The U.S. Gender Earnings Gap: A State- Level Analysis

Transcription

1 The U.S. Gender Earnings Gap: A State- Level Analysis Christine L. Storrie November 2013 Abstract. Although the size of the earnings gap has decreased since women began entering the workforce in large numbers, the gender earnings gap is still of concern since in many U.S. households, women s wages constitute a major source of family income. I estimate a model using fixed effects to estimate earnings for men and women to determine if state of residency can be linked to any size differences in the ratio of men s to women s earnings by state. I find state variations in average earnings for both men and women exist with no pattern emerging for the variations of states size deviations from the national mean ratio when control variables are added. The relative size of the state s gender earnings ratio often varies significantly between the unadjusted model and the model that controls for individual worker characteristics. Keywords: Gender wage gap, fixed effects model JEL classification: J16 Corresponding author: Christine L. Storrie, Department of Economics, University of Delaware, 338 Purnell Hall, Newark, DE 19716, USA, E- mail: storrie@udel.edu

2 1. Introduction The difference in pay between men and women is referred to as the gender earnings gap. In the United States today, women earn an average of 23% less than men earn (DeNavis-Walt, Proctor, and Smith 2012). The size of the overall earnings gap has decreased since 1960, when women earned only 60% of men s average pay. The narrowing of the gender earnings gap is due largely in part to women s increased education and workforce participation. Although the size of the earnings gap has decreased since women began entering the workforce in large numbers, the earnings gap is still of concern since in many U.S. households, women s wages constitute a major source of family income. Married women currently contribute over 36% of the total family income and in 34% of U.S. families, mothers are the sole wage earners (Hallman 2013). The gender earnings gap can be attributed mostly to an individual s choices as they pertain to education, career and lifestyle. Determining the causes of the earnings gap for men and women is challenging since many factors and varying individual characteristics, many of which are difficult to measure, attribute to one s educational and career path and ultimately one s salary. This makes identification of a clear-cut cause challenging. Existing studies on the gender earnings gap consider individual worker characteristics that potentially attribute to the overall gender earnings gap. Findings in recent literature indicate that earnings differences between men and women are impacted by gender differences in average number of hours worked per week, choice of occupation and number of career interruptions (Getz 2010; Goldin and Katz 2008; Hallman 2013). None of the prior studies of the gender earnings gap in the United States 2

3 however, have examined variations in the size of the earnings gap based on respondent s location, and particularly in relation to one s state of residence. In this paper, I use data from the Current Population Survey (CPS) March supplement to test for variation in the gender earnings gap across states in the United States. I attempt to determine if state of residency can be linked to any size differences in the ratio of men s to women s earnings. A report issued by the U.S. Census Bureau s American Community Survey (ACS) citing the average national gender earnings gap as well as the average size of the earnings gap for each state serves as motivation for this paper. The ACS report highlights differences in earnings between men and women at both the national and state levels. The report lists median earnings for men and women for all states as well as the national average and determines women s earnings as a percentage of men s earnings. This report does not, however, use any empirical tests to determine any significant state effects beyond presenting the median earnings values. The ACS reports annual earnings for men and women. The 2009 survey reports men s overall median earnings in 2008 for full-time, year round workers are $45,556. Women in the United States during the same period earned a median income of only $35,471 or only 77.9% of men s earnings. Additionally, women earned a lower salary than men on average in each of the 50 states (Semega 2009). Building upon the results presented in the ACS report, I test for any statistically significant differences in mean earnings for each gender by state. To determine the size of each 3

4 state s earnings gap, I measure the differences in earnings by estimating income in separate regressions for both men and women using three empirical specifications. Next, I determine if these differences in earnings can be attributed to a possible state effect. I compare the earnings differences based on state effects alone to those obtained when controlling for workplace and individual characteristics commonly used in labor economic studies for estimating earnings. The ratio of the differences in the state effects for men and women relative to the national mean gender earnings ratio (NGER) serves as the measure of the gender earnings gap for each state. I find state variations in average earnings for both men and women exist. Louisiana has the lowest gender earnings ratio (GER) in both estimated models. Maine and Massachusetts both have a gender earnings ratio that is 5.1% above the NGER when controlling for other variables. Many states display a GER that is similar to the NGER, but many states fall well below the NGER, meaning they have a larger earnings gap than the national average earnings gap. I find no pattern emerges for the variations of states size deviations from the national mean ratio when control variables are added to the unadjusted base model for state fixed effects. I find the relative size of the state s GER often varies significantly. 2. Review of Literature A report presented by the American Community Survey (ACS) highlights findings on men and women s earnings at the national and state levels based on data obtained from the 2009 ACS. Earnings are defined as the sum of an individual s wage and salary income plus any selfemployment income. The sample was restricted to include only full-time year-round workers 4

5 aged 16 years or older. A year-round worker is defined as an individual that worked 50 or more weeks in the past 12 months and included the individual s paid time off or sick time as weeks worked. They also consider full-time to mean 35 or more hours worked in a week (Getz 2010). i The ACS survey reports median earnings for both men and women. Overall, women s earnings as a percentage of men s for the United States were 77.9% (Semega 2009). Perhaps a more interesting result of this survey, however, was women s median earnings as a percentage of men s earnings varied substantially by state. Wyoming had the largest earnings gap among all states based on median earnings. According to the ACS report, women s earnings in Wyoming were only 64% of men s earnings, In Washington DC, women earned 88% of men s median income, the smallest earnings gap when measuring median earnings. These states represent the highest and lowest values for the gender earnings gap, but by no means are they outliers. The 2009 ACS report shows that women in Wyoming earn only 65.5% of a man s income and Utah women earn 66.4% of what men earn. In fact, four states show women s income as a percentage of men s to be under 70% while in nine states, women earn 80% of men s earnings or higher. Goldin and Katz (2008) discuss factors impacting earnings differences between men and women. Their results show the gender earnings gap and the correlation to some women s choices to have careers that fit with family-oriented goals as opposed to career goals. Examples of these choices could be choosing jobs that offer shorter workweeks or less travel and overall shorter time commitments. The tradeoff for these family-friendly benefits is typically a lowerpaying salary. They also found fewer men taking career interruptions than women do. A study of MBA s from the University of Chicago found three main causes of earnings differences of 5

6 men and women in corporate careers (Bertrand, Goldin, and Katz 2010). Women have slightly lower GPAs and took fewer finance courses than their male counterparts. They also found that women worked fewer hours compared to men. Women worked on average 52 hours per week in the first fifteen years of their careers compared to men who worked an average of 58 hours per week. Lastly, they found that only 10% of men in their study took a career interruption while 40% of the women went six months or more without working. Results obtained from prior studies reinforce the fact that there is in fact a gender earnings gap and seek to find sources of the gap. These studies do not however examine any geographical or state connection and a variance in the earnings gap. 3. Data Description To obtain a more robust sample size, I include data for two years. I use the Current Population Survey (CPS) March Supplement for the years 2008 and The original sample size of the CPS data set includes 138,018 observations with observations for all ages and employment categories. I use the Current Population Survey instead of the American Community Survey because the CPS provides measures of key explanatory variables that I use in my model to adjust for individual worker attributes. To test the comparability of the CPS data with the ACS, I replicate the results for median incomes as well as present mean incomes for men and women for each state for full-time year-round workers aged 16 and older. Table A1 of the appendix shows average earnings for each state for men and women measuring both median and mean. 6

7 The results for the median earnings of men and women for each state are comparable to the ACS s findings. In addition to median earnings, I also calculate the mean earnings for each state for both men and women. I use the means as the measure of averages for my analysis and all other estimates and comparisons in this paper use the mean as a measure of average earnings. I restrict the sample size to include only those full-time year round workers aged In addition, I further restrict the sample to include only individuals with one year of work experience or greater. These restrictions reduce the sample size to 136,268 observations for the fifty states as well as Washington DC. The number of observations per state varies, but each state has over 1,000 observations at a minimum, ensuring an adequate sample size from each state. 4. Empirical Methods I create dummy variables for education, marital status, race 1, Hispanic origin, metropolitan status and citizenship classification. In addition to the dummy variables, I use number of children under 18 in the household and years of work experience 2 as well as experience-squared. There are two common approaches for estimating fixed effects models. A within-groups method can be used to estimate the regression to net out the unobservable effects. A least squares 1 Race categories are white, black and Asian. Due to the low sample sizes of all other categories I treat all sub- categories of race as one group and together they act as the omitted variable. 2 I calculate experience by using age minus n minus 6. If the level of education was less than or equal to 10 th grade, n=10. For level of education equal to 11 th grade, n=11. If education=12 th grade n=12 regardless of whether they obtained a diploma or not. N=14 if there was some college and n=16 if they obtained a Bachelor s degree. For a Master s degree, n=18 and for a doctorate or professional degree, n=20. 7

8 dummy variable approach can also be used. This approach is done by creating dummy variables, which brings the unobserved effect explicitly into the model. Typically this method is not preferred over the within groups method since it adds extra variables to the regression equation and causes a loss in degrees of freedom. It is also common to have many fixed effect units with few observations per unit. The fixed effects may be of little interest or may be infeasible to recover. In my analysis however, the sample size is large, providing many observations for each state fixed effect. My goal was to capture and measure the state fixed effects since they can be interpreted as a measure of the differences in the average state earnings. I use fifty dummy variables to capture the state effects for the 50 states and DC 3. First, I examine any effect of the state of residence on earnings for each gender. I use OLS regression with fixed effects for state and year to test this (see equation (1) below). The state variable is a set of dummy variables for the individual state effects. Separate equations are estimated for men and women to capture the fixed effects for both genders for each state. Earnings i = α 0 + α 1 State ij + α 3 Year i + ε it (1) Next, I add variables to measure individual worker characteristics to the state effects model. Since the goal was to determine if state effects exist and not to determine the actual cause of the difference, I choose a parsimonious model adding variables commonly used in many labor economic studies. 3 In the regressions that only measure the fixed effects, the intercept can be interpreted as the mean for New Hampshire, which is the omitted variable for this study. 8

9 I re-estimate the model with these individual worker characteristic variables included. Equation (2) represents the regression equation for earnings based on state controlling for individual worker characteristics. The coefficient of interest is β 1, which is the coefficient on State j. This coefficient represents the estimates for the effect of state of residence of respondent on earnings controlling for education, race and ethnicity, marital and family status, citizenship, the two work experience measures as well as a dummy variable for year. Again, estimates are obtained for both men and women in separate OLS estimates. Earnings i = β 0 + β 1 State ij + β 2 Educ i + β 3 Race i + β 4 Marital i + β 5 Nund18 i + β 6 Hisp i + β 7 Metro i + β 8 Citizen i + β 9 Exper i + β 10 Exper 2 i + β 11 Year+ ε it (2) I also estimate an equation that does not include any state measure: Earnings i = γ 0 + γ 1Educ i + γ 2Race i + γ 3Marital i + γ 4Nund18 i + γ 5Hisp i + γ 6Metro i + γ 7Citizen i + γ 8 Exper i + γ 9Exper 2 i + γ 10 Year + ε it (3) Equation (3) represents earnings based on individual characteristics. This regression equation represents earnings estimated for the entire restricted sample controlling for personal attributes. Similarly, men and women s earnings based on these individual traits are estimated in separate regressions. 5. Empirical Results Summary statistics for the entire sample are found in Table 1. For the entire restricted sample, men on average, have more children in the home than women. More men are married than women and there are a higher percentage of white men than white women in the sample. Men also have higher percentages of those having only a high school diploma or less. Women on 9

10 the other hand, have a higher occurrence of US citizenship than men in the sample. Women also have a higher percentage of individuals who are black as well as those of Hispanic origin. The distribution of the sample for the other key variables tends to be fairly similar for men and women. Estimated coefficients for regression equations (1) and (2) are presented in Table 2. Estimates of the coefficients for the state fixed effects model are shown for men and women respectively. Columns (1) and (5) represent estimates the coefficient for the state dummy variable for equation (1). These columns represent results for the unadjusted model, which only estimates earnings using the state fixed effects with an added year fixed effect. The adjusted model, shown in Columns (3) and (7) of Table 2, represent the coefficient for the state dummy variables for equation (2) that include adjustments for individual worker traits. Since earnings are estimated using the natural log of earnings, the coefficients can be interpreted as percentage difference from the constant, which represents estimates for the mean log earnings for the omitted state New Hampshire. Most of the estimated coefficients are significantly different from zero at the 99% level of confidence and all were significant at the 90% confidence level. I use the results from the estimates from equations (2) and (3) to test the significance of state fixed effects on earnings 4. An F test concludes that the null hypothesis that the state effects are all equal to zero is rejected at all levels of significance indicating that state of residence impacts earnings. 4 Regression results for equation (3) in addition to the individual worker trait coefficients for equation (2) are presented in Table A2 in the appendix of this paper. 10

11 The raw regression results presented in Table 2 are estimates relative to the constant. For ease of interpretation, I re-scale the raw estimates as differences from the overall national mean. The overall average values are the mean of the states scaled mean earnings for each equation. I calculate the state effects for each sex. This value is used to determine the GER for each state. I compute the gender earnings ratio in state j as: GER = (!!!!") λ (4) (!!!!" ) where s equals the estimated state effect on earnings for men and women in state j (i.e. α 1 in equation 1 and β 1 in equation 2). NGER, as represented by λ in equation (4), equals the mean ratio for the average female to average male earnings. The national average gender earnings ratio is based on the mean earnings for men and women presented in Table 1. The results of the calculations are presented in Table 3. The GER columns are defined as the ratio of women s earnings to men s in each state relative to the overall mean ratio for women and men s earnings as determined from equation (4). Each state s GER for the unadjusted state effects only model is presented in columns (1) and (3). The state GER s for the adjusted model are found in columns (2) and (4). All states display an earnings gap based on gender. The overall size of the earnings gap decreases when controlling for individual and workplace characteristics. Additionally, the size of the earnings gaps does not remain constant when worker variables are added relative to the unadjusted model. The adjusted GER relative to the NGER tends to be smaller than the unadjusted GER in most states. Eight states have virtually no change in the size of their GER 11

12 relative to the NGER when comparing size of the adjusted GER to the unadjusted GER. Another 19 states only show a 1% decrease in the size relative to the national mean ratio between the models. This is not the case for all states however. The relative size decreases for most individual state earnings ratios, but the overall variation of the state effects remains for the adjusted model. Interestingly, the size of the GER for each state relative to the national GER does not remain constant between models. Figures 1 (a) and (b) illustrate some of the variations in size of the GER relative to the NGER for each model. I present the ten largest differences from the NGER for the unadjusted state fixed effect model in Figure 1(a). The composition of the states alters when adjustments for individual worker traits are included. The ten states with the largest deviation from the national mean ratio are presented in Figure 1(b). Only four states remain in the top ten largest deviations when adjusting for personal traits. Louisiana continues to have the largest deviation from the NGER for both models having only a slight decrease in size deviation for the adjusted model. For ease of presentation, only the largest deviations from the NGER are presented in Figure 1. The standard deviations for the state GER s are and for the unadjusted and adjusted models respectively. These results clearly show the presence of a state effect on the gender earnings gap. The ten largest differences between the unadjusted model and the adjusted model are presented in Figure 2. For example, Utah is 6.3% below the national mean ratio in the unadjusted model. Utah s GER relative to the NGER for the adjusted model, however, is insignificant. This implies that the variation in the state effect can be explained by individual 12

13 worker attributes for observations for that state. Minnesota, on the other hand, actually increases its positive difference from the national mean ratio in the adjusted model by approximately 4.5%. Still other states display a positive deviation from the NGER in the unadjusted model and find state GER s below the national mean ratio when adjusting for individual characteristics. 6. Conclusion This study tests for any significant differences in the size of the earnings gap based on state of residence. I find that the earnings gap as measured by the gender earnings ratio varies among states. I use regressions analysis using unadjusted and adjusted models and find GER variation among the states. The sizes of the earnings gaps vary by state when control variables are included, with no discernable patterns in the variations between the adjusted and unadjusted models. Due to the identification challenges of a generalized gender earnings gap, I hold individual worker characteristics constant for each state as well as assume that any possible omitted variables are held constant across all states since my goal is to determine the existence of a state varying gender earnings gap, not causes for the variations. The individual factors that vary across states could be a topic for future work. A possible pattern of mostly positive deviations from the national mean gender earnings ratio for the states in the northeast and mostly negative deviations for states in the south would suggest a possible connection between GER size and region of the country. Testing for a regional effect did not fall into the scope of this paper but could be subject of future research as well. 13

14 TABLE 1 Summary Statistics, Year- Round Full- time Workers Aged 18-65, Male Mean Std. Error of Mean Mean Earnings $61,098 $42,274 Female Std. Error of Mean Age Number of persons under age 18 - Household Metro status US Citizen Years of work experience Less than HS diploma HS Diploma Some College College Graduate Post- college education Married White Black Hispanic Asian Number of observations Data from CPS March supplement data obtained from DataFerrett. Means are calculated using restricted sample with observations from all states and Washington, D.C. The means for education, race, ethnicity as well as marital, metropolitan and citizenship statuses are calculated using dummy variables where the variable equals 1 if yes and 0 if no. 14

15 TABLE Men and Women s Earnings by State Results obtained from OLS regression estimates for equations (1) and (2) and depict raw results for the state dummy variables. Estimates are presented relative to New Hampshire. The Unadjusted columns represent estimates of log earnings from equation (1). Estimates for equation (2) are indicated in the Adjusted columns. State (1) Unadjusted (2) SE Men (3) Adjusted (4) SE (5) Unadjusted Women (6) SE (7) Adjusted Constant Alabama Alaska Arizona Arkansas California Colorado Connecticut DC Delaware Florida Georgia Hawaii Idaho Illinois Indiana Iowa Kansas Kentucky Louisiana Maine Maryland Massachusetts Michigan Minnesota Mississippi (8) SE 15

16 Men Women State (1) Unadjusted (2) SE (3) Adjusted (4) SE (5) Unadjusted (6) SE (7) Adjusted (8) SE Missouri Montana N. Carolina N. Dakota N. Mexico Nebraska Nevada New Jersey New York Ohio Oklahoma Oregon Pennsylvania Rhode Island S. Carolina S. Dakota Tennessee Texas Utah Vermont Virginia W. Virginia Washington Wisconsin Wyoming Dependent variable is log earnings for full- time year round workers aged with experience greater than zero. Sample period is and obtained from the CPS March supplement with a sample size of 136,268 observations. Standard Errors are reported. Adjusted R 2 for equation (1) was for the male regression and for the female regression. The adjusted R 2 for equation (2) is for men and for women. 16

17 TABLE 3 Measuring the Gender Earnings Gap by State Estimates for state s gender earnings ratio (GER) calculated using estimates obatined from equations (1) and (2) representing the unadjusted and adjusted models respectively. GER relative the the mean ratio indicate values above and below the national mean gender earnings ratio (NGER) for each state. Gender earnings ratio GER Relative to National Mean Ratio State (1) Unadjusted (2) Adjusted (3) Unadjusted (4) Adjusted Alabama Alaska Arizona Arkansas California Colorado Connecticut District of Columbia Delaware Florida Georgia Hawaii Idaho Illinois Indiana Iowa Kansas Kentucky Louisiana Maine Maryland Massachusetts Michigan Minnesota Mississippi

18 Gender earnings ratio GER Relative to National Mean Ratio State (1) Unadjusted (2) Adjusted (3) Unadjusted (4) Adjusted Missouri Montana N. Carolina N. Dakota N. Mexico Nebraska Nevada New Hampshire New Jersey New York Ohio Oklahoma Oregon Pennsylvania Rhode Island S. Carolina S. Dakota Tennessee Texas Utah Vermont Virginia W. Virginia Washington Wisconsin Wyoming

19 FIGURE 1 Variation in the Gender Earnings Ratio 10 largest deviations from national mean GER for the adjusted and unadjusted models ranked by absolute size differences. (a) Undjusted Model Louisiana District of Columbia Utah Wyoming Kansas N_Mexico Idaho California Alabama Vermont (b) Adjusted Model Louisiana Wyoming Mississippi Maine Massachusetts Alabama Minnesota Oregon Alaska District of Columbia

20 FIGURE 2 Unadjusted and Adjusted GER Size Variation by State States with the Largest Difference between the Unadjusted and Adjusted Models for the Gender Earnings Ratio for the State Relative to the National Mean Ratio Utah Idaho District of Columbia Minnesota Nevada Hawaii New_York Washington California Kansas Unadjusted Adjusted The Unadjusted figures represent the differences from the NGER using estimates obtained from equation (1). The Adjusted figures represent author calculations using values obtained using equation (2). Individual effects represent the differences in equation (2) for state effects controlling for individual worker characteristics. Figure is ranked based on the size of the difference between the unadjusted and adjusted models relative to the national mean GER. 20

21 Bibliography Bertrand, Marianne, Claudia Goldin, and Lawrence F. Katz "Dynamics of the Gender Gap for Young Professionals in the Financial and Corporate Sectors." American Economic Journal: Applied Economics 2 (2): DeNavis- Walt, Carmen, Bernadette D. Proctor, and Jessica C. Smith Income, Poverty and Health Insurance Coverage in the United States: Getz, David M Men s and Women s Earnings for States and Metropolitan Statistical Areas: 2009<br />: U.S. CENSUS BUREAU. Goldin, Claudia and Lawrence F. Katz "Transitions: Career and Family Life Cycles of the Educational Elite." American Economic Review: Papers and Proceedings 98 (2): Hallman, Linda D The Simple Truth about the Gender Pay Gap. Semega, Jessica Men s and Women s Earnings by State: 2008 American Community Survey: U.S. CENSUS BUREAU. 21

22 APPENDIX TABLE A1 Mean and Median Earnings By Gender and State Source: Author Calculations from CPS data for Full- Time Year- Round Workers Aged 16 and Older. ACS data replicated using data from 2008 ACS Survey results (Getz 2010, 1; Semega 2009, 1). 22

23 APPENDIX TABLE A2 Measuring the US Gender Earnings Gap Men Women State Fixed Standard Standard Standard Effects Yes No Yes No Error Error Error Included Constant Standard Error Metro Area US Citizen Self Employed Number of children Hispanic < HS Grad HS Grad College Grad Post College Married White Black Asian Experience Experience Year Estimates of log earnings obtained from OLS regression equation (2) and (3). Estimates for individual worker characteristic dummy variables for equation (2) for men and women respectively are presented in Columns (1) and (5). Columns 3 and 7 represent estimates for individual worker characteristic dummy variables for equation (3). The mean of the log earnings estimated for full- time year round workers aged with work experience greater than zero. Sample period is and obtained from the CPS March supplement. Adjusted R 2 for column (1) was and column (5) was Adjusted R 2 for column (5) is Adjusted R 2 for column (3) and (7) regressions are and respectively. All coefficients are significant at the 95% level of confidence. Number of observations is 136,268. i ACS Table 1 results were replicated and are included in the Appendix of this paper. 23