Transcript Findings from a Sample of Union Army Veterans.

Marriage, Health and Later-Life
Mortality:
Findings from a sample of Union Army Veterans
Sven Eric Wilson
Brigham Young University
&
University of Chicago
Prepared for the meetings of the Population Association Meetings of America
Philadelphia, PA
March 31, 2005
Motivations

Empirically, marital status is strongly
associated with health:


a variety of health measures (mortality, disease,
disability, mental health)
a wide array of populations and population
subgroups.
Theory is underdeveloped
 High policy-relevance: Do pro-marriage
policies promote the public health?

Persistent Puzzles
Although correlations are robust, we still
know little about the causal mechanisms.
 Furthermore, what do alternative
explanations say about questions such as:



How should the marriage effect vary across the
different categories of non-married—widowed,
divorced, separated, never-married?
How will the marriage effect vary with age?
Common Explanations

Categories of explanations
 Selection

Extensive margin
 Not necessarily positive selection


Capital Accumulation






Intensive margin
Behavioral modification
Increased return to investment
Monitoring
Caregiving
Social support
Bereavement Effects


Negative psychological effects of spousal loss
Positive “bliss effects?”
The Union Army Data

Extensive information has been collected on over
35,000 men who served in the Union Army
during the Civil War.


Linked to 1850, 1860, 1900, 1910 census records
Military records



individual medical records from the civil war hospitals
Military experiences—battles, wounds, location of service,
etc.
Pension records



Complete records of physician exams of pension applicants
Socioeconmic and demographic information on veterans
Family structure in later life: 1900, 1910 census records
Key Variables: Mortality

Mortality: Mortality is measured as the number of
days lived past the baseline date

Only those with death dates known are used


For those in the pension by 1900, over 95% have a reliable
death date
A small portion of cases have only year of death
known—months and days are imputed at the midpoint of
the period.
Key Variables: Health Status:

The health information comes primarily from the
most recent (prior to 12/31/1900) physical exam


These are physician recorded, not self-reported.
Two indicators of health status at the study
baseline are used:



BMI: Body Mass Index.
Disabling conditions: These are chronic conditions noted
by the physicians to be severe enough to merit disability
support.
Pension Amount: Best measure of severity
Key Variables: Marital Status



Marital status is reported in the 1900 census as
Married, Single, Divorced or Widowed
The Pension Bureau regularly surveyed
pensioners and kept dates on marriages and
death of spouses
Combination of census data and pension data
allows us to get at some marital history
information, such as whether it is a first or
second marriage.
Key Variables: Demographics

Age:


SES:




Occupation:
Home Ownership
Employment status (whether ever unemployed during
the past 12 months)
County Demographics




In the Union Army records, the age is reported at
numerous points across the life-cycle. We have
developed an algorithm to calculate the “best age” from
these disparate sources.
Population Density
Percent Foreign Born
Percent Urban
Region (9 census regions)
Sample Restrictions

35,570 recruits are in the entire sample



Sample used for analysis includes:





Over 17,000 actually enter the pension system at one
point.
Close to 12,000 of the veterans alive in 1900 were
linked to the U.S. pension schedules,
Only
Only
Only
Only
those
those
those
those
linked to Census
in pension system at baseline
alive as of June 1, 1900
aged 50-69 at baseline
Total analysis sample: 9,397
Age (1900)
.04
.06
.08
.1

0
.02
Density
Age distribution
50
60
70
age
80
90
Sample life table comparisons, 1900
US population numbers come from Haines (1998)
Marital Status, by Age
Sample Demographics
Sample Health Characteristics
Distribution of BMI, 1900, Age 50-69
BMI Distribution, By Marital Status
Mortality: Married v. Unmarried
10-year Mortality Rates
60.0%
50.0%
40.0%
Married
30.0%
Not Married
20.0%
10.0%
0.0%
All Ages
50-54
55-59
Age
60-64
65-69
Mortality: Marital Status Differences
10-year Mortality Rates, by Marital Status
70.0%
60.0%
50.0%
1st Marriage
Remarried
40.0%
Separated/Divorced
30.0%
Widowed
Never Married
20.0%
10.0%
0.0%
All Ages
50-54
55-59
60-64
65-69
Regression Methods

Three different methods explored thus far

Probit regression on 10-year mortality rates




recall there is no significant sample attrition to worry
about.
Cox Hazard model: censored after 10 years
Cox Hazard mode: uncensored
Methods yield quantitatively similar results

Probit results presented here.
Probit results; Married v. unmarried
Probit results: Marital status effects
Health Interactions

Marriage effects among those who are
underweight, normal and overweight/obese
(Not shown: Each model is “full model” containing all covariates)
Summary of Results
1.
Married men have consistently lower mortality
than unmarried men, even after controlling for
health and other demographics.
2.
Marriage effects are relatively small (though
important in a high mortality environment).
3.
The marital status effects are strongest for
divorced/separated men.
4.
The effects of being unmarried are stronger
among those who are underweight than among
the normal or the overweight.
Preliminary HRS Comparisons
•
Using a sample of white men from the 1992 HRS
cohort, it is possible to compare the marital
status effects between the Union Army (UA) and
HRS samples
•
The UA sample is restricted to follow the same
age range as the HRS
•
Method: A Cox proportional hazard model is used
to estimate marital status effects over a ten-year
period, with censoring occurring either at loss to
follow-up or the end of the period.
UA/HRS Comparison: Marital Status
•
Distributions of marital status: Males aged
51-61, 1900 and 1992
UA/HRS Comparison: 10-year mortality
(Descriptive Statistics)
UA/HRS Comparison: Model-Based
Marital Status Effects (Cox Model)
UA/HRS Comparison: Summary
•
Marital status effects are important in both
samples.
•
•
•
Simple (uncontrolled) effects are much higher
in the HRS than in the UA sample.
Divorce/Separation has the strongest
effect in both samples.
Health controls in HRS seem to have a
much stronger interaction with the marital
status effects than in the UA.