Hansen p val for levels 0.978 0.925 0.888
(6) (7) (8) (9) (10)
Women OLS FE System GMM
L.(wkly contact w/ relativ) 0.464*** -0.137*** 0.169*** 0.247*** 0.161***
(0.0177) (0.0226) (0.0296) (0.0325) (0.0294)
L2.(wkly contact w/ relativ) 0.139***
(0.0312)
spouse's wkly contact w/ relativ 0.183*** 0.117*** 0.212** 0.282**
(0.0165) (0.0232) (0.108) (0.126)
L.(spouse's wkly contact w/ relativ) 0.0309
(0.0258)
Observations 4500 4500 4500 3091 4535
Number of couples 1508 1508 1508 1188 1528
AR(1) test p val 5.92e-57 1.34e-39 2.16e-55
AR(2) test p val 0.00188 0.791 0.0975
Hansen J test p val 0.0133 0.478 0.00678
Di in Hansen p val for levels 0.518 0.813 0.696
Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05,
* p < 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living
daughters and sons, number of living siblings, respondent's and spouse's education dummies and
length of marriage; 2) log per capita de
ated household wealth.
90
Table 4.7 reports results from estimations on the intensity of smoking, measured by number
of cigarettes per day. On the intensive margin, there exist positive spillover eects in smoking for
both men and women, although such eects from wives are not signicant. Neither of the estimated
coecients for lagged spousal spillover eects is signicant in the reduced form GMM estimation,
and the validity of instruments are rejected in column (5) and (10).
Drinking. Table 4.8 shows that there is persistent inter-spousal concordance in the number
of days they drink in a week as demonstrated by OLS and FE results for both men and women.
This is intuitive since couples share their meals and drinks together. From causal specications,
both men and women have strong impact on their partners and this eect is present even in the
reduced form results, however, the specication tests are rejected in column (5) and (10). If one's
spouse drinks one more day a week, he/she will tend to drink about 0.47/0.58 more days if we
allow for further path dependence. Similar to what we have found for health status measures, the
instantaneous spillover eect is also larger than the eect from path dependence.
Exercise. We also nd strong causal spillover eects in physical exercise for both genders.
In Table 4.9, we see that the magnitudes of the causal estimates are similar to those from OLS
or FE estimations. Inclusion of more lags might be appropriate for studying exercise, based on
the specication tests in column (3) and (8). One more day of vigorous activities of one's spouse
in a week leads to 0.26 more days for husbands and 0.13 more days for wives, based on column
(4) and (8). In line with ndings for drinking, the instantaneous eects of wives' participation
in physical exercise are larger than husbands' own previous habit, though they are not signicant
in the reduced form estimations. The same conclusion can be drawn if we examine the spousal
spillover eects on whether the respondent participates in vigorous activities at least once a week,
as shown in Table C2 in appendix.
91
Social Contact. Table 4.10 demonstrates that for weekly contact with relatives in person,
both spouses have strong and positive in
uence on each other. The spillover eect of wives on
husbands is larger and more signicant. AR(2) tests and Hansen J tests suggest that it might be
necessary to allow for further lag dependence of women, though GMM estimates for instantaneous
spousal eects are similar across specications. Having a spouse contacting relatives weekly in
person implies an increase in the probability of the respondent doing so by 20 percentage points
for men and 21 to 28 percentage points for women.
4.7 Robustness Checks
In this section, we discuss the potential sources of bias of GMM estimates and perform robustness
checks.
4.7.1 Dierence GMM
There are concerns about the validity of the additional moment conditions in system GMM. How-
ever, most of the Dierence in Hansen tests results are consistent with exclusion restrictions of the
rst dierences as instruments in the level equations. In addition, we show that most of our ndings
are robust to the methods applied and assumptions imposed. Table 4.11 to 4.16 report the results
when we relax the stationarity assumption and utilize moment conditions only for rst dierenced
equations. We nd signicant estimates for spousal spillover eects in self-reported health, depres-
sive symptoms and exercise for both genders, which are in line with system GMM results. There
are also statistically signicant spillover eects from wives' smoking intensity and social contact
on their husbands, based on dierence GMM results. However, the estimated spillover eects in
smoking and drinking behaviors are insignicant in dierence GMM estimations.
92
Table 4.11: Dierence GMM Results on Good Self-reported Health (SRH)
(1) (2) (3) (4) (5) (6)
Men Women
L.(own good SRH) 0.0486 0.0679 0.0423 0.0941*** 0.0771 0.0896***
(0.0315) (0.0519) (0.0320) (0.0301) (0.0471) (0.0297)
L2.(own good SRH) 0.0350 0.00212
(0.0334) (0.0311)
spouse good SRH 0.219* 0.186 0.350*** 0.377***
(0.116) (0.125) (0.113) (0.109)
L.(spouse good SRH) -0.0190 -0.0252
(0.0243) (0.0231)
Observations 6718 4838 6759 6850 4916 6916
Number of couples 1881 1606 1886 1944 1636 1956
AR(1) test p val 1.14e-45 1.26e-25 7.34e-46 2.32e-45 1.45e-25 1.38e-49
AR(2) test p val 0.752 0.327 0.783 0.620 0.828 0.501
Hansen J test p val 0.00836 0.0103 0.00169 0.0342 0.0327 0.0135
Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05,
* p < 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living
daughters and sons, number of living siblings, respondent's and spouse's education dummies and
length of marriage; 2) log per capita de
ated household wealth.
93
Table 4.12: Dierence GMM Results on Depressive Symptoms (CES-D 8 Score)
(1) (2) (3) (4) (5) (6)
Men Women
L.(own CESD score) 0.00782 0.00816 0.0151 0.0947*** 0.111*** 0.0777***
(0.0279) (0.0548) (0.0290) (0.0266) (0.0429) (0.0270)
L2.(own CESD score) 0.0202 0.0167
(0.0361) (0.0299)
spouse CESD score 0.353*** 0.376*** 0.538*** 0.510***
(0.116) (0.124) (0.123) (0.124)
L.(spouse CESD score) -0.0252 -0.0257
(0.0187) (0.0293)
Observations 6641 4771 6680 6773 4857 6829
Number of couples 1872 1597 1877 1931 1626 1940
AR(1) test p val 5.46e-30 9.52e-19 5.52e-35 5.43e-49 2.27e-31 2.92e-41
AR(2) test p val 0.445 0.570 0.313 0.254 0.391 0.0436
Hansen J test p val 0.00926 0.00213 0.00391 0.109 0.0237 0.00471
Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05,
* p < 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living
daughters and sons, number of living siblings, respondent's and spouse's education dummies and
length of marriage; 2) log per capita de
ated household wealth.
94
Table 4.13: Dierence GMM Results on Smoking Status (= 1 if smokes now)
(1) (2) (3) (4) (5) (6)
Men Women
L.(own smoking status) 0.358*** 0.349*** 0.363*** 0.199** 0.195 0.186***
(0.0833) (0.0925) (0.0785) (0.0783) (0.129) (0.0702)
L2.(own smoking status) 0.0902** 0.0456
(0.0456) (0.0442)
spouse's smoking status 0.156* 0.0440 0.00611 0.0373
(0.0901) (0.0954) (0.0515) (0.0622)
L.(spouse's smoking status) 0.0298 0.0304
(0.0375) (0.0253)
Observations 6981 5048 6969 7174 5147 7138
Number of couples 1926 1665 1918 2019 1712 1997
AR(1) test p val 2.14e-08 0.0000136 8.25e-09 0.00000273 0.00332 0.00000200
AR(2) test p val 0.0828 0.136 0.102 0.716 0.897 0.780
Hansen J test p val 0.0259 0.0935 0.0158 0.0138 0.0355 0.0252
Note: Two-step corrected robust standard errors in the parenthesis. *** p< 0:01, ** p< 0:05, * p< 0:1. All
columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons, number
of living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per capita
de
ated household wealth.
95
Table 4.14: Dierence GMM Results on Smoking Intensity (# cigarettes/day)
(1) (2) (3) (4) (5) (6)
Men Women
L.(# cigarettes/day smokes) 0.215** 0.159 0.166 0.286*** 0.308** 0.253***
(0.0903) (0.121) (0.103) (0.0912) (0.124) (0.0932)
L2.(# cigarettes/day smoked) 0.0441 0.119**
(0.0448) (0.0472)
# cigarettes/day spouse smokes 0.0192 0.0200 0.0627 0.106*
(0.0449) (0.0358) (0.0796) (0.0617)
L.(# cigarettes/day spouse smokes) 0.000708 -0.00139
(0.0241) (0.0231)
Observations 6734 4839 6694 6528 4701 6524
Number of couples 1919 1632 1900 1824 1569 1819
AR(1) test p val 0.000226 0.00834 0.00222 0.000643 0.0151 0.000879
AR(2) test p val 0.401 0.177 0.665 0.0558 0.534 0.0553
Hansen J test p val 0.0351 0.106 0.0190 0.176 0.223 0.170
Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05, * p < 0:1. All
columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons, number of
living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per capita de
ated
household wealth.
96
Table 4.15: Dierence GMM Results on Drinking Frequency (# days/week drinks)
(1) (2) (3) (4) (5) (6)
Men Women
L.(# days/week drinks) 0.0336 0.0639 0.0553 0.129*** 0.212*** 0.179***
(0.0421) (0.0718) (0.0438) (0.0463) (0.0780) (0.0504)
L2.(# days/week drinks) 0.0567 0.0628*
(0.0469) (0.0379)
# days/week spouse drinks 0.255* 0.335** 0.294* 0.491***
(0.153) (0.146) (0.150) (0.158)
L.(# days/week spouse drinks) 0.0499 -0.0183
(0.0395) (0.0332)
Observations 4023 2678 4020 4083 2697 4090
Number of couples 1448 1185 1457 1483 1203 1496
AR(1) test p val 7.07e-24 1.64e-12 3.05e-22 6.55e-21 1.76e-11 1.26e-17
AR(2) test p val 0.864 0.432 0.894 0.785 0.410 0.456
Hansen J test p val 0.146 0.208 0.0723 0.646 0.870 0.0772
Note: Two-step corrected robust standard errors in the parenthesis. *** p< 0:01, ** p< 0:05, * p< 0:1.
All columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons,
number of living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per
capita de
ated household wealth.
97
Table 4.16: Dierence GMM Results on Frequency of Exercise (# days/week participating in
vigorous physical activities, VPA)
(1) (2) (3) (4) (5) (6)
Men Women
L.(# d/w w/ VPA) 0.0961*** 0.193*** 0.107*** 0.146*** 0.129*** 0.155***
(0.0245) (0.0336) (0.0252) (0.0230) (0.0329) (0.0234)
L2.(# d/w w/ VPA) 0.111*** 0.00999
(0.0224) (0.0231)
# d/w spouse w/ VPA 0.297** 0.283** 0.132 0.135
(0.116) (0.133) (0.0857) (0.0966)
L.(# d/w spouse w/ VPA) 0.0260 0.00953
(0.0222) (0.0184)
Observations 7020 5077 7013 7212 5182 7193
Number of couples 1934 1672 1931 2022 1717 2019
AR(1) test p val 1.83e-67 1.56e-55 4.07e-69 5.99e-64 1.29e-40 1.00e-68
AR(2) test p val 0.0628 0.124 0.150 0.279 0.562 0.357
Hansen J test p val 0.00000339 0.0500 0.00000413 0.346 0.0935 0.358
Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05, * p < 0:1.
All columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons,
number of living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per
capita de
ated household wealth.
98
Table 4.17: Dierence GMM Results on Social Contact (= 1 if any weekly contact with relatives
in person)
(1) (2) (3) (4) (5) (6)
Men Women
L.(wkly contact w/ relativ) 0.146*** 0.170*** 0.166*** 0.151*** 0.235*** 0.136***
(0.0352) (0.0523) (0.0372) (0.0344) (0.0497) (0.0343)
L2.(wkly contact w relativ) 0.0342 0.135***
(0.0383) (0.0335)
spouse's wkly contact w/ relativ 0.175* 0.216** 0.207* 0.285**
(0.0939) (0.102) (0.114) (0.144)
L.(spouse's wkly contact w/ relativ) 0.0128 0.0202
(0.0307) (0.0295)
Observations 4371 2991 4408 4500 3091 4535
Number of couples 1456 1131 1472 1508 1188 1528
AR(1) test p val 1.10e-41 1.04e-24 1.63e-41 3.39e-47 9.28e-25 5.46e-45
AR(2) test p val 0.363 0.858 0.641 0.00335 0.939 0.172
Hansen J test p val 0.00851 0.0132 0.000480 0.000512 0.156 0.0000957
Note: Two-step corrected robust standard errors in the parenthesis. *** p< 0:01, ** p< 0:05, * p< 0:1. All
columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons, number of
living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per capita de
ated
household wealth.
99
4.7.2 Path Dependence and Spousal In
uence
Comparing OLS, FE and GMM results, we nd that the causal estimates of spousal spillover eects
are usually larger than one's own path dependence, except for smoking behaviors. Concerns arise
as to whether the GMM estimates of spousal spillover eects are upward biased while those of state
dependence are downward biased in the structural model, because of wrong lag structures. We
test the robustness of system GMM estimates of state dependence by estimating the reduced form
equations and allowing for more own lags.
The self 1-period lag coecients are very similar in the reduced form results, while the lagged
spousal spillover eects are rarely signicant. With inclusion of further self-lags, the estimated state
dependence generally increases, while the instantaneous spousal spillover eect estimate decreases
for some outcomes, and increases for others. However, the extent of changes in the coecients are
limited such that the results still imply that path dependence is less important than instantaneous
spousal in
uence. This might be true for our subjective measures of health status which are
based on self reports: self reported good health and CES-D 8 index, for which state dependence
is relatively low. We also examined the spousal spillover eects in chronic conditions, measured
by the number of severe chronic conditions ever diagnosed by doctors
13
, where we show in Table
C3 in appendix that instantaneous spousal in
uence is insignicant and much smaller than state
dependence of one's own health
14
. For health behaviors, except for smoking which is proven to be
highly addictive, social in
uence and peer eects in exercising, drinking and contacting relatives
could possibly be more important than one's own past behaviors.
13
The severe diseases are diabetes, cancer, chronic lung disease, heart attack and stroke.
14
Here the spousal in
uence is not necessarily a spillover of health status, but is likely to be a spillover in diagnosis.
100
4.7.3 Sample Selection
As the estimation equation requires at least 3 consecutive (when m = 1) waves of observations for
each respondent, one potential source of bias for the estimated spousal spillover eects might be
sample selection or attrition in the panel data, if attrition is correlated with endogenous regressors.
By taking rst dierences, we can remove all selection bias which is related to the time-invariant
heterogeneity. However, there could remain selection bias on unobservables or shocks that are time-
varying. If couples with stabler relationships have stronger in
uence on each other, and more likely
to remain in more waves, our GMM estimates might serve as upper bounds of the true spillover
eects. Couples might separate due to growing care burden in the event of major illness to one
spouse or even death, which might cause our causal estimates to be upward biased as well. On
the other hand, if couples that are both healthier are more likely to move and hence attrite, our
health spillover estimates might be downward biased. Therefore, a priori, it is not clear what the
direction of the sample selection bias is if any. In ELSA, we nd that the cross-wave divorce rate
is very low and the main sample attrition is from the failure of follow-up. Therefore, we argue it is
unlikely that our analysis is biased from selective attrition due to marital dissolution.
4.8 Conclusion
This chapter studies the widely documented spousal concordance in health, albeit in a causal way.
We utilize data from the English Longitudinal Study of Aging and estimate in a dynamic panel
model the spousal spillover regarding health status and health related behaviors among elderly
couples in England. To address endogeneity biases from assortative mating and shared environment,
we apply the system GMM to an dynamic model of health with individual xed eects. GMM
estimations imply signicant and positive eects on health between elderly couples, with women's
101
mental health being aected more by their spouses. We also nd strong causal eects of spousal
smoking, drinking and physical activity, as well as social contact, which could be an important
channel leading to the inter-spousal concordance in health outcomes. While wives are a bigger
in
uence when it comes to their husband's smoking likelihoods, men aect more than women as to
how much they smoke, among the English older adults. However, the gender dierences in spillover
eects in both health and health behaviors are not substantial.
We explored both reduced form and structural estimations, as well as dierent lag structures.
We nd that allowing for state dependence with more distant lags is important for some health
measures but not for the others, implying a heterogeneity in the persistence of health status and
behaviors. The comparison between structural and reduced form estimations suggests the impor-
tance of instantaneous spillover eects, while we fail to nd evidence for the lagged spousal eects
in most of the health measures and behaviors considered. We nd the path dependence in health
and health behaviors (except smoking) is relatively low, in contrast with the correlation coecients
and the instantaneous spousal spillover eects. However, the state dependence also vary across
dierent dimensions of health status and behaviors. For health behaviors that are often conducted
with peers, and mental health, the spillover eects are larger than state dependence.
Sensitivity analysis shows that our ndings on the spousal spillover eects in self-reported
health, depressive symptoms, and exercise for both genders, as well as social contact for men are
robust to less restrictive methods (dierence GMM). This chapter has not explicitly addressed
the bias from sample selection or attrition by modeling its relation to health and marital status.
A more comprehensive framework on the mechanisms underlying the spillover eects, which also
incorporates the heterogeneity in such eects and the sample selection process, is warranted for
future research.
102
The analysis of this chapter has important policy implications. Understanding the extent to
which couples causally aect each other's health is relevant for comprehensive cost-benet calcu-
lations of health policies, since interventions towards improving individual heath have potentially
signicant spillovers along social ties. Health policies targeted at couple level might be more e-
cient to reduce health disparities in certain health dimensions and behaviors (e.g., mental health,
smoking, drinking and exercise), which involves strong social or spousal in
uence. Furthermore,
with aging population and declining familial support, individuals have to increasingly rely on their
spouses in later life, especially in the face of deteriorating health. As the spillovers from spouses
accumulate and the burden of care grows over time, targeting elderly couples, which have been
shown to manifest signicant spousal concordance in health, might be especially important for
better delivering health-care services.
103
Chapter 5
Conclusion
Population aging is a global reality, and it is projected to accelerate in the next three decades.
Bulk of this will be driven by increasing share of the elderly population in developing countries.
These populations not only grew up amidst low levels of economic development, but are also
likely to experience aging with relatively low access to resources. Compared to the more advanced
economies, relatively little is known about aging in environments of relative scarcity. My research
endeavors have been motivated by the aim of narrowing this gap. This dissertation explores three
aspects of health in later-life: its early-life origins, population level indicators of deterioration, and
the eect of closest peers' health on one's own health.
In Chapter 2, exploiting geographic variation in malaria endemicity and a plausibly exogenous
nationwide malaria eradication program, I use a dierence-in-dierence identication strategy to
study the eect of childhood exposure to malaria on later life cognition among a sample of Indian
older adults. The results show that malaria eradication, while aimed at reducing malaria-related
morbidity, also had long term eects on human capital via improved cognitive capacity among the
post-eradication cohorts in erstwhile malarious areas. These results suggest that improvements
in disease environments today can have potential gains in the future as well via improvements in
health.
104
Chapter 3 is an attempt to carefully study as to what is it that we measure when using height as a
covariate for measures of health pertinent to old age. Since human beings experience varying degrees
of height loss with aging, we rst document the extent of height loss in English and Indonesian
elderly populations. We then show that breaking down current measure of height into past height
and the loss in height over time is a useful exercise. Height loss emerges as a signicant covariate of
cognition, physical function and disability, as well as a predictor of mortality. Just as adult height
has long been used a measure of health at the population level, height loss could potentially be
useful as a population level measure of aging.
Results from Chapter 4 suggest that there are signicant spousal impacts in the health domain of
human capital among older couples in England. By controlling for endogeneity due to assortative
mating and shared environment, we nd evidence for strong spillover eects from spousal self-
reported health, depressive symptoms, smoking behaviors, drinking frequency, physical activities
and social contact. Gender dierence are found for some outcomes but not much. For instance,
women's mental health and smoking intensity seem to be more aected by their spouses, while the
opposite is true for men's smoking status.
105
Bibliography
Arellano, Manuel, and Olympia Bover. 1995. \Another look at the instrumental variable estimation
of error-components models." Journal of Econometrics 68(1): 29 { 51.
Arellano, Manuel, and Stephen Bond. 1991. \Some Tests of Specication for Panel Data: Monte
Carlo Evidence and an Application to Employment Equations." The Review of Economic Studies
58(2): 277{297.
Arokiasamy, P., Sulabha Parasuraman, T. V. Sekher, and H. Lhungdim. 2013. \Study on global
AGEing and adult health (SAGE) Wave 1: India National Report." WHO.
Auyeung, Tung Wai, Jenny Shun Wah Lee, Jason Leung, Timothy Kwok, Ping Chung Leung, and
Jean Woo. 2010. \Eects of height loss on morbidity and mortality in 3145 community-dwelling
Chinese older women and men: a 5-year prospective study." Age and Ageing 39(6): 699{704.
Banks, James, and Fabrizio Mazzonna. 2012. \The eect of education on old age cognitive abilities:
Evidence from a regression discontinuity design." The Economic Journal 122(560): 418{448.
Banks, James, Elaine Kelly, and James P. Smith. 2013. \Spousal Health Eects - the Role of
Selection." National Bureau of Economic Research Working Paper 19438.
Barker, David. 1992. \The fetal origins of diseases of old age." European journal of clinical nutrition
46(Suppl 3): 3{9.
Becker, Gary S. 1973. \A Theory of Marriage: Part I."Journal of Political Economy 81(4): 813{846.
Becker, Gary S. 1974. \A Theory of Marriage: Part II." Journal of Political Economy 82(2): S11{
S26.
Benjamins, Maureen Reindl, Robert A. Hummer, Isaac W. Eberstein, and Charles B. Nam. 2004.
\Self-reported health and adult mortality risk: An analysis of cause-specic mortality." Social
Science & Medicine 59(6): 1297{1306.
Bleakley, Hoyt. 2010. \Malaria Eradication in the Americas: A Retrospective Analysis of Childhood
Exposure." American Economic Journal: Applied Economics 2(2): 1{45.
Blundell, Richard, and Stephen Bond. 1998. \Initial conditions and moment restrictions in dynamic
panel data models." Journal of Econometrics 87(1): 115 { 143.
Briot, Karine, Erik Legrand, Denis Pouchain, St ephanie Monnier, and Christian Roux. 2010. \Ac-
curacy of patient-reported height loss and risk factors for height loss among postmenopausal
women." Canadian Medical Association Journal 182(6): 558{562.
106
Brown, Heather, Arne Risa Hole, and Jennifer Roberts. 2014. \Going the same `weigh': spousal
correlations in obesity in the United Kingdom." Applied Economics 46(2): 153{166.
Case, Anne, and Christina Paxson. 2008a. \Height, Health, and Cognitive Function at Older Ages."
American Economic Review 98(2): 463{67.
Case, Anne, and Christina Paxson. 2008b. \Stature and Status: Height, Ability, and Labor Market
Outcomes." Journal of Political Economy 116(3): 499{532.
Cawley, John, Johanna Catherine Maclean, and Asia Sikora Kessler. 2017. \Reporting error in
weight and height among older adults: Implications for estimating healthcare costs." The Journal
of the Economics of Ageing 9(Supplement C): 122{144.
Cawthon, Peggy M. 2011. \Gender Dierences in Osteoporosis and Fractures." Clinical Or-
thopaedics and Related Research 469(7): 1900{1905.
Chang, Simon, Belton Fleisher, Seonghoon Kim, and Shi yung Liu. 2014. \Long-Term Health Eects
of Malaria Exposure around Birth: Evidence from Colonial Taiwan." Economic Development and
Cultural Change 62(3): 519{536.
Chiappori, Pierre-Andr e, Sonia Orece, and Climent Quintana-Domeque. 2012. \Fatter Attraction:
Anthropometric and Socioeconomic Matching on the Marriage Market." Journal of Political
Economy 120(4): 659{695.
Christakis, Nicholas A., and James H. Fowler. 2007. \The Spread of Obesity in a Large Social
Network over 32 Years." New England Journal of Medicine 357(4): 370{379. PMID: 17652652.
Clark, Andrew E., and Fabrice" Etil e. 2006. \Don't give up on me baby: Spousal correlation in
smoking behaviour." Journal of Health Economics 25(5): 958 { 978.
Clark, Andrew E., and Fabrice" Etil e. 2011. \Happy house: Spousal weight and individual well-
being." Journal of Health Economics 30(5): 1124 { 1136.
Clark, Andrew E., and Youenn Loh eac. 2007. \\It" wasn't me, it was \them!" Social in
uence in
risky behavior by adolescents." Journal of Health Economics 26(4): 763 { 784.
Cline, Martha G., Keith E. Meredith, John T. Boyer, and Benjamin Burrows. 1989. \Decline of
height with age in adults in a general population sample: estimating maximum height and distin-
guishing birth cohort eects from actual loss of stature with aging." Human Biology 61(3): 415{
425.
Cunha, Flavio, and James J. Heckman. 2008. \Formulating, Identifying and Estimating the
Technology of Cognitive and Noncognitive Skill Formation." The Journal of Human Resources
43(4): 738{782.
Currie, Janet. 2009. \Healthy, Wealthy, and Wise: Socioeconomic Status, Poor Health in Child-
hood, and Human Capital Development." Journal of Economic Literature 47(1): 87{122.
Currie, Janet, and Douglas Almond. 2011. \Chapter 15 - Human capital development before age
ve." In Handbook of Labor Economics, edited by David Card and Orley Ashenfelter, 4: 1315 {
1486. Elsevier.
107
Currie, Janet, and Tom Vogl. 2013. \Early-Life Health and Adult Circumstance in Developing
Countries." Annual Review of Economics 5(1): 1{36.
Cutler, David, Winnie Fung, Michael Kremer, Monica Singhal, and Tom Vogl. 2010. \Early-Life
Malaria Exposure and Adult Outcomes: Evidence from Malaria Eradication in India." American
Economic Journal: Applied Economics 2(2): 72{94.
Davillas, Apostolos, and Stephen Pudney. 2017. \Concordance of health states in couples: Analysis
of self-reported, nurse administered and blood-based biomarker data in the UK Understanding
Society panel." Journal of Health Economics 56: 87 { 102.
Deaton, Angus. 2007. \Height, health, and development." Proceedings of the National Academy of
Sciences 104(33): 13232{13237.
Dohmen, Thomas, Armin Falk, David Human, and Uwe Sunde. 2012. \The Intergenerational
Transmission of Risk and Trust Attitudes." The Review of Economic Studies 79(2): 645{677.
Eisenberg, Daniel, Ezra Golberstein, and Janis L. Whitlock. 2014. \Peer eects on risky behaviors:
New evidence from college roommate assignments." Journal of Health Economics 33: 126{138.
Fernihough, Alan, and Mark E. McGovern. 2015. \Physical stature decline and the health status
of the elderly population in England." Economics & Human Biology 16: 30{44.
Fletcher, Jason M. 2009. \All in the family: Mental health spillover eects between working
spouses." The BE journal of economic analysis & policy 9(1).
Fletcher, Jason M. 2010. \Social interactions and smoking: evidence using multiple student cohorts,
instrumental variables, and school xed eects." Health Economics 19(4): 466{484.
Fletcher, Jason M., and David E. Frisvold. 2009. \Higher education and health investments: Does
more schooling aect preventive health care use?" Journal of Human Capital 3(2): 144{176.
Fletcher, Jason M., and Ryne Marksteiner. 2017. \Causal Spousal Health Spillover Eects and
Implications for Program Evaluation." American Economic Journal: Economic Policy 9(4): 144{
66.
Fogel, Robert. 1986. \Physical Growth as a Measure of the Economic Well-being of Populations:
The Eighteenth and Nineteenth Centuries." In Human growth: a comprehensive treatise, edited
by Falkner Frank and James M. Tanner, 3: 263{81. New York: Plenum Press.
Fowler, James H, and Nicholas A Christakis. 2008. \Dynamic spread of happiness in a large social
network: longitudinal analysis over 20 years in the Framingham Heart Study." BMJ 337.
Frankenberg, Elizabeth, and Duncan Thomas. 2000. \The Indonesia Family Life Survey (IFLS):
Study Design and Results from Waves 1 and 2." DRU-2238/1-NIA/NICHD. Santa Monica:
RAND.
Glymour, M. M., I. Kawachi, C. S. Jencks, and L. F. Berkman. 2008. \Does childhood schooling
aect old age memory or mental status? Using state schooling laws as natural experiments."
Journal of Epidemiology & Community Health 62(6): 532{537.
Guven, Cahit, and Wang Sheng Lee. 2013. \Height and cognitive function at older ages: is height
a useful summary measure of early childhood experiences?" Health Economics 22(2): 224{233.
108
Guven, Cahit, and Wang-Sheng Lee. 2015. \Height, aging and cognitive abilities across Europe."
Economics & Human Biology 16: 16{29.
Hillier, Teresa A., Li-Yung Lui, Deborah M. Kado, E. S. LeBlanc, Kimberly K. Vesco, Douglas C.
Bauer, Jane A. Cauley, Kristine E. Ensrud, Dennis M. Black, Marc C. Hochberg, and Steven R.
Cummings. 2012. \Height loss in older women: Risk of hip fracture and mortality independent
of vertebral fractures." Journal of Bone and Mineral Research 27(1): 153{159.
Holding, Penny A., and Patricia K. Kitsao-Wekulo. 2004. \Describing the burden of malaria on
child development: what should we be measuring and how shoud we be measuring it?" The
American Journal of Tropical Medicine and Hygiene 71(2): 71{79.
Huang, T. L., M. C. Carlson, A. L. Fitzpatrick, L. H. Kuller, L. P. Fried, and P. P. Zandi. 2008.
\Knee height and arm span: A re
ection of early life environment and risk of dementia." Neu-
rology 70(19 Part 2): 1818{1826.
Huang, Wei, and Yi Zhou. 2013. \Eects of education on cognition at older ages: Evidence from
China's Great Famine." Social Science & Medicine 98(Supplement C): 54{62.
Huang, Wei, Xiaoyan Lei, Geert Ridder, John Strauss, and Yaohui Zhao. 2013. \Health, Height,
Height Shrinkage, and SES at Older Ages: Evidence from China." American Economic Journal:
Applied Economics 5(2): 86{121.
Kalmijn, Matthijs. 1994. \Assortative Mating by Cultural and Economic Occupational Status."
American Journal of Sociology 100(2): 422{452.
Krege, John H., David Kendler, Kelly Krohn, Harry Genant, Jahangir Alam, Pierre-Yves Berclaz,
Barbara Coey, and Corina Loghin. 2015. \Relationship Between Vertebral Fracture Burden,
Height Loss, and Pulmonary Function in Postmenopausal Women With Osteoporosis." Journal
of Clinical Densitometry 18(4): 506{511.
LaFave, Daniel, and Duncan Thomas. 2016. \Height and cognition at work: Labor market produc-
tivity in a low income setting." Economics & Human Biology 25: 52{64.
Llewellyn, David J., Iain A. Lang, Kenneth M. Langa, Graciela Muniz-Terrera, Caroline L. Phillips,
Antonio Cherubini, Luigi Ferrucci, and David Melzer. 2010. \Vitamin d and risk of cognitive
decline in elderly persons." Archives of Internal Medicine 170(13): 1135{1141.
Lucas, Adrienne M. 2010. \Malaria Eradication and Educational Attainment: Evidence from
Paraguay and Sri Lanka." American Economic Journal: Applied Economics 2(2): 46{71.
Manski, Charles F. 1993. \Identication of endogenous social eects: The re
ection problem." The
review of economic studies 60(3): 531{542.
Mare, Robert D. 1991. \Five Decades of Educational Assortative Mating." American Sociological
Review 56(1): 15{32.
Masunari, Naomi, Saeko Fujiwara, Fumiyoshi Kasagi, Ikuno Takahashi, Michiko Yamada, and
Toshitaka Nakamura. 2012. \Height loss starting in middle age predicts increased mortality in
the elderly." Journal of Bone and Mineral Research 27(1): 138{145.
109
Maurer, J urgen. 2010. \Height, education and later-life cognition in Latin America and the
Caribbean." Economics & Human Biology 8(2): 168{176.
McArdle, John J., James P. Smith, and Robert Willis. 2011. \Cognition and Economic Outcomes
in the Health and Retirement Survey." In Explorations in the Economics of Aging, edited by
David A. Wise, 209{233. Chicago: University of Chicago Press.
McGovern, Mark E. 2014. \Comparing the relationship between stature and later life health in six
low and middle income countries." The Journal of the Economics of Ageing 4: 128{148.
Meyler, Deanna, Jim P. Stimpson, and M. Kristen Peek. 2007. \Health concordance within couples:
a systematic review." Social Science & Medicine 64(11): 2297{2310.
Michaud, Pierre-Carl, and Arthur van Soest. 2008. \Health and wealth of elderly couples: Causality
tests using dynamic panel data models." Journal of Health Economics 27(5): 1312 { 1325.
Mosca, Irene, and Robert E. Wright. 2016. \Height and cognition at older ages: Irish evidence."
Economics Letters 149: 98{101.
Old, Jerry L., and Michelle Calvert. 2004. \Vertebral Compression Fractures in the Elderly." Amer-
ican Family Physician 69(1): 111{116.
Phillips, Drystan, Yu-Chen Lin, Jenny Wight, Sandy Chien, and Jinkook Lee. 2017. Harmonized
ELSA Documentation. Version E. Los Angeles: Center for Global Aging Research, University of
Southern California. https://g2aging.org/?section=pagepageid=18.
Powdthavee, Nattavudh. 2009. \I can't smile without you: Spousal correlation in life satisfaction."
Journal of Economic Psychology 30(4): 675 { 689.
Powell, Lisa M., John A. Tauras, and Hana Ross. 2005. \The importance of peer eects, cigarette
prices and tobacco control policies for youth smoking behavior." Journal of Health Economics
24(5): 950 { 968.
Rantanen, T., J.M. Guralnik, D. Foley, K. Masaki, S. Leveille, J. D. Curb, and L. White. 1999.
\Midlife hand grip strength as a predictor of old age disability." JAMA 281(6): 558{560.
Raspe, Heiner, Angelika Hueppe, and Hannelore Neuhauser. 2008. \Back pain, a communicable
disease?" International Journal of Epidemiology 37(1): 69{74.
Rawlings, Samantha B. 2016. \Gender, Race, and Heterogeneous Eects of Epidemic Malaria on
Human Capital and Income." Economic Development and Cultural Change 64(3): 509{543.
Roodman, David. 2009. \How to do xtabond2: An introduction to dierence and system GMM in
Stata." Stata Journal 9(1): 86{136.
Schultz, T. Paul, and John Strauss, ed. 2008. Handbook of Development Economics. Amsterdam:
North-Holland Press.
Schulz, Richard, and Paula R Sherwood. 2008. \Physical and mental health eects of family care-
giving." Journal of Social Work Education 44(sup3): 105{113.
Sohn, Kitae. 2015. \Is leg length a biomarker of early life conditions? Evidence from a historically
short population." American Journal of Human Biology 27(4): 538{545.
110
Sorkin, John D., Denis C. Muller, and Reubin Andres. 1999. \Longitudinal Change in Height
of Men and Women: Implications for Interpretation of the Body Mass Index: The Baltimore
Longitudinal Study of Aging." American Journal of Epidemiology 150(9): 969{977.
Steckel, Richard H. 1995. \Stature and the Standard of Living." Journal of Economic Literature
33(4): 1903{1940.
Steckel, Richard H. 2009. \Heights and human welfare: Recent developments and new directions."
Explorations in Economic History 46(1): 1{23.
Strauss, John, and Duncan Thomas. 2008. \Health over the Life Course." In Handbook of De-
velopment Economics, edited by T. Paul Schultz and John Strauss, 4: 3375{3474. Amsterdam:
North-Holland Press.
Strauss, John, Firman Witoelar, and Bondan Sikoki. 2016. \The Fifth Wave of the Indonesia Family
Life Survey (IFLS5): Overview and Field Report." WR-1143/1-NIA/NICHD. Santa Monica:
RAND.
Strauss, John, Firman Witoelar, Bondan Sikoki, and Anna Marie Wattie. 2009. \The Fourth
Wave of the Indonesian Family Life Survey (IFLS4): Overview and Field Report." WR-675/1-
NIA/NICHD. Santa Monica: RAND.
Strauss, John, Kathleen Beegle, Bondan Sikoki, Agus Dwiyanto, Yulia Herawati, and Firman
Witoelar. 2004. \The Third Wave of the Indonesia Family Life Survey (IFLS): Overview and
Field Report." WR-144/1-NIA/NICHD. Santa Monica: RAND.
Thomas, Duncan, and John Strauss. 1992. \Prices, infrastructure, household characteristics and
child height." Journal of Development Economics 39(2): 301{331.
Thomas, Duncan, and John Strauss. 1997. \Health and wages: Evidence on men and women in
urban Brazil." Journal of Econometrics 77(1): 159{185.
Thomas, Duncan, and John Strauss. 1998. \Health, Nutrition, and Economic Development." Jour-
nal of Economic Literature 36(2): 766{817.
Thomas, Duncan, John Strauss, and Maria-Helena Henriques. 1990. \Child survival, height for age
and household characteristics in Brazil." Journal of Development Economics 33(2): 197{234.
Thomas, Duncan, John Strauss, and Maria-Helena Henriques. 1991. \How does mother's education
aect child height?" Journal of human resources 26(2): 183{211.
Venkataramani, Atheendar S. 2012. \Early life exposure to malaria and cognition in adulthood:
Evidence from Mexico." Journal of Health Economics 31(5): 767 { 780.
Wannamethee, S., Shaper A., Lennon L., and Whincup P. H. 2006. \Height loss in older men:
Associations with total mortality and incidence of cardiovascular disease." Archives of Internal
Medicine 166(22): 2546{2552.
Weir, David, Margaret Lay, and Kenneth Langa. 2014. \Economic development and gender inequal-
ity in cognition: A comparison of China and India, and of SAGE and the HRS sister studies."
The Journal of the Economics of Ageing 4: 114{125.
111
Wilson, Sven E. 2002. \The health capital of families: an investigation of the inter-spousal corre-
lation in health status." Social Science & Medicine 55(7): 1157{1172.
Windmeijer, Frank. 2005. \A nite sample correction for the variance of linear ecient two-step
GMM estimators." Journal of Econometrics 126(1): 25 { 51.
112
Appendix A
Chapter 2 Appendix
Tables
113
Table A1: Additional birth-year weather controls
Dependent variable: Immediate Delayed Digit span Digit span Verbal
recall recall forward backward
uency
(1) (2) (3) (4) (5)
Post x malaria index 0.150 0.268*** 0.0895 0.180** 0.179*
(0.106) (0.0966) (0.0656) (0.0762) (0.0964)
Observations 3,960 3,960 3,960 3,960 3,960
Notes: Standard error clustered at the district level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1.
Dependent variable are standardized scores.
Sample consists of individuals from rural areas, born between 1938-1952, and 1960-1974.
Estimations control for average rainfall and temperature in respondent's birth year, include state-post xed
eects, district-specic linear trends.
Monthly rainfall and temperature data at the district level was obtained from India water portal (sourced
from the Indian Meteorological Department), downloaded for the time range 1901-2002
1
.
114
Appendix B
Chapter 3 Appendix
Data Cleaning
We set the lower limit for measured height at 100cm. We did not need to drop any observations in ELSA
as per this rule. Measurement error seemed highest in ELSA Wave 4, and lowest in Wave 6 (Latest Height
Measurement), as seen by the extent of outliers for both men and women. In the IFLS, measurement error
seemed highest in Wave 4 (Height 7 years ago). We dropped 17 respondents from Wave 4 and 2 respondents
from Wave 5 for whom measured height is less than 100cm.
For the extent of change in height, we dropped respondents for whom height gain over the years is more
than 10cm (shrinkage<-10cm), or height loss is more than 10cm, trimming away the outliers. We dropped 18
respondents in the ELSA sample as per this rule. In the IFLS sample, we dropped the following respondents:
30 for whom height gain over 7 years is more than 10cm, 2 each for whom height gain over 14 and 17 years
is more than 10cm. 45, 38 and 20 respondents exhibit height shrinkage more than 10cm over 7, 14 and 17
years respectively- we dropped these as well.
115
Tables
Table B1: Limb Length and Shrinkage in IFLS
Male sample (1) (2) (3) (4) (5) (6)
shrinkage shrinkage shrinkage shrinkage shrinkage shrinkage
VARIABLES over 7 yrs over 7 yrs over 14 yrs over 14 yrs over 17 yrs over 17 yrs
Upper arm length 0.0295 0.00726 0.00169
(0.0181) (0.0300) (0.0253)
Knee Height 0.0263 0.0127 0.0107
(0.0207) (0.0196) (0.0207)
Observations 970 970 970 970 970 970
R-squared 0.054 0.055 0.079 0.080 0.099 0.099
Female sample (1) (2) (3) (4) (5) (6)
shrinkage shrinkage shrinkage shrinkage shrinkage shrinkage
VARIABLES over 7 yrs over 7 yrs over 14 yrs over 14 yrs over 17 yrs over 17 yrs
Upper arm length -0.0565* -0.00245 -0.00265
(0.0331) (0.0311) (0.0375)
Knee Height -0.0279 -0.0210 -0.00522
(0.0203) (0.0239) (0.0256)
Observations 1,086 1,086 1,086 1,086 1,086 1,086
R-squared 0.049 0.046 0.096 0.097 0.107 0.107
Note: Robust standard errors clustered at district level in parentheses. *** p< 0.01, ** p<0.05,
* p<0.1. Age, Education, SES controls suppressed. IFLS results control for district xed eects.
116
Table B2: Height Shrinkage and ADL in ELSA
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Panel A: Male sample Panel B: Female sample
Dependent Variable: ADL Index 4 years 8 years 4 years 8 years
Current Height -0.00138 -0.00701**
(0.00309) (0.00323)
Height 4/8 yrs ago 0.000107 0.000737 0.000489 0.000472 -0.00590* -0.00575* -0.00541* -0.00521
(0.00313) (0.00313) (0.00315) (0.00315) (0.00327) (0.00326) (0.00328) (0.00328)
Height loss over 4/8 yrs 0.0287** 0.0338*** 0.0352*** 0.0338***
(0.0124) (0.0126) (0.0126) (0.0123)
Height loss: 1-1.9cm -0.0492 -0.0235 0.0724 -0.0923*
(0.0534) (0.0511) (0.0488) (0.0498)
Height loss: 2-2.9cm -0.0291 0.00762 0.186*** 0.0374
(0.0645) (0.0608) (0.0661) (0.0580)
Height loss: 3cm + 0.189** 0.188*** 0.116 0.219***
(0.0803) (0.0656) (0.0847) (0.0669)
Observations 1,042 1,042 1,042 1,042 1,042 1,166 1,166 1,166 1,166 1,166
R-squared 0.065 0.050 0.052 0.053 0.056 0.058 0.064 0.065 0.065 0.075
Note: Robust standard errors in parentheses. *** p< 0.01, ** p<0.05, * p<0.1. Age, Education, SES controls suppressed.
117
Table B3: Height Shrinkage and IADL in ELSA
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Panel A: Male sample Panel B: Female sample
Dependent Variable: IADL Index 4 years 8 years 4 years 8 years
Current Height -0.00334 -0.00689***
(0.00226) (0.00220)
Height 4/8 yrs ago -0.00285 -0.00287 -0.00223 -0.00222 -0.00649*** -0.00653*** -0.00595*** -0.00570**
(0.00228) (0.00229) (0.00229) (0.00229) (0.00223) (0.00223) (0.00224) (0.00224)
Height loss over 4/8 yrs 0.00999 0.0299*** 0.0176** 0.0235***
(0.00902) (0.00918) (0.00859) (0.00839)
Height loss: 1-1.9cm 0.00261 0.00743 0.0348 -0.0479
(0.0390) (0.0371) (0.0333) (0.0341)
Height loss: 2-2.9cm -0.00827 0.0643 0.0953** 0.00790
(0.0471) (0.0442) (0.0451) (0.0397)
Height loss: 3cm + 0.101* 0.149*** 0.125** 0.139***
(0.0586) (0.0477) (0.0578) (0.0458)
Observations 1,042 1,042 1,042 1,042 1,042 1,166 1,166 1,166 1,166 1,166
R-squared 0.049 0.041 0.043 0.052 0.052 0.073 0.075 0.078 0.077 0.083
Note: Robust standard errors in parentheses. *** p< 0.01, ** p<0.05, * p<0.1. Age, Education, SES controls suppressed.
118
Table B4: Height Shrinkage and ADL in IFLS
(1) (2) (3) (4) (5) (6) (7)
Panel A: Male sample
Dependent Variable: ADL Index 7 years 14 years 17 years
Current Height -0.00249
(0.00302)
Height 7/14/17 yrs ago -0.00290 -0.00291 -0.000798 -0.000887 -9.57e-05 -0.000368
(0.00302) (0.00300) (0.00308) (0.00309) (0.00330) (0.00335)
Height loss over 7/14/17 yrs: 0.0120 0.0257 0.0382**
(0.0121) (0.0167) (0.0143)
Height loss: 1-1.9cm 0.0475 0.0263 0.0959
(0.0652) (0.0618) (0.0642)
Height loss: 2-2.9cm 0.0165 0.128 0.194**
(0.0768) (0.0811) (0.0856)
Height loss: 3cm + 0.146 0.177* 0.214**
(0.103) (0.0923) (0.0861)
Observations 612 614 614 614 614 614 614
R-squared 0.063 0.062 0.064 0.066 0.073 0.075 0.081
(1) (2) (3) (4) (5) (6) (7)
Panel B: Female sample
Dependent Variable: ADL Index 7 years 14 years 17 years
Current Height -0.00613
(0.00422)
Height 7/14/17 yrs ago -0.00537 -0.00543 -0.00521 -0.00488 -0.00356 -0.00334
(0.00439) (0.00442) (0.00424) (0.00418) (0.00394) (0.00382)
Height loss over 7/14/17 yrs: 0.0160 0.0195 0.0289**
(0.0132) (0.0131) (0.0141)
Height loss: 1-1.9cm 0.0162 -0.0590 0.0187
(0.0491) (0.0625) (0.0514)
Height loss: 2-2.9cm 0.151* -0.102* -0.00226
(0.0818) (0.0564) (0.0521)
Height loss: 3cm + 0.118 0.145 0.169**
(0.0935) (0.0888) (0.0671)
Observations 896 897 897 897 897 897 897
R-squared 0.032 0.033 0.038 0.033 0.050 0.037 0.041
Note: Robust standard errors clustered at district level in parentheses. *** p< 0.01, ** p<0.05, * p<0.1. Age,
Education, SES controls suppressed. IFLS results control for district xed eects.
119
Table B5: Height Shrinkage and IADL in IFLS
(1) (2) (3) (4) (5) (6) (7)
Panel A: Male sample
Dependent Variable: IADL Index 7 years 14 years 17 years
Current Height -0.0126
(0.00795)
Height 7/14/17 yrs ago -0.00846 -0.00790 -0.00691 -0.00606 -0.00478 -0.00486
(0.00773) (0.00808) (0.00824) (0.00825) (0.00814) (0.00800)
Height loss over 7/14/17 yrs: 0.0888*** 0.0739*** 0.0927***
(0.0256) (0.0247) (0.0259)
Height loss: 1-1.9cm 0.126 0.0605 0.143
(0.143) (0.142) (0.108)
Height loss: 2-2.9cm 0.263 0.149 0.266*
(0.248) (0.126) (0.141)
Height loss: 3cm + 0.751*** 0.517*** 0.541***
(0.249) (0.184) (0.111)
Observations 967 970 970 970 970 970 970
R-squared 0.083 0.093 0.096 0.090 0.094 0.095 0.097
(1) (2) (3) (4) (5) (6) (7)
Panel B: Female sample
Dependent Variable: IADL Index 7 years 14 years 17 years
Current Height -0.0240***
(0.00783)
Height 7/14/17 yrs ago -0.0188** -0.0186** -0.0188** -0.0185** -0.0168** -0.0164**
(0.00814) (0.00809) (0.00845) (0.00837) (0.00726) (0.00744)
Height loss over 7/14/17 yrs: 0.0721** 0.0767** 0.0818***
(0.0275) (0.0287) (0.0224)
Height loss: 1-1.9cm 0.0868 -0.0248 0.00571
(0.0814) (0.100) (0.108)
Height loss: 2-2.9cm 0.240 -0.0320 0.0894
(0.144) (0.126) (0.0912)
Height loss: 3cm + 0.745*** 0.411*** 0.479***
(0.218) (0.136) (0.127)
Observations 1,088 1,089 1,089 1,089 1,089 1,089 1,089
R-squared 0.090 0.096 0.111 0.096 0.103 0.100 0.103
Note: Robust standard errors in parentheses. *** p< 0.01, ** p<0.05, * p<0.1. Age, Education, SES controls suppressed.
IFLS results control for district xed eects, cluster errors at the same level.
120
Table B6: Height Shrinkage and Diculty in Lifting in ELSA
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Panel A: Male sample Panel B: Female sample
Dep Var: Diculty in Lifting 4 years 8 years 4 years 8 years
Current Height -0.00320** -0.00250
(0.00149) (0.00210)
Height 4/8 yrs ago -0.00286* -0.00304** -0.00274* -0.00277* -0.00190 -0.00179 -0.00186 -0.00181
(0.00150) (0.00150) (0.00151) (0.00151) (0.00213) (0.00212) (0.00214) (0.00215)
Height loss over 4/8 yrs 0.0122** 0.0147** 0.0154* 0.0136*
(0.00594) (0.00607) (0.00820) (0.00801)
Height loss: 1-1.9cm 0.0258 0.0163 0.0266 -0.0333
(0.0256) (0.0245) (0.0318) (0.0326)
Height loss: 2-2.9cm 0.0621** 0.000873 0.115*** 0.0244
(0.0309) (0.0292) (0.0430) (0.0380)
Height loss: 3cm + 0.105*** 0.0902*** -0.0375 0.0709
(0.0385) (0.0315) (0.0551) (0.0438)
Observations 1,042 1,042 1,042 1,042 1,042 1,166 1,166 1,166 1,166 1,166
R-squared 0.067 0.066 0.071 0.068 0.071 0.091 0.094 0.098 0.094 0.096
Note: Robust standard errors in parentheses. *** p< 0.01, ** p<0.05, * p<0.1. Age, Education, SES controls suppressed.
121
Table B7: Height Shrinkage and Diculty in Getting up from Chair in ELSA
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Panel A: Male sample Panel B: Female sample
Dep Var: Diculty in getting up 4 years 8 years 4 years 8 years
Current Height -0.000262 0.00104
(0.00186) (0.00211)
Height 4/8 yrs ago 0.000220 -5.69e-05 -6.18e-05 -0.000164 0.00191 0.00229 0.00251 0.00244
(0.00187) (0.00187) (0.00189) (0.00188) (0.00213) (0.00213) (0.00214) (0.00214)
Height loss over 4/8 yrs 0.00693 0.00708 0.0200** 0.0256***
(0.00741) (0.00757) (0.00820) (0.00800)
Height loss: 1-1.9cm -0.00128 -0.0296 0.0803** 0.0663**
(0.0319) (0.0305) (0.0319) (0.0326)
Height loss: 2-2.9cm 0.0372 -0.0317 0.0196 0.0852**
(0.0385) (0.0364) (0.0431) (0.0379)
Height loss: 3cm + 0.125*** 0.0695* 0.0506 0.123***
(0.0480) (0.0392) (0.0552) (0.0438)
Observations 1,042 1,042 1,042 1,042 1,042 1,166 1,166 1,166 1,166 1,166
R-squared 0.046 0.044 0.051 0.046 0.052 0.041 0.047 0.048 0.053 0.053
Note: Robust standard errors in parentheses. *** p< 0.01, ** p<0.05, * p<0.1. Age, Education, SES controls suppressed.
122
Table B8: Height Shrinkage and Diculty in Carrying a Heavy Load in IFLS
(1) (2) (3) (4) (5) (6) (7)
Panel A: Male sample
Dep Var: Diculty in carrying 7 years 14 years 17 years
Current Height -0.00445
(0.00295)
Height 7/14/17 yrs ago -0.00409 -0.00440 -0.00278 -0.00243 -0.00345 -0.00335
(0.00305) (0.00301) (0.00320) (0.00325) (0.00315) (0.00316)
Height loss over 7/14/17 yrs: 0.00846 0.0184** 0.0139*
(0.00872) (0.00897) (0.00691)
Height loss: 1-1.9cm -0.00560 -0.00712 0.00830
(0.0321) (0.0298) (0.0377)
Height loss: 2-2.9cm 0.108** -0.00597 -0.0136
(0.0463) (0.0384) (0.0425)
Height loss: 3cm + 0.147** 0.126** 0.0867**
(0.0673) (0.0501) (0.0335)
Observations 967 970 970 970 970 970 970
R-squared 0.068 0.072 0.079 0.076 0.080 0.072 0.073
(1) (2) (3) (4) (5) (6) (7)
Panel B: Female sample
Dep Var: Diculty in carrying 7 years 14 years 17 years
Current Height -0.00500*
(0.00257)
Height 7/14/17 yrs ago -0.00387 -0.00405 -0.00262 -0.00245 -0.00269 -0.00258
(0.00270) (0.00268) (0.00282) (0.00282) (0.00270) (0.00273)
Height loss over 7/14/17 yrs: 0.0147* 0.0277*** 0.0218***
(0.00869) (0.00831) (0.00647)
Height loss: 1-1.9cm 0.0772** -0.00222 -0.0736**
(0.0356) (0.0404) (0.0344)
Height loss: 2-2.9cm 0.130** -0.0317 0.0182
(0.0617) (0.0365) (0.0372)
Height loss: 3cm + 0.151** 0.137*** 0.0944*
(0.0655) (0.0410) (0.0478)
Observations 1,088 1,089 1,089 1,089 1,089 1,089 1,089
R-squared 0.082 0.083 0.093 0.088 0.094 0.086 0.092
Note: Robust standard errors clustered at district level in parentheses. *** p< 0.01, ** p<0.05, * p<0.1. Age,
Education, SES controls suppressed. IFLS results control for district xed eects.
123
Table B9: Height Shrinkage and Diculty in Getting up from Chair in IFLS
(1) (2) (3) (4) (5) (6) (7)
Panel A: Male sample
Dep Var: Diculty in getting up 7 years 14 years 17 years
Current Height -0.000526
(0.00102)
Height 7/14/17 yrs ago -0.000714 -0.000703 -0.000250 -0.000184 9.32e-06 -2.68e-05
(0.000956) (0.000949) (0.000913) (0.000904) (0.000904) (0.000884)
Height loss over 7/14/17 yrs: 0.000794 0.00401 0.00810**
(0.00312) (0.00383) (0.00329)
Height loss: 1-1.9cm 0.0115 -0.00941 0.0334**
(0.0212) (0.0175) (0.0149)
Height loss: 2-2.9cm -0.0296 0.00932 0.0342
(0.0255) (0.0185) (0.0252)
Height loss: 3cm + 0.0212 0.0329 0.0367*
(0.0352) (0.0288) (0.0183)
Observations 967 970 970 970 970 970 970
R-squared 0.057 0.054 0.057 0.057 0.060 0.060 0.061
(1) (2) (3) (4) (5) (6) (7)
Panel B: Female sample
Dep Var: Diculty in getting up 7 years 14 years 17 years
Current Height -0.00335*
(0.00171)
Height 7/14/17 yrs ago -0.00284 -0.00290 -0.00233 -0.00235 -0.00203 -0.00192
(0.00193) (0.00194) (0.00190) (0.00183) (0.00195) (0.00191)
Height loss over 7/14/17 yrs: 0.00736* 0.0127** 0.0134***
(0.00428) (0.00587) (0.00459)
Height loss: 1-1.9cm 0.0266 -0.000244 -0.00393
(0.0175) (0.0150) (0.0174)
Height loss: 2-2.9cm 0.0496** -0.0159 0.00361
(0.0234) (0.0194) (0.0205)
Height loss: 3cm + 0.0828** 0.0873*** 0.0799***
(0.0379) (0.0306) (0.0250)
Observations 1,088 1,089 1,089 1,089 1,089 1,089 1,089
R-squared 0.050 0.054 0.061 0.055 0.068 0.057 0.062
Note: Robust standard errors clustered at district level in parentheses. *** p< 0.01, ** p<0.05, * p<0.1. Age, Education,
SES controls suppressed. IFLS results control for district xed eects.
124
Appendix C
Chapter 4 Appendix
Tables
125
Table C1: Results on Self-reported Health (1{3 Scale)
(1) (2) (3) (4) (5)
Men OLS FE System GMM
L.(own SRH) 0.569*** -0.125*** 0.171*** 0.224*** 0.177***
(0.0176) (0.0206) (0.0311) (0.0426) (0.0335)
L2.(own SRH) 0.0823**
(0.0357)
spouse's SRH 0.122*** 0.147*** 0.344*** 0.287***
(0.0138) (0.0203) (0.0890) (0.0907)
L.(spouse's SRH) 0.00146
(0.0253)
Observations 6718 6718 6718 4838 6758
Number of couples 1881 1881 1881 1606 1886
AR(1) test p val 2.32e-48 1.13e-33 4.79e-46
AR(2) test p val 0.486 0.395 0.263
Hansen J test p val 0.000601 0.00481 0.00000797
Di in Hansen p val for levels 0.0255 0.00627 0.00410
(6) (7) (8) (9) (10)
Women OLS FE System GMM
L.(own SRH) 0.597*** -0.121*** 0.146*** 0.178*** 0.146***
(0.0182) (0.0229) (0.0285) (0.0384) (0.0288)
L2.(own SRH) 0.0294
(0.0302)
spouse's SRH 0.115*** 0.124*** 0.324*** 0.310***
(0.0118) (0.0169) (0.0662) (0.0669)
L.(spouse's SRH) -0.00642
(0.0229)
Observations 6849 6849 6849 4915 6915
Number of couples 1944 1944 1944 1636 1956
AR(1) test p val 1.28e-49 1.48e-35 2.74e-49
AR(2) test p val 0.727 0.949 0.786
Hansen J test p val 0.0351 0.177 0.000901
Di in Hansen p val for levels 0.0744 0.0898 0.00455
Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05,
* p < 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living
daughters and sons, number of living siblings, respondent's and spouse's education dummies and
length of marriage; 2) log per capita de
ated household wealth.
126
Table C2: Results on Frequency of Exercise (=1 if any weekly vigorous physical activities)
(1) (2) (3) (4) (5)
Men OLS FE System GMM
L.(weekly vig phy activ) 0.433*** -0.112*** 0.141*** 0.220*** 0.144***
(0.0141) (0.0153) (0.0207) (0.0229) (0.0217)
L2.(weekly vig phy activ) 0.116***
(0.0201)
spouse's weekly vig phy activ 0.201*** 0.232*** 0.289*** 0.315***
(0.0133) (0.0170) (0.0730) (0.0833)
L.(spouse's weekly vig phy activ) 0.0400**
(0.0188)
Observations 7020 7020 7020 5077 7013
Number of couples 1934 1934 1934 1672 1931
AR(1) test p val 2.75e-97 2.71e-75 2.16e-94
AR(2) test p val 0.00825 0.419 0.0127
Hansen J test p val 0.0000129 0.116 0.00000267
Di in Hansen p val for levels 0.263 0.681 0.185
(6) (7) (8) (9) (10)
Women OLS FE System GMM
L.(weekly vig phy activ) 0.425*** -0.107*** 0.177*** 0.187*** 0.189***
(0.0152) (0.0169) (0.0215) (0.0269) (0.0217)
L2.(weekly vig phy activ) 0.0469**
(0.0214)
spouse's weekly vig phy activ 0.168*** 0.204*** 0.124* 0.112
(0.0120) (0.0152) (0.0666) (0.0716)
L.(spouse's weekly vig phy activ) -0.00964
(0.0162)
Observations 7212 7212 7212 5182 7193
Number of couples 2022 2022 2022 1717 2019
AR(1) test p val 4.72e-88 1.63e-62 5.94e-89
AR(2) test p val 0.00457 0.280 0.00395
Hansen J test p val 0.00481 0.0215 0.00438
Di in Hansen p val for levels 0.00279 0.00821 0.00519
Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05,
* p < 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living
daughters and sons, number of living siblings, respondent's and spouse's education dummies and
length of marriage; 2) log per capita de
ated household wealth.
127
Table C3: Results on Chronic Conditions (# of severe diseases)
(1) (2) (3) (4) (5)
Men OLS FE System GMM
L.(own # of severe diseases) 0.984*** 0.497*** 0.964*** 0.969*** 0.964***
(0.00540) (0.0172) (0.0119) (0.0132) (0.0120)
L2.(own # of severe diseases) 0.00476
(0.0174)
spouse's # of severe diseases 0.00776 0.0252 -0.0148 -0.00933
(0.00622) (0.0168) (0.0159) (0.0145)
L.(spouse's # of severe diseases) -0.0169
(0.0155)
Observations 7044 7044 7044 5097 7044
Number of couples 1936 1936 1936 1675 1936
AR(1) test p val 1.93e-49 4.90e-39 1.86e-49
AR(2) test p val 0.659 0.928 0.643
Hansen J test p val 0.202 0.157 0.213
Di in Hansen p val for levels 0.206 0.189 0.371
(6) (7) (8) (9) (10)
Women OLS FE System GMM
L.(own # of severe diseases) 0.981*** 0.475*** 0.978*** 0.982*** 0.978***
(0.00460) (0.0188) (0.0121) (0.0137) (0.0121)
L2.(own # of severe diseases) -0.0174
(0.0155)
spouse's # of severe diseases 0.00824* 0.0287** -0.00161 0.000775
(0.00433) (0.0117) (0.0107) (0.00974)
L.(spouse's # of severe diseases) -0.00251
(0.0106)
Observations 7237 7237 7237 5193 7237
Number of couples 2028 2028 2028 1719 2028
AR(1) test p val 7.58e-44 2.31e-32 7.70e-44
AR(2) test p val 0.513 0.510 0.517
Hansen J test p val 0.0282 0.0556 0.0302
Di in Hansen p val for levels 0.00125 0.00490 0.00164
Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05,
* p < 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living
daughters and sons, number of living siblings, respondent's and spouse's education dummies and
length of marriage; 2) log per capita de
ated household wealth.
128
Abstract (if available)
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Essays on health and aging with focus on the spillover of human capital
PDF
Essays on development and health economics
PDF
Essays on health economics
PDF
Three essays on health economics
PDF
Essays on health insurance programs and policies
PDF
Three essays in health economics
PDF
Three essays on economics of early life health in developing countries
PDF
Essays on the empirics of risk and time preferences in Indonesia
PDF
Selection and impacts of early life events on later life outcomes
PDF
Essays on education and institutions in developing countries
PDF
Essays on family planning policies
PDF
Essays on economics of education and private tutoring
PDF
Essays on development and health economics: social media and education policy
PDF
The determinants and measurement of human capital
PDF
Essays in health economics and provider behavior
PDF
Life transitions, leisure activity engagement, and cognition among older adults
PDF
Essays on health and well-being
PDF
Essays on human capital accumulation -- health and education
PDF
Three essays on behavioral economics approaches to understanding the implications of mental health stigma
PDF
Essays on education: from Indonesia to Los Angeles, CA
Asset Metadata
Creator
Jain, Urvashi
(author)
Core Title
Three essays on health & aging
School
College of Letters, Arts and Sciences
Degree
Doctor of Philosophy
Degree Program
Economics
Publication Date
07/03/2018
Defense Date
04/09/2018
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
aging,cognition,height,OAI-PMH Harvest
Format
application/pdf
(imt)
Language
English
Contributor
Electronically uploaded by the author
(provenance)
Advisor
Strauss, John (
committee chair
), Angrisani, Marco (
committee member
), Lee, Jinkook (
committee member
), Nugent, Jeffrey B. (
committee member
)
Creator Email
urvashi.jain29@gmail.com,urvashij@usc.edu
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-c40-508678
Unique identifier
UC11266831
Identifier
etd-JainUrvash-6372.pdf (filename),usctheses-c40-508678 (legacy record id)
Legacy Identifier
etd-JainUrvash-6372.pdf
Dmrecord
508678
Document Type
Dissertation
Format
application/pdf (imt)
Rights
Jain, Urvashi
Type
texts
Source
University of Southern California
(contributing entity),
University of Southern California Dissertations and Theses
(collection)
Access Conditions
The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...
Repository Name
University of Southern California Digital Library
Repository Location
USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA
Tags
cognition