Data
This study is based on data from the Utah Population Database (UPDB). The central component of UPDB is an extensive set of Utah family histories, in which family members are linked to demographic and medical information. The UPDB contains information on more than 11 million individuals, including the genealogies of the founders of Utah and their descendants, sourced from the Genealogical Society of Utah. The genealogy records for early migrants, their families, and their descendants represent birth cohorts that date back to near 1760. These early records provide basic demographic information on almost 200,000 families (more than 1.6 million individuals). These early records have been linked across generations, and in some instances, the records encompass as many as 11 generations. Individuals in these data may or may not have an affiliation to the Church of Jesus Christ of Latter-day Saints and may have lived in other states or countries. The genealogy records have been linked to other data sets, including Utah birth and death certificates, cancer records, the Social Security Death Index, and U.S. Censuses from 1880 to 1940. Multiple sources of death information are available in UPDB for parents and their children, including genealogy records, Utah death certificates beginning in 1904, and the Social Security Death Index. The analysis for this study is limited to individuals born up to 1900. We follow individuals from the point at which they marry until the time of death.
Data from the UPDB indicate that the frequency of polygamous men as a fraction of all married men was highest among men born in 1833, at 17.8% (Moorad et al.
2011; Smith and Kunz
1976). Decennial censuses showed that the proportion of the population in polygamous families reach a peak around 1860, when 43.6% men, women, and children were in a polygamous family of some sort, although this proportion was heavily dominated by children (Bushman
2008). After this peak, the proportion declined over time, until the equivalent figure was 25% in 1880 and 7.1% in 1900 (Bushman
2008).
Marital Status
Our key explanatory variable is marital status, which we treat as a time-varying covariate. This is important because we distinguish between monogamous and polygamous marriages in our analyses, and a monogamous marriage does not become polygamous until at least a second woman joins that marriage. Furthermore, it is critical to model widowhood, the death of more than one wife, the death of a sister wife, and remarriage after widowhood using a time-varying approach. As a result, the relative risks estimated in our event history analyses reflect the hazard of mortality of an individual in a given marital state, and the same individual may contribute exposure to multiple different categories of our marital status variables (e.g. married, widowed, remarried). We conduct separate analyses for men and women. For men, we top-code the number of wives at 4+ (few men had more than four wives). The sample size for the analytical population is 110,890 for women and 106,979 for men.
Our analyses for women include women who had up to three husbands, which is explained by remarriage following the death of the partner, but only a tiny fraction of women had more than two husbands because that would require the death of two husbands as well as a second remarriage. As a result, we collapse first and second monogamous remarriages, and second and third husband deaths in monogamous marriages, into the same categories. We model not only the death of husbands but also the deaths of sister wives. We use 20 categories (states) for marital status for our analyses, and these categories should be read from the perspective of the index person in the marriage:
1.
First monogamous marriage
2.
Widowed in first monogamous marriage
3.
Remarried (second or third monogamous marriage)
4.
Widowed again (second or third monogamous marriage)
5.
Polygamous marriage with two wives where the husband and sister wife are both alive
6.
Polygamous marriage with two wives where the sister wife has died, and the husband is alive
7.
Polygamous marriage with two wives where the sister wife is alive, and the husband is dead
8.
Polygamous marriage with two wives where the husband and sister wife are both dead
9.
Polygamous marriage with three wives where the husband and sister wives are all alive
10.
Polygamous marriage with three wives where one sister wife has died, one sister wife is alive, and the husband is alive
11.
Polygamous marriage with three wives where both sister wives have died, and the husband is alive
12.
Polygamous marriage with three wives where both sister wives are alive, and the husband is dead
13.
Polygamous marriage with three wives where one sister wife is dead, one sister wife is alive, and the husband is dead
14.
Polygamous marriage with three wives where both sister wives are dead, and the husband is dead
15.
Polygamous marriage with four or more wives where the husband and sister wives are all alive
16.
Polygamous marriage with four or more wives where some of (not all) the sister wives are dead, and the husband is alive
17.
Polygamous marriage with four or more wives where all the sister wives are dead, and the husband is alive
18.
Polygamous marriage with four or more wives where all the sister wives are alive, and the husband is dead
19.
Polygamous marriage with four or more wives where some of (not all) the sister wives are dead, and the husband is dead
20.
Polygamous marriage with four or more wives where all the sister wives are dead, and the husband is dead
Because very few women have more than one polygamous husband, which would require being widowed in the first polygamous marriage and then remarrying into another polygamous marriage, we collapse all higher-order polygamous marriages into the same categories. For example, if a women joins a polygamous marriage, and is then widowed, and then remarried into a second polygamous marriage, she could move from state 5 to state 7, and back to state 5 again in the second marriage. We have also conducted analyses where we model second and third marriages as completely separate states, but the patterns are qualitatively similar across first, second, and third marriages, and so we collapse the categories to increase precision.
For our interaction analyses, we also run models with a simplified set of categories for polygamous marriages to reduce some of the inherent complexity, collapsing polygamous marriages into those where (1) all sister wives and the husband are alive; (2) some, but not all, of the marriage group are deceased; and, (3) all members of the marriage group, except for the index person, are deceased. Among women, we also study the relationship between marriage order and mortality.
Our analyses for men include a marital status variable with 18 categories:
1.
First monogamous marriage
2.
Widowed from first monogamous marriage
3.
First monogamous remarriage
4.
Widowed from second monogamous marriage
5.
Second monogamous remarriage
6.
Widowed from second monogamous remarriage
7.
Polygamous marriage with two wives, where both wives are alive
8.
Polygamous marriage with two wives, where one wife has died
9.
Polygamous marriage with two wives, where both wives are deceased
10.
Polygamous marriage with three wives, where all wives are alive
11.
Polygamous marriage with three wives, where one wife has died
12.
Polygamous marriage with three wives, where two wives have died
13.
Polygamous marriage with three wives, where all three wives are deceased
14.
Polygamous marriage with four wives, where all wives are alive
15.
Polygamous marriage with four wives, where one wife has died
16.
Polygamous marriage with four wives, where two wives have died
17.
Polygamous marriage with four wives, where three wives have died
18.
Polygamous marriage with four wives, where all four wives are deceased
As mentioned earlier, an index individual may move between these different states, and may contribute exposure to each state that he occupies in the event history analysis. For example, a man in a monogamous marriage who then takes a second wife will move from state 1 to state 7; and if he then takes on a third wife, and none of the wives have died, he will move to state 10. If he outlives all three of his wives, he would then move from state 10 through states 11, 12, and 13.
For our interaction analyses, we also run models with a simplified set of categories for polygamous marriages. We collapse polygamous marriages into those where (1) all sister wives are alive; (2) some of, but not all, the sister wives have died; and (3) all the sister wives have died.
Statistical Analyses
To study the relationship between marital status and mortality, we use survival analysis in the form of Cox proportional hazard models (Cox
1972). The general proportional hazards model is expressed as:
$$ h\left(\left.t\right|{X}_1,\dots, {X}_k\right)={h}_0(t)\exp \left({\sum}_{j=1}^k{\upbeta}_j{X}_j(t)\right), $$
where
h(
t|X1, . . . ,
Xk) is the hazard rate for individuals with characteristics
X1, . . . ,
Xk at time
t;
h0(
t) is the baseline hazard at time
t; and β
j,
j = 1, . . . ,
k are the estimated coefficients. The failure event in our analysis is the death of the index person, and the baseline hazard in our model
h0(
t) is time since first marriage. We censor on the time when the person is lost to follow-up or outmigration from Utah, which is recorded in the UPDB.
We also estimate a series of stratified Cox models, which allows us to stratify the baseline hazard by different groups, based on the assumption that there are unobserved factors particular to each category that may lower or elevate the shared hazard. This allows us to effectively adjust for factors that are shared within strata to the extent that they are time-invariant (Allison
2009). The stratified Cox model takes the following form, where the hazard for an individual from stratum
s is
$$ {h}_s\left(\left.t\right|{X}_1,\dots, {X}_k\right)={h}_{0s}(t)\exp \left({\sum}_{j=1}^k{\upbeta}_j{X}_j(t)\right), $$
where
h0s(
t) is the baseline hazard for stratum
s,
s = 1, . . . ,
S. Our general strategy is to stratify by age at first marriage in single-year age groups because the baseline hazard in our models is time since first marriage. This effectively allows us to adjust for age differences in mortality. We also employ stratified Cox models to examine the effect of marriage order on the mortality of women in polygamous marriages; in these models, we stratify by the shared husband ID. We estimate the following models for women:
$$ \log \kern0.3em h(t)={\upbeta}_1{Marital}_{ij}+{\upbeta}_2{A}_{ij}+{\upalpha}_j $$
(1)
$$ \log \kern0.5em h(t)={\upbeta}_1{Marital}_{ij}+{\upbeta}_2{A}_{ij}+{\upbeta}_3{B}_{ij}+{\upalpha}_j, $$
(2)
where log
hi(
t) is the log hazard of mortality, α
j is the shared group by age at first marriage
j, and the index
ij refers to the individual
i in that group
j;
Maritalij is entered into the model as a series of 20 dummy variables based on the categories of marital status for women described in detail in the preceding section; and
Aij is a set of control variables including birth cohort and church affiliation. Birth cohort is a categorical variable that groups together those who are born in the 1700s, who are few, and 10-year groupings for those born from 1800 to 1900. We adjust for affiliation with the Church of Jesus Christ of Latter-day Saints because it is linked directly to the likelihood of entering a plural marriage in the first place and is also related to mortality risk. Adherents have a proscription from alcohol and tobacco use; the emphasis placed on community and social integration, as well as monthly fasting, means that health outcomes for both the men and women in our sample are related to affiliation. This variable has three categories: (1) no affiliation, (2) inactive, and (3) active. Those who are active pledged to abide by the doctrine of the religion. Inactive members are those who were baptized, but we have no evidence that they later pledged to abide by the doctrine of the religion.
In Model 2, we introduce an additional set of control variables,
Bij, for whether the woman was born in Utah, the occupational status of the husband, whether the husband was a farmer, age difference from the husband, biological parity, and child deaths. Our measure for the socioeconomic status (SES) of the husband is a NPSES (Nam-Powers-Boyd) occupational status score for a measure of occupational status, where scores range from 0 to 99 (Nam and Boyd
2004). We include this variable in the models split into 11 categories (1–9, 10–19, . . . , 90–99, and missing). The variable for whether the husband was a farmer is entered as a separate dummy variable. We control for the SES of the husband in both our analyses of men and of women because household SES was best determined by the husband’s occupational status during this period. Age difference between the husband and wife is operationalized as husband age minus wife age and is split into seven categories (<–9; –9 to –5; –4 to –1; –1 to 1; 1 to 4; 5 to 9; >9). We control for age difference between the husband and wife because research suggests that the age difference between partners affects their hazard of mortality (Drefahl
2010). If a woman is widowed, she retains the values from the deceased husband for the variables husband occupation, whether the husband was a farmer, and age difference from the husband until the time at which she might remarry, at which point those variables are updated to reflect the characteristics of the new husband. The variables for parity (0, 1, . . . , 9, 10+) and child deaths (0, 1, 2, 3, 4, 5+) are time-varying covariates, updating with each subsequent birth and death. We also repeat Models 1 and 2 using a simplified version of marital status, with only seven categories, detailed earlier.
We also estimate several models examining the statistical interaction between the simplified version of our marital status variable for women (seven categories) and several key variables:
$$ \log \kern0.3em h(t)={\upbeta}_1{Marital}_{ij}\times {ChildDeaths}_{ij}+{\upbeta}_2{A}_{ij}+{\upbeta}_3{B}_{ij}+{\upalpha}_j $$
(3)
$$ \log\;h(t)={\upbeta}_1{Marital}_{ij}\times {Parity}_{ij}+{\upbeta}_2{A}_{ij}+{\upbeta}_3{B}_{ij}+{\upalpha}_j $$
(4)
$$ \log \kern0.3em h(t)={\upbeta}_1{Marital}_{ij}\times {HusbandFarmer}_{ij}+{\upbeta}_2{A}_{ij}+{\upbeta}_3{B}_{ij}+{\upalpha}_j, $$
(5)
where
ChildDeathsij is a time-varying covariate for experience of child deaths (0, 1, 2+),
Parityij is a time-varying covariate for number of children ever born (0, 1–3, 4–6, 7–9, 10+), and
HusbandFarmerij is a binary variable for whether the husband was a farmer. In Models 3–5, control variable set B omits the variable directly associated with the interaction variable of interest.
Finally for women, we also estimate models to examine how marriage order affects the mortality of women in first marriages that were polygamous:
$$ {\displaystyle \begin{array}{l}\log\;h(t)={\upbeta}_1{Order}_{ij}+{\upbeta}_2{BirthYear}_{ij}+{\upbeta}_3{HusbandDead}_{ij}+{\upbeta}_4{HusbandOcc}_{ij}\kern0.3em +\\ {}{\upbeta}_5{HusbandFarmer}_{ij}+{\upalpha}_j\end{array}} $$
(6)
$$ \log \kern0.3em h(t)={\upbeta}_1{Order}_{ik}+{\upbeta}_2{BirthYear}_{ik}+{\upbeta}_3{HusbandDead}_{ik}+{\upbeta}_4{AFM}_{ik}+{\upzeta}_k, $$
(7)
where
Order refers to marriage order (1, 2, 3, 4, 5+),
BirthYear is a continuous variable for birth year,
HusbandDead is a time-varying covariate for whether the husband is alive or deceased,
HusbandOcc is husband occupational status, and
HusbandFarmer is whether the husband is a farmer. Model 6 stratifies by age at first marriage, and Model 7 stratifies by shared husband ID, ζ
k. Nested within each husband, the wives share the same baseline hazard, which adjusts for factors that are shared by wives to the extent that they are time-invariant (Allison
2009). In Model 7, we also estimate cluster-adjusted robust standard errors clustered at the level of the shared husband ID (Lin and Wei
1989), and we explicitly control for age at first marriage (16
, 17, . . . , 59
, 60+),
AFMik. We conduct separate analyses by the number of sister wives (2, . . . , 5+), as well as a pooled analysis. In these analyses, we condition on all preceding sister wives being alive at the time that the index person joins the marriage. We introduce this condition because higher-order wives might be replacements for previously deceased wives in polygamous marriages, and this produces a statistical artifact whereby earlier wives have a relative mortality risk much greater than higher-order wives.
In our analyses of men, we estimate the following models:
$$ \log \kern0.3em h(t)={\upbeta}_1{Marital}_{ij}+{\upbeta}_2{A}_{ij}+{\upalpha}_j $$
(8)
$$ \log \kern0.3em h(t)={\upbeta}_1{Marital}_{ij}+{\upbeta}_2{A}_{ij}+{\upbeta}_3{C}_{ij}+{\upalpha}_j, $$
(9)
where
Maritalij is entered into the model as a series of 18 dummy variables based on the categories of marital status for men described in detail in the preceding section; control set
Aij refers to birth cohort and church affiliation, and control set
Cij includes variables for whether the man was born outside Utah, the occupational status of the man, whether the man was a farmer, biological parity (0, 1, . . . , 10
, 11–14
, 15–19
, 20+), and child deaths (0, 1, 2, 3, 4, 5+). The variables for parity and child deaths are time-varying covariates, updating with each subsequent birth and death. Polygamous men whose wives die generally will have many children, and rearing these children would be a stressor. We do not control for age difference in our analyses of men because there is no single value for age difference between husband and wife for men who have multiple wives. We prefer to omit the covariate for spousal age difference rather than to assume the mean age difference across multiple wives.
We also estimate models examining the statistical interaction between the simplified version of our marital status variable for men (seven categories) and key variables:
$$ \log \kern0.3em h(t)={\upbeta}_1{Marital}_{ij}\times {ChildDeaths}_{ij}+{\upbeta}_2{A}_{ij}+{\upbeta}_3{C}_{ij}+{\upalpha}_j $$
(10)
$$ \log\;h(t)={\upbeta}_1{Marital}_{ij}\times {Parity}_{ij}+{\upbeta}_2{A}_{ij}+{\upbeta}_3{C}_{ij}+{\upalpha}_j $$
(11)
$$ \log\;h(t)={\upbeta}_1{Marital}_{ij}\times {Farmer}_{ij}+{\upbeta}_2{A}_{ij}+{\upbeta}_3{C}_{ij}+{\upalpha}_j, $$
(12)
where
ChildDeathsij is a time-varying covariate for experience of child deaths (0, 1, 2+),
Parityij is a time-varying covariate for number of children ever born (0, 1–3, 4–6, 7–9, 10+), and
Farmerij is a binary variable for whether the man was a farmer. For Models 10–12, control variable set C omits the variable directly associated with the interaction variable of interest.