Vous êtes sur la page 1sur 7

Divorce Rate and its Determinants

Anisa Kosta Dinesh Vatani

Hypothesis: The Divorce Rate is dependent on Education, Income, Religion and the existence of No-Fault Divorce Laws. According to divorce statistics, the divorce rate in the United States for first marriage is between 45-50%, for second marriage about 65%, and for third marriage about 70%. Divorce in the United States seems like an unavoidable phenomenon, however it possible that some individuals/groups are more prone to divorce than others. This paper will analyze different variables, in search for correlation between these variables (explanatory variables) and divorce rates (depended variable). Although there are a myriad of factors that play a role in the divorce rates, only four explanatory variables were selected for this paper, based on theoretical relevance and data availability. Table 1: Variables (by US state) and their expected effects Variable(s) DIV (D) EDU (E) INC (E) REL (E) LAW (E) Definition(s) Expected Effect(s)

The number of divorces, per 1,000 population Percentage population, age 25+, with a bachelors degree or higher Negative Median Household Income (in current dollars) Ambiguous The combined percentage of Christian and Jewish adherents Negative The no. of years since the No-Fault Divorce Laws were enacted Positive

Divorce Rate (DIV) variable is measured by the number of divorces, per 1,000 population, by state. This is the depended variable, which we will attempt to explain in relation to the below listed explanatory variables. Education (EDU) variable is measured by the percentage of population, by state, age 25+, with a bachelors degree, or higher. It is expected that the effect on divorce rates would be negative as highly educated individuals would most likely take more time to examine all the costs and benefits of getting married and getting divorced, when compared to less educated individuals. Income (INC) variable is measured by the median household income, in current dollars, by state. The relation between income and divorce rates is a bit ambiguous as one could argue that income could have a positive or a negative effect on divorce rates. Higher income means financial independency, which could cause divorce rates to go up, however higher income is most likely correlated to higher education, which it is expected to be negatively correlated to divorce rates. Religion (REL) variable is measured by the combined percentage of Christian and Jewish adherents, by state. It is expected that the effect on divorce rates would be negative, as these religions are strict on the issues of marriage and divorce, and would present themselves as an additional barrier to individuals who contemplate divorce. No-Fault Divorce Laws (LAW) variable is measured by the number of years since the No-Fault Divorce Laws were enacted, by state, as of 2009. It is expected that the effect on divorce rates would be positive, as the presence of these laws allows for quicker divorces. It is reasoned that

the longer these laws have been in place, the more people become aware of them, and therefore utilize them. The data for all variables were collected for year 2009, as it was the year with the most complete data (see Appendix A). There were a few noteworthy exceptions. (1) Divorce rates for the states of California and Indiana were not available for year 2009. Per WJS, data for the two states is over 20 years old; therefore the two states were excluded from the model. (2) Divorce rates, for year 2009, were not available for a number of additional states. The most recent data was used for the states of Georgia (2003), Hawaii (2002), Louisiana (2003) and Minnesota (2004), as excluding these states made little difference in the R^2 value (0.47) and reduced the sample size by a significant amount (8%). (3) The census data for the percentage of Christian adherents was not available for year 2009; therefore the most recent available census data (2000) was used. Table 2: Description of Statistics (Maximum, Mean and Minimum for each) Variable(s) DIV (D) EDU (E) INC (E) REL (E) LAW (E) MAX Nevada (6.7 divorcees / 1,000) D.C (48.5%) Connecticut ($64,851) Utah (74.5%) Oklahoma (56 years) MEAN 3.6 divorcees / 1,000 26.5% $49,098 50.7% 36 years MIN Massachusetts (2.2 divorcees / 1,000) West Virginia (17.3%) Mississippi ($35,078) Oregon (31.4%) Utah (22 years)

Using the 2009 data, a few generalizations can be made about the entire nation. Roughly 3.6 individuals of every 1,000 population filed for divorce; 26.5% of the population of age 25 and up, had at least a bachelors degree; the average household income was $49,098; 50.7% of the population defined themselves as Christian or Jewish adherents, and the average length of time that No-Fault Divorce Laws had been in place, as of 2009, was 36 years. Prior to running the regression, multicollinearity testing was performed in order to ensure no high multicollinearity was detected between the explanatory variables. Intuitively, a high linear relationship between the education and income variables was expected, therefore the need for remedial measures might have been necessary. One way to detect multicollinearity is to observe the pair wise correlations among the explanatory variables. Table 3: The Correlation Matrix Variable(s) EDU (E) INC (E) REL (E) LAW (E) EDU (E) 1 0.6921 0.0758 0.1319 INC (E) 1 (0.0291) 0.1688 REL (E) LAW (E)

1 (0.2913)

As Table 3 shows, the suspicion of a high relationship between education and income was correct. The relationship, however, was less than 0.7, which was not considered to be severe. To eliminate any doubts of high multicollinearity between the explanatory variables, auxiliary regressions were performed. In other words, each explanatory variable was regressed on the remaining explanatory variables and R^2 values were observed (see Appendix D). As Table 4 shows, all auxiliary regressions had low R^2 values, therefore it was concluded that no high multicollinearity existed between the explanatory variables. Table 4: Auxiliary Regressions Variable(s) EDU (E) INC (E) REL (E) LAW (E) Intercept (2.3143) 23694.3673 67.6265 37.7539 EDU (E) 927.7798 0.3710 0.0872 INC (E) 0.0005 (0.0002) 0.0001 REL (E) 0.0571 (44.9056) (0.1603) LAW (E) 0.0458 76.1285 (0.5475) R^2 0.49 0.49 0.10 0.11

Given the nature of the data (cross-sectional) the model was based upon, it was feared that heteroscedasticity was be present. To test for it, Whites General Test of Heteroscedasticity was be used. For the test, sets of hypothesis were set. The null hypothesis stated that homoscedasticity existed, whereas the alternate hypothesis stated the presence of heteroscedasticity.

In contrast to heteroscedasticity, which is generally associated with cross-sectional data, autocorrelation is usually associated with time series data; therefore the existence of autocorrelation within the data was very small. A test for autocorrelation was, however, performed in order to ensure its nonexistence. To test for autocorrelation, the Durbin-Watson d Test was used. For the test, sets of hypothesis were set. The null hypothesis stated that no autocorrelation existed, while the alternate hypothesis inferred the presence of autocorrelation. Using the regression residuals (Table X), i.e. the difference between observed and predicted values of the dependent variable, the sum of squared differences in successive residuals to the RSS, popularly known as the Durbin-Watson d statistic, was obtained.
4

Table X: Regression Residuals (observations 1-10 only, in the interest of space) Observation 1 2 3 4 5 6 7 8 9 10 Predicted Divorce Rate 4.0391 4.4130 4.1595 4.1564 3.2937 2.8334 3.9505 1.4343 4.1869 3.8302 Residuals 0.3609 (0.0130) (0.6595) 1.5436 0.9063 0.2666 (0.3505) 1.1657 0.0131 (0.6302)

The Durbin-Watson d statistic was calculated to be 2.0312 (see Appendix E). The number of explanatory variables in the model, excluding the constant term, was 4, while the sample size was 49; hence the following statistics were obtained from the Durbin-Watson d Statistic table: dL 1.378 dU 1.721 4 - dU 2.2790 4 - dL 2.6220

The Durbin-Watson d statistic value of 2.0312 fell between dU and 4 dU, which meant that the null hypothesis was accepted, in other words, just as expected, no autocorrelation was present. Given that the tests for multicollinearity, heteroscedasticity and autocorrelation presented no issues, the regression of divorce rates on education, income, religion and no-fault divorce laws, yielded the following equation (see Appendix B): DIV = 7.0323 - 0.0914EDU - 0.0291REL + 0.0104LAW The above equation could be interpreted as follows: while keeping all else constant, divorce rates per 1,000 population (1) decrease by 0.0914 as the percentage population, age 25+, with a bachelors degree or higher increases by 1 unit (%); (2) are not affected by income; (3) decrease by 0.0291 as the combined percentage of Christian and Jewish adherents increase by 1 unit (%), and (4) increase by 0.0104 as the number of years since the No-Fault Divorce Laws were enacted increases by 1 unit (year). The variable coefficients were consistent with the expectations established earlier in the paper, however the value of R^2, which is a measure of the overall goodness of fit, was not particularly high (0.46). Prior to any testing of the statistical significance of each explanatory variable, the global significance was run. The F-test, which is a measure of the overall significance of the estimated regression line, is also a test of significance of R^2. As calculated below, the F ratio equaled 9.3454, where as the critical F value equaled
5

3.83, 2.61 and 2.09 for 1%, 5% and 10% levels of significance, respectively. This concluded that the model was globally significant. R^2 0.4593 F-test (k - 1) 4 9.3454 (1 R^2) 0.5407 (n k) 44

Given that the model was globally significant, at least one of the variable coefficients was significant, i.e. different than zero. As the regression provided the t-statistics for each variable, each of them were tested using the t test of significance, in order to determine whether the variable coefficients were significant. The significance levels chosen for this comparison were 1%, 5% and 10%. The 1% critical value, using the two-tailed test and 44 d.f, was near 2.704, the 5% critical value was near 2.021 and the 10% critical value was near 1.684. When comparing the t-statistics to the critical values, the education and religion variables were statistically significant, while the income and no-fault divorce laws variables were statistically insignificant. Table XX: Regression Results and Variable Significance Variable(s) Specification Intercept EDU (E) INC (E) REL (E) LAW (E) R^2 7.0323 (0.0914) 0.0000 (0.0291) 0.0104 0.4593 t-statistics 6.6183 (3.6914) 0.2883 (2.9986) 0.5830 Significant at 10% (1.684)
X X

Significant at 5% (2.021)
X X

Significant at 1% (2.704)
X X

In conclusion, the purpose of the paper was to find variables that affected divorce rates, using cross-sectional data from 49 states in year 2009. According to the variable coefficients, divorce rates are negatively affected by education and religion variables, are not affected by the income variable, and are positively affected by the presence of no-fault divorce laws variable. The F-test, found the model including education, income, religion and no-fault divorce laws variables, to be globally significant, however only two of the variables (education and religion) were statistically significant and the model R^2 value was low (0.46). R^2 value of 0.46 meant that only 46% of the variation in divorce rates could be explained by the variation in education, income, religion and the existence of no-fault divorce laws. One reason for the low R^2 value could be model under fitting as only four variables were taken into consideration, while the variables that affect divorce rates are large in number. Adding two additional variables, namely the number of total children under the age of 18 and unemployment rate, increased R^2 value to 0.54 (see Appendix F). In the interest of time and because of the relatively small change in R^2 value, re-testing of multicollinearity, heteroscedasticity and autocorrelation was not performed. Other values that could have significantly improved the overall fit of the regression

equation would be age at marriage, the number of females who participate in the labor force, location, etc. The paper seems to suggest ways to worsen or improve the divorce rates relative to the chosen explanatory variables. In other words promoting higher education and religious beliefs would lower divorce rates, whereas increasing the ease of the divorce process would increase divorce rates. However, given that the variable coefficients are very small for all three explanatory variables, a substantial improvement would be needed in order for the divorce rates to be significantly affected.

Vous aimerez peut-être aussi