Stefan Sampaleanu

ECO 375 Winter 2017

Prof. Lee

Marriage Trends in Young American Adults

Introduction

        Evolutionarily speaking, the concept of monogamy may be more ancient than the human race itself. Gibbons, French Angelfish and swans are examples of animals who keep a single mate for life. Humans too have celebrated and suffered this behavior, but in doing so have also blurred the lines that distinguish monogamy. Today, marriage represents a leap of faith that may or may not end happily. Indecisive sorts even have the opportunity to divorce and remarry, rinse and repeat. Entire sectors of the economy are dedicated to divvying up assets and making nice after messy divorces. People are marrying for different reasons, too - maybe that person doesn’t quite give your stomach butterflies, but their stable job and Roth IRA will come in handy down the road. Perhaps somebody jumped the gun and there’s a kid on the way; you better buckle down and walk the aisle! Maybe you’re unsure, but committed to “making it work.” Hopefully, you’ve struck gold and found your “best friend” you wish to grow old with. There are multiple factors that motivate us to marry - this paper serves to numerically examine their effect on that decision.

        The notion of matrimony prefers to be portrayed as sacred, eternal and full of passion. In the scope of our lives, however, care must be taken to the economic and legal ramifications of betrothal. Our hope of finding love is contingent on our ability to align a serious relationship alongside professional growth, financial obligations, our own willingness (or lack thereof), and our cultural background. With that said, we can look to a gamut of personal traits for their effect on marriage behavior. The research question that motivates this study is: what are the economic and demographic factors that motivate adults in their 20s to make the marriage decision? We hypothesize that income, education level and financial stability will be positive indicators of marriage behavior. We will use log of nominal income for the first factor income, presence of a college degree and student status for the second factor education, and stability indicators like dividend income and government employment to assess stability’s effect on marriage behavior. From early analysis of the data, we postulate that female status and being a caucasian will positively predict marriage, while being African American will do so negatively.

Literature Review

        A cornucopia of quantitative research has been conducted with regard to marriage behavior and the trends contained within it. With regard to statistical techniques, many works that pertain to marriage behaviors use a hazard regression, a form of survival analysis wherein each observation represents a span of time over which a subject was observed. In this presentation, each independent variable’s coefficient adjusts the probability of a certain event occurring. Marriage behaviors, particularly an individual’s first marriage (which defines my dependent variable triedit) represent seismic life transitions, so a hazard model is appropriate. Unfortunately, hazard models require time-series data collected in waves spaced several years apart. I will instead use the cross-sectional data at my disposal to construct a logistic regression towards marriage behavior.

        Research from Dew and Price offers a look at the factors that predict marriage among never-married young adult couples age 18-35, dividing their observations between cohabiting and non-cohabiting individuals (cohabitation refers to sharing a residence with a romantic partner). Cohabitation represents a transitionary period, often in preparation for marriage, in which marriage indicators may differ from non-cohabiting individuals. Cohabitation also represents an emerging alternative to marriage that may in part explain lower marriage rates. Dew constructs a hazard model predicting the likelihood of marriage, based on the descriptive characteristics observed in each wave, and how they differ among those who become married or begin cohabitation between waves. Their study looks mainly to hours worked per week and their quantification of “occupational prestige” as positive predictors of marriage, particularly in individuals that don’t begin the study in cohabitation. Alongside prestige and hours clocked, they include income, debt, education, and savings, all as positive predictors of marriage hazard. They also point to “visible assets” like car value and how they positively predict marriage hazard - this brings in a psychological component that may work to explain differences in marriage behaviors between men and women.

        Qualitative research by evolutionary psychologist David Buss assesses the relationship between personal finances and mate selection, a relationship that will be useful to harness in predicting the presence of marriage behavior. He demonstrates that generally women value economic prospects in a partner more highly than men, and men, cognizant of this, become distressed when compared with another suitor of greater wealth. This work helps to explain the differences in marital behaviors between genders described below.

Data

        To accomplish our research objective, cross-sectional data from the 2016 Current Population Survey (CPS) was used. We will restrict the data to individuals aged 20-30. This restriction denotes the period in which many US individuals marry.

In this dataset, marital status is listed as Married, Divorced, Widowed, Separated, or Never Married. This is a useful and clear classification, but offers no discretion towards multiple marriages. Adults that frequently marry, divorce, and remarry can jump from “Married” to “Divorced” and back again continuously, blurring the significance of those two classifications. To combat this, we will use a dependent variable indicating presence of marital behavior “triedit”. “Triedit” is a dummy variable which indicates marital status is contained in {Married, Divorced, Widowed, Separated}. Since all four of those classifications must entail a prior decision to marry, we can safely draw that as the proportion of “triedit” to “!triedit” increases, a higher portion of young adults are marrying. The proportion of adults in the sample who have “tried it” is 6,706/21086 = 31.80%. The following graphical presentation illustrates the movement of these two mutually exclusive groups with respect to age and gender:

In this presentation, the right histogram represents female and the left male participants. Clear trends in marital behavior are visible as we analyze progressively older individuals.  At age 30, 57.92% of subjects had married (62.56% of women and 52.78% of men) compared to 5.84% at age 20 (7.34% of women and 4.43% of men). At every age, we consistently see higher marriage participation in women. This would imply that women generally marry earlier than men, and, due to lack of transitivity, we may also assume that women generally marry older partners that may not be contained in our 20-30 year-old sample.

Independent variables

Analysis of literature in the topic points to a number of economic indicators that can be tied to marriage behavior. The following table outlines the descriptive statistics with regard to our independent variables. Most, such as “female”, “white”, “black”, and “degree” are dummy variables. Their effect on our results are contingent upon that variable’s condition being satisfied. Income, age and usual hours worked per week are numerical measures. We use the log of income to better grasp the effect of a per-unit change. If nominal income = 0, our log variable does not return a value, so we drop those cases and hone our study on employed adults. Missing values of the uhours variable were replaced with 0, indicating the individual is not participating in the labor force.

The proxy variable uni_student denotes individuals under 25 who have completed some college (educ==3).  This classification is important: “Some college” has the lowest triedit rate of all educational categories at 27.62% of individuals, but this effect is severely concentrated in individuals at the lower-aged half of our dataset. At age 25 and above, “some college” does not appear to significantly affect triedit, so we may presume that the rate is low due to early-20s adults still in the process of attending university. We anticipate that uni_student will be a strong negative predictor of marital behavior.

Research Methodology

        In order to analyze marriage behavior, we shall employ a logistic regression. This predicts the likelihood of the triedit condition being satisfied - that is, based on an individual’s demographic and financial information, we can deduce the probability that that individual has married. Ordinary logistic regressions that predict P (probability) are subject to error in that the resultant P value may be less than zero or greater than one. Furthermore, variance is expressed as a function of p, so the condition of constant variance needed for a linear regression is not met. A Logit regression uses the same independent variables to predict a test statistic in the form of ln(P/(1-P)), so that regardless of the test statistic’s value, we may extrapolate a probability between 0 and 1. The Logit regression follows the form:

ln(P/(1-P)) = β0 + β1*(age) + β2*(female) + β3*(white) + β4*(black) + β5*(uhours) + β6*(log_inc) + β7*(govt_wrkr) + β8*(eq_owner) + β9*(degree) + β10*(uni_student) + εi

Wherein each coefficient (represented by the symbol β) represents the effect of a one-unit change in the corresponding independent variable’s value. For dummy variables, who have value 0 or 1, their effect on the test statistic is limited to the size of β.  For numerical values like age and uhours, their effect on the test statistic is equal to β*(value). Β0 may be interpreted as a starting value against which all the other coefficients are valid - in a more simple, linear context, this would be easily portrayed as the y-intercept of an OLS regression line. The error term ε denotes the natural error that occurs with any predictive model, that is, the difference between actual and predicted probabilities.

        Once the test statistic is interpreted, we can use stata’s mfx command to assess the marginal effects of each of the variables. It reports a mean probability, as well as the size and direction of each independent variable’s effect on our predicted probability. From there, we may use the at(_) command to gather other probabilities, based on different permutations of our independent variables.

Results

The data proved all independent variables to be significant in predicting marital behaviors, however, my expectations with respect to trend direction proved mistaken. The regression offered a pseudo-R2 value of .1572, with output as follows:

        From this we see that even the least significant predictor, equity ownership, was significant at a 99% confidence level, so we may deduce that all included independent variables are correlated to marital behavior. To further assess the probabilities generated by this regression, let’s look at marginal effects. The data yields:

         Let’s break down the meaning of each dy/dx to determine how that variable is affecting our prediction. Dy/dx(age) yields .0564, meaning that at mean age 25.25, the addition of a year corresponds to predicted likelihood increasing by .0564. This adjustment would change according to input - when we pose a subject with age = 30, the mfx table shows a larger dy/dx. The condition of being female improves predicted likelihood by .0762. Contrary to numerical values, the effect of binary variables is of constant size (other things the same) and limited to whether or not that variable is satisfied. The condition of being white also improves predicted likelihood, in this case by .0599. The condition of being black, however, would lower predicted likelihood (indicated by the negative sign) by a staggering .1619. We see that being black has a strong negative correlation to marital behavior in one’s 20s. Individuals of other races, like Asians, Hispanics and Mixed-race adults, would be interpreted as satisfying neither the positive-trending white condition nor the negative-trending black condition, so, other things the same, their predicted likelihood would fall between a white person and black person’s. So far, our findings are consistent with the changes in triedit frequency when pinned against age and race variables.

        Earlier research by Dew and Price indicates that hours worked per week can be a strong predictor of marriage - that is, when individuals were working full-time, they had a much higher likelihood of marrying in the following period. From that, we would expect hours worked to be a positive predictor of marriage likelihood, and the data is consistent with that finding. We see a per-unit effect of .021 on our likelihood, meaning working one hour per week more than the mean of 33.17 hours would increase your predicted likelihood by .0021. This is only a moderately significant finding, as we can imagine 5 additional hours to have an approximate impact of .01 on our marriage likelihood. Another key employment-related predictor, log_income, appears to have a negative impact on likelihood. This is puzzling relative to earlier findings - on the intuitive level as well, it seems that adults would sooner marry if they have a degree of financial security. Just the same, wealth does not necessarily indicate security. Furthermore, we know that married individuals make more, but are uncertain if individuals who make more necessarily marry. As the data stands, it indicates that at the mean log_inc = 9.71, a log increase of one corresponds to a .0163 fall in predicted likelihood. The positive trend we predicted may be counteracted by an unknown negative one. This is a significant effect, but its impact is small on our eventual likelihood. Further research must be administered to better understand the effect of income.

        In the case of government employees, whose employer has a reputation of stability, we see that individuals that work for state, federal, or municipal governments are positively inclined to marry, and satisfying the govtwrkr condition increases our predicted likelihood by a sizeable .0964. Equity owners had a similar effect of .0379, however, in the marginal effects model, this variable was no longer significant at the 99% level. Our p value falls at .011 still, so we can reject the null hypothesis of no correlation at up to 98.9% significance.

        With regard to education, the data yields another puzzling indicator. We see that, as expected, university students are far less likely to marry, with that condition lowering likelihood by .0878. Yet, the presence of a college degree also lowers marriage likelihood, countering our initial assumptions, and we may understand this mismatch in a similar manner to income. Our regression found a college degree or higher to be predictive of a .0731 fall in likelihood. This could be due to higher labor force participation that results from higher education, in turn leaving less time for serious relationships. The data’s “z” statistics represents the quotient of dy/dx to standard error, and the “P>|z|” value indicates the likelihood that the data would yield such a result under a null hypothesis, that is, the mentioned variable is unassociated with our dependent variable triedit.

Conclusion

        The data shines an interesting light on our initial hypotheses. We see that our demographic variables of race and gender were clear in the direction of their effect, and this effect was consistent with our hypothesis. Insofar as stability, our measures of dividend income presence and government employment were both indicative of greater marriage likelihood. This is consistent with the notion that secure income and employment conditions are conducive of marriage. Secure income, however, clearly has different implications than high income, as we see that higher income negatively predicts marriage. Our data yields negative prediction for current students, and, against our assumptions, presence of a college degree or higher also negatively predicts marriage.

We conclude that, particularly with regard to college degrees and income, there are omitted variables that would better explain marriage behavior. This could include a more nuanced presentation of income, indicating income’s makeup from earned, unearned, and and outside sources like social welfare and child support. With regard to a college degree, perhaps a delineation between a bachelors degree and a masters or PhD would yield better predictions. In actuality, there are innumerable intermediary forces that bolster and mitigate education’s, as well as income’s, effect on marriage behavior.

Further research should seek to explain this effect in greater detail. Additionally, future research should be employed using time-series, rather than cross-sectional data. In this manner, we may study the changes in individual subjects, and with much greater specificity than our current presentation, which shows age trends across tens of thousands of subjects that lead different lives and act for different reasons. Additional questions we may pursue include: in what manner do changes in income predict life transitions like marriage and divorce? How do the presence of previous marriage(s) affect the likelihood of further marriage decisions? What are the more specific income indicators that predict marriage, both positively and negatively? Do particular indicators of a marriage decision also act to predict a potential divorce decision? The data holds numerous indicators, many of which are yet unknown, that can help us explain this.

Bibliography

Buss, David M. and David P. Schmitt. "Sexual Strategies Theory: An Evolutionary Perspective on Human Mating." Psychological Review, vol. 100, no. 2, Apr. 1993, p. 204. EBSCOhost, ezproxy.depaul.edu/login?url=http://search.ebscohost.com/login.aspx?direct=true&db=a9h&AN=9306105983&site=ehost-live&scope=site.

Dew, Jeffrey and Joseph Price. "Beyond Employment and Income: The Association between Young Adults' Finances and Marital Timing." Journal of Family and Economic Issues, vol. 32, no. 3, Sept. 2011, pp. 424-436. EBSCOhost, doi:link.springer.com/journal/volumesAndIssues/10834.

Gicheva, Dora. "Student Loans or Marriage? A Look at the Highly Educated." Economics of Education Review, vol. 53, Aug. 2016, pp. 207-216. EBSCOhost, doi:www.sciencedirect.com/science/journal/02727757.

Ost, Cecilia Enstrom. "Housing and Children: Simultaneous Decisions?--A Cohort Study of Young Adults' Housing and Family Formation Decision." Journal of Population Economics, vol. 25, no. 1, Dec. 2011, pp. 349-366. EBSCOhost, doi:link.springer.com/journal/volumesAndIssues/148.

Rotz, Dana. "Why Have Divorce Rates Fallen? The Role of Women's Age at Marriage." Journal of Human Resources, vol. 51, no. 4, Fall, pp. 961-1002. EBSCOhost, doi:jhr.uwpress.org/content/by/year.

Santos, Cezar and David Weiss. "'Why Not Settle Down Already?' a Quantitative Analysis of the Delay in Marriage." International Economic Review, vol. 57, no. 2, May 2016, pp. 425-452. EBSCOhost, doi:onlinelibrary.wiley.com/journal/10.1111/%28ISSN%291468-2354/issues.

Appendix

use http://condor.depaul.edu/jlee141/econdata/cps_data/cepr_march_2016.dta

/*

restrict to ages 20-30

*/

keep if age<=30 & age>=20

drop if incp_all==0  /*removes 0 income cases to ensure real log_inc values */

replace uhours = 0 if mi(uhours) /*replaces missing values of usual hours worked with 0*/

/*

VARIABLES GENERATED

RACE*/

gen white = wbhaom==1        /* caucasian dummy variable   */

label var white "White individual"

gen black = wbhaom==2        /* black dummy variable           */

label var black "Black individual"

gen other = wbhaom>2

label var other "Hispanic, Asian, Other, or Mixed individual"

/* EDUCATION */

gen degree = educ>3         /*Bachelors' degree or higher dummy variable*/

label var degree "Presence of Bachelors' Degree or Higher"

gen uni_student = educ==3 & age<25 /*proxy variable indicating the subject intends on completing 4-year university*/

label var uni_student "educ = "Some college & age<25""

/* INCOME */

gen log_inc = log(incp_all) /* Log of Personal Nominal Income */

gen govtwrkr = clslyr>1 & clslyr<5

label var govtwrkr "Employee of state, federal, or local government"

/* dummy variable indicating those who worked in federal(2),

state(3) or local government(4) last year*/

gen eq_owner = rincp_div>0

label var eq_owner "Presence of Dividend Income"

/* dummy variable indicating the presence of income from dividends, approximating

those individuals who own equity */

gen triedit = marstat!=5

label var triedit "Marriage indicator"

/* dummy variable indicating those whose marital status is Married,

Widowed, Divorced, or Separated. All of these individuals made a

decision to marry in their lifetime. */

twoway(histogram age if triedit, discrete color(green) frequency)(histogram age if !triedit, discrete fcolor(none) lcolor(black) frequency), by(female) legend(order(1 "Tried it" 2 "Never Married"))

/* two-way histogram included on pg. 4 */

/* DESCRIPTIVE STATISTICS used in table, pg. 5-6*/

tab triedit

summ age

tab female

tab triedit if female

tab white

tab triedit if white

tab black

tab triedit if black

summ uhours

summ log_inc

tab govtwrkr

tab triedit if govtwrkr

tab eq_owner

tab triedit if eq_owner

tab degree

tab triedit if degree

tab uni_student

tab triedit if uni_student

logit triedit age female white black uhours log_inc govtwrkr eq_owner degree uni_student, robust

/*yields regression used on pg.8 */

mfx /* yields marginal effects info (dy/dx, etc.) contained in pg. 8-9 table */