The Econometric Study of Determinants of Divorce Rate in United States: Economic Approach

Introduction

Divorce has become one of the most discussed social phenomena in the recent years. Increasing number of divorce cases among couples make it interesting for researchers from various fields such as sociology, econometrics, and psychology to investigate main factors leading to socio-economic issues. Different research works, such as Kalmjin and Poortman, Amano

and Beattie, as well as Nakonezny et al. look at distinct sets of divorce determinants. These authors utilized mixed set of variables, precisely speaking economic and social ones, such as presence of children, age difference at marriage and dating period. In contrast to other studies this paper aims to present another angle of research, which is solely economic. We analyzed the effect of female labor force participation, female income, men unemployment, total household income, and aggregate unemployment and education level on the divorce rate. In order to estimate our model we use four estimations of panel data.

Literature review

There are plenty of sources on determinants of divorce. Mainly the determinants can be divided into two groups, the first are the variables

that are determined to have exact positive or negative effect, and the second are the variables with ambiguous effect on divorce rate. One of the key determinants affecting the decision to dissolve marriage is participation of women in

the labor force. According to Kalmijn and Poort-

man (2006) the female labor force participation has positive effect on divorce rate. It is believed that, on one hand, a woman with a job threatens

the breadwinner status of the husband. They argue that a successful career of a woman undermines the occupational position of the husband. On the other hand, working women tend to have a strong belief that they are independent and competent. Hence, they are economically independent and have small financial costs of end

ing the marriage. Moreover, according to Trent and South (1989), high female income decreases the dependence on their partner's earnings, and economic opportunities of a woman gives more incentives to break up the marriage.

Another important variable that certainly affects the decision to divorce is education. Nonetheless, the effect of the education is ambiguous. Some scholars believe that highly educated people are more motivated to divorce, because they are sure in their ability to form a new family, and to find out a better partner. Moreover, according

to Kalmjin and Poortman (2006) educated people

have «fewer moral objections to divorce» (211).

Hence, people with good education background have fewer moral attachments towards the marriage and divorce. In contrast, the education may also have negative effect on the divorce rate. It is believed that people with good education background are better at solving the problems that arise between the spouses. According to Na- konezny et al. (1995) people with high education

are at lower risk of marriage dissolution. They state that educated people «may embody greater interpersonal skills, maturity and resources, that benefit a marital relationship» (478). Further

more, educated people are more likely to consult psychologist that may help to solve the family and marriage issues.

Additionally, unemployment is one more factor that affects the divorce rate, but its effect is controversial. Male unemployment is expected to have positive effect on divorce because generally accepted cultural perception of a man as a breadwinner is put under threat (Kalmjin, Poortman,2006). An unemployed man is unable to provide good living conditions for his family. Moreover, Amato and Beattie believe that initially when husband loses a job both partners are

quite optimistic about career prospects. However, later on situation in family gets worse due to the absence of the fundamental source of income, and more tensions might arise within the spouse. The stress after losing a job appears to be

the main reason for having distruptive harmony within the partners (Amato, Beattie, 2011). Also, there is an opposite effect of unemployment on the divorce rate. Unemployment can have indirect effect, which is mostly connected with costs associated with divorce. During hard times people knowing the costs of divorce, which include alimony, taxes and divorce fees, try to improve their relationships (Amato, Beattie, 2011). Hence, economic disincentives force people to reconsider their relationships.

Furthermore, such factors as poverty and household income are considered to have ambiguous effect on the divorce rate. On the one hand, according to Kalmjin and Poortman (2006) when the finance and economic opportunities are

limited, the family is tend to have more worries and less of harmony as well as marital satisfaction. The scholars believe that people living in bad financial conditions divorce more than couples living in better economic conditions. Low income and poverty are factors leading to dissolution of a marriage (Ambert, 2009). On the other hand, the greedy human nature may create problems in the families with good financial standing. These people are more concerned with financial issues, such as where to invest and what to buy. This in turn may generate growing tension between partners. In contrast, families with medium or low income are focused on providing the necessities first. As it was mentioned before, people in bad economics standing are less likely to divorce due to financial costs associated with divorce process.

Summarizing the evidences and hypothesis of the scholars we can construct a table identifying the signs of the variables.

Data

Dependent Variable

Each state including District of Columbia is taken into consideration for the purpose of this research project. Data on divorce rate is obtained

from annual reports of NDHS National Vital Statistics System for the years 2000 – 2010. This variable is measured by number of divorces per 1000 population. Divorce rate for some states such as California, Indiana, Hawaii, Louisiana and Georgia are not reported at NDHS. Even though we attempted to obtain lacking observations from bureau of statistics of each state, some observations were not found, which resulted in having unbalanced panel data. However, as long as the reason for missing the data is not affected by the factors included in the idiosyncratic error u_it, there is no need to worry about unbalanced panel data.

Independent Variables

Education (educ)

In this regression analysis education is expressed as a percentage of state population over the age 25 with a bachelor’s degree. Data on education is obtained from annual reports of U.S. National Center for Education Statistics. While observing all 51 units, we can see that education rate was fluctuating around the same range indi

vidually for each state within a 10-year period.

Female weekly earnings (lf.inc)

Data on female weekly earnings is derived from U.S. Department of Labor Statistics annual reports from 2000-2010. This variable measures average income in dollars women receive per week. In the regression analysis this variable is transformed into logarithmic function in order to have better behaved distributions and to reduce effect of outliers.

Unemployment (tot.unem)

Data on total unemployment is obtained from US Census Bureau of Statistics. Total unemployment measures percentage of unemployed people per labor force in each state. It worth pointing out that unemployment rate has considerably increased during financial crisis for some states.

Male Unemployment (male.unem)

This variable measures percentage of males per labor force, and it is derived from Bureau of Labor Statistics. We expect for male unemployment to be highly correlated with total unemployment rate, but as long as there is no perfect collinearity, assumptions are not violated.

Poverty (pov)

Data on poverty is reported annually in US Census Bureau. This variable shows percentage of state population living under poverty line in each state.

Female Labor Force Participation Rate (fem.lf)

This variable is expressed in terms of percentage of females over the age 16 who participate in the labor force of each state. Data on this variable is acquired from Bureau of Labor Statistics.

Household Income (lh.inc)

This variable is included in the model in logarithmic form in order to get rid of strongly positively skewed distribution of residuals. Household income is average income in dollars households receive per year. Data on average household income is obtained from Annual Social and Economic Supplements Report of Current Population Survey conducted by US Census Bureau of Statistics.

Data was available for 50 states and Washington DC for each of the 10 years from 2000 to 2010. Total number of observations for almost every variable equals to 561 (51x11). The concerned data set was sorted in a manner to meet panel dataset sorting requirements. All information related to data have been published online and are available for reference.

Empirical Model

Our unobserved effects model using panel data includes eight explanatory variables and dummy variables for each year except the base year. The year dummy variables were included because states may have different distributions in different time periods. We will estimate the coefficients using four methods for panel data:

Pooled Ordinary Least Square (POLS), First – Difference Estimator (FD), Fixed Effects Estimator (FE) and Random Effects Estimator (RE). We collected data for eleven year from 2000 to 2010 and used the year 2000 as base year. The Stata software package was utilized to obtain the results of four estimators.

The basic unobserved effects model:

Yit= βo +δ1d01t + δ2d02t + δ3d03t +δ4d04t

+ δ5d05t + δ6d06t + δ7d07t + δ8d08t + δ9d09t +

δ10d010t + β1educit + β2 lf.incit + β3tot.unemit

+ β4male.unemit + β5povit + β6fem.lfit + β7lh.

incit + β8lh.incsqit + ai + uit

where Y = number of divorces per 1000 population

i = states (panel) denoted using numeric key t = time period from 2000 to 2010

d01t = year dummy equals one if year = 2001 and zero otherwise, the same year dummies were applied for d02t to d10t.

(Explanatory variables of household income and female income were used in logged form)

The error term consists of two parts: fixed effect error, ai, and timevarying error, uit. The unobserved effects or unobserved heterogeneity captures time-constant factors of the state that do not change over time. Such factors might include the state-specific features such as general attitude

toward divorce. There are other determinants that may not be exactly constant, but they might be approximately constant over period, these include constituents such as demographic features (race, age), religious heterogeneity or historical reasons for divorce rate. Time-varying error or idiosyncratic error, uit includes unobserved factors that affect the dependent variable and change over time.

Pooled OLS is the first method that we explored to estimate the parameters of interest. In order for the estimates to be consistent we would have to make an assumption that there is zero correlation between unobserved effect ai and independent variables xitj (where j=1,..,k).

Whilst the issue of serial correlation or cluster correlation is dealt using stata command, consistency issue still leaves room for bias and inconsistency of estimates. In other words if fixed ef

fect, ai is correlated with any of the explanatory variables, which is most likely to be the case, then pooled OLS results will produce inconsistent results.

One of the main reasons for using panel data in our model is to permit the fixed effect, ai to be correlated with independent variables. The first-difference estimator (FD) allows for such correlation since it rules out time-constant error by differencing it from the equation of previous period. Idiosyncratic error, uit must be uncorrelated with independent, xitj variables in each

time period. The key assumption that makes FD produce consistent estimators is strict exogeneity assumption. However, one of the potential disadvantages of using FD is reduced variation in regressors. In our model some explanatory variables such as higher education level, poverty rate or female labor force participation rate tend to stay constant especially taking into consideration adjacent years. Differencing these variables further depresses variation in explanatory variables and most likely will produce large standard errors.

The fixed effects estimator (FE) like firstdifference also eliminates the fixed effect, ai and uses time demeaning on each explanatory variable. Strict exogeneity assumption on the explanatory variables implies that the estimated coefficients are unbiased: there should be zero

correlation between idiosyncratic error term uit and regressors xitj across all time periods. Stata

software provides heteroskedasticity and serial correlation robust coefficients so there should be no worry in that sense. Given that serial correlation problem is properly dealt, fixed effect is better estimator than first differencing and heteroskedasticity robust standard errors are valid. Therefore, we consider the fixed effects estima tor to be more valid compared to FD.

The random effects estimator (RE) is different from fixed effects because unlike FE it does not

allow zero correlation between unobserved effects ai and explanatory variables xitj. As well as FE, coefficient obtained from random effect esti mators derived from stata are heteroskedasticity and serial correlation robust. We will be comparing the results of random effects with fixed

effects estimators, but practically FE is widely believed to be a more cogent tool in interpreting the results in term of ceteris paribus. On the other hand, some of the explanatory variables that we employ in our model are roughly timeconstant such as higher education level and poverty rate, therefore along with fixed effects random effects

estimator is also a convenient method of estimation. Nevertheless, bearing in mind that our units of observations are big geographic areas (states) fixed effects estimator better suits aggregated data sets.

Discussion of Empirical Results

Drawing from the estimated results of POLS, education has negative effect on divorce rate and significant at 5% level against one-sided alterna tive. The signs of average female income, total unemployment and male unemployment are as were expected with female income being significant at 5% level. Interestingly household income has negative effect on divorce rate, but after it reaches certain point household income is predicted to have positive effect on divorce rate. This might be explained as follows: increase in the average household income in states is predicted to decrease divorce rate up to the point when the average household income will be high enough to allow more independence. The F-test for joint significance of yearly dummy variables showed that these are jointly not significant.

Estimating the unobserved effects model with fixed effects estimator produced statistically insignificant estimators at any reasonable significance level. The signs of several explana

tory variables such as average female weekly earnings by states and total unemployment rate do not match the expected effects. As it was discussed above since these variables do not show significant variation across years and therefore,

differencing those explanatory variables depresses variation even more and might lead to misleading results.

As it was predicted in empirical model results obtained from the fixed effects estimators matched with their expected signs. Higher education level has positive effect on divorce rate, but it is not statistically significant. Random effects estimator predicts negative effect of higher education level on divorce rate. Other variables of FE and RE show similar magnitude of effect of each explanatory variable. The RE coefficients of average female weekly earnings and male unemployment rate are significant at 5%

level against one-sided alternative.

We also estimated the same model but using only data from 2006 to 2010. Estimating determinants of divorce rate for shorter time period allows state demographics of race and sex ratio and also religious heterogeneity to stay roughly constant and provide more convenient results. POLS for five years produced nearly the same results in terms of signs of the coefficients with only poverty rate having different sign compared to full set of eleven years. FE estimators of eleven years have the same magnitude as FE estimators of only five years and in the latter case

the coefficient on higher education level being significant at 5%.

Having two unobserved effects model: one

for eleven years and another for last five years

helps us interpret the potential important variables that are missing in our model. The results of model obtained from five years estimations

of RE and FE shows that assumption of zero correlation between unobserved effect, ai and control variables xitj is violated which makes

Pooled OLS and Random Effect results biased and inconsistent. Results from First-Difference estimators are also inconsistent because explanatory variables that we used do not vary signifi

cantly over time and differencing them further reduces variation. Therefore, fixed effect esti

mators are more reliable in terms of the direction of the relationship of regressors on divorce rate.

The difference in the direction of the relationship obtained from four methods shows that our model suffers from omitted variable bias. Such important factors as number of children and laws that affect divorce-processing time and filing prices which vary by each state were

not included in our model.

This leads us to the considerations for the areas of further research on the determinants of divorce rate where we should include important missing variables to reach consistent results.

Conclusion

This paper examined factors of divorce rate using the data on fifty states of the USA and the

District of Columbia from 2000 to 2010. The given research work differs from previous researches, because it considers only economic correlates of the divorce rate such as female labor force participation, female income, poverty rate, total unemployment level, male unemployment, educational attainment and average household income. Our assumption relates to the level of education, aggregate and male unemployment rates, poverty and household income, which we categorized as ambiguous ones, due to their questionable effect on divorce rate. Only female weekly earnings were assumed to change in opposite direction with divorce rate.

Regression results from panel data, particularly Fixed Effect estimator reveals that similar to our hypothesis, increasing female labor force participation leads to higher levels of divorce rate in the society, which is also very significant estimator. Taking into account that Fixed Effect produces more contingent values particularly for panel data than other estimations, it can be said that our initial characterization of the female variable is consistent with the results. As for the female weekly earnings, the results of POLS, FE and RE estimations illustrate that whenever female income rises, divorce rate will decline. These results also confirm results from the previous research works.

In addition, our findings demonstrate that male unemployment rate is negatively correlated with divorce rate, which means that decreasing number of men in labor force decreases divorce rate. Overall, important time-variant variables such as number of children and divorce laws can affect explained variable and therefore should be considered in the model.

REFERENCES:

1. Kalmijn M. and Anne-Rigt Poortman, «His or Her Divorce? The Gendered
2. Nature of Divorce and Its Determinants», European Sociological Review 22 (2006): 201-214, accessed: 06-04-2015, http://www.jstor.org/stable/3559577
3. Trent K. and Scott J. South, «Structural Determinants of the Divorce Rate: A Cross-Societal Analysis», Journal of Marriage and Family 51(1989): 391-404, accessed: 06-04-2015, http://www.jstor.org/stable/352502
4. Amato P. and Bret Beattie, «Does the unemployment rate affect the divorce rate», Social Science Research 40 (2011): 705-715, accessed: 11-04-2015, www. elsevier.com/locate/ssresearch
5. Annie-Marie Ambert, «Divorce: Facts, Causes and Consequences», Report Contemporary Family Trends, accessed: 19-04-2015, http://www.thefamilywatch. org/doc/doc-0073-es.pdf
6. Nakonezny P. Rebert Sh., and Joseph Lee Rodgers, «The effect of no-fault divorce law on the divorce rate across the 50 states and its relation to income, education, and religiosity», Journal of Marriage and Family 57 (1995), accessed: 19-04-2015, https://www.questia.com/library/journal/1P3-6663604/the-effect-of-no-fault-divorce-law-on-the-divorce
7. Divorce Rate: NDHS National Vital Statistics System. United State Census Bureau.
8. Female Labor Force Participation Rate: Labor Force Participation Rates by Marital Status, Sex and Age. United States Census Bureau.http://www.census.gov/compendia/statab/2012/tables/12s0597.pdf
9. Female weekly earnings: Highlights of Women’s earning in 2000 (up to 2010). Median weekly earnings of full-time wage and salary workers by State and sex, 2000 (up to 2010) annual averages. Bureau of Labor Statistics. http:// www.bls.gov/cps/earnings.htm
10. Household Income: Median Household Income by State: 1984 to 2013. Table H-8. United States Census Bureau. link
11. Male Unemployment: Geographic Profile of Employment and Unemploy ment. Bureau of Labor Statistics. http://www.bls.gov/gps/tables.htm
12. Poverty: Number of and poverty rate, by state: 1980 to 2013. Table-21. United States Census Bureau. http://www.census.gov/hhes/www/poverty/data/ historical/hstpov21.xls
13. Total Unemployment: Total Unemployed and Insured Unemployed by State: 2000 to 2010. Table 629. Un