And Type the TITLE of YOUR WORK in All Caps

ESSAYS ON MARKET OUTCOMES AND RELIGIOSITY

NEIL R. MEREDITH

(Under the Direction of DAVID B. MUSTARD)

ABSTRACT

In one subject area of economics, the economics of religion, a growing body of research explores the relationships between market and religious outcomes. The essays add to this

research by investigating links between market outcomes and religiosity.

In the first essay with David Mustard, we consider differential enrollment growth for religiously affiliated postsecondary institutions relative to private secular institutions in the

United States from 1991-2005. We uncover evidence that enrollment growth in religiously affiliated institutions is higher for total, whites, blacks, Hispanics, and males than private secular institutions. We also address whether the religious intensity of an institution also affects enrollment gains. After controlling for other factors we find that intensely Protestant institutions are growing faster for total, blacks, Hispanics, and females relative to other Protestant institutions, which in turn are growing faster than their private secular counterparts.

For the second essay, I revisit the relationship between labor income and religiosity to explore whether an omitted variable creates an endogeneity bias. I conduct robustness checks using wages in place of labor income. With the exception of the relationship between labor income and religiosity for men, I find that ordinary least squares of the relationship between labor income and religiosity and between wages and religiosity are reliable due to weak

instrument problems. Panel estimation is also attempted for both genders. A lack of variation in frequency of attendance and frequency of prayer within individuals yields insignificant results.

In the third essay, I use count data estimation to evaluate the relationship between unemployment and the frequency of religious service attendance and the relationship between being out of the labor force and the frequency of religious service attendance for individuals of working age. I also address the duration of unemployment and time spent out of the labor force.

Results reveal that unemployment and the frequency of religious service attendance are uncorrelated. When out of the labor force, younger men, older men, and older women are predicted to attend less frequently.

INDEX WORDS: labor, religion, education, unemployment, income

ESSAYS ON MARKET OUTCOMES AND RELIGIOSITY

NEIL R. MEREDITH

B.A., Indiana University of Pennsylvania, 2006

A Dissertation Submitted to the Graduate Faculty of The University of Georgia in Partial

Fulfillment of the Requirements for the Degree

DOCTOR OF PHILOSOPHY

ATHENS, GEORGIA

2011

NEIL R. MEREDITH

ESSAYS ON MARKET OUTCOMES AND RELIGIOSITY

NEIL R. MEREDITH

Major Professor: David B. Mustard

Committee: Ronald S. Warren Christopher M. Cornwell

Electronic Version Approved:

Maureen Grasso Dean of the Graduate School The University of Georgia August 2011

DEDICATION

To my beautiful bride Amy, I would not have made it through this without your love, respect, and encouragement.

To my mentors Jim and Stephanie, the Drs. J., none of this would have been possible without the solid foundation, guidance, and encouragement you provided.

To my friends, thank you for all of your support and encouragement. I could not have done this without you.

ACKNOWLEDGEMENTS

I would like to thank David Mustard for the time, guidance, and encouragement he has provided throughout my graduate studies and research. As members of my dissertation committee, I would also like to thank Ron Warren and Chris Cornwell for their time, patience, comments, and suggestions throughout the development of this dissertation.

I would also like to thank my colleagues Chadi Abdallah and Tom McGahee who patiently listened to me and offered helpful suggestions as I developed my research ideas.

I gratefully acknowledge faculty and students in the UGA Department of Economics who attended research seminars at which I presented earlier drafts of this work. Your suggestions and insightful comments improved the quality of this research.

I am also grateful for comments I received from participants in the 2009 and 2011 Association for the Study of Religion, Economics, and Culture (ASREC) Conferences and participants in the

American Christian Economist (ACE) sessions at the 2010 Allied Social Science Association

(ASSA) Conference.

TABLE OF CONTENTS

Page

ACKNOWLEDGEMENTS ...... v

LIST OF TABLES ...... viii

LIST OF FIGURES ...... xi

CHAPTER

1 INTRODUCTION ...... 1

2 A POSTSECONDARY REVIVAL: THE IMPORTANCE OF RELIGIOSITY FOR

POSTSECONDARY ENROLLMENT GROWTH ...... 5

2.1 INTRODUCTION ...... 5

2.2 HYPOTHESES ...... 6

2.3 DATA AND VARIABLE DEFINITIONS ...... 12

2.4 ECONOMETRIC MODEL ...... 16

2.5 RESULTS ...... 18

2.6 CONCLUSIONS...... 23

3 LABOR INCOME AND RELIGIOSITY: EVIDENCE FROM SURVEY DATA ....36

3.1 INTRODUCTION ...... 36

3.2 DATA AND THEORY ...... 38

3.3 ECONOMETRIC MODEL ...... 46

3.4 RESULTS ...... 52

3.5 CONCLUSIONS...... 58

4 RELIGION AND LABOR: AN EXAMINATION OF RELIGIOUS SERVICE

ATTENDANCE, UNEMPLOYMENT, AND LABOR FORCE STATUS USING

COUNT DATA METHODS ...... 73

4.1 INTRODUCTION ...... 73

4.2 THEORY ...... 76

4.3 DATA AND ECONOMETRIC MODEL ...... 87

4.4 RESULTS ...... 96

4.5 CONCLUSIONS...... 103

BIBLIOGRAPHY ...... 126

APPENDICES

A APPENDICES TO CHAPTER 3 ...... 140

A.1 EXCLUSION OF OBSERVATIONS ...... 140

A.2 TREATMENT OF LABOR AND NON-LABOR INCOME ...... 141

A.3 INSTRUMENTAL VARIABLES ...... 143

B APPENDICES TO CHAPTER 4 ...... 148

B.1 NATIONAL LONGITUDINAL SURVEY OF YOUTH 1979 COHORT ..148

B.2 HEALTH AND RETIREMENT STUDY...... 150

B.3 COUNT DATA ESTIMATION METHODS ...... 153

vii

LIST OF TABLES

Page

Table 2.1: Descriptive Statistics-Raw Data ...... 27

Table 2.2: Descriptive Statistics-Whole Sample ...... 28

Table 2.3: Descriptive Statistics-CCCU ...... 29

Table 2.4: Total Enrollment Estimation Results ...... 30

Table 2.5: White Enrollment Estimation Results ...... 31

Table 2.6: Black Enrollment Estimation Results ...... 32

Table 2.7: Hispanic Enrollment Estimation Results ...... 33

Table 2.8: Male Enrollment Estimation Results ...... 34

Table 2.9: Female Enrollment Estimation Results ...... 35

Table 3.1: Descriptive Statistics for Weighted Pooled Cross Section Dataset ...... 60

Table 3.2: Descriptive Statistics for Weight Panel Dataset ...... 61

Table 3.3: First Stage Cross Section Results for Labor Income and Labor Income Squared ...... 62

Table 3.4: First Stage Cross Section Results for Wages and Wages Squared ...... 63

Table 3.5: OLS and Second Stage Cross Section Results for Attendance and Labor Income ...... 64

Table 3.6: OLS and Second Stage Cross Section Results for Prayer and Labor Income ...... 65

Table 3.7: OLS and Second Stage Cross Section Results for Attendance and Wages ...... 66

Table 3.8: OLS and Second Stage Cross Section Results for Prayer and Wages ...... 67

Table 3.9: OLS Cross Section Results for Spouse Education, Labor Income, and Labor Income

Squared ...... 68

viii

Table 3.10: OLS Cross Section Results for Spouse Education, Wages, and Wages Squared ...... 69

Table 3.11: Frequency of Prayer Transition Matrix ...... 70

Table 3.12: Frequency of Attendance Transition Matrix ...... 70

Table 3.13: First Differences of Prayer and Attendance ...... 70

Table 3.14: Descriptive Statistics for Unrestricted and Weighted Pooled Cross Section

Dataset...... 71

Table 3.15: Descriptive Statistics for Unrestricted and Weighted Panel Dataset ...... 72

Table 4.1: NLSY79 Descriptive Statistics Unrestricted and Weighted Sample ...... 107

Table 4.2: NLSY79 Descriptive Statistics Restricted and Weighted Sample ...... 108

Table 4.3: HRS Descriptive Statistics Unrestricted and Weighted Sample ...... 109

Table 4.4: HRS Descriptive Statistics Restricted and Weighted Sample ...... 110

Table 4.5: NLSY79 Pooled Negative Binomial Results for Frequency of Attendance and

Unemployment ...... 111

Table 4.6: NLSY79 Pooled Negative Binomial Results for Frequency of Attendance and Labor

Force Status ...... 112

Table 4.7: NLSY79 Poisson Fixed Effects Results for Frequency of Attendance and

Unemployment ...... 113

Table 4.8: NLSY79 Poisson Fixed Effects Results for Frequency of Attendance and Labor Force

Status ...... 114

Table 4.9: HRS Pooled Negative Binomial Results for Frequency of Attendance and

Unemployment ...... 115

Table 4.10: HRS Pooled Negative Binomial Results for Frequency of Attendance and Labor

Force Status ...... 116

Table 4.11: HRS Poisson Fixed Effects Results for Frequency of Attendance and

Unemployment ...... 117

Table 4.12: HRS Poisson Fixed Effects Results for Frequency of Attendance and Labor Force

Status ...... 118

Table 4.13: HRS First Difference of Attendance ...... 118

Table 4.14: Distribution of Frequency of Attendance Per Year in the NLSY79

Restricted Sample ...... 119

Table 4.15: NLSY79 Descriptive Statistics Restricted and Weighted 1979 ...... 120

Table 4.16: NLSY79 Descriptive Statistics Restricted and Weighted 1982 ...... 121

Table 4.17: NLSY79 Descriptive Statistics Restricted and Weighted 2000 ...... 122

Table 4.20: HRS Descriptive Statistics Restricted and Weighted 2004 ...... 123

Table 4.19: HRS Descriptive Statistics Restricted and Weighted 2006 ...... 124

Table 4.20: HRS Descriptive Statistics Restricted and Weighted 2008 ...... 125

LIST OF FIGURES

Page

Figure 2.1: Percent of 12th Graders who Report that Religion Plays a “Very Important” Role in

Their Lives ...... 25

Figure 2.2: Percent of 12th Graders who Expect to Complete a College Degree and Say that

Religion Plays a “Very Important” Role in Their Lives ...... 25

Figure 2.3: Average Enrollment Share for Blacks ...... 26

Figure 2.4: Average Enrollment Share for Hispanics ...... 26

Figure 4.1: Percent of Americans Who Attend Worship Services at Least Weekly and the U.S.

Unemployment Rate ...... 106

CHAPTER 1

INTRODUCTION

The economics of religion is a growing field that seeks to explore where economic

analysis can be applied to religious outcomes. Within this set of three essays I add to the growing body of economics of religion literature by using economic analysis to examine religion and education, religion and income, and religion and unemployment. In Chapter 2 with David B.

Mustard, we use institutional and state-level data from the National Center for Education

Statistics to explore the degree to which enrollment grows from 1991 to 2005 in religiously

affiliated postsecondary institutions relative to their private secular counterparts. After

controlling for institutional characteristics, we find that enrollment in religiously affiliated

colleges and universities grows 12, 26, 21,12, and 11 percentage points more for total, whites,

blacks, Hispanics, and males, respectively, than private secular institutions. Because simply

having a religious affiliation can have little or no bearing on an institution's policies and mission,

we evaluate whether the intensity of an institution's attachment also affects enrollment gains.

Enrollment gains in institutions in the Council for Christian Colleges and Universities (CCCU),

for whom Protestant faith is a direct determinant of institutional mission, are significant. For

example, after controlling for other factors total enrollment grows 11 percentage points, black

enrollment grows 35 percentage points, Hispanic enrollment grows 19 percentage points, and

female enrollment grows by 14 percentage points relative to other Protestant institutions, which

in turn are growing faster than their private secular counterparts.

Using data from the 1996-2004 General Social Survey, in Chapter 3 I examine the relationship between labor income and religiosity for individuals. I argue that the labor income and religiosity relationship is affected by endogeneity creating a bias. To test my hypothesis, I employ instrumental variables estimation for men and women separately. I expect findings for men and women to differ as established by research on labor market outcomes and religious behavior. Initial ordinary least squares results reveal that labor income and frequency of prayer have a statistically significant negative correlation for men. A $25,000 increase in labor income reduces frequency of prayer by once per week. There is no statistically significant relationship between labor income and frequency of attendance for men. Empirical results with instrumental variables estimation suggest that labor income is a poor predictor of frequency of prayer and frequency of attendance for men. A statistically significant negative correlation between labor income and frequency of attendance does not exist when utilizing instrumental variables.

Downward bias attributed to noncognitive factors accounts for these results. Using wages for a robustness check indicates similar results. Labor income for women is a statistically significant negative correlate of frequency of prayer and frequency of religious service attendance regardless of estimation technique. Estimates using ordinary least squares show that a $100,000 increase in labor income reduces frequency of religious service attendance by once per month and a $20,000 increase in labor income reduces prayer frequency by once per week for women.

Instrumental variables results reveal that a $9,090.91 increase in labor income decreases frequency of religious service attendance by once per month and a $2,083.33 increase in labor income decreases prayer frequency by once per week for women. A robustness check using wages in place of labor income reveals similar results for women. Instrumental variables diagnostic tests indicate a weak instruments problem for men when using wages as a predictor.

Weak instruments are a problem for women in all cases. A two-period, panel estimation is also carried out for both genders to explore the endogenous relationship. I do not obtain statistically significant relationships using the panel dataset regardless of estimation technique due in part to a lack of variation in frequency of attendance and frequency of prayer within respondents. I conclude that ordinary least squares provides the best estimates of the relationship between labor income and the frequency of prayer. It also yields the best estimates of the relationship between labor income and the frequency of religious service attendance for women.

For Chapter 4, I use count data estimation with data from the National Longitudinal

Survey of Youth 1979 cohort (NLSY79) and the Health and Retirement Study (HRS) to evaluate the relationship between the duration of unemployment and the frequency of religious service attendance and the relationship between time spent out of the labor force and the frequency of religious service attendance for individuals of working age. I also address whether being unemployed and being out of the labor force are associated with the frequency of religious service attendance.

The empirical results for men and women under age 50 in the NLSY79 using poisson fixed effects estimation show that there is no significant relationship between unemployment and the frequency of religious service attendance. The amount of time spent unemployed has no additional implications. Men who are under age 50 and out of the labor force at any point in the previous calendar year are predicted to attend religious services less than younger men who are never out of labor force in the previous calendar year. The magnitude fluctuates from 14 percent when not controlling for student status to 16 percent when controlling for student status. Each additional month out of the labor force, meanwhile, is positively correlated with a 2 percent increase in the frequency of religious service attendance for younger men when not controlling

for student status. This correlation is insignificant when controlling for student status. There are no significant relationships between labor force status and the frequency of religious service attendance or time spent out of the labor force and the frequency of religious service attendance for women under age 50.

Pooled negative binomial results for individuals between ages 50 and 65 in the HRS display no significant correlation between unemployment and the frequency of religious service attendance or the length of unemployment and the frequency of religious service attendance. For labor force status, men between ages 50 and 65 who are out of the labor force at the of interview attend religious services less often relative to men between ages 50 and 65 in the labor force at time of interview. The size of the correlation varies from 11 percent when controlling for overall health to 17 percent when not controlling for health. Women between ages 50 and 65 who are out of the labor force at time of interview, meanwhile, attend religious services 8 percent less often than women between ages 50 and 65 who are not out of the labor force when not controlling for overall health. A negative statistically insignificant relationship is present when controlling for overall health.

These findings suggest that unemployment and religious service attendance are not correlated. For labor force status, men under age 50 appear to attend religious services less frequently when out of the labor force. I ascribe this finding to younger men’s religious service attendance being related to having work or the pursuit of work. Men and women between ages

50 and 65 attend religious services less frequently when out of the labor force, which I attribute to serious health problems in later age forcing labor market exiting and reduced frequency of religious service attendance.

CHAPTER 2

A POSTSECONDARY REVIVAL: THE IMPORTANCE OF RELIGIOSITY FOR

POSTSECONDARY ENROLLMENT GROWTH

2.1 INTRODUCTION

Research consistently finds that private postsecondary schools have grown more slowly than their public counterparts since at least 1890 (Goldin and Katz; 1999). This chapter provides greater insight into this differential growth pattern by allowing private institutions to have heterogeneous enrollment growth. After controlling for a wide array of factors, we document that private schools with a religious affiliation grew substantially faster than private secular institutions between 1991 and 2005. Because such institutions exhibit a substantial difference in the extent of adherence to their religious commitment, binary measures of religious affiliation mask the role of religion. We find that the institutions with the strongest commitments to their religious heritage grew about 250 percentage points faster on average than did their private secular counterparts.

This chapter is also unique in that it is the first and only study to examine whether religiously affiliated higher educational institutions experience differential enrollment growth.

More broadly, there is a paucity of systematic quantitative research about how religious affiliation and intensity affects various outcomes of postsecondary institutions. This is striking given that there is a large literature about the role of private religious primary and secondary schools and that about 40 percent of four-year postsecondary institutions in the U.S. have a

religious affiliation compared with about 30 percent apiece for public and private secular institutions.

This analysis examines determinants of enrollment growth in four year postsecondary institutions from 1991 to 2005. Institution-level data suggest that private religious institutions grew between 11 to 26 percentage points more than private secular institutions. Institutions with

Catholic control or affiliation grew between 8 and 30 percentage points faster while Protestant institutions grew 9 to 29 percentage points faster than private secular institutions. Also, the schools with the most intensive religious commitment, such as those in the Council for Christian

Colleges and Universities (CCCU) grew between 11 and 35 percentage points faster than even the other Protestant institutions.

The chapter is divided into six sections. Section 2.2 outlines the theoretical relationship between religiosity and enrollment. Section 2.3 examines the data while section 2.4 develops the

Econometric model. Results appear in section 2.5. Section 2.6 concludes and highlights the most important findings.

2.2 HYPOTHESES

2.2.1 THEORY

There are two central reasons why enrollment in religious colleges and universities may have experienced relatively fast growth. First, the number of students who go to religious high schools and are more likely to attend a religious college may have grown. These students are more likely to transition to religiously affiliated colleges for at least three reasons. A greater share of their peers may attend such colleges, they may get more information about religious colleges, and having already experienced a religious high school may make them more likely to

choose a religious college. Recent evidence on Catholics supports this hypothesis. For example,

D’Antonio et al. (2007), who use a 2005 Gallup produced random sample of 875 self-identified

Catholics, report that 65 percent of Catholic college or university attendees also attended a

Catholic high school. They also report that 79 percent of Catholic college or university attendees also attended a Catholic elementary school.1 Regnerus et al. (2004), meanwhile, using data from the National Longitudinal Study of Adolescent Health, which is a nationally representative sample of adolescents in grades 7 to 12, establish that the church attendance habits of youths’ schoolmates influence church attendance of youth. In other words, peers influence religious attendance. Hence, peers also influence attendance at religious colleges. We explicitly test this hypothesis as part of our empirical strategy described in section 4 of this paper.

Second, we expect that institutions with stronger religious intensity have higher enrollment growth rates over the period we consider. Measures of religious intensity have much greater explanatory power than measures of religious affiliation because they better gauge the underlying commitment of an individual or organization to religious principles. For instance,

Brooks (2007) finds that people who attend a house of worship every week give more to charity and volunteer more often than those who do not attend, even after controlling for nonreligious differences such as race, education, or gender. Brooks identifies similar differences in charitable giving and volunteering for people who pray every day or simply belong to a congregation or say they “devote a great deal of effort” to their spiritual lives. In each case, the intensity matters more than affiliation in determining charitable giving or the amount they volunteer. As another example, Sander (2005) uses data from that General Social Survey to show that parents with higher levels of religiosity have children with a higher probability of attending primary and

1 To our knowledge, there are no publications reporting a similar or different finding for Protestant, Jewish, or other religious groups.

secondary Catholic schools. Using data from the World Values Survey, Torgler (2005) uses several measures of religiosity such as church attendance, religious education, religious guidance, and others to show that tax morale has a positive relationship with religiosity. Loury

(2004), Park and Smith (2004), Kraft and Neimann (2009), and Gerber et al. (2008), among others, establish the importance of religiosity.

Institutions with a strong commitment to religious principles may have experienced greater enrollment growth over this period because there has been an increase in the number of college-going students who indicate that religion is important to them. The Child Trends Data

Bank (2009) uses survey data on twelfth graders from the Monitoring the Future Survey.2 Figure

2.1 shows that they find that the percentage of religiously intense twelfth graders increased from

27.7 percent to 31.7 percent between 1991 and 2005. Furthermore, Figure 2.2 demonstrates that the percentage of religiously intense twelfth graders planning to complete a four-year college degree increased from 29.4 percent to 33.4 percent during the sample period.3

Another reason why institutions with strong religious commitments may have grown faster is that these institutions or their public and private secular counterparts have changed in ways that make attending a religiously affiliated college relatively more attractive than it was previously. Riley (2004) cites professors with experience at both religious and secular colleges who state that secular institutions have greater limits on academic freedom. For example, Alan

Jacobs, who teaches English at Wheaton College and taught at the University of Virginia as a graduate student states the following:

2 The Monitoring the Future Survey is a nationwide sample of 16,000 students in approximately 133 public and private high schools conducted each year by the University of Michigan Survey Research Center. For more details, see http://monitoringthefuture.org/purpose.html. 3 Because the survey data are only reported at the national level, we cannot use variations over geographic areas in the regressions.

“…the difference between teaching at UVA and teaching at Wheaton ‘is the experience

of academic freedom. It was being able to explore my [Christian] convictions and my

[Christian] concerns in the classroom and have people in the classroom willing to explore

with me” (Riley, 2004, pp. 225).

Other faculty members report that religiously related institutions have a more academically

oriented climate. Professors assert that the average student at a religious institution devotes more

time and effort to academics than at a secular institution. Political scientist Robert Stacey at

Patrick Henry College argues:

“…[Stacey’s] students at Patrick Henry are ‘every bit as good as his best students’ at the

University of Virginia and the University of Richmond, where he taught previously. ‘I

can hold these students to a higher standard than I have ever done in my life.’ They

always come to class prepared, Stacey claims…” (Riley, 2004, pp. 216).

Religious institutions, finally, experience less problems with alcohol, drugs, sexual activity, and violence than their secular counterparts, according to Riley (2004).4

2.2.2 RELIGION MEASURES

To examine how the religious nature of the institution affects enrollment growth, we use

a variety of different measures to characterize a college’s religious commitment. We start by

analyzing a basic measure that accounts for whether a school has a religious affiliation or

control. Sacerdote and Glaeser (2001) contend that we should expect a positive relationship with

enrollment growth. According to the authors, religious attendance increases with education

across individuals, meaning that religiosity should be higher for those attending postsecondary

institutions and that there may be a significant preference for more religious institutions.

4 Unfortunately, we cannot test this explicitly because the National Survey of Student Engagement (NSSE), which collects data on student engagement at college, does not identify university affiliations in its data.

However, this simple measure of religious affiliation may not adequately capture the true extent

to which religiosity influences the institution.

The second set of religion measures uses different religious affiliations and divides

religious higher educational institutions into three types—Catholic, Protestant, and Jewish. But

even this type of religious affiliation may provide little information about the intensity of the

religious commitment. For example, Messiah College and Emory University are both classified

as Protestant institutions with a religious affiliation, but the former makes a much more

concerted effort to integrate faith and academics than the latter institution.

Therefore, our third and final measure considers intensity of the religious commitment by

institutions. We identify religious intensity only with Protestant institutions that are part of the

Council for Christian Colleges and Universities (CCCU).56 We use 120 CCCU member

institutions for Protestant intensity because they require members to assent to more intense

spiritual statements. For example, CCCU institutions must hire only Christians for all full-time

faculty and administrative positions and exhibit a commitment to Christ-centered higher

education.7 These schools report stronger growth between 1990 and 2004 compared to all other

4-year peer institutions.8 The former President of the CCCU in a brief article on CCCU growth from 1990 to 1998 believed there were four reasons for the strong growth of the CCCU: high

5 We attempted to identify Catholic institutional intensity but found no reliable measure. One possibility is the mandatum, which is a commitment certified by the local bishop for theology faculty to teach in accordance with the Catholic Church, to indicate Catholic intensity. However, this measure had problems. One is that there are only a small number of schools that have a mandatum, some of which did not report data to IPEDS, thus leaving us with a very small sample to identify the effect. There is also some secrecy associated with the mandatum. A school can have a mandatum but choose not to reveal it publicly. We explored alternatives such as being listed on NCRegister.com or being included on a list developed by the Newman Society, however, these lists largely overlapped the mandatum schools. We also tried to obtain the fraction of faculty who were Catholic, but did not find consistent data on this variable across institutions. 6 We also attempted to identify Jewish institutional intensity. We were unable to do so since there is no set of consistent information available on intensity for Jewish institutions. 7 See www.cccu.org/members for more details about the membership requirements 8 See the http://www.cccu.org/news/pressroom link for ``Explosive Growth for CCCU Schools”.

academic quality, commitment to whole-person development, a growing market of the evangelical community, and financial assistance.9

2.2.3 ENROLLMENT GROWTH

As part of our investigation, we examine enrollment growth by gender, race, and

ethnicity. We look at enrollment by gender because research such as Miller and Hoffman (1995),

Thompson (1991), and Collett and Lizardo (2009) among others consistently show that we

should expect a difference in religiosity between males and females. Females, on average,

exhibit higher levels of religiosity than males. Debate continues on whether this difference is due

to biological or environmental factors.10 We expect that females’ higher levels of religiosity will lead to more enrollment growth in religiously intense institutions for females than males.

We examine enrollment growth by ethnicity and race to determine the effectiveness of

religious schools’ efforts to diversify their campuses. Historically, the United States has seen

significant levels of segregation in religion. For example, according to statistics from the

Hartford Seminary, 90 percent of churchgoing African Americans belong to black-controlled religious organizations.11 Religious colleges and universities also have histories of segregation.

For example, Baylor University did not admit its first black student until 1965 while Bob Jones

University did not admit its first black student until 1974. With the challenge of segregation in religion to overcome, Riley (2004, pp.158) notes that some religious schools, such as Baylor and

Notre Dame, increase their financial aid to make their schools more affordable for economically disadvantaged minorities. Some religious institutions, such as Calvin College, also have

9 See www.cccu.org/filefolder/Enrollment\_Trends\_Handout\_doc..doc. Extracted on July 20, 2009. 10 See Bradshaw and Ellsion (2009), Cornwall (2009), and Collett Lizardo (2009) for recent developments on this debate. 11 See http://hirr.hartsem.edu/ency/african.htm. Extracted on September 15, 2009.

affirmative action programs in place to recruit minority faculty to make nonwhite students more

comfortable on campus and increase faculty diversity.

Figures 2.3 and 2.4 present the average ratio of black enrollment to total enrollment and

Hispanic enrollment to total enrollment for the years 1991 and 2005, respectively.12 We report

the average share for our entire sample of institutions, religious institutions, and CCCU

institutions.13 The figures show that the average share of enrollment for blacks and Hispanics is

lowest for CCCU institutions in both years. Additionally, the average share of enrollment for

blacks and Hispanics grows for the entire sample, religious institutions, and CCCU institutions

from 1991 to 2005. While the magnitude of change in each share from 1991 to 2005 is similar

for religious institutions and the entire sample, CCCU institutions show gains larger than

religious institutions and the entire sample. The evidence, in essence, suggests that CCCU

institutions start less diverse but increase their diversity to a greater extent.

Given diversity efforts Riley (2004) describes and our preliminary analysis from Figures

2.3 and 2.4, it appears that schools with higher levels of religiosity will show the strongest

growth for minorities. Moreover, we also expect significant enrollment growth in general for

minorities due to widespread efforts at institutions to diversify enrollment.

2.3 DATA AND VARIABLE DEFINITIONS

To examine enrollment growth from 1991 to 2005, we use data from the National Center for Education Statistics' Integrated Postsecondary Education Data System (IPEDS), Common

Core of Data, and Private School Universe Survey. We include only 4-year public and private

12 We compute the average share by taking the average of the ratios of black and Hispanic enrollment to total enrollment for each institution. 13 We discuss our sample in detail in Section 3.

colleges and universities within the fifty states and the District of Columbia.14 We begin with

1991, which is the first year that all of the variables of interest are available from all three data

sources.15 The year 2005 is the other end point since it is the most current year for which all

measures are available.16 We extract our dataset using 1991 and 2005 as collection years in the

IPEDS database. The dataset contains 3,014 4-year post-secondary colleges and universities.

However, only 1,845 report their enrollment data for both years through IPEDS.17 Of these 1,845

colleges and universities, 97 schools do not report tuition and fee data, which leaves a sample of

1,748 observations.18

The dependent variables are the growth between 1991 and 2005 in full-time first-time

(FTFT) degree/certificate seeking students for six groups—total, white, black, Hispanic, male, and female enrollment. FTFT students are generally 18-22 years of age and continue education right after high school. Growth is measured by approximating the percentage change in enrollment.19

Measures of religious affiliation include whether an institution has a religious, Catholic,

or Protestant affiliation or control.20 To account for Protestant religiosity (religious intensity), we

examine members in the CCCU.

Control variables include whether the institution is public, whether it offers Master’s and

Doctoral degrees, and whether it is a Historically Black Colleges and Universities (HBCU). We

include the change in real tuition and fees from 1991 to 2005 in thousands of 2005 dollars, 2005

14 Very few two-year institutions have a religious affiliation. 15 More specifically, the Private School Universe Survey is only available every two years beginning in 1989. Race data are not available, however, in IPEDS until 1990. 16 More specifically, 2005 is the last year the Private Universe Survey is available. 17 Schools reporting any enrollment data measurement of 0 are adjusted to 1 to ensure that all variables are identified when we estimate the regressions in logarithms. 18 Using different years besides 1991 and 2005 makes little difference to the number of excluded observations. To our knowledge, there is no alternative data source that does better than IPEDS for the measures we consider. 19 For more details see Section 4. 20 Control or affiliation is defined by the IPEDS database.

core revenue in hundreds of millions of 2005 dollars, the initial enrollment level to control for

institution size, and own group initial size.21

To determine the degree to which changes in the composition of high school students

affects enrollment growth, we collect data on twelfth grade enrollment through using the

Common Core of Data and the Private School Universe Survey. Raw twelfth grade enrollment

by school is collected for the years 1991 and 2005. We then compute the total twelfth grade

enrollment by state for Public, Private, Religious, Catholic, Protestant, and Association of

Christian Schools International schools (ACSI).22 These state enrollment measures are used to compute the growth in twelfth grade enrollment by state by approximation.23

Table 2.1 provides descriptive statistics of the raw data we use for postsecondary

institutions. The first half of the table displays descriptive statistics for the entire 1,748

observation sample while the second half of the table shows descriptive statistics only for CCCU

institutions. We report enrollment data in hundreds of students while real tuition and fee data are

in thousands of 2005 dollars. The average enrollment at CCCU institutions is lower than average

enrollment for the entire sample for all groups considered. In 1991, CCCU institutions enrolled

365 fewer students than the average institution and 454 fewer students in 2005. Average white

enrollment is 254 less in 2005 and 248 less in 1991 for CCCU institutions. The average black

enrollment is 71 less in 2005 and 55 less in 1991 for CCCU schools. Average Hispanic

enrollment is 45 less in 2005 and 26 less in 1991. Males and females also have lower averages

21 Two other measures that change significantly over time were considered for inclusion: number of faculty and average SAT and ACT scores. Data for SAT and ACT scores are not available before 2002 and the inclusion of the number faculty resulted in too many lost observations and a sample where most CCCU schools were not present. Trying to include SAT and ACT data just for 2005 also resulted in too many lost observations. 22 ACSI enrollment by state is greater than zero in all states but three in 1991 and in all states but one in 2005. We change the 0 enrollments for the three states in 1991 and the one state in 2005 to 1. The next footnote provides further details. The average ACSI enrollment per state in 1991 is 244 students and is 600 students in 2005. 23 Just as with colleges and universities reporting enrollments of 0, states reporting 0 for these twelfth grade enrollments are set equal to 1 due to our use of logarithms to approximate growth rates. For more details see Section 4.

for the CCCU with males being 216 less in 2005 and 182 less in 1991 while females are 238 less

in 2005 and 183 less in 1991. CCCU schools also have higher tuition and fees on average than

the entire sample, with the exception of out of state tuition and fees in 1991. Specifically, CCCU

schools have $2550 more in state tuition and fees in 2005 and $920 more in state tuition and fees

in 1991. Out of state tuition and fees for CCCU schools is $210 higher in 2005 but $560 lower in

1991. In summary, CCCU schools tend to have lower enrollments for all groups and higher

costs of attendance, on average.

Table 2.2 provides descriptive statistics for the variables in this study. Total enrollment at institutions grows by an average of 22 percent. Male enrollment grows by 22 percent while female enrollment grows by 28 percent. The fastest growth is experienced by Hispanics whose enrollment increases by 71 percent compared to 39 percent growth for blacks, and 8 percent for whites.

Among the religion measures, 39 percent of the sample is religiously controlled or

affiliated. The religiously affiliated institutions include Catholic (11 percent of the total),

Protestant (27 percent), and Jewish (1 percent). Approximately 6 percent of all institutions are

CCCU members. Public institutions comprise about 32 percent of the sample and private secular

schools account for about 29 percent of the sample. For the twelfth grade state enrollment

measures, Public growth is an average of 19 percent, Private growth an average of 19 percent,

Religious growth an average of 5 percent, Catholic growth an average of 10 percent, Protestant

growth an average of 39 percent, and American Christian Schools International an average of

100 percent. The highest Public growth is the state of Arizona at 76 percent while the highest

Private growth is 205 percent in Wyoming.

Table 2.3 provides the same descriptive statistics for variables in this study but only considers CCCU institutions. Total enrollment grows by an average of 40 percent while white, black, and Hispanic enrollments grow by an average of 32 percent, 72 percent, and 98 percent, respectively. These growth averages are all higher than the averages for the entire sample. Since

CCCU institutions are private schools where in and out of state status makes no difference for tuition, the tuition and fees change variables are the same for in and out of state changes with an average increase of $6,660.00 in 2005 dollars. This increase is higher than the increases for tuition and fees associated with the entire sample. CCCU institutions offer Masters degrees more frequently than the entire sample with 84 percent of CCCU schools offering Masters degrees.

There is less frequency of Doctoral degrees being offered at CCCU institutions in comparison to the entire sample.

2.4 ECONOMETRIC MODEL

The equations we estimate use ordinary least squares to measure enrollment growth. The following model summarizes the estimation strategy.

Yij(t,t−k ) = α + β Relij + γX ij +ηEij(t−k ) + δ∆ t Z ij(t−k ) +ψTG j(t,t−k ) + ε ij (1)

Yij(t,t−k ) = α + β Relij + γX ij +ηEij(t−k ) + δ∆ t Z ij(t−k ) +φS j + ε ij (2)

Yij(t,t−k ) is a vector of dependent variables for institution i in state j with t being 2005 and t−k being 1991. The six dependent variables are total enrollment growth, enrollment growth for each gender (males and females), enrollment growth for race (white and black) and ethnicity

(Hispanic).24 Since we estimate a growth model controlling for various types of institutions, our

constant α represents average growth for the omitted group, which is private secular institutions.

Relij is a vector of religion measures used in separate regressions. For each dependent

variable one regression simply considers whether the presence of religious control or affiliation

affects growth in enrollment. A second regression divides the affiliation into Catholic, Protestant,

and Jewish institutions. Finally, a third regression considers the affiliation divisions along with

the Protestant and Catholic measures for religiosity (religious intensity).

These three regressions are run for each dependent variable with twelfth grade enrollment

growth effects TG j(t,t−k ) for equation (1), and state fixed effects S j in equation (2). We explore the twelfth grade enrollment effects to determine the relationship between high school enrollment growth and post secondary enrollment growth.25 We also evaluate how much of the

state effects are attributable to twelfth grade growth patterns while testing our first hypothesis

that the composition of secondary schools a student attends affects enrollment in religious higher

education institutions. We test our first hypothesis specifically by comparing the coefficients for

Relij in equation (1) to equation (2) to evaluate whether the inclusion of twelfth grade state

effects matters.

As a vector of fixed institutional characteristics, X ij includes whether the institution is

Public, offers Master’s and Doctoral degrees, and is an HBCU. For the initial level of total enrollment in hundreds, which serves as a measure to control for the initial size of the institution,

24 These dependent variables are computed as the difference between log of the type of enrollment in 2005 and log of the type of enrollment in 1991. 25 The twelfth grade enrollment growth measures are computed by taking the difference between the log of the type of total twelfth grade enrollment in the state in 2005 and the log of the type of total twelfth grade enrollment in the state in 1991.

we include vector Eij(t−k ) . Z ij(t,t−k ) , finally, includes the changing independent variable measures

in state tuition fees change and out of state tuition fees change in thousands of 2005 dollars.

2.5 RESULTS

Tables 2.4-2.9 present the regression results and follow the same organization. In each table, Column 1 shows the results for the simple religious affiliation. Column 2 reports the results for the regressions that divide the affiliation into different groups—Catholic, Protestant, and Jewish. Column 3 shows the results when including the intensity variables with the affiliation division measures to measure the intensity of belief. We only report results for state fixed effects in all of our tables since none of the coefficient estimates differ significantly whether using state fixed effects or twelfth grade enrollment effects. We, therefore, lack evidence to support our hypothesis that the change in the composition of primary and secondary schools a student attends affects enrollment in religious higher education institutions.

Accordingly we dismiss the hypothesis and only report state fixed effects results.

The results for our intensity measure show a positive sign for all six dependent variables

we consider. The positive effect, however, is only significant for total, black, Hispanic, and

female enrollment growth. Overall results show that the religious intensity of an institution

matters.

Table 2.4 indicates that religiously affiliated institutions grow 12 percentage points more

than their private secular counterparts with statistical significance at the 1 percent level. Catholic

institutions grow 8 to 9 percentage points more than private secular institutions at the 10 percent

threshold of statistical significance. Protestant institutions, meanwhile, experience growth 10 to

12 percentage points more than private secular institutions with statistical significance at the 5

percent threshold or lower. Jewish institutions have 51 percentage points more growth than their private secular counterparts with statistical significance at the five percent level. After controlling for other factors, CCCU institutions grow 11 percentage points higher than their

Protestant peers, a result that is statistically significant at the 10 percent level.

Tables 2.5 through 2.7 disaggregate the results by race and ethnicity and report the results for whites, blacks, and Hispanics, respectively. Table 2.5 shows that white enrollment at institutions with a religious affiliation grows 26 percentage points more than the enrollment at private secular institutions. Catholic institutions experience growth of 15 percentage points higher than their private secular counterparts while Protestant institutions have 27 to 29 percentage points more growth than private secular schools. Jewish institutions grow 73 to 74 percentage points higher than their private secular peers. Each of these results is statistically significant at least at the 5 percent level or lower. The CCCU coefficient is statistically insignificant.

The enrollment growth for blacks exhibits some similarities and some differences in

Table 2.6. The Religious and Catholic estimates are statistically significant in all cases at the 1 percent level or lower. Religiously affiliated institutions experience 21 percentage points greater enrollment growth among blacks than do their private secular counterparts. The growth among

Catholic institutions is 29 to 30 percentage points higher than private secular institutions. With statistical significance at the ten percent level or lower, the growth among Protestant colleges is

11 to 19 percentage points more than private secular schools. Jewish schools, meanwhile, do not grow faster than private secular schools for blacks. The CCCU schools, which are characterized by the strongest religious commitments, experience 35 percentage points more growth than their other Protestant counterparts with significance at the 1 percent level.

The results for Hispanics, provided in Table 2.7, also show that the CCCU institutions

grow relatively quickly—19 percentage points faster than their Protestant counterparts with

statistical significance at the 5 percent level. The coefficient estimates for Catholic and Protestant

institutions are also positive, but only the Protestant institutions are statistically significant

showing growth 13 to 18 percentage points higher than their private secular peers. Jewish

schools grow 56 percentage points less than private secular peer institutions with statistical

significance at the 1 percent level.

Table 2.8 examines the change in enrollment for males. Three religion measures are

consistently statistically significant at the 10 percent level or lower—Jewish, Catholic, and

Religious schools. Male enrollment in Catholic institutions grows 15 percentage points faster

than private secular institutions while Religious institutions have 10 percentage points more

growth for males than private secular institutions. Jewish institutions display 43 percentage

points more growth with statistical significance at the 10 percent level. The Protestant measure,

meanwhile, has statistical significance in column 2 with growth of 9 percentage points more than

private secular schools. The CCCU measure is statistically insignificant.

Table 2.9 measures the change in enrollment for females. The only estimate for an effect of religion that is statistically different from zero is for CCCU schools. In other words, only institutions featuring Protestant religious intensity matter for increasing female enrollment growth. Female enrollment in CCCU institutions grows 15 percentage points more than other

Protestant institutions with significance at the 5 percent level or lower.

The magnitudes of the constant in the various tables show growth and shrinking for

enrollment in private secular institutions. Except for total and males, all of the constants are

significant at the 5 percent threshold or lower.26

The Religious variable coefficients have some statistical significance for total, white,

black, Hispanic, and male enrollment growth with magnitudes ranging from 11 to 26 percentage

points. In other words, just having a religious control or affiliation matters. The Protestant

measure is statistically significant for total, male, white, black, and Hispanic enrollment growth

with values of the coefficient showing 9 to 29 percentage points differential growth. The

coefficient estimates for Protestant are typically positive and statistically significant except for a

few examples when they are not different from 0. The Catholic measure, meanwhile, has

statistical significance for its correlation with total, male, white, and black enrollment growth

with size going from 8 to 30 percentage points. All Catholic coefficients are positive. For the

Jewish measure, statistical significance is present for Hispanic, white, and total enrollment

growth with positive magnitudes from 43 to 74 percentage points for total, male, and white

enrollment and a negative magnitude of 56 percentage points for Hispanics. The religious control

or affiliation measures Religious, Protestant, and Catholic are largest for whites and blacks.

The CCCU coefficients are positive in all cases and statistically significant for total,

black, Hispanic and female enrollment growth. The statistically significant coefficients for

CCCU vary from 11 to 35 percentage points. The evidence, in essence, shows that religiosity

26 Hispanics have the largest constants with magnitudes ranging from 47 to 49 percentage point growth for Hispanics. Blacks and females experience similar growth with blacks showing 16 to 21 percentage point growth while females have 17 to 18 percentage point growth. White enrollment in private secular institutions, meanwhile, shrinks 19 to 21 percentage points according to the constants. The constants for the various races and ethnicities suggest that enrollment in public four year postsecondary institutions within the U.S. diversified racially and ethnically to a greater extent from 1991 to 2005. The higher constants that females have in comparison to males also shows that, on average, females have experienced greater enrollment growth than males in private secular institutions from 1991 to 2005.

matters in a significant and positive way for the groups we expect it to matter for—minorities

(blacks and Hispanics), females, and total. Females respond more to religiosity than males and

CCCU schools have higher enrollment growth rates for blacks and Hispanics than their public counterparts through higher levels of growth for minorities, which Figures 2.3 and 2.4 support.

Our hypothesis that religiosity matters is supported by the evidence for four of the six groups we investigate.

For the non-state effect measures in the regression, magnitudes and statistical significance fluctuate, but most of the variables have signs we expect. In State Tuition Fees

Change coefficients show that enrollment growth drops 3 to 7 percentage points for every $1,000 increase in state tuition and fees. Statistical significance at the 10 percent threshold or lower is present for all groups except blacks. The Out of State Tuition Fees Change coefficients display a

4 to 7 percentage points increase for every $1,000 increase in out of state tuition and fees with statistical significance at the 10 percent level or lower for all groups. This could be because the higher out of state tuition and fees is an indicator of a higher quality institution that an average out of state student will pay extra to attend.

Being a public school implies 10 to 28 percentage points more growth than private secular schools. Our public school coefficients are statistically significant for total, white, black, and Hispanic enrollment. Schools designated as Historically Black Colleges and Universities

(HBCU), meanwhile, show between 45 and 15 percentage points less growth than non-HBCU institutions. Statistical significance is present for total, Hispanic, male, and female enrollment.

The degree availability measure Masters reveals that schools offering masters degrees have 11 to

17 percentage points more enrollment growth than schools who do not offer masters degrees with significance at the 5 percent level for all groups. Having doctoral degrees available,

meanwhile, leads to 10 to 16 percentage points more enrollment growth than institutions without

doctoral degrees available. Statistical significance at the 10 percent level or lower is present for all groups except blacks and Hispanics.

The 1991 total enrollment coefficients depict that there is 1 to 3 percent perctange points less growth per 100 additional initial total students for total, white, black, and Hispanic enrollment with 10 percent significance or lower. Results by gender for 1991 total enrollment disclose that there is 2 to 4 percentage points more growth per 100 additional total students with statistical significance at the 1 percent level. Looking at the own group 1991 coefficients, magnitudes indicate between 8 and 11 percentage points less growth for every 100 additional own group students with statistical significance at the 1 percent level for all groups except for whites.

2.6 CONCLUSIONS

The results of this study provide empirical evidence that religion matters for enrollment growth of postsecondary institutions. Evidence shows that just being a Religious, Catholic, or

Protestant controlled or affiliated school is generally correlated positively with growth. There are differences, as the evidence suggests, in how enrollment by gender, race, and ethnicity correlate with religion and other control variables. Enrollment growth for women, for example, has no correlation with Protestant, Catholic, Jewish, and Religious controlled or affiliated institutions. In general, growth patterns for race (white and black) are more strongly correlated with the religion variables than ethnicity or gender. The estimates also stress the importance of

institutional religiosity. For minority, female, and total enrollment, intensity is associated with

higher levels of enrollment growth. Diversity expands to a greater extent for our religiously

intense schools while female enrollment only responds to religious intensity and not mere affiliation. Our findings coincide with previous research that stresses the importance of religiosity as a determinant and higher levels of religiosity for females.

While this study offers evidence supporting a hypothesis of stronger growth for religious postsecondary institutions, it is unable to explore some possible reasons for stronger growth. The stronger growth in religious institutions and religiously intense institutions could be due to resurgence in religious participation among young adults. It may also be that religious colleges and universities are systematically different in some other aspect besides associated religious beliefs that this study fails to measure. Should data become available, future work should explore these alternatives.

Figure 2.1: Percent of 12th Graders who Report that Religion Plays a “Very Important” Role in Their Lives

Figure 2.2: Percent of 12th Graders who Expect to Complete a College Degree and Say that Religion Plays a “Very Important” Role in Their Lives

Figure 2.3: Average Enrollment Share for Blacks

Figure 2.4: Average Enrollment Share for Hispanics

Table 2.1: Descriptive Statistics-Raw Data Raw Measure Obs. Mean Std. Dev. Min Max Entire Sample Enrollment 2005 1748 7.96 12.90 0.01 306.65 Enrollment 1991 1748 5.97 7.88 0.02 62.60 White 2005 1748 5.33 8.65 0.01 161.56 White 1991 1748 4.51 6.44 0.01 52.66 Black 2005 1748 0.91 2.19 0.01 53.26 Black 1991 1748 0.66 1.56 0.01 18.70 Hispanic 2005 1748 0.60 1.96 0.01 42.32 Hispanic 1991 1748 0.33 1.09 0.01 24.79 Male 2005 1748 3.58 5.38 0.01 78.95 Male 1991 1748 2.83 3.93 0.01 32.44 Female 2005 1748 4.39 7.92 0.01 227.70 Female 1991 1748 3.14 4.09 0.01 34.33 Real In State Tuition and Fees 2005 1748 13.85 8.43 0 62.55 Real In State Tuition and Fees 1991 1748 8.83 5.84 0 27.82 Real Out of State Tuition and Fees 2005 1748 16.19 6.82 0 62.55 Real Out of State Tuition and Fees 1991 1748 10.31 4.85 0 27.82 CCCU Sample Enrollment 2005 104 3.42 2.23 0.47 14.16 Enrollment 1991 104 2.32 1.44 0.24 8.53 White 2005 104 2.79 1.92 0.32 11.86 White 1991 104 2.03 1.32 0.16 7.47 Black 2005 104 0.20 0.21 0.01 1.38 Black 1991 104 0.11 0.15 0.01 1.13 Hispanic 2005 104 0.15 0.18 0.01 0.81 Hispanic 1991 104 0.07 0.09 0.01 0.44 Male 2005 104 1.42 0.90 0.01 5.06 Male 1991 104 1.01 0.64 0.01 3.67 Female 2005 104 2.01 1.39 0.31 9.10 Female 1991 104 1.31 0.86 0.16 4.86 Real In State Tuition and Fees 2005 104 16.40 3.24 9.25 22.92 Real In State Tuition and Fees 1991 104 9.75 2.69 0.14 16.35 Real Out of State Tuition and Fees 2005 104 16.40 3.24 9.25 22.92 Real Out of State Tuition and Fees 1991 104 9.75 2.69 0.14 16.35

Table 2.2: Descriptive Statistics-Whole Sample Variable Obs. Mean Std. Dev. Min Max College Enrollment Growth by Students and Group Total 1748 0.22 0.58 -3.91 5.92 White 1748 0.08 0.70 -4.84 5.93 Black 1748 0.39 0.91 -3.74 6.97 Hispanic 1748 0.71 0.89 -3.89 6.15 Male 1748 0.22 0.76 -5.11 5.41 Female 1748 0.28 0.64 -3.43 6.48 Religion Variables Religious 1748 0.39 0.49 0 1 Catholic 1748 0.11 0.31 0 1 Protestant 1748 0.27 0.44 0 1 Jewish 1748 0.01 0.08 0 1 CCCU 1748 0.06 0.24 0 1 Control Variables Public 1748 0.32 0.47 0 1 HBCU 1748 0.04 0.20 0 1 In State Tuition Fees Change 1748 5.02 3.51 -7.97 48.35 Out of State Tuition Fees Change 1748 5.88 3.25 -7.97 48.35 Masters 1748 0.70 0.46 0 1 Doctoral 1748 0.27 0.44 0 1 High School Enrollment Growth by State and Type Public 1748 0.19 0.16 -0.26 0.76 Private 1748 0.19 0.22 -0.23 2.05 Religious 1748 0.05 0.20 -1.39 0.59 Catholic 1748 0.10 0.28 -1.16 0.93 Protestant 1748 0.39 0.23 -0.32 1.30 ACSI 1748 1.00 0.73 -0.45 5.40

Table 2.3: Descriptive Statistics-CCCU Variable Obs. Mean Std. Dev. Min Max College Enrollment Growth by Students and Group Total 104 0.40 0.41 -0.47 2.13 White 104 0.32 0.41 -0.64 2.24 Black 104 0.72 0.94 -2.20 3.22 Hispanic 104 0.98 0.83 -1.10 3.43 Male 104 0.36 0.46 -0.51 2.45 Female 104 0.44 0.42 -0.56 1.92 Control Variables In State Tuition Fees Change 104 6.66 1.98 1.89 15.04 Out of State Tuition Fees Change 104 6.66 1.98 1.89 15.04 Masters 104 0.84 0.37 0 1 Doctoral 104 0.19 0.40 0 1

Table 2.4: Total Enrollment Estimation Results (1) (2) (3) Religion Variables Religious 0.12*** (0.04) Catholic 0.08* 0.09* (0.05) (0.05) Protestant 0.12*** 0.10** (0.04) (0.04) Jewish 0.51*** 0.51*** (0.17) (0.17) CCCU 0.11* (0.06) Control Variables Public 0.10* 0.10* 0.10* (0.05) (0.05) (0.05) HBCU -0.16** -0.16** -0.15** (0.07) (0.07) (0.07) In State Tuition Fees Change -0.04*** -0.04*** -0.04*** (0.01) (0.01) (0.01) Out of State Tuition Fees Change 0.05*** 0.05*** 0.05*** (0.01) (0.01) (0.01) Enrollment 1991 -0.02*** -0.01*** -0.01*** (0.00) (0.00) (0.00) Masters 0.18*** 0.19*** 0.18*** (0.03) (0.03) (0.03) Doctoral 0.12*** 0.11*** 0.11*** (0.04) (0.04) (0.04) Constant 0.01 -0.01 0.00 (0.04) (0.04) (0.04) Observations 1748 1748 1748 R-squared 0.13 0.13 0.13 F-Statistic 15.50 13.08 12.18 Standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1%

Table 2.5: White Enrollment Estimation Results (1) (2) (3) Religion Variables Religious 0.26*** (0.04) Catholic 0.15** 0.15** (0.06) (0.06) Protestant 0.29*** 0.27*** (0.05) (0.05) Jewish 0.74*** 0.73*** (0.20) (0.20) CCCU 0.09 (0.08) Control Variables Public 0.17*** 0.17*** 0.17*** (0.06) (0.06) (0.06) HBCU 0.01 0.01 0.01 (0.09) (0.09) (0.09) In State Tuition Fees Change -0.04** -0.03** -0.03** (0.01) (0.01) (0.01) Out of State Tuition Fees Change 0.04*** 0.04*** 0.04*** (0.01) (0.01) (0.01) Enrollment 1991 -0.02** -0.02** -0.02** (0.01) (0.01) (0.01) White 1991 0.01 0.01 0.01 (0.01) (0.01) (0.01) Masters 0.15*** 0.17*** 0.17*** (0.04) (0.04) (0.04) Doctoral 0.16*** 0.15*** 0.15*** (0.05) (0.05) (0.05) Constant -0.19*** -0.21*** -0.20*** (0.05) (0.05) (0.05) Observations 1748 1748 1748 R-squared 0.09 0.09 0.09 F-Statistic 11.53 10.28 9.52 Standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1%

Table 2.6: Black Enrollment Estimation Results (1) (2) (3) Religion Variables Religious 0.21*** (0.06) Catholic 0.29*** 0.30*** (0.08) (0.08) Protestant 0.19*** 0.11* (0.06) (0.07) Jewish -0.19 -0.19 (0.27) (0.27) CCCU 0.35*** (0.10) Control Variables Public 0.28*** 0.27*** 0.27*** (0.08) (0.08) (0.08) HBCU 0.00 0.01 0.04 (0.15) (0.15) (0.15) In State Tuition Fees Change -0.03 -0.03* -0.04* (0.02) (0.02) (0.02) Out of State Tuition Fees Change 0.04** 0.04** 0.04** (0.02) (0.02) (0.02) Enrollment 1991 -0.01** -0.01** -0.01* (0.00) (0.00) (0.00) Black 1991 -0.10*** -0.10*** -0.10*** (0.02) (0.02) (0.02) Masters 0.15*** 0.14*** 0.11** (0.05) (0.05) (0.05) Doctoral 0.01 0.02 0.02 (0.06) (0.06) (0.06) Constant 0.17** 0.19*** 0.22*** (0.07) (0.07) (0.07) Observations 1748 1748 1748 R-squared 0.09 0.10 0.10 F-Statistic 9.55 8.21 8.54 Standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1%

Table 2.7: Hispanic Enrollment Estimation Results (1) (2) (3) Religion Variables Religious 0.12** (0.05) Catholic 0.11 0.11 (0.08) (0.08) Protestant 0.18*** 0.13** (0.06) (0.06) Jewish -0.56** -0.56** (0.26) (0.26) CCCU 0.19** (0.10) Control Variables Public 0.24*** 0.25*** 0.24*** (0.08) (0.08) (0.08) HBCU -0.44*** -0.45*** -0.44*** (0.11) (0.11) (0.11) In State Tuition Fees Change -0.06*** -0.07*** -0.07*** (0.02) (0.02) (0.02) Out of State Tuition Fees Change 0.07*** 0.07*** 0.07*** (0.02) (0.02) (0.02) Enrollment 1991 -0.01** -0.01** -0.01** (0.00) (0.00) (0.00) Hispanic 1991 -0.11*** -0.11*** -0.11*** (0.02) (0.02) (0.02) Masters 0.16*** 0.16*** 0.14*** (0.05) (0.05) (0.05) Doctoral 0.05 0.08 0.08 (0.06) (0.06) (0.06) Constant 0.47*** 0.48*** 0.49*** (0.06) (0.06) (0.07) Observations 1748 1748 1748 R-squared 0.11 0.12 0.12 F-Statistic 12.60 11.24 10.65 Standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1%

Table 2.8: Male Enrollment Estimation Results (1) (2) (3) Religion Variables Religious 0.11** (0.05) Catholic 0.14** 0.14** (0.07) (0.07) Protestant 0.09* 0.08 (0.05) (0.06) Jewish 0.43* 0.43* (0.22) (0.22) CCCU 0.07 (0.08) Control Variables Public 0.04 0.04 0.04 (0.07) (0.07) (0.07) HBCU -0.20** -0.19** -0.19** (0.09) (0.09) (0.09) In State Tuition Fees Change -0.05*** -0.05*** -0.05*** (0.02) (0.02) (0.02) Out of State Tuition Fees Change 0.06*** 0.06*** 0.06*** (0.02) (0.02) (0.02) Enrollment 1991 0.04*** 0.04*** 0.04*** (0.01) (0.01) (0.01) Male 1991 -0.11*** -0.11*** -0.11*** (0.02) (0.02) (0.02) Masters 0.16*** 0.16*** 0.15*** (0.04) (0.04) (0.04) Doctoral 0.12** 0.10** 0.10** (0.05) (0.05) (0.05) Constant 0.05 0.04 0.05 (0.06) (0.06) (0.06) Observations 1748 1748 1748 R-squared 0.09 0.09 0.10 F-Statistic 10.88 9.24 8.53 Standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1%

Table 2.9: Female Enrollment Estimation Results (1) (2) (3) Religion Variables Religious 0.02 (0.04) Catholic 0.02 0.02 (0.05) (0.05) Protestant 0.02 -0.01 (0.04) (0.05) Jewish 0.04 0.04 (0.18) (0.18) CCCU 0.14* (0.07) Control Variables Public 0.02 0.02 0.01 (0.06) (0.06) (0.06) HBCU -0.21*** -0.21*** -0.20** (0.08) (0.08) (0.08) In State Tuition Fees Change -0.05*** -0.05*** -0.05*** (0.01) (0.01) (0.01) Out of State Tuition Fees Change 0.06*** 0.06*** 0.06*** (0.01) (0.01) (0.01) Enrollment 1991 0.02** 0.02** 0.02** (0.01) (0.01) (0.01) Female 1991 -0.08*** -0.08*** -0.08*** (0.02) (0.02) (0.02) Masters 0.16*** 0.16*** 0.15*** (0.04) (0.04) (0.04) Doctoral 0.10** 0.10** 0.10** (0.04) (0.04) (0.04) Constant 0.17*** 0.17*** 0.17*** (0.05) (0.05) (0.05) Observations 1748 1748 1748 R-squared 0.13 0.13 0.13 F-Statistic 11.18 9.13 8.70 Standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1%

CHAPTER 3

LABOR INCOME AND RELIGIOSITY: EVIDENCE FROM SURVEY DATA

3.1 INTRODUCTION

Research on the relationship between income and religious determinants has been ongoing for several years. Modern findings begin with Weber (2001 [1905]), who argues that income and Protestant adherence are positively related. Azzi and Ehrenberg (1975) explore a backward bending relationship between income and frequency of religious attendance, where attendance initially increases then declines with income, in their work. Lipford and Tollison

(2003), meanwhile, control for simultaneity between income and frequency of religious service attendance and show a negative relationship between the two using state level data. I present new empirical evidence in this paper by examining endogeneity between labor income and religiosity with the measures of prayer frequency and attendance frequency. I also argue that there is a negative relationship between income and religiosity that represents a tradeoff between religious and secular activity.

This study is helpful for four reasons. First, this study employs a new instrumental variables approach in addition to ordinary least squares in an attempt to identify the best possible estimate of the relationship between religiosity and income. Second, it is the first attempt to control for endogeneity when examining the relationship between frequency of prayer and income. I contend that previous estimates of the correlation between frequency of prayer and income and between frequency of attendance and income are affected by bias by not accounting for noncognitive skills. This bias leads to imprecise estimates of the magnitude of the

relationship between religiosity and income. Third, this study reviews the relatively understudied measure of prayer frequency. I examine prayer to better understand the relationship between income and religiosity and because I consider it a better measure of religiosity. I argue that prayer is a more private religious activity that is less tainted by ulterior motives. Fourth, the empirical strategy is novel in that it is the first to explore panel data to examine the effect of income on religiosity. All other studies rely solely on cross-section data.1 Exploring variation of religiosity within individuals rather than between individuals may improve estimates of the income and religiosity relationship by eliminating individual heterogeneity. Furthermore, panel estimation provides information on variation of religiosity within individuals which adds additional insight into individuals’ religious behavior over time and consequently the income and religiosity relationship.

I estimate the relationship between income and religiosity for men and women separately because I expect findings by gender to differ as research on labor market outcomes and religious behavior confirms.2 The empirical results suggest that labor income is a poor predictor of both frequency of religious service attendance and prayer frequency for men when accounting for endogeneity. Ordinary least squares results indicate that a $25,000 increase in labor income reduces frequency of prayer by once per week. This magnitude is 3.63 times greater than the estimates reported by Brown (2009).3 Using instrumental variables leads to statistical insignificance for the coefficient on income. Robustness checks using wages reveal similar findings. Weak identification is a problem for men when using wages as a correlate with first stage F-test statistics on the instrumented variables below 10.

1 Panel data on income, prayer frequency, and attendance frequency from the General Social Survey became available in April 2010. 2 See Section 3.1 for a further discussion with specific references. 3 This study closely mirrors Brown (2009). Brown (2009) estimates the impact of wages on frequency of prayer. Assuming a 2000 hour work year, he reports magnitudes for the effect of labor income on frequency of prayer.

Results for women show that labor income is negatively correlated with frequency of

religious service attendance and prayer frequency regardless of estimation technique. The

magnitude of the effect of labor income on religiosity increases when employing an instrumented

variable. While a $100,000 increase in labor income reduces frequency of religious service

attendance by once per month and a $20,000 increase in labor income reduces prayer frequency

by once per week using ordinary least squares, a $9,090.91 increase in labor income decreases

frequency of religious service attendance by once per month and a $2,083.33 increase in income

decreases prayer frequency by once per week using instrumental variables. The ordinary least

squares magnitude for frequency of prayer is greater in size by a factor of 4.36 in comparison to

estimates reported by Brown (2009). Similar results are obtained when conducting a robustness

check using wages. A weak identification problem is indicated for women with all first stage F-

test statistics on the instruments below 10.

This chapter is divided into 5 sections. Section 3.2 discusses the measures I use, the

theoretical relationship between income and religiosity, and the endogeneity problem. It also

reviews the data. Section 3.3 develops the econometric model, Section 3.4 presents the empirical

results, and Section 3.5 concludes.

3.2 DATA AND THEORY

The data are from the General Social Survey (GSS), a nationally representative sample of

American adults conducted by the National Opinion Research Center at the University of

Chicago. Two separate datasets are constructed—a pooled cross section dataset with data

generated every two years for 14,130 respondents from 1996 to 2004 and a panel dataset of

1,536 respondents from 2006 to 2008.4 After restricting to individuals who are religious and

eliminating individuals with missing or unreported data of interest, the final datasets consists of

4,240 respondents in the pooled cross section and 628 respondents in the panel dataset.5

Weighted descriptive statistics of the final datasets are similar to the full range of data available

for both weighted datasets before restrictions.6 Therefore, I conclude that the final weighted

datasets are nationally representative. The pooled cross section dataset is weighted to account for

the number of persons over 18 in the household for all years in addition to adjusting for non-

response in 2004. The panel dataset is weighted to account for the panel sample design, number

of persons over 18 in the household, and adjusts for non-response. I use these weights to make

the estimates as nationally representative as possible in addition to correcting for non-response.

3.2.1 DEPENDENT VARIABLES

The dependent variables in this study are frequency of prayer and frequency of religious service attendance. Frequency of prayer is originally a categorical variable broken down into six categories in this study. I convert frequency of prayer into a numerical variable following Brown

(2009). The transformation leaves frequency of prayer measured in number of prayers per week.7

According to Gill (2005), prayer is defined as human communication with divine and

spiritual entities. I closely examine the determinants of prayer frequency in this study for two

4 Approximately 1,500-4,500 individuals are surveyed each time the GSS is administered for the pooled cross section. I do not use cross section data before 1996 or after 2004 to follow Brown (2009) as closely as possible to obtain his baseline ordinary least squares results for regressions using wages as a determinant. Similar results to Brown’s are obtained when using all available GSS data, which is limited to 1994-2008. Data on weeks worked in the previous year is not available until 1994 and 2008 data are the most recently available data when this study started. 5 See Appendix A.1 for detailed information on lost observations due to missing or non-response data. I restrict to religious individuals only to follow Brown (2009). 6 Tables 3.14 and 3.15 contain weighted descriptive statistics for the unrestricted pooled cross section and panel datasets. Comparison of these tables to Tables 3.1 and 3.2 reveals relatively similar descriptive statistics. Therefore, I conclude that my estimates are approximately nationally representative. 7 For the numerical conversion I set “never” to 0, “less than once a week” to 0, “once a week” to 1, “several times a week” to 3, “once a day” to 7, and “several times a day” to 21. Altering this conversion slightly where appropriate does not significantly alter the results I obtain.

reasons. First, frequency of prayer has received relatively little attention in previous work exploring the relationship between religion and income. Only Brown (2009), Iannaccone (1990), and Branas-Garza and Neuman (2004) explore the relationship between frequency of prayer and income.

Second, I contend that frequency of prayer is a superior measure of religiosity. To describe why frequency of prayer may be a better measure of religiosity, I develop a formal theoretical approach that adds to Brown (2009). Let r,a,and p represent religiosity, frequency of

8 religious service attendance, and frequency of prayer, respectively. Let θ a andθ p , meanwhile, stand for noise that falsely portrays religiosity. Observed religiosity in the form of frequency of attendance and frequency of prayer is defined as follows:

2 a = r +θ a , θ a > 0 and θ a ~ N(0,σ a ) 2 p = r +θ p , θ p > 0 and θ p ~ N(0,σ p )

If θ a > θ p , then frequency of prayer more accurately proxies the underlying religiosity than does

attendance. It is likely that θ a > θ p because individuals who attend religious services do so for social benefits while giving little credence to a religious tradition.9 Individuals who attend for social reasons tend to pray less, showing less commitment to religious principles than their frequency of attendance implies. Therefore, frequency of prayer is a more private religious activity that is less tainted by ulterior motives such as the private benefit from social interactions.

Similar to frequency of prayer, frequency of attendance is a categorical variable with nine categories in this study. I convert these categories to numerical measurements so that frequency

8 I assume that r,a,and p are all greater than or equal to zero. 9 Sacerdote and Glaeser (2001), for instance, suggest that churches act as civic organizations. More frequent church attendance may increase social capital through networking, thus leading to higher incomes.

of attendance is measured in number of services attended per month.10 Prior studies affirm

frequency of religious service attendance as a measure for religiosity. Azzi and Ehrenberg

(1977), for example, examine religious service attendance in relation to a household’s allocation

of time. Lipford and Tollison (2003) analyze simultaneity between religiosity and income

through considering religious service attendance and income. Gruber (2004), finally, reviews the

impact of subsidized charitable giving on religious service attendance.11 Noting the importance

of religious service attendance as an indicator of religiosity, I also utilize it within this paper.

3.2.2 INCOME

The primary explanatory variable of this study, labor income, is measured as a

respondent’s annual real labor income in thousands of year 2004 dollars.12 As an alternative to

this measure for robustness checks, I construct wages in tens of year 2004 dollars following

Brown (2009) by using data on labor income, number of hours worked last week, and number of

weeks worked in the past year.13 I also include annual non-labor income, which is measured in

thousands of year 2004 dollars. 14

As a market variable that influences market and non-market behavior, I focus on labor

income, the amount earned annually from working, as a determinant of religiosity within this

study. Research on religiosity using income or a similar determinant such as wages, the amount

earned per hour from working, has yielded mixed results. For example, Iannaccone (1990) uses

10 For the conversion, I set “never” equal to 0, “less than once a year” to 0, “once a year” to (1/12), “several times a year” to (1/2), “once a month” to 1, “2-3 times a month” to (5/2), “nearly every week” to 4, “every week” to (13/3), and “more than once a week” to (13/2). Altering this conversion slightly where appropriate does not significantly alter the results I obtain. 11 Other studies looking at frequency of attendance include Smith and Sawkins (2003), Gruber and Hungerman (2008), and Sullivan (1985) among others. 12 See Appendix A.2 for detailed information on the labor income measure in this study. 13 I compute wages by taking labor income and dividing it by the product of number of hours worked last week and number of weeks worked last year. Wages are set to 0 when labor income is 0 or weeks worked in the previous year is 0 or number of hours worked last week is 0. 14 See Appendix A.2 for detailed information on the non-labor income measure. I compute non-labor income by subtracting the respondent’s labor income from family income. It stems primarily from two sources—spousal income and income from assets.

General Social Survey (GSS) data from 1983 to 1987 and finds no correlation between family

income and frequency of prayer. Branas-Garza and Neumann (2004) also find no correlation

using 1998 data on Catholics in Spain from the Center for Sociological Research. As a general

finding, Iannaconne (1998) contends that while income strongly predicts religious contributions,

it is a poor predictor of other measures of religious activity such as church attendance, church

membership, and rates of religious belief. Estimates he presents using 1990 General Social

Survey data on family income substantiate his claims that income is a weak correlate of religious

attendance. Brown (2009), meanwhile, with GSS data from 1996 to 2004, finds a statistically

significant negative correlation between wages and frequency of prayer. Following Brown

(2009), I also utilize wages as part of my empirical strategy with GSS data from 1996 to 2004.15

While findings are mixed, I contend that there is a negative relationship between labor

income and religiosity that represents a tradeoff between religious and secular activity.

Iannaccone (1988) and Azzi and Ehrenberg (1977) support this argument by developing utility

maximization models where there is a tradeoff between time spent in religious activity and time

spent in secular activity. In other words, time spent at work, a secular activity, must be traded off

against time spent in religious activities such as prayer or religious service attendance. Therefore,

a negative relationship between labor income and religiosity exists because labor income depends upon time spent at work.

3.2.3 CONTROL VARIABLES

For the religious upbringing control variables, I use the religion the respondent was raised in at age 16 and how fundamentalist the religion was in which the respondent was raised at age

16. Other control variables include employment status, age, geographic mobility, sex, marital

15 For more details see Section 3.3.

status, race, ethnicity, educational attainment, and region where the respondent was

interviewed.16

3.2.4 ENDOGENEITY

Research on the relationship between income and religious determinants does not consider potential problems stemming from unobservable characteristics such as noncognitive skills. A growing body of research such as Heckman, Stixrud, and Urzua (2006) and Fortin

(2008) explores noncognitive skills, which can be defined as skills stemming from motivation, personality traits, and persistence, and estimate their effects on market outcomes.17 These

noncognitive skills are typically measured using questionnaires such as the Rotter (1966) Locus

of Control Scale and Rosenberg (1965) Self-Esteem Scale. The Rotter scale indicates the degree

of control individuals feel they possess over their life, also known as locus of control. It is used

in studies such as Heckman, Stixrud, and Urzua (2006) and Groves (2005) to evaluate the impact of noncognitive skills on labor outcomes. Locus of control measures individuals on a continuum from external to internal. Individuals with a higher internal locus of control tend to believe that control of their life comes from within themselves while individuals with a higher external locus of control tend to believe that control of their life comes from outside themselves.

Psychological studies such as Fiori, Hays, and Meador (2004), Spilka, Shaver, and

Kirkpatrick (1985), and McIntosh, Silver, and Wortman (1993) note that religion offers extrinsic

control, such as prayer and attendance, and intrinsic control, such as belief that everything will

16 Data are unavailable for the state and the local place where respondents were interviewed to protect the identity of GSS respondents. This data may be obtained by special permission, however, the process for approval requires a lead researcher holding a Ph.D. and a $750 fee. 17 There is no generally accepted definition for the term noncognitive skills. The definition I mention here comes from Heckman, Stixrud, and Urzua (2006).

turn out well if an individual trusts in a spiritual or divine deity. In this study I only consider

religion measures related to extrinsic control.18

Within this paper, I argue that, by not accounting for noncognitive skills, endogeneity is

present in the religiosity and income relationship. This omitted variable leads to downward

biased estimates.19 The following three propositions summarize the argument to be investigated in this paper:

(1) On average, higher levels of internal perceived control are associated with higher wages

that increase labor income. Heckman, Stixrud, and Urzua (2006), for instance, find that as

individuals obtain a higher internal locus of control, their wages are higher through direct

effects on productivity and indirect effects on schooling and work experience.

(2) On average, as individuals increase outward forms of religious behavior such as prayer or

attending religious services, it indicates a higher external locus of control. Kahoe (1974)

finds that extrinsic religious orientation correlates negatively with an internal locus of

control. In essence, internal locus of control will tend to be associated with lower levels

of extrinsic religious participation such as religious service attendance or prayer.

(3) Propositions (1) and (2) present an endogeneity problem when estimating the correlation

between labor income and religious activity. Specifically, a statistically significant

negative relationship between labor income and religious activity can be obtained using

ordinary least squares due to bias from the omission of noncognitive skills as measured

18 Data on religion measures related to intrinsic control such as belief that everything will turn out well are rarely collected in national surveys and are not easily quantified. Religion measures related to extrinsic control such as financial contributions, frequency of prayer, and frequency of attendance are more easily measured and frequently included in national surveys containing religion data. 19 I do not quantify the bias because of weak instrument problems as detailed in Section 5. In cases where there is no evidence of weak instruments, coefficients estimated using ordinary least squares shift from statistical significant to being statistical insignificance when using two stage least squares. In these cases it is not possible to quantify the downward bias given the shift from statistically significant to statistically insignificant.

by locus of control. Through the use of instrumental variables, I will test to see if this

argument is valid.

The instrumental variable in the study is highest year of education completed by the

respondent’s spouse. This measure is set equal to 0 when the respondent has no spouse. I discuss

the requirements of a good instrumental variable and the choice of spouse education as the

instrumental variable in Section 3.3.2

3.2.5 DESCRIPTIVE STATISTICS

Table 3.1 presents weighted descriptive statistics for the cross section sample while Table

3.2 presents weighted descriptive statistics for the panel sample. Both tables are divided among

four types of variables—dependent, primary variables of interest (labor income, wages, non- labor income, and the instrumental variable spouse education), religious upbringing, and control.

Since most of the variables exhibit similar descriptive statistics in Tables 3.1 and 3.2, I will discuss only Table 3.1 in its entirety and measures in Table 3.2 that differ significantly from

Table 3.1.

Table 3.1 reveals that average frequency of prayer is 8.86 times per week. The average

respondent attends religious services 2.10 times per month. Average labor income per year is

$27,970, average wage is $15.20, and average non-labor income per year is $26,740.

The religious upbringing variables in Table 3.1 show that 97 percent of the sample is

religious at age 16, with 33 percent Catholic, 61 percent Protestant, 2 percent Jewish, and 1

percent Other. It also shows that the sample is 34 percent Fundamentalist, 44 percent is

Moderate, and 22 percent is Liberal. The control variables reveal that 73 percent is employed, 27

percent is not employed, average age is 46.63 years, and 66 percent are living in the state in

which they lived at age 16. The gender breakdown is 55 percent women and 45 percent men.

Marital status decomposes into 50 percent married, 9 percent widowed, and 20 percent divorced or separated. The racial composition is 14 percent black and 16 percent other race or ethnicity.

For educational attainment, 13 percent of the sample is less than high school, 55 percent is high school graduates, 7 percent associates degree holders, 16 percent bachelors degree holders, and 8 percent graduate degree holders.

Only the measures of religiosity and labor markets for this study have descriptive statistics in Table 3.2 that differ significantly from Table 3.1.20 Averages for frequency of prayer, frequency of attendance, labor income, wages, and non-labor income are higher in Table 3.2 than in Table 3.1. Average weekly frequency of prayer is 10.21 times, average monthly frequency of religious service attendance is 2.37 times, average annual labor income is $28,980 and average wage is $18.00. The average annual non-labor income, finally, is $32,570.

3.3 ECONOMETRIC MODEL

3.3.1 ESTIMATING EQUATIONS

The equation I estimate with the pooled cross section employs ordinary least squares and two stage least squares with weighting. The equations below describe the model specification.

Prayit = α + βYit +ωRel16it +δ Xit + ρdt +φsik + ε it (1) Attend = θ +γY +κRel16 + µX +υd +οs +η (2) it it it it t ik it

The dependent variables frequency of prayer and frequency of attendance of person in year

are represented by Prayit and Attendit .Endogenous labor income and its square are denoted

by Yit . The vector Rel16it , meanwhile, contains six dummies, Catholic, Protestant, Jewish, Other,

Fundamentalist, and Moderate for religion raised in at age 16.

20 Spouse education is omitted from Table 2 because it is only used as an instrumental variable.

The vector X it contains the control variables, which include employment status, age, geographic mobility, marital status, race, ethnicity, and educational attainment. Time dummies

and region of interview dummies correspond to dt and sik , respectively. The error term is

represented by ε it in the estimating equation for prayer and ηit in the estimating equation for attendance.

The equations I estimate for the panel sample include individual fixed effects estimation with weighting but are otherwise very similar to equations (1) and (2). The estimating equations are:

Prayit = ς +ψYit + ςRel16it + σ Xit +ιdt +τsik + ci + ϕit (3) Attend = χ +ιY + λRel16 +νX +πd +ϖs + a + ζ (4) it it it it t ik i it

The time invariant, individual-specific variable ci represents unobserved individual fixed effects

such as family background, religious background, and non-cognitive skills.

The two estimation methods I use provide different information on the relationship

between religiosity and income. The pooled cross section allows for estimation of variation in

the dependent variable between individuals. As Wooldridge (2000, p.409) describes, using a

pooled cross section is relatively useful provided that the relationship between the dependent

variable and some of the independent variables remains constant over time. When I use pooled

cross section estimation, I am evaluating the relationship between religiosity and income

between individuals and assuming that the relationship remains fairly constant over time. The

panel dataset, meanwhile, is used for fixed effects estimation, also known as the within

estimator. The estimation uses the time variation in the dependent variable and independent

variables within each cross-sectional observation, according to Wooldridge (2000, p.442). In

essence, it provides information on the variation of religiosity within individuals while

accounting for the possible variation in the relationship between religiosity and income over

time. To summarize, the pooled cross section estimation explores variation in religiosity between

individuals that is treated as relatively constant over time and panel fixed effects estimation

examines variation in religiosity within individuals while controlling for time.21

I run estimates separately for men and women as in Brown (2009) and Branas-Garza and

Neumann (2004). I perform the estimates separately because of expected differences in religious participation and labor market outcomes for men and women. It is well established in studies such as Miller and Hoffman (1995), Thompson (1991), and Collett and Lizardo (2009) that men

have lower levels of religiosity than women. The precise reasons for this trend are debated and

usually ascribed to physiological, environmental, or behavioral differences as discussed in Stark

(2002), Bradshaw and Ellison (2009), and Cornwall (2009) among others. Research on labor

market outcomes also notes significant differences by gender. Altonji and Blank (1999), for

instance, use the Current Population Survey to show that average wages and labor force

participation rates for men are higher than women over time after controlling for various factors

such as education, experience, occupation, industry, and job characteristics. Pencavel (1986) and

Killingsworth and Heckman (1986) conduct thorough surveys of the male labor supply and

female labor supply, respectively, in the United States, Canada, Great Britain, and Germany to

reveal significant differences in labor force participation rates, hours worked, and occupational

distribution. Reasons for the differences by gender in labor market outcomes, as uncovered in

theoretical research, are preference differences, comparative advantage, human capital

investments, and discrimination, according to Altonji and Blank (1999).

21 For further technical details and discussion on these methods, see Wooldridge (2002).

3.3.2 INSTRUMENTS

Given an ordinary least squares estimation of a dependent variable on a set of

independent variables where one is endogenous, a good instrument requires strong correlation

between the instrument and the endogenous variable. It also requires no correlation between the

instrument and the residuals of the originally specified estimating equation. Furthermore, the

instrumental variable should not be directly correlated with the dependent variable and only be

correlated with the dependent variable through correlation with the endogenous variable. A weak

instrument problem arises when correlation between the instrument and the residuals of the

original estimation equation is relatively stronger than the correlation between the endogenous

variable and the instrument.22 To test for weak instruments in this study, I conduct F-tests on the

instruments in the first stage regression. F-statistic values less than 10 indicate a weak instrument

problem according to Staiger and Stock (1997).

Requirements for a good instrument can be generalized in cases where there are multiple

endogenous variables and multiple instruments. In cases where there are more instruments than

endogenous variables, instruments can be jointly evaluated to determine if they are correlated

with the residuals of the originally specified estimating equation. This is known as an

overidentification test. To evaluate overidentification in this paper I utilize Hansen’s J-test which

tests the null hypothesis that the instruments are not correlated with the residuals of the originally

specified estimating equation. 23 Rejection of the null hypothesis indicates that the instruments

22 An example of this occurs in Bound, Jaeger, and Baker (1995), where the authors reconsider an instrumental variable approach from Angrist and Krueger (1991). Specifically, quarter of birth is used as an instrument for years of education in an estimating equation where log wages is the dependent variable. Bound, Jaeger, and Baker (1995) argue that quarter of birth and years of education are weakly correlated relative to the correlation between the quarter of birth and the stochastic error of the log wages estimating equation. They argue that the correlation between quarter of birth and the stochastic error stems from quarter of birth’s association with performance in school and differences in physical and mental health of individuals born at different times of the year. 23 The Hansen’s J-test is used instead of the Sargan-Hansen test because sample weights are applied to the two-stage least squares estimates. Generalized method of moments (GMM) is used to obtain the Hansen’s J-statistic by

are jointly correlated with the residuals of the originally specified estimating equation. In

essence, a weak instrument problem is present for one or more of the instruments.

To address the endogeneity problem with the model I estimate with the pooled cross section data, I utilize two-stage least squares with spouse education and interactions of spouse education with non-labor income and not being employed as instruments. Research comments extensively on the relationship between earnings and spouse education. Loh (1996) uses National

Longitudinal Survey of Youth data on married white males to find that a husband’s wage increases as a wife’s education increases. He ascribes the positive effect on husband’s wages to

unobserved heterogeneity of men in his sample in addition to more highly educated wives

improving resource allocation decisions of husbands. Huang et al. (2009) use data on Chinese

twins to identify a positive correlation between spouse education and earnings. They argue and

present evidence that the correlation stems from cross productivity between couples and positive

assortative mating. Cross-productivity refers to spousal education helping individuals to

accumulate human capital and increase earnings while positive assortative mating means that

individuals tend to marry individuals with similar socioeconomic status.24 Evidence reveals that spouse education is positively correlated with earnings through cross-productivity and positive assortative mating. Groothuis and Gabriel (2010) use U.S. Census and Current Population

Survey data to show that the marriage premium, the positive earnings differential married individuals have over nonmarried individuals, for husbands and wives is directly related to

computing the minimized value of the GMM criterion function. The Hansen’s J-statistic has degrees of freedom equal to the number of overidentifying restrictions and has chi-squared distribution. The process by which this statistic can be obtained is described in detail and by example in Wooldridge (2002, p.123-124). 24 This socioeconomic status includes education and professional background. According to Mare (1991), who uses Census and Current Population Survey data, positive assortative mating has been an emergent trend in the U.S. from the 1930s to the 1990s.

spouse education. They argue that positive assortative mating contributes to this outcome through the effect of intellectual interaction between spouses that raises productivity.

Lefgren and McIntyre (2006), finally, use 2000 Census data to show that women’s education appears to have a positive causal effect on husband’s earnings.

Other studies such as Benham (1974), Neuman and Ziderman (1992), and Tiefenhalter

(1997) confirm a positive relationship between spouse education and earnings. In essence, research establishes a positive correlation between spouse education and earnings. This positive correlation is attributed to a variety of aforementioned reasons such as positive assortative mating, improved resource allocation, unobserved heterogeneity, intellectual interaction, and cross-productivity. I expect spouse education to be positively correlated with labor income and wages in this study.

Concerns arise over using spouse education as an instrument when considering whether spouse education is directly correlated with frequency of religious service attendance or frequency of prayer. No studies directly consider the effect of spouse education on religiosity.

Positive assortative mating implies that spouse education is correlated with an individual’s own education. Since an individual’s own education is positively correlated with religiosity, as in

Sacerdote and Glaeser (2001), spouse education could be positively correlated with religiosity.

Furthermore, studies such as O’Connor et al (2002) and Petersen (1994) show that spouses influence individual religious behavior. I conduct ordinary least squares regressions of equations

(1) and (2) with spouse education included as an explanatory variable to determine if spouse education directly impacts frequency of religious service attendance or frequency of prayer.

Results are discussed in Section 4.

I also utilize C-tests as described by Baum et al (2007) to evaluate endogeneity.25 The test is performed by computing the difference of two Hansens J-statistics, one where the endogenous regressors are treated as endogenous and one where the endogenous regressors are treated as exogenous.26 The C-statistic has degrees of freedom equal to the number of

endogenous regressors with a chi-squared distribution. The null hypothesis considered by the test

is that the endogenous regressors can be treated as exogenous.

3.4 RESULTS

First stage regressions with sample weights appear in Tables 3.3 and 3.4. Ordinary least

squares and second stage regressions with sample weights are presented in Tables 3.5-3.8. All tables are similarly organized. In Tables 3.5-3.8, the first and third columns represent results from running ordinary least squares with sample weights. The remaining columns present the second stage regressions for instrumental variables with sample weights. Tables 3.9 and 3.10 present ordinary least squares regressions that include spouse education as a determinant with sample weights. Tables 3.11, 3.12, and 3.13 present information rather than estimation results from the panel dataset. 27 I conduct Chow tests using both datasets and find that each estimating

equation estimated should be done separately for men and women with all F-statistics≥1.95 and

p<0.01.28 A weakness of the Chow test is that it allows for no differences at all between

coefficient estimates for men and women and does not reveal the exact sources of the disparity.

25 This test is also known as a GMM distance test. 26 The Hansens’s J-statistic is computed using two step efficient GMM. It is computed as the sum of the squared residuals weighted by the sample weights multiplied by the sample size. In the case where the endogenous variables are treated as exogenous, two step efficient GMM reduces to OLS. For more details and an example of a Hansen’s J- statistic computation see Wooldridge (1995), Wooldridge (2002, p.123-124), and Hayashi (2000, p. 217-221). 27 Panel estimation results are not presented due to their being no statistical significance in the labor income or wages and prayer frequency or attendance frequency relationships regardless of estimation technique. I specifically conducted pooled ordinary least squares and fixed effects estimation. None of the wages or labor income coefficients are significant in any of the techniques employed. 28 These results reject the null that there is no difference between male and female findings.

It only estimates an overall difference in coefficient estimates. Even so, the tests indicate that

there is a divergence between men and women. The distinction is most likely due to differences

in religiosity such as in Miller and Hoffman (1995), Thompson (1991), and Collett and Lizardo

(2009) and differences in labor market outcomes such as in Altonji and Blank (1999), Pencavel

(1986), and Killingsworth and Heckman (1986).

Table 3.3 displays first stage regressions for labor income and its square. The three

instrumental variables are statistically significant at the one percent threshold for men and at or

below the ten percent threshold for women in all cases. F-tests on the instruments reveal F- statistics greater than 10 for men and less than 10 for women in all cases. The evidence, therefore, indicates a weak instrument problem for women when instrumenting for labor income and its squares.

Table 3.4 presents first stage regressions for wages and its square. All three instrumental variables are statistically significant at the five percent threshold or lower for men’s wages. For men’s wages squared, spouse education and its interaction not being employed are statistically significant at the ten percent level or lower. In the case of women’s wages, only the interaction of spouse education and not being employed is significant at the one percent threshold. F- statistics on the instruments are 13.88, 2.93, 2.64, and 0.77 for men’s wages, men’s wages squared, women’s wages, and women’s wages squared, respectively. In essence, the instruments are weak in all cases except men’s wages because their F-statistics are less than 10.

Table 3.5 shows that labor income has a very small negative effect on the frequency of religious service attendance for women with statistical significance at the one percent threshold using ordinary least squares. The magnitude of the coefficient indicates that a $100,000 increase in labor income reduces frequency of religious service attendance by once per month.

Instrumenting leads to the coefficient on labor income becoming statistically significant at the

five percent level. A $9,090.91 increase in labor income decreases frequency of religious service

attendance by once per month, according to the estimate. Men, meanwhile, display no statistical

significance regardless of estimation technique. Instrument diagnostic tests indicate no evidence that the instruments are not exogenous with both Hansen’s J-statistics showing p>0.05.29 The

null hypothesis that the instruments are jointly exogenous is not rejected. I also perform C-tests

as described by Baum et al (2007).30 I reject the null hypothesis that the endogenous regressors

can be treated as exogenous in both cases.

Table 3.6 presents that men and women both have statistically significant coefficients for

the ordinary least squares effect of labor income on frequency of prayer. Specifically, a $25,000

increase in labor income leads to a reduction of one prayer per week for men while a $20,000

increase leads to a reduction of one prayer per week for women. These magnitudes are greater

than findings by Brown (2009), who contends that $90,700 and $87,160 increases in labor

income are required to reduce frequency of prayer by one per week for men and women,

respectively.31 When instrumenting, the coefficient for labor income becomes statistically

indistinguishable from 0 for men. The coefficient for women, however, is statistically significant

at the one percent level and indicates that a $2,083.33 increase in labor income decreases prayer

frequency by once per week for women. There is no evidence that the instruments are not

exogenous with both Hansen’s J-statistics showing p>0.05. The null hypothesis that the

instruments are jointly exogenous is not rejected. For the C-tests, I am able to reject the null

hypothesis that the endogenous regressors can be treated as exogenous for women’s prayer. I

29 This test is robust to heteroskedasticity. 30 This test is also known as a GMM distance test. 31 Brown (2009) computes these magnitudes from his findings using wages while assuming 2000 hours of work per year. He does not directly estimate the effect of labor income on frequency of prayer.

cannot, however, reject the null hypothesis that the endogenous regressors can be treated as

exogenous for men’s prayer.

Table 3.7 shows that the wage is a poor predictor of frequency of attendance for men.

Women are initially estimated to have wages negatively affect frequency of attendance with

statistical significance at the one percent level. An increase of $142.86 decreases frequency of

religious service attendance by once per month. Instrumenting for wages leads to statistical

significance in the coefficient on wages at the five percent level. Raising wages by $12.99 lowers

frequency of religious service attendance by once per month. There is no evidence that the

instruments are not exogenous with both Hansen’s J-statistics showing p>0.05. The null hypothesis that the instruments are jointly exogenous is not rejected. For the C-tests, I am able to reject the null hypothesis that the endogenous regressors can be treated as exogenous in both cases.

Table 3.8 displays wages as negatively affecting frequency of prayer for men and women using ordinary least squares with statistical significance at the one percent level. A $22.22 increase in wages for men and a $37.04 increase for women are associated with a reduction of one prayer per week. These magnitudes are greater than Brown (2009), who predicts that a

$45.35 increase in wages for men and a $43.58 increase in wages for women reduce frequency of prayer by one per week. In the case of men, instrumenting leads to the coefficient of wages being statistically insignificant. For women, meanwhile, a $3.46 increase in wages implies a reduction in frequency of prayer by once per week with statistical significance at the five percent level when instrumenting. There is evidence that the instruments are not exogenous with the Hansen’s

J-statistic for women’s prayer showing p<0.05. The other Hansen’s J-statistic for men’s prayer displays no evidence that the instruments are not exogenous. For the C-test, I reject the null

hypothesis that the endogenous regressors can be treated as exogenous in the case of women’s

prayer. The C-test for men’s prayer indicates no rejection of the null hypothesis that the endogenous regressors can be treated as exogenous.

Tables 3.9 and 3.10 reveal that spouse education only affects men’s attendance directly.

F-tests from Tables 3.3-3.4 reveal a weak instrument problem in 4 of 4 cases for women and 1 of

4 cases for men. Hansen’s J-statistic tests in Tables 5-8 show evidence that the instruments are

not exogenous in 1 of 4 cases for women and 0 of 4 cases for men. The C-tests imply that the

endogenous regressors can be treated as exogenous in 0 of 4 cases for women and 2 of 4 cases

for men. Together these results show that the instruments for women are weak while instruments

for men are questionable. Despite problematic trends with the instruments, alternative

instruments provide results that are less reliable.32 Given these findings on the instruments, I

conclude that the ordinary least squares provides the best estimates of the labor income and

religiosity relationship.

The ordinary least squares results indicate that labor income is a negative correlate of

frequency of prayer for men and women and a negative correlate of frequency of religious

service attendance for women. These results hold when I conduct a robustness check using

wages. A tradeoff between religious and secular activity appears to exist. This finding coincides

with theories and findings from Azzi and Ehrenberg (1975) and Iannaccone (1988).

For the other variables I present in Tables 3.5-3.8, being raised Fundamentalist is

consistently positive with statistical significance at the one percent level in all cases except when

instrumenting for wages with men. Being raised Jewish is consistently negative and statistically

significant at the one percent threshold except when instrumenting for wages. In essence,

respondents who are raised Jewish exhibit lower frequency of prayer and frequency of religious

32 See Appendix A.3 for details concerning other instruments that I attempt to use as part of this study.

service attendance. Respondents raised in a Fundamentalist religion have higher frequencies of prayer and religious service attendance.33

Tables 3.11, 3.12, and 3.13 present information on the prayer frequency and attendance frequency measures. They also show first differences for prayer frequency and attendance frequency in the panel dataset. Panel estimation results are not presented since pooled ordinary least squares, random effects, and fixed effects estimation did not produce any statistically significant coefficients for labor income or wages. Tables 3.1, 3.2, 3.11, 3.12, and 3.13 indicate that the panel results differ from the cross section results due to sample differences and lack of variation in prayer frequency and attendance frequency. Key measures in this study, which include labor income, wages, attendance frequency, and prayer frequency, have different means for the panel sample in comparison to the cross section. Standard deviations are also dissimilar.

The panel sample, furthermore, only contains 628 respondents with two years of data. The transition matrix for the frequency of prayer measure in Table 3.11, for example, shows that 87 percent of respondents report the same frequency of prayer or the next category up or below their original frequency of prayer category for years 2006 and 2008. A similar trend is also present in

Table 3.12 for frequency of religious service attendance. In essence, there is a lack of variation within individuals for frequency of prayer and frequency of religious service attendance over a two year period. Table 3.12, meanwhile, displays first difference descriptive statistics for frequency of prayer and frequency of religious service attendance. The means for the variables show that frequency of prayer falls by less than one prayer per week and frequency of religious service attendance is nearly unchanged on a per week basis. The relatively small means imply a

33 Iannaccone (1998) highlights that I should expect a positive correlation between any measure of religious commitment such as prayer or attendance and the conservatism, strictness, or sectarianism of a religious group.

lack of variation, thus it is not surprising that the panel dataset yields no statistically significant results.

3.5 CONCLUSIONS

The results of this study provide evidence using ordinary least squares that labor income is a negative correlate of frequency of prayer for men and women. It is also a negative correlate of frequency of religious service attendance for women. These results hold when I conduct a robustness check using wages. Magnitudes indicate that frequency of prayer is more sensitive to changes in labor income and wages than in Brown (2009). I estimate that a $25,000 increase in labor income leads to a reduction of one prayer per week for men while a $20,000 increase leads to a reduction of one prayer per week for women. For wages, I find that a $22.22 increase in wages for men and a $37.04 increase for women are associated with a reduction of one prayer per week. Magnitudes for frequency of attendance and income or wages indicate that an increase of $142.86 decreases frequency of religious service attendance by once per month for women and a $9,090.91 increase in labor income decreases frequency of religious service attendance by once per month for women.

I use instrumental variables estimation to account for the impact of noncognitive measure locus of control. The estimation technique is problematic due to weak identification and exogeneity tests determining that the endogenous regressors can be treated as exogenous in some cases. While panel data estimation is considered and undertaken, the panel suffers from lack of within variation in key measures to yield any meaningful results.

This study offers insight into the relatively understudied measure of prayer frequency by being the first to consider endogeneity in the frequency of prayer and labor income or frequency of prayer and wages relationships. By examining endogeneity with the General Social Survey, I

find that ordinary least squares estimation provides the best estimates of the frequency of prayer

and labor income relationship. I also find additional evidence that ordinary least squares

estimates of frequency of religious service attendance and labor income are the best estimates for

women.

Further research should revisit endogeneity and instrumental variables if better

instruments become available. Extensive panel data on frequency of prayer and frequency of

religious service attendance should also be employed if it becomes available. Restricted data on

the location of GSS respondents should also be obtained to employ clustering and other grouped- data methods. If information becomes available on the average length of time per prayer, it should be utilized.34 Time use studies should also be incorporated as part of future work.35 The possibility of exploring the relationship between noncognitive traits and other measures of religious commitment should also be carried out.

34 I carried out an exhaustive search for this information. To my knowledge it does not exist. 35 The American Time Use Survey is one specific example of a data source on time use. Unfortunately prayer is not listed as separate category for the use of time in this survey.

Table 3.1: Descriptive Statistics for Weighted Pooled Cross Section Dataset Variable Obs Mean Std. Dev. Min Max Dependent Variables Frequency of Prayer Per Week 4240 8.72 8.10 0 21 Frequency of Attendance Per Month 4240 2.12 2.19 0 6.5 Primary Variables Labor Income (Thousands of 2004 dollars) 4240 28.85 37.78 0 258.58 Wages (Tens of 2004 dollars) 4240 1.61 4.39 0 129.29 Non-labor Income (Thousands of 2004 dollars) 4240 31.13 36.88 0 182.57 Spouse Education 4240 7.92 6.99 0 20 Religious Upbringing Variables Raised Catholic 4240 0.33 0.47 0 1 Raised Protestant 4240 0.60 0.49 0 1 Raised Jewish 4240 0.02 0.15 0 1 Raised Other 4240 0.01 0.10 0 1 Raised None 4240 0.03 0.17 0 1 Raised Fundamentalist 4240 0.34 0.47 0 1 Raised Moderate 4240 0.45 0.50 0 1 Raised Liberal 4240 0.21 0.41 0 1 Control Variables Employed 4240 0.75 0.43 0 1 Not Employed 4240 0.25 0.43 0 1 Age 4240 45.67 16.24 18 89 Lived in Same State as When Aged 16 4240 0.66 0.47 0 1 Women 4240 0.53 0.50 0 1 Men 4240 0.47 0.50 0 1 Never Married 4240 0.20 0.40 0 1 Married 4240 0.59 0.49 0 1 Widowed 4240 0.06 0.24 0 1 Divorced/Separated 4240 0.15 0.35 0 1 White 4240 0.81 0.39 0 1 Black 4240 0.13 0.33 0 1 Other Race/Ethnicity 4240 0.07 0.25 0 1 Less Than High School 4240 0.12 0.33 0 1 High School Graduate 4240 0.56 0.50 0 1 Associates Degree 4240 0.07 0.26 0 1 Bachelors Degree 4240 0.16 0.37 0 1 Graduate Degree 4240 0.08 0.28 0 1

Table 3.2: Descriptive Statistics for Weighted Panel Dataset Variable Obs Inds Mean Std. Dev. Min Max Dependent Variables Frequency of Prayer Per Week 1256 628 9.92 8.33 0 21 Frequency of Attendance Per Month 1256 628 2.39 2.28 0 6.5 Primary Variables Labor Income 1256 628 32.08 55.71 0 476.79 (Thousands of 2004 dollars) Wages 1256 628 1.91 8.81 0 305.63 (Tens of 2004 dollars) Non-labor Income 1256 628 40.58 43.73 0 195.57 (Thousands of 2004 dollars) Religious Upbringing Variables Raised Catholic 1256 628 0.37 0.48 0 1 Raised Protestant 1256 628 0.56 0.50 0 1 Raised Jewish 1256 628 0.02 0.14 0 1 Raised Other 1256 628 0.01 0.09 0 1 Raised None 1256 628 0.04 0.18 0 1 Raised Fundamentalist 1256 628 0.36 0.48 0 1 Raised Moderate 1256 628 0.45 0.50 0 1 Raised Liberal 1256 628 0.19 0.39 0 1 Control Variables Employed 1256 628 0.66 0.47 0 1 Not Employed 1256 628 0.34 0.47 0 1 Age 1256 628 47.77 15.39 18 89 Lived in Same State as When Aged 16 1256 628 0.68 0.47 0 1 Women 1256 628 0.55 0.50 0 1 Men 1256 628 0.45 0.50 0 1 Never Married 1256 628 0.15 0.36 0 1 Married 1256 628 0.70 0.46 0 1 Widowed 1256 628 0.05 0.21 0 1 Divorced/Separated 1256 628 0.11 0.31 0 1 White 1256 628 0.82 0.39 0 1 Black 1256 628 0.11 0.32 0 1 Other Race/Ethnicity 1256 628 0.07 0.25 0 1 Less Than High School 1256 628 0.08 0.27 0 1 High School Graduate 1256 628 0.52 0.50 0 1 Associates Degree 1256 628 0.08 0.28 0 1 Bachelors Degree 1256 628 0.21 0.41 0 1 Graduate Degree 1256 628 0.10 0.30 0 1

Table 3.3: First Stage Cross Section Results for Labor Income and Labor Income Squared Variable IV(1) IV(1) IV(1) IV(1) Men Men Women Women Labor Income Labor Income Labor Income Labor Income Squared Squared Spouse Education 3.19*** 6.46*** 1.58*** 3.55** (0.57) (1.37) (0.51) (1.41) Spouse Education x -0.02*** -0.05*** -0.01* -0.02** Nonlabor Income (0.00) (0.01) (0.00) (0.01) Spouse Education x -1.70*** -1.95*** -0.45*** -0.41* Not Employed (0.25) (0.50) (0.13) (0.21) Non-labor Income -0.07** -0.05 -0.09*** -0.03 (0.03) (0.05) (0.02) (0.04) Raised Catholic -0.07** -0.05 -0.09*** -0.03 (0.03) (0.05) (0.02) (0.04) Raised Protestant 7.28 18.29* 3.81 2.99 (5.24) (10.95) (3.48) (6.95) Raised Jewish 3.36 7.98 1.30 3.65 (4.44) (9.44) (2.43) (3.39) Raised Other 24.49** 62.01** 2.30 6.56 (9.61) (26.16) (5.15) (8.91) Raised -7.18 -12.35 3.84 -1.95 Fundamentalist (8.79) (16.76) (5.31) (6.89) Raised Moderate -7.88*** -14.20** -0.54 0.58 (2.67) (6.52) (1.66) (3.09) Constant -8.92*** -18.11** 1.96 7.42 (3.19) (7.08) (2.73) (6.63) Control Variables Yes Yes Yes Yes Year Effects Yes Yes Yes Yes Region Effects Yes Yes Yes Yes Observations 1926 1926 2314 2314 Adjusted R-squared 0.43 0.24 0.37 0.15 F-Statistic 45.45 10.53 53.91 7.47 Instruments’ F- 31.82 18.45 5.87 2.49 Statistic Robust standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1% Control Variables include: Not Employed, Age, Age Squared, Lived in Same State as When Aged 16, Married, Widowed, Divorced/Separated, High School Graduate, Associates Degree, Bachelors Degree, Graduate Degree, Black, and Other Race

Table 3.4: First Stage Cross Section Results for Wages and Wages Squared Variable IV(1) IV(1 ) IV(1) IV(1) Men Men Women Women Wages Wages Squared Wages Wages Squared Spouse Education 0.22*** 0.49* 0.32 3.29 (0.06) (0.27) (0.23) (3.08) Spouse Education x -0.00** -0.00 -0.00 -0.02 Non-labor Income (0.00) (0.00) (0.00) (0.02) Spouse Education x -0.12*** -0.24** -0.07*** -0.32 Not Employed (0.02) (0.11) (0.03) (0.22) Non-labor Income -0.00 -0.00 0.00 0.06 (0.00) (0.01) (0.01) (0.07) Raised Catholic -0.37 -2.60 -2.19 -17.58 (0.73) (3.28) (2.52) (18.75) Raised Protestant -0.57 -2.68 -2.51 -20.86 (0.71) (3.13) (2.40) (17.76) Raised Jewish -0.62 -5.65 -2.99 -28.72 (0.88) (3.75) (2.60) (20.37) Raised Other -1.32* -4.41 -2.50 -23.19 (0.75) (3.25) (2.35) (17.79) Raised Fundamentalist -0.28 0.25 0.36 4.46 (0.25) (1.13) (0.31) (3.21) Raised Moderate -0.30 0.45 0.39 4.08 (0.31) (1.56) (0.29) (2.95) Constant -0.73 -1.77 2.76 18.23 (0.76) (3.27) (4.12) (32.29) Control Variables Yes Yes Yes Yes Year Effects Yes Yes Yes Yes Region Effects Yes Yes Yes Yes Observations 1926 1926 2314 2314 Adjusted R-squared 0.19 0.02 0.11 0.04 F-Statistic 28.71 2.05 18.88 0.18 Instruments’ F-Statistic 13.88 2.93 2.64 0.77 Robust standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1% Control Variables include: Not Employed, Age, Age Squared, Lived in Same State as When Aged 16, Married, Widowed, Divorced/Separated, High School Graduate, Associates Degree, Bachelors Degree, Graduate Degree, Black, and Other Race

Table 3.5: OLS and Second Stage Cross Section Results for Attendance and Labor Income Variable OLS IV(2) OLS IV(2) Men Men Women Women Attendance Attendance Attendance Attendance Labor Income -0.00 0.01 -0.01*** -0.11** (0.00) (0.02) (0.00) (0.05) Labor Income Squared 0.00 0.00 0.00 0.06*** (0.00) (0.01) (0.00) (0.02) Non-labor Income 0.00 0.00** -0.00** -0.00 (0.00) (0.00) (0.00) (0.00) Raised Catholic -0.18 -0.29 0.24 0.42 (0.34) (0.37) (0.34) (0.42) Raised Protestant -0.22 -0.27 -0.14 -0.25 (0.30) (0.32) (0.29) (0.34) Raised Jewish -1.39*** -1.99*** -1.24*** -1.46*** (0.36) (0.50) (0.34) (0.46) Raised Other -0.93** -0.76* -0.28 0.20 (0.41) (0.43) (0.49) (0.56) Raised Fundamentalist 0.48*** 0.62*** 0.67*** 0.58*** (0.15) (0.17) (0.16) (0.18) Raised Moderate 0.11 0.27 0.11 -0.13 (0.18) (0.20) (0.19) (0.26) Constant 0.45 0.87 -0.28 0.23 (0.60) (0.64) (0.57) (0.73) Control Variables Yes Yes Yes Yes Year Effects Yes Yes Yes Yes Region Effects Yes Yes Yes Yes Observations 1926 1926 2314 2314 Adjusted R-squared 0.11 0.09 F-Statistic 8.88 8.46 8.36 5.32 First Stage F-Statistic 31.82 5.87 Labor Income First Stage F-Statistic 18.45 2.49 Labor Income Squared Hansen’s J Statistic for 3.45 0.74 Overidentification Hansen’s J Statistic 0.06 0.39 p-value ( ) C-Test Statistic 11.16 12.45 C-Test Statistic 0.00 0.00 p-value ( ) Robust standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1% Control Variables include: Not Employed, Age, Age Squared, Lived in Same State as When Aged 16, Married, Widowed, Divorced/Separated, High School Graduate, Associates Degree, Bachelors Degree, Graduate Degree, Black, and Other Race

Table 3.6: OLS and Second Stage Cross Section Results for Prayer and Labor Income Variable OLS IV(2) OLS IV(2) Men Men Women Women Prayer Prayer Prayer Prayer Labor Income -0.04*** -0.01 -0.05*** -0.48*** (0.01) (0.08) (0.01) (0.18) Labor Income Squared 0.01** 0.02 0.01 0.18** (0.01) (0.04) (0.01) (0.08) Non-labor Income -0.01 0.00 -0.01** -0.03*** (0.01) (0.01) (0.01) (0.01) Raised Catholic -0.41 -0.70 0.03 1.14 (1.31) (1.35) (1.23) (1.56) Raised Protestant -0.57 -0.69 0.74 0.65 (1.16) (1.16) (1.06) (1.31) Raised Jewish -4.11*** -5.71*** -3.03** -3.08 (1.45) (1.94) (1.37) (1.93) Raised Other 2.15 2.58 -2.41 -0.34 (2.36) (2.33) (1.66) (2.13) Raised Fundamentalist 2.25*** 2.62*** 2.52*** 2.18*** (0.59) (0.64) (0.55) (0.65) Raised Moderate 0.60 1.02 1.53** 1.12 (0.68) (0.74) (0.70) (0.90) Constant -0.82 0.29 -2.61 -3.45 (2.09) (2.26) (2.07) (2.56) Control Variables Yes Yes Yes Yes Year Effects Yes Yes Yes Yes Region Effects Yes Yes Yes Yes Observations 1926 1926 2314 2314 Adjusted R-squared 0.10 0.13 F-Statistic 7.45 6.64 12.16 7.21 First Stage F-Statistic 31.82 5.87 Labor Income First Stage F-Statistic 18.45 2.49 Labor Income Squared Hansen’s J Statistic for 1.10 3.53 Overidentification Hansen’s J Statistic 0.30 0.07 p-value ( ) C-Test Statistic 4.34 8.71 C-Test Statistic 0.11 0.01 p-value ( ) Robust standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1% Control Variables include: Not Employed, Age, Age Squared, Lived in Same State as When Aged 16, Married, Widowed, Divorced/Separated, High School Graduate, Associates Degree, Bachelors Degree, Graduate Degree, Black, and Other Race

Table 3.7: OLS and Second Stage Cross Section Results for Attendance and Wages Variable OLS IV(2) OLS IV(2) Men Men Women Women Attendance Attendance Attendance Attendance Wages 0.01 -1.64 -0.07*** -0.77** (0.04) (3.62) (0.02) (0.34) Wages Squared -0.01 0.91 0.00** 0.09* (0.01) (1.68) (0.00) (0.05) Non-labor Income 0.00 -0.00 -0.00 -0.00 (0.00) (0.01) (0.00) (0.00) Raised Catholic -0.18 1.64 0.13 0.06 (0.34) (2.67) (0.33) (0.73) Raised Protestant -0.23 1.32 -0.23 -0.24 (0.30) (2.45) (0.28) (0.69) Raised Jewish -1.40*** 2.71 -1.34*** -1.05 (0.36) (6.99) (0.32) (0.78) Raised Other -0.93** 0.98 -0.41 -0.21 (0.41) (2.49) (0.50) (0.77) Raised Fundamentalist 0.49*** -0.21 0.68*** 0.56*** (0.15) (1.22) (0.16) (0.18) Raised Moderate 0.12 -0.81 0.11 0.05 (0.18) (1.71) (0.19) (0.22) Constant 0.44 0.79 -0.06 0.17 (0.59) (2.35) (0.55) (1.13) Control Variables Yes Yes Yes Yes Year Effects Yes Yes Yes Yes Region Effects Yes Yes Yes Yes Observations 1926 1926 2314 2314 Adjusted R-squared 0.11 0.08 F-Statistic 8.90 3.44 14.77 5.86 First Stage F-Statistic 13.88 2.64 Labor Income First Stage F-Statistic 2.93 0.77 Labor Income Squared Hansen’s J Statistic for 0.09 0.58 Overidentification Hansen’s J Statistic 0.77 0.44 p-value ( ) C-Test Statistic 12.48 12.25 C-Test Statistic 0.00 0.00 p-value ( ) Robust standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1% Control Variables include: Not Employed, Age, Age Squared, Lived in Same State as When Aged 16, Married, Widowed, Divorced/Separated, High School Graduate, Associates Degree, Bachelors Degree, Graduate Degree, Black, and Other Race

Table 3.8: OLS and Second Stage Cross Section Results for Prayer and Wages Variable OLS IV(2) OLS IV(2) Men Men Women Women Prayer Prayer Prayer Prayer Wages -0.45*** 8.72 -0.27*** -2.89** (0.14) (15.45) (0.09) (1.20) Wages Squared 0.09*** -3.89 0.02** 0.25** (0.03) (7.22) (0.01) (0.12) Non-labor Income -0.01 0.02 -0.01 -0.01 (0.01) (0.03) (0.01) (0.01) Raised Catholic -0.47 -7.31 -0.46 -2.15 (1.33) (12.05) (1.19) (2.70) Raised Protestant -0.65 -6.05 0.32 -1.43 (1.18) (10.78) (1.02) (2.53) Raised Jewish -4.29*** -21.25 -3.49*** -4.59 (1.45) (30.49) (1.32) (2.90) Raised Other 2.11 -3.29 -2.94* -4.05 (2.35) (11.39) (1.67) (2.77) Raised Fundamentalist 2.28*** 5.85 2.58*** 2.48*** (0.59) (5.48) (0.55) (0.65) Raised Moderate 0.61 5.13 1.53** 1.64** (0.68) (7.44) (0.70) (0.82) Constant -0.56 -1.68 -1.59 1.20 (2.10) (8.89) (2.02) (4.21) Control Variables Yes Yes Yes Yes Year Effects Yes Yes Yes Yes Region Effects Yes Yes Yes Yes Observations 1926 1926 2314 2314 Adjusted R-squared 0.10 0.12 F-Statistic 7.34 2.38 15.65 9.02 First Stage F-Statistic 13.88 2.64 Labor Income First Stage F-Statistic 2.93 0.77 Labor Income Squared Hansen’s J Statistic for 0.00 5.18 Overidentification Hansen’s J Statistic 0.98 0.02 p-value ( ) C-Test Statistic 4.87 7.70 C-Test Statistic 0.09 0.02 p-value ( ) Robust standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1% Control Variables include: Not Employed, Age, Age Squared, Lived in Same State as When Aged 16, Married, Widowed, Divorced/Separated, High School Graduate, Associates Degree, Bachelors Degree, Graduate Degree, Black, and Other Race

Table 3.9: OLS Cross Section Results for Spouse Education, Labor Income, and Labor Income Squared Variable OLS OLS OLS OLS Men Men Women Women Attendance Prayer Attendance Prayer Spouse Education 0.09*** 0.03 0.02 -0.01 (0.03) (0.03) (0.10) (0.09) Labor Income -0.00 -0.01*** -0.04*** -0.05*** (0.00) (0.00) (0.01) (0.01) Labor Income Squared 0.00 0.00 0.01** 0.01 (0.00) (0.00) (0.01) (0.01) Non-labor Income -0.00 -0.00** -0.01 -0.01** (0.00) (0.00) (0.01) (0.01) Raised Catholic -0.13 0.24 -0.40 0.03 (0.34) (0.34) (1.31) (1.23) Raised Protestant -0.21 -0.14 -0.56 0.74 (0.30) (0.29) (1.16) (1.06) Raised Jewish -1.41*** -1.26*** -4.11*** -3.02** (0.36) (0.34) (1.45) (1.37) Raised Other -0.90** -0.30 2.16 -2.41 (0.41) (0.50) (2.36) (1.66) Raised Fundamentalist 0.48*** 0.67*** 2.25*** 2.52*** (0.15) (0.16) (0.59) (0.55) Raised Moderate 0.10 0.11 0.60 1.53** (0.18) (0.19) (0.68) (0.70) Constant 0.48 -0.23 -0.81 -2.61 (0.59) (0.57) (2.09) (2.07) Control Variables Yes Yes Yes Yes Year Effects Yes Yes Yes Yes Region Effects Yes Yes Yes Yes Observations 1926 2314 1926 2314 Adjusted R-squared 0.11 0.09 0.10 0.13 F-Statistic 9.14 8.27 7.23 11.82 Robust standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1% Control Variables include: Not Employed, Age, Age Squared, Lived in Same State as When Aged 16, Married, Widowed, Divorced/Separated, High School Graduate, Associates Degree, Bachelors Degree, Graduate Degree, Black, and Other Race

Table 3.10: OLS Cross Section Results for Spouse Education, Wages, and Wages Squared Variable OLS OLS OLS OLS Men Men Women Women Attendance Prayer Attendance Prayer Spouse Education 0.09*** 0.03 0.01 -0.01 (0.03) (0.03) (0.10) (0.09) Wages -0.00 -0.07*** -0.45*** -0.27*** (0.04) (0.02) (0.14) (0.09) Wages Squared -0.00 0.00** 0.09*** 0.02** (0.01) (0.00) (0.03) (0.01) Non-labor Income -0.00 -0.00* -0.01 -0.01 (0.00) (0.00) (0.01) (0.01) Raised Catholic -0.14 0.12 -0.46 -0.46 (0.34) (0.33) (1.33) (1.19) Raised Protestant -0.22 -0.24 -0.65 0.33 (0.30) (0.28) (1.18) (1.02) Raised Jewish -1.43*** -1.37*** -4.29*** -3.48*** (0.36) (0.32) (1.46) (1.32) Raised Other -0.91** -0.43 2.12 -2.93* (0.41) (0.50) (2.35) (1.67) Raised Fundamentalist 0.49*** 0.69*** 2.28*** 2.58*** (0.15) (0.16) (0.59) (0.56) Raised Moderate 0.10 0.11 0.61 1.53** (0.18) (0.19) (0.68) (0.70) Constant 0.48 -0.00 -0.56 -1.61 (0.59) (0.55) (2.10) (2.01) Control Variables Yes Yes Yes Yes Year Effects Yes Yes Yes Yes Region Effects Yes Yes Yes Yes Observations 1926 2314 1926 2314 Adjusted R-squared 0.11 0.08 0.10 0.12 F-Statistic 9.17 12.73 7.13 15.40 Robust standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1% Control Variables include: Not Employed, Age, Age Squared, Lived in Same State as When Aged 16, Married, Widowed, Divorced/Separated, High School Graduate, Associates Degree, Bachelors Degree, Graduate Degree, Black, and Other Race

Table 3.11: Frequency of Prayer Transition Matrix Frequency of Prayer 2008 Frequency of Prayer 2006 1 2 3 4 5 6 1 142 81 16 2 4 0 2 48 102 28 6 5 0 3 9 18 28 11 3 1 4 2 8 9 11 16 0 5 3 5 2 14 20 8 6 0 3 0 4 7 12 Note: Never=6, Less than once a week=5, Once a week=4, Several times a week=3, Once a day=2, and Several times a day=1

Table 3.12: Frequency of Attendance Transition Matrix Frequency of Attendance 2008 Frequency of Attendance 2006 1 2 3 4 5 6 7 8 1 6 4 5 2 1 0 1 0 2 8 32 11 4 2 3 2 1 3 6 21 25 12 10 4 2 2 4 1 7 11 5 7 2 5 2 5 0 2 10 7 17 6 11 2 6 0 2 0 2 1 9 17 1 7 1 1 3 3 11 13 92 20 8 1 2 0 3 3 6 25 34 Note: Never=1, Less than once a year=2, Once a year to several times a year=3, Once a month=4, 2-3 times a month=5, Nearly every week=6, Every week=7, and More than once a week=8

Table 3.13: First Differences of Prayer and Attendance Variable Obs Mean Std. Dev. Min Max First Difference Prayer 628 -0.99 7.94 -21 21 First Difference Attendance 628 0.01 1.60 -6.5 6.42

Table 3.14: Descriptive Statistics for Unrestricted and Weighted Pooled Cross Section Dataset Variable Obs Mean Std. Dev. Min Max Dependent Variables Frequency of Prayer Per Week 6481 8.11 8.16 0 21 Frequency of Attendance Per Month 13892 1.88 2.16 0 6.5 Primary Variables Labor Income 13233 26.21 35.69 0 258.58 (Thousands of 2004 dollars) Wages 12284 1.58 5.01 0 191.31 (Tens of 2004 dollars) Non-labor Income 12271 30.99 36.77 0 182.57 (Thousands of 2004 dollars) Spouse Education 14014 7.56 7.06 0 20 Religious Upbringing Variables Raised Catholic 14049 0.31 0.46 0 1 Raised Protestant 14049 0.56 0.50 0 1 Raised Jewish 14049 0.02 0.14 0 1 Raised Other 14049 0.04 0.19 0 1 Raised None 14049 0.07 0.25 0 1 Raised Fundamentalist 13589 0.32 0.47 0 1 Raised Moderate 13589 0.43 0.49 0 1 Raised Liberal 13589 0.25 0.43 0 1 Control Variables Employed 14128 0.68 0.46 0 1 Not Employed 14128 0.32 0.46 0 1 Age 14089 44.43 16.65 18 89 Lived in Same State as When Aged 16 14082 0.65 0.48 0 1 Women 14130 0.54 0.50 0 1 Men 14130 0.46 0.50 0 1 Never Married 14127 0.23 0.42 0 1 Married 14127 0.56 0.50 0 1 Widowed 14127 0.06 0.24 0 1 Divorced/Separated 14127 0.14 0.35 0 1 White 14130 0.80 0.40 0 1 Black 14130 0.13 0.34 0 1 Other Race/Ethnicity 14130 0.07 0.26 0 1 Less Than High School 14089 0.14 0.35 0 1 High School Graduate 14089 0.54 0.50 0 1 Associates Degree 14089 0.07 0.26 0 1 Bachelors Degree 14089 0.17 0.37 0 1 Graduate Degree 14089 0.08 0.27 0 1

Table 3.15: Descriptive Statistics for Unrestricted and Weighted Panel Data Variable Obs Inds Mean Std. Dev. Min Max Dependent Variables Frequency of Prayer Per Week 3053 1533 8.77 8.37 0 21 Frequency of Attendance Per Month 3063 1534 2.00 2.20 0 6.5 Primary Variables Labor Income 2819 1507 25.74 50.88 0 476.79 (Thousands of 2004 dollars) Wages 2805 1505 1.63 9.69 0 414.13 (Tens of 2004 dollars) Non-labor Income 2532 1446 37.46 41.87 0 195.57 (Thousands of 2004 dollars) Religious Upbringing Variables Raised Catholic 2732 1366 0.37 0.48 0 1 Raised Protestant 2732 1366 0.52 0.50 0 1 Raised Jewish 2732 1366 0.02 0.15 0 1 Raised Other 2732 1366 0.02 0.15 0 1 Raised None 2732 1366 0.06 0.23 0 1 Raised Fundamentalist 2454 1227 0.30 0.46 0 1 Raised Moderate 2454 1227 0.49 0.50 0 1 Raised Liberal 2454 1227 0.21 0.41 0 1 Control Variables Employed 3070 1536 0.62 0.48 0 1 Not Employed 3070 1536 0.38 0.48 0 1 Age 3045 1533 45.49 16.64 18 89 Lived in Same State as When Aged 16 3063 1536 0.64 0.48 0 1 Women 3044 1522 0.56 0.50 0 1 Men 3044 1522 0.44 0.50 0 1 Never Married 3070 1536 0.23 0.42 0 1 Married 3070 1536 0.62 0.49 0 1 Widowed 3070 1536 0.04 0.20 0 1 Divorced/Separated 3070 1536 0.11 0.32 0 1 White 2876 1438 0.78 0.41 0 1 Black 2876 1438 0.12 0.33 0 1 Other Race/Ethnicity 2876 1438 0.09 0.29 0 1 Less Than High School 3072 1536 0.13 0.33 0 1 High School Graduate 3072 1536 0.51 0.50 0 1 Associates Degree 3072 1536 0.08 0.27 0 1 Bachelors Degree 3072 1536 0.19 0.39 0 1 Graduate Degree 3072 1536 0.10 0.30 0 1

CHAPTER 4

RELIGION AND LABOR: AN EXAMINATION OF RELIGIOUS SERVICE ATTENDANCE,

UNEMPLOYMENT, AND LABOR FORCE STATUS USING COUNT DATA METHODS

4.1 INTRODUCTION

The Great Recession has sparked a renewed interest in the relationship between unemployment and religion.1 Specifically, the correlation between unemployment and religious service attendance has come to the forefront. Existing empirical evidence on the relationship between unemployment and religiosity is scarce. Preliminary analysis of monthly national data on the U.S. unemployment rate and U.S. weekly religious service attendance, such as in Figure

4.1, suggests that there may be no link between unemployment and that dimension of religiosity.

To provide unique evidence on this topic, I conduct the first count data estimates of the relationship between frequency of religious service attendance and unemployment and the relationship between frequency of religious service attendance and labor force status. This study is also the first to consider the relationship between the duration of unemployment and the frequency of religious service attendance. I also examine the relationship between the length of time spent out of the labor force in relation to the frequency of religious service attendance.

More broadly, I seek to add to a growing body of literature that explores the relationship between religious activities and labor market outcomes. Additionally, this study incorporates two panel datasets to evaluate younger and older working age individuals. Individuals in different life stages may have disparate relationships between the frequency of religious service attendance

1 See recent press articles such as Miller (2010), Pew Forum (2009), and Economist (2009).

and unemployment and between the frequency of religious service attendance and labor force

status.

Research on the economics of happiness suggests that unemployment negatively affects

life satisfaction and that this effect on life satisfaction varies by the duration of unemployment.

Therefore, I evaluate whether being unemployed and the duration of unemployment are

correlated with frequency of religious service attendance through their effect on life satisfaction.

In line with other research, I expect that frequency of religious service attendance is negatively

correlated with unemployment for men while uncorrelated with unemployment for women. I also

predict that individuals under age 50 will have a larger negative correlation with frequency of

religious service from unemployment in comparison to individuals between ages 50 and 65. I

make this argument on the basis that there is a U-shaped relationship between life satisfaction and age. I investigate whether these hypotheses hold when considering the duration of unemployment. Meanwhile, I also argue that being out of the labor force and longer durations out of the labor force are positively associated with the frequency of religious service attendance.

Research suggests that women are both more likely to be out of the labor force and attend religious services more frequently than men. It is also apparent that the relationship between labor force status and frequency of religious service attendance varies by life stage. I specifically contend that being out of the labor force is negatively correlated with frequency of religious service attendance for individuals between ages 50 and 65 as a result of the increasing incidence of serious health problems with age. For individuals under age 50, I argue that being a student is a significant mitigating factor in the labor force status and frequency of religious attendance relationship. Students are both more likely to be out of the labor force and attend religious

services more frequently than non-students. I explore whether these predictions are accurate

when controlling for time spent out of the labor force.

The empirical results for men and women under age 50 in the National Longitudinal

Survey of Youth 1979 cohort dataset using poisson fixed effects estimation show that there is no significant relationship between unemployment and the frequency of religious service attendance. The amount of time spent unemployed has no additional implications. Men under age 50 and out of the labor force at any point in the previous calendar year are predicted to attend religious services less than men who are under age 50 and never out of labor force in the previous calendar year. The magnitude fluctuates from 14 percent when not controlling for student status to 16 percent when controlling for student status. Each additional month out of the labor force, meanwhile, is positively correlated with a 2 percent increase in the frequency of religious service attendance for men under age 50 when not controlling for student status. This correlation is insignificant when controlling for student status. There are no significant relationships between labor force status and the frequency of religious service attendance or time spent out of the labor force and the frequency of religious service attendance for women under age 50.

Pooled negative binomial results for individuals between ages 50 and 65 in the Health and Retirement Study dataset reveal that there is no significant correlation between unemployment and the frequency of religious service attendance or the length of unemployment and the frequency of religious service attendance. For labor force status, men between ages 50 and 65 who are out of the labor force at the of interview attend religious services less often relative to men between ages 50 and 65 in the labor force at time of interview. The size of the correlation varies from 11 percent when controlling for overall health to 17 percent when not

controlling for health. Women between ages 50 and 65 out of the labor force at time of interview, meanwhile, attend religious services 8 percent less often than women between ages 50 and 65 who are not out of the labor force when not controlling for overall health. A negative statistically insignificant relationship is present when controlling for overall health.

This paper is split into 5 sections. Section 4.2 discusses the theoretical framework.

Section 4.3 reviews the dataset, the measures utilized, and covers the econometric estimation.

Section 4.4 presents the results. Section 4.5 contains conclusions and recommendations for further research.

4.2 THEORY

4.2.1 UNEMPLOYMENT AND THE ECONOMICS OF HAPPINESS

Theoretical arguments of the relationship between the frequency of religious service attendance and unemployment duration begin with a discussion of the relationship between unemployment and happiness. A number of stylized facts emerge from previous research.

4.2.1.1 BEING UNEMPLOYED MAKES INDIVIDUALS WORSE OFF

Carroll (2007) uses the Household, Income, and Labour Dynamics in Australia 2001-

2003 surveys to find that the unemployed report lower life satisfaction than the employed. Grun et al. (2010), similarly, employ data from the 1990-2006 German Socio-Economic Panel and report that having any job is better than having no job when it comes to life satisfaction.

Blanchflower and Oswald (2004) utilize data from the 1972-1998 General Social Survey and the

1972-1998 Eurobarometer Surveys to show that unemployment negatively affects life satisfaction in the U.S. and Britain. Clark et al. (2008) use German Socio-Economic Panel data from 1984-2003 to evaluate whether life satisfaction adapts to life events. Their evidence reveals

that unemployment negatively affects life satisfaction, with men adapting less than women. In

other words, the negative effect of unemployment on life satisfaction is larger for men than for

women.

Mental health is also negatively affected by unemployment. Murphy and Athenasou

(1999) review sixteen longitudinal studies on mental health and unemployment. They contend

that mental health is negatively affected by unemployment. Flatau et al. (2000) use data from the

1995 National Health Survey and 1997 National Survey of Mental Health and Wellbeing of

Adults. The authors report that unemployed people exhibit poorer mental health and well-being

outcomes than do the full-time employed.

4.2.1.2 THE IMPACT OF UNEMPLOYMENT VARIES BY DURATION

Research explores the relationship between the duration of unemployment and life

satisfaction. With data from the 1996-1999 British Household Panel Survey (BHPS), 1984-1998

German Socio-Economic Panel (GSOEP), and the 1994-1997 European Community Household

Panel (ECHP), Clark (2006) shows that the negative effect of unemployment on life satisfaction is worst at the beginning but attenuates over time as individuals adapt to being unemployed.

Clark (2006) notes that in the ECHP data the negative effect of unemployment worsens as unemployment duration increases. However, panel estimation with all three datasets reveals a strong, precisely estimated negative effect of unemployment on life satisfaction that is mostly independent of the duration of unemployment. Clark and Oswald (1994) use data from the 1991

BHPS to discover that being unemployed reduces utility. Yet, people who have been unemployed a long time show less distress than those who are unemployed short-term. This result coincides with Clark (2003), where individuals in the BHPS who experience the greatest

decreases in well-being upon becoming unemployed, as measured by mental stress scores, are also the fastest to find new jobs.

Research establishes that unemployment makes individuals worse off, with the size of the negative effect varying with the duration of unemployment. I investigate whether the negative effect of unemployment on life satisfaction is associated with alterations in religious service attendance. I will also test whether the duration of unemployment is correlated with religious service attendance through its effect on life satisfaction. I develop hypotheses for the expected relationship by gender in Section 4.2.1.4.

4.2.1.3 RESEARCH ON THE EFFECT OF HAPPINESS ON RELIGIOUS BEHAVIOR IS

SPARSE

Research on the effect of life satisfaction on religious behavior is sparse. Childs (2010), for example, challenges the prevailing wisdom that more frequent religious attendance increases self-reported happiness. With data from the 1987 and 1988 National Survey of Families and

Households, she uncovers evidence suggesting that happiness does not affect religious attendance. Ferraro and Kelley-Moore (2000), meanwhile, with 1986 and 1988 data from the

Americans’ Changing Lives Survey, find that chronic, non-serious health conditions, which can decrease life satisfaction, lead to more frequent seeking of religious comfort and support.2 In essence, negative changes to life satisfaction, such as unemployment and its duration, may impact religious activity.

2 The survey question on seeking of religious comfort and support is ambiguous on what religious comfort and support means. The question is: “When you have problems or difficulties in your work, family or personal life, how often do you seek spiritual comfort and support”—almost aways (5), often (4), sometimes (3), rarely (2), or never (1).

4.2.1.4 MEN AND WOMEN MAY RESPOND DIFFERENTLY TO UNEMPLOYMENT AND

ITS DURATION

Becker and Hofmeister (2001) use a random sample from upstate New York to show that

men’s religious involvement is associated with marriage, children, and full-time employment.

Women’s religious involvement, in contrast, is related to the presence of school-aged children at home and whether a particular religion coincides with values and lifestyle. It does not depend on full-time employment. In other words, being employed full-time is important to men than to women for religious involvement. Devaus (1984) uses 1972 to 1980 General Social Survey data to find that employment and religious service attendance are not correlated for women. There is also evidence, such as in Blanchflower and Oswald (2004) using Eurobarometer data, showing that life satisfaction decreases more for men than women as a result of unemployment. Men also historically exhibit lower levels of religious behavior than women.3 The lower level of

commitment to religion men have may alter their religious participation disproportionately

because of the unhappiness resulting from unemployment. Based on previous work, I expect a

negative and significant correlation between unemployment and religious service attendance for

men.4 I also anticipate no relationship between unemployment and religious service attendance for women. I test these two hypotheses empirically in Section 4.4.

3 This pattern is well documented and largely unquestioned. See Miller and Hoffman (1995), Thompson (1991), and Collett and Lizardo (2009) for a few of many examples confirming this phenomenon. Stark (2002) explores the physiological reasons normally debated for the difference in religious activity between men and women. 4 Another group that may exhibit lower levels of religious behavior is what Putnam and Campbell (2010) call the working class. The authors claim that the working class is less likely to attend church in the past thirty years because it has been disproportionately hit with unemployment. Because the authors do not define what constitutes the working class, I am unable to test their hypothesis.

4.2.1.5 THE CAUSAL DIRECTION OF UNEMPLOYMENT AFFECTING HAPPINESS IS

MORE PRONOUNCED THAN THE REVERSE DIRECTION

Several studies address the possibility of reverse causality in which greater unhappiness

leads to longer and more frequent spells of unemployment. Using the first six waves of the

German Socio-Economic panel, Winkelmann and Winkelmann (1998) uncover that unemployed individuals are, to some extent, dissatisfied before becoming unemployed. The effect, however, is small compared to the drop in satisfaction from becoming unemployed. Marks and Fleming

(1999) use Australian Youth in Transition Panel Data to show that unemployment is associated with lower levels of well-being, and that lower levels of well-being are correlated with

unemployment. Similar to Winklemann and Winklemann (1988), the effect of unemployment on

life satisfaction is more pronounced than the effect of life satisfaction on unemployment.

4.2.1.6 LIFE SATISFACTION IS U-SHAPED IN AGE

A number of studies substantiate a U-shaped relationship between life satisfaction and age. Fritjers and Beatton (2008), for example, use German Socio-Economic Panel Data to reveal this U-shaped relationship. Blanchflower and Oswald (2007) use data from the General Social

Survey, Eurobarometer, and World Values Survey to create an 800,000 respondent dataset covering several countries and estimate a U-shaped relationship between life satisfaction and age. Others, such as Clark (2006), Powdthavee (2005), Senik (2004), and Di Tella et al. (2001), use various datasets to reach the same conclusion. Across all of these studies, the nature of the relationship between life satisfaction and age is such that life satisfaction generally declines from ages 18 to 50 and rises thereafter. Given this finding, I utilize two distinct datasets in this study to examine individuals under age 50 and between ages 50 and 65. Because life satisfaction declines before age 50, I argue that unemployment should negatively affect younger individuals

proportionately more than older individuals at or over age 50 who are experiencing increases in

life satisfaction. This disproportionate negative effect on life satisfaction may impact the

religious service attendance of individuals under age 50 while having no effect on individuals

between ages 50 and 65.

4.2.2 LABOR FORCE STATUS AND TIME ALLOCATION

The relationship between time spent out of the labor force and religious service

attendance is largely a matter of time allocation. Initiated by Becker (1965), economists have

widely studied how households and individuals allocate their time. While the choice to be out of

the labor force is multidimensional, individuals who are out of the labor force generally have

more time to pursue religious activity. Therefore, I test the hypothesis that both being out of the

labor force and increased time out of the labor force raise the frequency of religious service

attendance.

I expect gender to play a significant role in the relationship between labor-force nonparticipation and frequency of religious service attendance. Between 1967 and 2003, women’s labor force participation rates increased substantially but at a decreasing rate, according to Brusentsev (2006). Recent Bureau of Labor Statistics data reported by Toosi (2010) show that labor force participation rates for men remain higher than for women and are projected to stay that way for at least the next eight years. In 2008, specifically, labor force participation rates were 73.0 percent for men and 59.5 percent for women. From these data and projection, I conclude that women are more likely to be out of the labor force in the samples considered in this study.

Research on the correlation between being out of the labor force and religious service attendance for women is divided. DeVaus and McAllister (1987), for example, use data from the

1983 Australian Values Survey to examine whether social structure patterns of women explain the fact that women are more religious than men. They conclude that the lower labor force participation rate of females is an important factor determining the difference between men and women’s church attendance. They also highlight the finding that women working full-time are less religious than full-time housewives. Ulbrich and Wallace (1984) conduct a study with 1980

National Opinion Research Data on women who either work full time or do not work at all. The authors report a statistically significant difference in religious attendance between full-time employed and non-working women, with non-working women attending religious services more frequently. The difference disappears, however, when religious attendance is adjusted by age, self-identified religious intensity, having a spouse of the same denomination, and labor force status.

I also expect differences in health to impact the relationship between labor-force nonparticipation and frequency of religious services attendance. Specifically, individuals between ages 50 and 65 tend to more frequently develop physical limitations that make it more difficult for them to work and attend religious services. Meanwhile, individuals under age 50 usually have less physical limitations that do not affect their ability to work or attend religious services. Kelley-Moore and Ferraro (2001) use 1986 and 1989 longitudinal data on persons aged

60 or older from the Americans’ Changing Lives Survey to establish that physical limitations constitute a barrier mechanism that inhibits regular or frequent activity within a religious group.

Research by Bound et al. (1998) using Health and Retirement Study data indicates that declines in health help explain labor force withdrawal by older workers. This finding is substantiated by a survey of literature on health and labor market status discussed in Currie and Madrian (1999).

Overall health and frequency of religious service attendance, meanwhile, are positively linked

according to Soydemir et al. (2004), Deaton (2009), Musick et al., and Lehrer (2004), among

others. While the direction of the relationship has only been studied as frequency of religious

service attendance impacting health, a positive association does exist. Therefore, I expect that

being out of the labor force will be negatively correlated with frequency of religious service

attendance of individuals between ages 50 and 65 as a result of the increasing incidence of

serious health problems.

For individuals under age 50, I expect the positive relationship between being out of the

labor force and religious service attendance to be partly explained by student status. Moreover, I

contend that secondary and postsecondary students will attend religious services more frequently

than non-students. For secondary students, research by Glanville et al. (2008) using National

Longitudinal Study of Adolescent Health shows that religious service attendance is negatively

associated with dropping out of school. Muller and Ellison (2001) also demonstrate that religious

involvement is positively correlated with high school graduation. In other words, students

attending secondary schooling are predicted to attend religious services more frequently than

secondary school dropouts. Secondary students are also predicted to attend religious services

more frequently than all secondary graduates. With data from the 1993-2002 Monitoring the

Future Surveys, Presser and Chaves (2007) report that the percentage of individuals attending religious services once per week declines as individuals transition from their senior year of high school to 9-10 years post-graduation. High school seniors, in essence, are predicted to attend more frequently than young post-graduation adults. Considering postsecondary students,

Sacerdote and Glaeser (2001) use 1972-2008 General Social Survey data to find that religious service attendance rises sharply with education across individuals. Postsecondary students, therefore, who are attaining a higher level of education, should attend religious services more

frequently than individuals who are not seeking additional education. Smith (2009), additionally,

uses 2007-2008 National Survey of Youth and Religion data on emerging adults age 18-23 to

find a statistically significant difference between the percent of devoted religious emerging

adults and the percent of disengaged religious emerging adults who have completed some

college or more. There appears to be a positive correlation between religiosity and postsecondary

education, according to Smith (2009). To address the relationship between being a student and

religious service attendance in addition to its possible impact on the relationship between labor

force nonparticipation and religious service attendance, I control for being a secondary student

and being a postsecondary student as part of the empirical estimation in Section 4.3.

4.2.2.1 POTENTIAL ENDOGENEITY, OMITTED VARIABLES, AND POSSIBLE

INSTRUMENTS

When estimating the relationship between labor force status and the frequency of

religious service attendance for both genders, endogeneity concerns arise from specific reasons

why individuals are out of the labor force. Some of the reasons for being out of the labor force

are by individual choice and may be accounted for by the inclusion of control variables while

others cannot. Some reasons for dropping out of the labor force include health deterioration or handicaps, children, additional education, marriage, retirement, racial and ethnic patterns, and

high non-wage income.5 Evidence exists that health deterioration or handicaps, children,

additional education, marriage, retirement, racial and ethnic patterns, and non-wage income are

5 Other reasons for dropping out of the labor force such as moving to a new state exist. Studies such as Smith-Lovin and Tickmayer (1979), Pencavel (1986), Killingsworth and Heckman (1986), Currie and Madrian (1999), Altonji and Blank (1999), and Clark and Summers (1979) substantiate the specific reasons I mention for dropping out of the labor force with the exception of high non-wage income and retirement. Retired individuals by definition are out of the labor force while I argue that high non-wage income earners, on average, are more likely to drop out of the labor force than individuals without high non-wage incomes.

also correlated with the frequency of religious service attendance.6 To mitigate possible

endogeneity problems, I control for overall health, number of children, education, student status,

marital status, race, and ethnicity.7 This study is designed to only consider working-age

individuals between ages 16 and 65, therefore, retirement is not a factor for consideration.

Further consideration will be needed in subsequent work to address other possible sources of

endogeneity such as discouraged workers and non-wage income.8

Pregnancy and fertility constitute an omitted variable that may impact the labor force

status and frequency of religious service attendance relationship for women. Killingsworth and

Heckman (1986) find a negative relationship between birth rates of women and labor force

participation while Smith-Lovin and Tickamyer (1979) explore a negative relationship between

labor force participation and fertility. In essence, any estimated relationship of labor force status

and the frequency of religious service attendance for women may be accounting for fertility and

pregnancy decisions rather than an exogenous relationship. While I do not directly address

pregnancy and fertility, I control for the number of children a respondent has in this study.9

As a significant omitted predictor of frequency of religious service attendance, religious

human capital is a determinant to consider when estimating the relationship between the

frequency of religious service attendance and labor force status and between the frequency of

religious service attendance and unemployment. Iannaccone (2006) defines religious human

6 Studies such as Brown (2009), Sacerdote and Glaeser (2001), Lehrer (2004), Becker and Hofmeister (2001), Johnson et al. (1991), Musick et al. (2004), and Johnson et al. (1991) provide evidence substantiating this claim with the exception of retirement. Because individuals tend to retire at a later age and age and the frequency of religious service attendance are positively correlated as in Iannaccone (1998), I contend that there is a positive relationship between retirement and the frequency of religious service attendance. 7 I discuss these controls further in Section 4.3.1. 8 Specific reasons are available in the National Longitudinal Survey of Youth dataset I use for years 1979 and 1982 that may provide information on discouraged workers. In the case of non-wage income, enough information exists for a non-wage income measure to be derived in the two datasets I use within this study. 9 Data is available on the number of pregnancies in the National Longitudinal Survey of Youth dataset but not the Health and Retirement Study. Future work for women should include controlling for being pregnant.

capital as the “…accumulated stock of skills, sensitivities, and relationships that alter a person’s

(real or perceived) benefits from subsequent religious activity” (pp. 25). Empirical examples of

religious human capital, as Iannaccone (1990) denotes, include parents’ frequency of religious service attendance, spouse with the same religion, multi-item standardized scale of strength of belief, standardized scale of religious experiences, religious instruction scale score, and the fraction of neighbors sharing the same religion when growing up. As the definition and examples show, religious human capital is dynamic and builds over time. While I do not control for religious human capital in this study, future research should utilize it.10

To provide better exogenous estimates of the relationship between labor force status and

the frequency of religious service attendance and between unemployment and the frequency of

religious service attendance, future research should include important omitted variables and, if necessary, use instrumental variables for labor force status and unemployment. Examples of instruments that may account for exogenous labor force status and unemployment include cost of living changes, tax law changes making it more or less favorable to work, changes to unemployment or labor laws, size of birth cohort, size of graduation cohort when education is completed, diversity of the local workforce, fertility, fecundity, pregnancy, and number of siblings.11

10 For the two datasets in this study, the National Longitudinal Survey of Youth reports data on the religious identity of the spouse and religion in which the respondent was raised. The Health and Retirement Study, meanwhile, contains data on religious identity of the spouse and the importance of religion to the respondent. These data may be used in future work to generate measures for religious human capital. 11 The examples of instruments are my own except for fertility, fecundity, pregnancy, and number of siblings which come from Smith-Lovin and Tickamyer (1979).

4.3 DATA AND ECONOMETRIC MODEL

4.3.1 NATIONAL LONGITUDINAL SURVEY OF YOUTH 1979

To evaluate working-age individuals under age 50, I use data from the National

Longitudinal Survey of Youth 1979 (NLSY79), a sample of 12,686 young men and women who were 14-22 years old when they were first surveyed in 1979. The 1979, 1982, and 2000 waves are the only ones utilized because questions on religious affiliation and attendance are only asked in these years. I also remove the oversampled subsamples of Blacks, Hispanics, economically disadvantaged Whites, and military personnel, which cuts the initial number of young men and women to 6,111 in the nationally representative cross-sectional sample.12 The initial panel dataset contains 18,333 observations. After removing observations where an individual is under age 16, has a missing or zero sample weight, has missing data, or has non-response data, the final dataset contains 10,511 observations for 5,275 individuals.13 Comparing the descriptive statistics of the final dataset in Table 4.2 and the full range of data available for respondents age 16 and over in Table 4.1, I find that the sample is nationally representative.14

The dependent variable in this study is the annual frequency of religious service attendance. In this dataset it is originally a categorical variable broken into six categories. I convert these categories so that frequency of religious service attendance is measured in number of services attended per year.15 The key explanatory variables in this paper as measured in the

NLSY79 are being unemployed at any point in the past calendar year, the number of months unemployed in the past calendar year, being out of the labor force at any point in the past

12 For more details see Appendix B.1.1. By nationally representative I mean accurately reflecting the structure of the entire United States population fitting the age range of individuals in the dataset. 13For more details see Appendix B.1.2. The panel is left unbalanced because balancing would result in an additional loss of 6,272 observations. 14 See Tables 1 and 2. 15 For the conversion, I set “not at all” equal to 0, “infrequently” equal to 6, “once per month” equal to 12, “2-3 times per month” equal to 30, “once per week” equal to 52, and “greater than once per week” equal to 78. Varying this conversion where appropriate does not significantly impact the results of this study.

calendar year, and the number of months out of the labor force in the past calendar year.16 As the only religious control, I control for current religious identity which is sorted into five groups—

Protestant, Catholic, Jewish, Other, and None.17 I control for religious identity to see if differential patterns of the frequency of religious service attendance exist between religious groups.

Other control variables include log real income in 2008 dollars, race, ethnicity, marital status, age, highest grade completed, number of children, region of residence, and type of metropolitan statistical area where the residence is located.1819 I include income because it is considered a significant determinant of the frequency of religious service attendance by Brown

(2009), Branas-Garza and Neumann (2004), and Lipford and Tollison (2003) among others.

Ethnicity and race are controlled for because Johnson et al. (1991), Taylor et al. (1996), and

Antunes and Gaitz (1975), among others, reveal differences in religious service attendance by race and ethnicity where minorities attend more frequently than whites. Neuman (1986), Brown

(2009), and Musick et al. (2004) find that being married is a significant determinant of the frequency of religious service attendance. Therefore, I control for marital status. I control for age because it is consistently identified as a positive and important determinant of the frequency of religious service attendance in studies such as Iannaccone (1998), Azzi and Ehrenberg (1975), and Branas-Garza and Neumann (2004). I control for highest grade completed because level of education is a positive predictor of the frequency of religious service attendance, according to

16 Data on duration of unemployment and time spent out of the labor force is originally in number of weeks in the past calendar year. I convert the measure to number of months in the past calendar year for convenience of comparison to coefficients obtained using the Health and Retirement Study. 17 See Appendix B.1.3 for a discussion of why I do not also control for childhood religious upbringing. 18 Data on state, county, and metropolitan statistical area of residence exist but are restricted. They may be obtained by filing a Geocode Program Information and Application form. Details are available at ftp://ftp.bls.gov/pub/special.requests/nls/geocodea.zip. I convert income to real 2008 dollars to match real income data I obtain from the Health and Retirement Study. 19 I replace Income values of “0” with “1” to use the log of income.

Sacerdote and Glaeser (2001). Ulbrich and Wallace (1984) present evidence that the number of preteen children is positively correlated with the frequency of religious service attendance. To further explore the relationship between children and an adult’s religiosity, I include the number of children as a control. Region of residence and type of metropolitan statistical area are controlled for because frequency of religious service attendance may differ based on location of a respondent and the type of location. I also control for student status to determine if being a secondary or postsecondary student matters when estimating the relationship between frequency of religious service attendance and labor force status.

Table 4.2 provides weighted descriptive statistics for the restricted dataset. Average attendance at religious services for the entire weighted sample is 26.90 times per year, which is approximately once every two weeks. The standard deviation is 27.42 which implies a variance of 751.86. Because the variance of the dependent variable is significantly greater than the mean, there is evidence of overdispersion.20

Approximately 28 percent of the weighted sample is unemployed at some point in the past calendar year. The average number of months unemployed in a calendar year is 0.72. The percent of respondents out of the labor force at any point in the past calendar year is 47. Average number of months out of the labor force in a calendar year is 3.07. Averages of these measures in

1979 and 1982 are relatively similar, with average attendance at religious services lower than the entire sample and averages for the other measures higher than the entire sample. 21 In 2000, the

20 According to Cameron and Trivedi (2010), overdispersion refers to when the conditional variance of the dependent variable is greater than the conditional mean. Because the conditional mean and variance do not usually vary substantially from the unconditional mean and variance, I evaluate the unconditional mean and unconditional variance of the dependent variable for evidence of overdispersion as in Cameron and Trivedi (2010). Overdispersion is a violation of the equidispersion property of the Poisson distribution where the conditional variance and conditional mean of the dependent variable are assumed to be equal. To address overdispersion, I utilize pooled negative binomial and fixed effects poisson estimation with clustered-robust standard errors. I discuss this further in Section 4.3.3 and Appendix B.3.1. 21 See Tables 4.13-4.15 for NLSY79 descriptive statistics for the restricted dataset for each individual year.

opposite pattern is observed. Average attendance at religious services is higher than the entire

sample while the other measures exhibit lower averages than the entire sample.22

Approximately 32 percent of the weighted sample is students, which decomposes into 15

percent secondary students and 17 percent postsecondary students. Of the 17 percent that are

postsecondary students, 13 percent are full-time and 4 percent are part-time. The percent that are students are higher in 1979 and 1982 than in 2000.

The current religious identity variables for the entire weighted sample in Table 4.2 show that 55 percent is Protestant, 29 percent is Catholic, 1 percent is Jewish, 4 percent is other, and

11 percent has no religious identity. These measures are fairly constant in each of the three years in the sample.

Other control variables reveal that 13 percent of the entire weighted sample is black and 6 percent are Hispanic. The gender breakdown for the entire weighted sample is 49 percent men and 51 percent women. The racial, ethnic, and gender composition of the dataset is fairly constant over time. For marital status in the weighted sample, 56 percent have never been married, 35 percent are married, 9 percent are divorced or separated, and less than 1 percent are widowed. In 1979 and 1982, a larger part of the weighted sample has never been married while a smaller part of the sample are married, divorced or separated, and widowed in comparison to the entire weighted sample. The opposite trend occurs in 2000 by construction of the dataset.

Average age in the dataset is 26.63, average highest year of school completed is 12.34 years, and average number of children is 0.63. The averages of these three measures move from smaller to larger from 1979 to 1982 to 2000 by construction of the dataset.

22 See Appendix B.1.4 for a discussion concerning the substantially higher average attendance at religious services in 2000 in comparison to 1979 and 1982.

4.3.2 HEALTH AND RETIREMENT STUDY

To analyze working-age individuals between ages 50 and 65, I use the Health and

Retirement Study (HRS), a national longitudinal study of individuals age 50 and over that oversamples Blacks, Hispanics, and residents of Florida.23 I specifically use the birth-year cohorts War Babies (WB), who are born between 1942-1947, and Early Baby Boomers (EBB), who are born between 1948-1953.24 I only use years 2004, 2006, and 2008 because they are the

only years that frequency of religious service attendance is available for each cohort by birth

year.25 I begin with a dataset containing the WBB and EBB birth year cohorts for years 2004,

2006, and 2008 containing 21,249 observations for 7,038 individuals. After removing

observations where an individual is over age 65, has a missing or zero respondent level weight,

missing data, unknown data, don’t know data, the final dataset contains 16,260 observations for

5,993 individuals.26 Comparing descriptive statistics of the final weighted dataset in Table 4.4 to

the full range of weighted data available for respondents age 65 and younger in Table 4.3, I

conclude that the sample is nationally representative.

As with the NLSY79, the dependent variable is the annual frequency of religious service

attendance. For the HRS it is originally a categorical variable broken into five categories. I

convert these categories so that frequency of religious service attendance is measured in number

of services attended per year.27 The key explanatory variables in this paper as measured by the

23 See Appendix B.2.1 for details on the data source. 24 See Appendix B.2.2 for details on the Health and Retirement study and reasons why I use the War Babies and Early Baby Boomers cohorts. 25 See Appendix B.2.3 for more details. The details include a discussion of the sample weights I use which are designed for the sample to match the Current Population Survey, a nationally representative sample of the entire United States population overseen by the U.S. Census Bureau. By nationally representative I mean accurately reflecting the structure of the entire United States population fitting the age range of individuals in the dataset. 26 See Appendix B.2.4 for details. 27 For the conversion, I set “not at all” equal to 0, “1 or more times a year” equal to 9, “2-3 times a month” equal to 30, “once per week” equal to 52, and “greater than once per week” equal to 78. Varying this conversion where appropriate does not significantly impact the results of this study.

HRS are being unemployed at the time of interview, the number of months unemployed since

last job, being out of the labor force at the time of interview, and the number of months out of the

labor force since last job. 28

Control variables are chosen to mirror the NLSY79 controls.29 I control for current

religious identity. Other control variables include log real income in 2008 dollars, race, ethnicity,

marital status, age, highest year of education, number of children, and census division of

residence.3031 Unlike the NLSY79, I also control for self-reported overall health to determine if

health influences the relationship between labor force status and the frequency of religious

service attendance.32

Table 4.4 provides weighted descriptive statistics for the entire pooled HRS sample. With

the exception of months unemployed since last job and months out of the labor force since last

job, the averages of all measures in the HRS are relatively constant over time. I only report the

entire sample averages for all measures and note the fluctuation in months unemployed since last

job and months out of the labor force since last job over the three years in the dataset.

The average number of months unemployed since last job increases from 2004 to 2006 and decreases from 2006 to 2008. The entire weighted sample average indicates that 3 percent of the sample is unemployed at the time of interview. Average duration of unemployment for the

28Appendix B.2.5 provides details on the out of labor force and unemployed since last job designations. See Appendix B.2.6 for details on the construction of months out of the labor force and months unemployed. 29 Control variables in the HRS dataset mirror those in the NLSY79 dataset. See Section 4.3.1 for a discussion of why I use these control variables. 30 Data on state, county, and metropolitan statistical area of residence exist but are restricted. To obtain the restricted data, the institution a researcher is affiliated with must have an Assurance of Compliance from the Office for Human Research Protections (OHRP) of the Department of Health and Human Services (DHHS). Only researchers with permanent faculty-level appointments at such institutions may receive HRS Restricted Datasets. For more details see http://hrsonline.isr.umich.edu/index.php?p=resappreq. 31 I replace Income values of “0” with “1” to use the log of income. 32 A similar measure of overall health does not exist in the NLSY79 for the years 1979, 1982, and 2000.

entire sample is 0.51 months since last job.33 Average number of months out of the labor force increases from 2004 to 2008. Estimates for the entire sample show that 30 percent of the sample is out of the labor force at the time of interview. The average amount of time spent out of the labor force for the entire samples is equal to 22.29 months.

In the entire weighted sample, average frequency of religious service attendance is 26.53

times per year, an equivalent of about 2 times a month. The standard deviation is 26.94 which

translates into a variance of 725.76. Overdispersion appears to be an issue because the variance

of the dependent variable is significantly greater than the mean.

Current religious identity indicators show that 58 percent of the weighted sample is

Protestant, 27 percent is Catholic, 2 percent is Jewish, 1 percent is other, and 11 percent has no religious identity. Other control variables reveal that 11 percent of the weighted sample is black and 8 percent are Hispanic. The gender breakdown for the entire weighted sample is 48 percent men and 52 percent women. Current marital status for the weighted sample decomposes into 5 percent never married, 69 percent married, 20 percent divorced or separated, and 6 percent widowed. Average age is 57.47, average highest year of education is 13.37 years, and average number of children is 2.74. The average self-reported healthiness level is between very good and good.

4.3.3 ECONOMETRIC MODEL

The empirics of this paper incorporate count data estimation because the dependent variable, frequency of religious service attendance per year, only contains nonnegative integer

33 This number is relatively low since only 3 percent of the sample is unemployed at the time of interview. Therefore, 97 percent of the sample reports a value of “0” for unemployment duration.

values with a nontrivial portion of those values being 0 in the two datasets within this study.34

The equations I estimate utilize pooled negative binomial and poisson fixed effects estimation,

with sample weighting for the pooled estimates and clustered-robust standard errors for both sets

of estimates.35 I cluster by individual to adjust for non-independence of observations within

individuals. I use pooled negative binomial and poisson fixed effects estimation to address

overdispersion.36 The following specification displays the estimating equations:

Attendance = α + βU + σRelig + δX + a + ε (1) it it it it i it Attendanceit = θ + κLit +τReligit +υX it + bi +ηit (2)

Equations (1) and (2) are estimated using both datasets. Attendanceit represents the religious

service attendance of individual i at time t. The primary explanatory variables are being

unemployed, the duration of unemployment, being out of the labor force, and time spent out of

the labor force and are denoted by U it and Lit , respectively.

The vector Religit represents dummies for current religious identity. The four categories I include are Catholic, Jewish, Other, and None. I omit the Protestant category. Control variables including log of income, race, ethnicity, marital status, age, education, and number of children

are contained in vector X it . As an additional control in equation (2), I also include and exclude

student status in estimates for the NLSY79 dataset with three dummies for status as a secondary

student, full-time postsecondary student, and part-time postsecondary student. For the HRS

dataset, I include and exclude overall health as an additional control variable in equation (2) to

evaluate whether it impacts the correlation between labor force participation and the frequency

34 Specifically, 1,895 of 10,511 observations in the final NLSY79 dataset show a value of “0” for frequency of religious service atteandance and 4,227 of 16,260 observations in the final HRS dataset show a value of “0” for frequency of religious service attendance. 35 I use NLSY 1979 cross section weights to adjust for sample design. I use HRS 2004 respondent level weights to adjust the sample to mirror the nationally representative Current Population Survey. 36 See Appendix B.3.1 for a discussion of why I use pooled negative binomial and poisson fixed effects to address overdispersion.

of religious service attendance.37 Type of Metropolitan Statistical Area, region of interview,

census division of interview, and time dummies are also included as controls. The error terms are

decomposed into the time invariant ai andbi , which represent unobserved heterogeneity, and

idiosyncratic ε it andηit .

The two estimation methods I use provide different information on the relationship between frequency of religious service attendance and labor market outcomes. The poisson fixed effects estimator, also known as a within estimator, uses the time variation in the dependent

variable and the explanatory variables for each cross-sectional observation, according to

Wooldridge (2000, p.442). In essence, it provides information on the variation over time of the

frequency of religious service attendance within individuals, while accounting for the

relationship between frequency of religious service attendance and labor market outcomes over

time. It also controls fully for unobserved and observable time invariant determinants. Because

poisson fixed effects is able to control for unobservable time invariant determinants, also known

as unobserved heterogeneity, I prefer it as an estimator in comparison to pooled negative

binomial estimation. I report the results from poisson fixed effects as better than pooled negative binomial results whenever statistically reliable poisson fixed effects estimates are available.

Unobserved heterogeneity could stem from personal background and family background characteristics that are unobserved and may significantly explain variation in the frequency of religious service attendance.

Pooled negative binomial estimation allows for variation in the dependent variable between individuals. As Wooldridge (2000, p.409) describes, using a pooled estimator is

37 I also conduct regressions estimates of equation (3) with overall health included as a determinant. While the coefficient on overall health is statistically significant, I do not report or include it in the tables I present because the relationship between the other variables and the frequency of religious service attendance is unaffected by including overall health as a determinant.

relatively useful provided that the relationship between the dependent variable and some of the

explanatory variables remains constant over time. When I use pooled cross section estimation, I

am evaluating the relationship between frequency of religious service attendance and labor

market outcomes between individuals and assuming that the relationship remains fairly constant

over time. I am also assuming that unobserved heterogeneity is either not a serious issue or is

adequately controlled for by time-invariant determinants. I only report pooled negative binomial

estimates as more reliable than poisson fixed effects estimates when there is a lack of within

variation in the dependent variable that results in poor estimation using poisson fixed effects.

4.4 RESULTS

The empirical results appear in Tables 4.5-4.13. All tables report clustered-robust standard errors that are clustered by individual. I conduct Wald tests to determine if there are differences between coefficient estimates for men and women overall. If I instead estimate both genders together and include a binary measure for gender, then I implicitly assume that all other variables in the estimating equations have the same effect across genders. This assumption is inappropriate given that some of the explanatory variables, such as labor force status, may have different effects by gender on the frequency of religious service attendance. A weakness of an overall Wald test is that it does not reveal the exact sources of the disparity. It only estimates an overall difference in coefficient estimates. Even so, the tests indicate that there is a divergence between men and women. Specifically, the Wald tests show that estimation should be carried out separately for men and women (Chi-Squared statistics ≥ 127.92 and p < 0.01). The distinction is most likely due to differences in religiosity such as in Miller and Hoffman (1995), Thompson

(1991), and Collett and Lizardo (2009) and differences in labor market outcomes such as in

Toosi (2010), Altonji and Blank (1999), Pencavel (1986), and Killingsworth and Heckman

(1986). Key variables are listed at the top of each table and include unemployed, months unemployed, out of the labor force, months out of the labor force, and religious identity dummies. Student status dummies are included and excluded for NLSY79 labor force status estimates and overall health is included and excluded for HRS labor force status estimates.

Table 4.5 reports results using pooled pooled negative binomial estimation for the

NLSY79 dataset and shows that being unemployed is correlated with the frequency of religious service attendance for men under age 50. The coefficient on unemployed, which is statistically significant at the one percent level, indicates that respondents who were unemployed at any point in the past calendar year attend religious services 14 percent less often than respondents who are not unemployed at any point in the past calendar year. For women under age 50, there is no statistically significant relationship between being unemployed and the frequency of religious service attendance. There is no statistically significant relationship between the length of unemployment and frequency of religious service attendance for both genders.

Table 4.6 shows results using pooled negative binomial estimation for labor force status and frequency of religious service attendance for the NLSY79. When not controlling for student status, men and women under age 50 are predicted to attend religious services 2 and 1 percent more, respectively, for each additional month they spend out of the labor force with statistical significance at the five and ten percent levels, respectively. Controlling for student status yields statistically insignificant coefficients for being out of the labor force and time spent out of the labor force.38 The coefficients for secondary and full-time postsecondary student status indicate that secondary and full-time postsecondary students attend religious services more frequently

38 I also obtain statistically insignificant results for out of the labor force and time spent out of the labor force when running the pooled negative binomial estimates for non-students only.

than non-students with statistical significance at the one percent level. Specifically, men under

age 50 who are secondary students attend 43 percent more services per year and women under

age 50 who are secondary students attend 39 percent more services per year compared to non-

students. For full-time postsecondary students, men and women under age 50 attend 40 percent

and 20 percent more services per year compared to non-students, respectively.

Tables 4.7 and 4.8 report poisson fixed effects estimates using the NLSY79 dataset. The

poisson fixed effects estimator requires that there be at least two years of data for each individual

and that frequency of attendance be nonzero in at least one period. For these two reasons, the

number of individuals and the number of observations in Tables 4.7 and 4.8 are lower than in

Tables 4.5 and 4.6. Also, a constant term is not reported for any of the poisson fixed effects

results in this study due to its omission to normalize the unobserved fixed effect term.

Table 4.7 indicates no statistical significance for any of the unemployment measures.

Table 4.8, meanwhile, reveals with ten percent statistical significance that, when not controlling

for student status, men under age 50 who are out of the labor force at any point in the last year

attend religious services 14 percent less often than men under age 50 who are in the labor force

the entire year. Additionally, frequency of religious service attendance is 2 percent higher for

each additional month younger men spend out of the labor force in the past calendar year, with

statistical significance at the ten percent level. Controlling for student status, men who are under

age 50 and out of the labor force in past calendar year attend religious services 16 percent less

than men under age 50 who are in the labor force the entire year with statistical significance at

the five percent threshold. Coefficient estimates for full-time postsecondary and secondary students are statistically significant at the five percent level or lower for men and women under age 50. Men and women under age 50 who are secondary students are predicted to attend

religious services 20 percent and 30 percent more often, respectively, than non-students. Full- time postsecondary students attend religious services 30 percent and 21 percent more often, respectively, than non-students, according to the estimates. 39

Tables 4.9 and 4.10 display pooled negative binomial estimation results using the HRS dataset. None of the unemployment coefficients in Table 4.9 are statistically significant. When not controlling for health in Table 4.10, men and women between ages 50 and 65 who are out of the labor force at the time of interview attend religious services less frequently than men and women between ages 50 and 65 who are in the labor force at the time of interview. Men between ages 50 and 65 attend religious services 17 percent less often and women between ages 50 and

65 attend religious services 8 percent less often than men and women between ages 50 and 65 who are in the labor force at the time of interview, respectively, with statistical significance in both cases at the ten percent threshold or lower. Controlling for health still indicates a negative relationship between being out of the labor force and the frequency of religious service attendance for men and women between ages 50 and 65. The magnitudes of the coefficients are smaller and statistical significance is only obtained for men between ages 50 and 65 at the five percent level. Men between the ages of 50 and 65 who are out of the labor force at the time of interview attend religious services11 percent less often than men between the ages of 50 and 65 who are in the labor force at the time of interview, according to the estimates. The coefficients for health indicate that as health deteriorates, religious service attendance is predicted to be lower for men and women between ages 50 and 65. For example, men and women between ages 50 and 65 with good overall health are predicted to attend religious services 8 percent less per year

39 I also run estimates restricted to non-students for men and women. In each case I obtain statistically insignificant coefficients for being in the labor force and time spent out of the labor force. In the case of men, the coefficient estimates are still the same sign whereas for women the coefficient estimates are positive instead of negative.

than men and women between ages 50 and 65 with excellent overall health, according to the estimates. Statistical significance for the health coefficients is attained at the one percent level.

Tables 4.11 and 4.12 show poisson fixed effects estimates for the HRS dataset. Wald chi- squareds and corresponding p-values indicate a poor goodness-of-fit for all regressions I estimate except for labor force status of women in Table 4.12. None of the coefficients in Table 4.11 are statistically significant. Estimates for women between ages 50 and 65 in Table 4.12, when not controlling for health, indicate that women between ages 50 and 65 who are out of the labor force at the time of interview attend religious services 5 percent more often than women between ages 50 and 65 who are in the labor force at the time of interview. When controlling for health, women between ages 50 and 65 who are out of the labor force at the time of interview attend religious services 6 percent more than women between ages 50 and 65 who are in the labor force at the time of interview. The poor fit results in Tables 4.11 and 4.12 are most likely due to lack of within variation in the dependent variable. Table 4.13 substantiates this claim by presenting the first difference of frequency of religious service attendance. The average indicates that frequency of religious service attendance changes by less than one unit per year within an individual between 2004 and 2006 and between 2006 and 2008.

Because I argue that unobservable heterogeneity matters, I view NLSY79 results in

Tables 4.7 and 4.8 as more reliable than those in Table 4.5 and 4.6 and only make inferences using Tables 4.7 and 4.8 for individuals under age 50. For the HRS estimates, failure of joint significance tests in four of the six regressions run using poisson fixed effects estimation leads me to rely on the pooled negative binomial results in Tables 4.9 and 4.10 for inference.

Overall, the results in Tables 4.7 and 4.8 indicate no statistically significant relationship between being unemployed and the frequency of religious service attendance for men and

100

women under age 50. This result adds to Childs (2010), who highlights that changes in happiness do not affect the frequency of religious service attendance. Results for labor force status do reveal some significant relationship. For men under age 50, being out of the labor force is negatively associated with the frequency of religious service regardless of student status.

Additional nonlabor time, therefore, is allocated away from attending religious services for men under age 50, which is contrary to my hypothesis. I add to existing research by showing that men under age 50 who are out of the labor force are predicted to attend religious services less often.

Combined with the effect of time spent out of the labor force, men under age 50 who are out of the labor force are predicted to have lower frequency of religious service attendance that remains lower over time. This finding may be explained by frequency of religious service attendance being related to work or the active pursuit of work for men under age 50. Student status indicators for men and women under age 50 also show that secondary and full-time postsecondary students attend religious services more frequently than non-students. This result adds to research on education and the frequency of religious service attendance such as

Sacerdote and Glaeser (2001).

Tables 4.11-4.13 show a lack of variation in the frequency of religious service attendance within individuals between ages 50 and 65 over a four year period. Looking at variation in the frequency of religious service attendance between individuals who are ages 50 to 65, Table 4.9 reveals no significant relationship between unemployment and the frequency of religious service attendance. Table 4.10 also examines variation in the frequency of religious service attendance between individuals and shows that health status appears to be a significant correlate of the frequency of religious service in accordance with my hypothesis. Being out of the labor force, meanwhile, is negatively associated with frequency of religious service attendance for men and

101

women between ages 50 and 65. The magnitude of the correlation is reduced for both groups

when controlling for health status. It is feasible that an increased incidence of serious health

problems among individuals who are between ages 50 and 65 reduces both the frequency of

religious service attendance and increases exiting from the labor force.

The distinctions present in the evidence are best explained by gender, life stage, and

measurement differences between the HRS and NLSY79 datasets. It is well documented that

disparities in religious behavior exist by gender.40 Hence, it is not surprising that the magnitude and significance of key variables, such as labor force status, vary in this study by gender. Life stage appears to have an impact on religious attendance through changes in health and student status. Higher religious service attendance appears on average for men under age 50 who are

working or actively pursuing work. Individuals under age 50 are often attending secondary or postsecondary school full time, which is associated with higher frequency of religious service attendance. Individuals between ages 50 and 65 tend to be subjected more frequently to serious health problems and adjust their frequency of religious attendance behavior and labor force status accordingly. Moreover, a number of economic models such as in Ghez and Becker (1975) show that time allocation changes over the life cycle. Measurement dissimilarities may also matter. For example, the NLSY79 measures all individuals who are out of the labor force at any point in a given previous calendar year while the HRS only captures individuals who are currently out of the labor force. The share of individuals out of the labor force is partly higher in the NLSY79 because of this measurement difference. In essence, I obtain more information on individuals’ labor force status history in a survey year in the NLSY79, which could contribute to the results being different between the NLSY79 and HRS.

40 See Miller and Hoffman (1995), Thompson (1991), and Collett and Lizardo (2009) for a few of many examples.

102

For the other measures I include in this study in Tables 4.7-4.10, being part of the other religious group implies between 60 percent and 73 percent lower frequency of religious service attendance for men and women under age 50 and 30 to 32 percent lower frequency of religious service attendance for men between ages 50 and 65 compared to the omitted Protestant group.

Blacks between ages 50 and 65 have between 58 percent and 61 percent higher frequency of religious service attendance than older respondents who are not black or Hispanic. Hispanics between ages 50 and 65 have between 39 percent and 42 percent higher frequency of religious service attendance than respondents who are not black or Hispanic. An additional year of education is associated with frequency of religious service attendance that is 6 percent to 14 percent lower for individuals under age 50, while it is correlated with frequency of religious service attendance that is 3 percent to 5 percent higher for individuals between ages 50 and 65.

Each additional child is associated with 19 to 27 percent less frequent religious service attendance for individuals under age 50 and 3 to 4 percent more frequent religious service attendance for individuals between ages 50 and 65. Marriage is correlated with frequency of religious service attendance that is between 16 percent and 30 percent higher for individuals between ages 50 and 65 relative to never married individuals of the same age group.

4.5 CONCLUSIONS

The empirical results of this study present a variety of findings. Evidence indicates that the frequency of religious service attendance and unemployment are not correlated. The duration of unemployment provides no further evidence. Being out of the labor force is negatively correlated with frequency of religious service attendance for men under age 50 and men and women between ages 50 and 65. Men under age 50 out of the labor force at any point in the past calendar year have frequency of religious service attendance that is 14 percent less than men of

103

the same age group who are in the labor force throughout the past calendar year. The magnitude is 16 percent when controlling for status as a student. Men who are between ages 50 and 65 and out of the labor force at the time of interview have frequency of religious service attendance that is 17 percent less than men in the same age group who are in the labor force at the time of interview when not controlling for health. The magnitude is 11 percent when controlling for health. Women between ages 50 and 65 who are out of the labor force at the time of interview have frequency of religious service attendance that is 8 percent less than older women in the same age group who are out of the labor force at the time of interview when not controlling for health. The relationship is still negative but statistically insignificant when controlling for health.

This study is the first to use count data estimation methods for examining the frequency of religious service attendance, unemployment, and labor force status. It is also the first to explore the relationship between the duration of unemployment and the frequency of religious service attendance and between time spent out of labor force and the frequency of religious service attendance. It also examines being unemployed and being out of the labor force in relation to the frequency of religious service attendance. Changes in frequency of religious service attendance correlated with being out of the labor force are viewed in the context of differences in time allocation for working and nonworking men and women.

Further research should more precisely distinguish between the effects of life stage, measurement, and gender. The present study presents these effects as important but is unable to isolate and quantify all three separately for analysis. Future research should also include obtaining restricted data on geographic location of respondents to employ additional clustering.

Additionally, if software and higher computation power become available, pooled negative binomial estimation with individual dummies should be undertaken to obtain fixed effects

104

negative binomial estimation where fixed effects are accounted for in the traditional sense.

Omitted variables such as religious human capital and fertility should also be reviewed. If necessary, a fixed effects instrumental variables approach using instruments such as recessions and size of birth cohort should be used.

105

Figure 4.1: Percent of Americans Who Attend Worship Services at Least Weekly and the U.S. Unemployment Rate

Sources: Worship attendance data are monthly and come from the Pew Research Center for the People & the Press surveys. Question wording: Aside from weddings and funerals, how often do you attend religious services... more than once a week, once a week, once or twice a month, a few times a year, seldom, or never? Unemployment rate data are monthly from the Bureau of Labor Statistics, Series ID: LNS14000000.

106

Table 4.1: NLSY79 Descriptive Statistics Unrestricted and Weighted Sample Variable Obs Inds Mean Std. Dev. Min Max Time Varying Variables Frequency of Attendance Per Year 14496 5690 26.60 27.30 0 78 Unemployed 13493 5664 0.28 0.45 0 1 (in past calendar year) Months Unemployed 13493 5664 0.74 1.91 0 12 (in past calendar year) Out of Labor Force 13493 5664 0.49 0.50 0 1 (in past calendar year) Months Out of Labor Force 13493 5664 3.27 4.43 0 12 (in past calendar year) Student 14543 5696 0.34 0.47 0 1 Secondary Student 14543 5696 0.17 0.37 0 1 Postsecondary Student 14489 5696 0.18 0.38 0 1 Full-Time Postsecondary Student 14489 5696 0.14 0.34 0 1 Part-Time Postsecondary Student 14489 5696 0.04 0.19 0 1 Protestant 14518 5696 0.55 0.50 0 1 Catholic 14518 5696 0.29 0.45 0 1 Jewish 14518 5696 0.01 0.11 0 1 Other 14518 5696 0.03 0.18 0 1 None 14518 5696 0.12 0.32 0 1 Log Income (thousands of 2008 dollars) 11739 5440 10.62 1.44 0 12.94 Married 14543 5696 0.32 0.47 0 1 Never Married 14543 5696 0.59 0.49 0 1 Divorced/Separated 14543 5696 0.08 0.27 0 1 Widowed 14543 5696 0.00 0.05 0 1 Age 14543 5696 25.84 9.29 16 43 Education (highest grade completed) 14465 5685 12.21 2.25 0 20 Number of Children 14543 5696 0.58 1.02 0 9 Time Invariant Variables Black 14543 5696 0.15 0.35 0 1 Hispanic 14543 5696 0.07 0.25 0 1 Men 14543 5696 0.49 0.50 0 1 Women 14543 5696 0.51 0.50 0 1 Note: The unrestricted dataset is restricted to respondents age 16 and older who have a sample weight that is greater than zero

107

Table 4.2: NLSY79 Descriptive Statistics Restricted and Weighted Sample Variable Obs Inds Mean Std. Dev. Min Max Time Varying Variables Frequency of Attendance Per Year 10511 5275 26.90 27.42 0 78 Unemployed 10511 5275 0.28 0.45 0 1 (in past calendar year) Months Unemployed 10511 5275 0.72 1.89 0 12 (in past calendar year) Out of Labor Force 10511 5275 0.47 0.50 0 1 (in past calendar year) Months Out of Labor Force 10511 5275 3.07 4.33 0 12 (in past calendar year) Student 10511 5275 0.32 0.47 0 1 Secondary Student 10511 5275 0.15 0.36 0 1 Postsecondary Student 10511 5275 0.17 0.38 0 1 Full-Time Postsecondary Student 10511 5275 0.13 0.34 0 1 Part-Time Postsecondary Student 10511 5275 0.04 0.19 0 1 Protestant 10511 5275 0.55 0.50 0 1 Catholic 10511 5275 0.29 0.45 0 1 Jewish 10511 5275 0.01 0.10 0 1 Other 10511 5275 0.04 0.19 0 1 None 10511 5275 0.11 0.32 0 1 Log Income (thousands of 2008 dollars) 10511 5275 10.64 1.43 0 12.94 Married 10511 5275 0.35 0.48 0 1 Never Married 10511 5275 0.56 0.50 0 1 Divorced/Separated 10511 5275 0.09 0.28 0 1 Widowed 10511 5275 0.00 0.05 0 1 Age 10511 5275 26.66 9.45 16 43 Education (highest grade completed) 10511 5275 12.34 2.27 0 20 Number of Children 10511 5275 0.63 1.05 0 9 Time Invariant Variables Black 10511 5275 0.13 0.34 0 1 Hispanic 10511 5275 0.06 0.25 0 1 Men 10511 5275 0.49 0.50 0 1 Women 10511 5275 0.51 0.50 0 1

108

Table 4.3: HRS Descriptive Statistics Unrestricted and Weighted Sample Variable Obs Inds Mean Std. Dev. Min Max Time Varying Variables Frequency of Attendance Per Year 16790 6117 26.38 26.95 0 78 Unemployed 16818 6120 0.03 0.16 0 1 (at time of interview) Months Unemployed 16818 6120 0.58 12.76 0 494 (since last job) Out of Labor Force 16818 6120 0.30 0.46 0 1 (at time of interview) Months Out of Labor Force 16818 6120 22.55 80.30 0 599 (since last job) Health (1=excellent…5=poor) 16806 6118 2.67 1.13 1 5 Log Income (thousands of 2008 dollars) 16818 6120 8.50 4.00 0 15.83 Married 16812 6120 0.69 0.46 0 1 Never Married 16812 6120 0.05 0.22 0 1 Divorced/Separated 16812 6120 0.20 0.40 0 1 Widowed 16812 6120 0.06 0.23 0 1 Age 16818 6120 57.47 3.70 50 65 Number of Children 16485 6061 2.75 1.82 0 17 Time Invariant Variables Protestant 16758 6098 0.58 0.49 0 1 Catholic 16758 6098 0.27 0.44 0 1 Jewish 16758 6098 0.02 0.15 0 1 Other 16758 6098 0.01 0.12 0 1 None 16758 6098 0.11 0.32 0 1 Education (highest year) 16711 6082 13.36 2.88 0 17 Black 16815 6119 0.11 0.31 0 1 Hispanic 16817 6119 0.08 0.27 0 1 Men 16818 6120 0.48 0.50 0 1 Women 16818 6120 0.52 0.50 0 1 Note: The unrestricted dataset is restricted to respondents age 65 and younger who have a sample weight that is greater than zero.

109

Table 4.4: HRS Descriptive Statistics Restricted and Weighted Sample Variable Obs Inds Mean Std. Dev. Min Max Time Varying Variables Frequency of Attendance Per Year 16260 5993 26.53 26.94 0 78 Unemployed 16260 5993 0.03 0.16 0 1 (at time of interview) Months Unemployed 16260 5993 0.51 11.18 0 494 (since last job) Out of Labor Force 16260 5993 0.30 0.46 0 1 (at time of interview) Months Out of Labor Force 16260 5993 22.30 79.37 0 599 (since last job) Health (1=excellent…5=poor) 16260 5993 2.67 1.13 1 5 Log Income (thousands of 2008 dollars) 16260 5993 8.49 4.01 0 15.83 Married 16260 5993 0.69 0.46 0 1 Never Married 16260 5993 0.05 0.22 0 1 Divorced/Separated 16260 5993 0.20 0.40 0 1 Widowed 16260 5993 0.06 0.23 0 1 Age 16260 5993 57.47 3.69 50 65 Number of Children 16260 5993 2.74 1.82 0 17 Time Invariant Variables Protestant 16260 5993 0.58 0.49 0 1 Catholic 16260 5993 0.27 0.44 0 1 Jewish 16260 5993 0.02 0.15 0 1 Other 16260 5993 0.01 0.12 0 1 None 16260 5993 0.11 0.32 0 1 Education (highest year) 16260 5993 13.37 2.88 0 17 Black 16260 5993 0.11 0.31 0 1 Hispanic 16260 5993 0.08 0.28 0 1 Men 16260 5993 0.48 0.50 0 1 Women 16260 5993 0.52 0.50 0 1

110

Table 4.5: NLSY79 Pooled Negative Binomial Results for Frequency of Attendance and Unemployment Men Women Labor Market Variables Unemployed -0.14*** -0.05 (in past calendar year) (0.05) (0.04) Months Unemployed 0.01 -0.00 (in past calendar year) (0.01) (0.01) Religion Variables Catholic 0.04 -0.01 (0.04) (0.03) Jewish -0.18 -0.11 (0.20) (0.17) Other -0.16 -0.16* (0.12) (0.09) None -0.32*** 0.03 (0.06) (0.05) Constant 3.97*** 3.83*** (0.20) (0.17) Control Variables Yes Yes MSA Effects Yes Yes Year Effects Yes Yes Region Effects Yes Yes Observations 5076 5435 Individuals 2581.00 2694.00 Wald Chi-Squared 609.85 217.29 Prob. Chi-Squared 0.00 0.00 Log Pseudolikelihood -1.14e+10 -1.22e+10 Alpha (dispersion parameter) 1.79 1.64 Clustered-robust standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1% Control Variables include: Log Income, Black, Hispanic, Married, Divorced/Separated, Widowed, Age, Education, Number of Children

111

Table 4.6: NLSY79 Pooled Negative Binomial Results for Frequency of Attendance and Labor Force Status Men Men Women Women Labor Market Variables Out of Labor Force -0.04 -0.08 -0.01 -0.03 (in past calendar year) (0.06) (0.06) (0.04) (0.05) Months Out of Labor Force 0.02** 0.00 0.01* 0.00 (in past calendar year) (0.01) (0.01) (0.00) (0.00) Student Variables Secondary Student 0.43*** 0.30*** (0.07) (0.05) Full-Time Postsecondary Student 0.40*** 0.20*** (0.06) (0.05) Part-Time Postsecondary Student 0.15 0.01 (0.09) (0.08) Religion Variables Catholic 0.04 0.04 -0.01 -0.01 (0.04) (0.04) (0.03) (0.03) Jewish -0.17 -0.19 -0.13 -0.13 (0.21) (0.18) (0.17) (0.17) Other -0.18 -0.18 -0.16* -0.17* (0.11) (0.12) (0.09) (0.09) None -0.32*** -0.31*** 0.02 0.03 (0.06) (0.06) (0.05) (0.05) Constant 3.66*** 3.28*** 3.65*** 3.32*** (0.21) (0.23) (0.18) (0.19) Control Variables Yes Yes Yes Yes MSA Effects Yes Yes Yes Yes Year Effects Yes Yes Yes Yes Region Effects Yes Yes Yes Yes Observations 5076 5076 5435 5435 Individuals 2581.00 2581.00 2694.00 2694.00 Wald Chi-Squared 603.53 653.08 224.23 249.45 Prob. Chi-Squared 0.00 0.00 0.00 0.00 Log Pseudolikelihood -1.14e+10 -1.14e+10 -1.22e+10 -1.21e+10 Alpha (dispersion parameter) 1.80 1.78 1.64 1.63 Clustered-robust standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1% Control Variables include: Log Income, Black, Hispanic, Married, Divorced/Separated, Widowed, Age, Education, Number of Children

112

Table 4.7: NLSY79 Poisson Fixed Effects Results for Frequency of Attendance and Unemployment Men Women Labor Market Variables Unemployed -0.10 -0.03 (in past calendar year) (0.07) (0.05) Months Unemployed -0.01 0.01 (in past calendar year) (0.02) (0.01) Religion Variables Catholic 0.00 -0.02 (0.10) (0.10) Jewish -0.45 0.09 (0.66) (0.56) Other -0.72*** -0.60*** (0.18) (0.14) None -0.01 0.26*** (0.08) (0.10) Control Variables Yes Yes MSA Effects Yes Yes Year Effects Yes Yes Region Effects Yes Yes Observations 4208 4648 Individuals 1780 1943 Wald Chi-Squared 502.66 240.80 Prob. Chi-Squared 0.00 0.00 Log Likelihood -36047.86 -43431.56 Clustered-robust standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1% Control Variables include: Log Income, Married, Divorced/Separated, Widowed, Age, Education, Number of Children

113

Table 4.8: NLSY79 Poisson Fixed Effects Results for Frequency of Attendance and Labor Force Status Men Men Women Women Labor Market Variables Out of Labor Force -0.14* -0.16** 0.09 0.08 (in past calendar year) (0.07) (0.07) (0.06) (0.06) Months Out of Labor Force 0.02* 0.01 -0.01 -0.01 (in past calendar year) (0.01) (0.01) (0.01) (0.01) Student Variables Secondary Student 0.20** 0.30*** (0.09) (0.07) Full-Time Postsecondary Student 0.32*** 0.21*** (0.08) (0.07) Part-Time Postsecondary Student 0.18 0.02 (0.12) (0.10) Religion Variables Catholic 0.01 0.03 -0.02 -0.02 (0.10) (0.10) (0.10) (0.10) Jewish -0.49 -0.50 0.07 0.08 (0.67) (0.71) (0.56) (0.55) Other -0.73*** -0.72*** -0.60*** -0.60*** (0.18) (0.18) (0.14) (0.14) None -0.00 0.00 0.25*** 0.27*** (0.08) (0.08) (0.10) (0.10) Control Variables Yes Yes Yes Yes MSA Effects Yes Yes Yes Yes Year Effects Yes Yes Yes Yes Region Effects Yes Yes Yes Yes Observations 4208 4208 4648 4648 Individuals 1780 1780 1943 1943 Wald Chi-Squared 505.80 517.47 245.76 263.22 Prob. Chi-Squared 0.00 0.00 0.00 0.00 Log Likelihood -36041.89 -35824.51 -43410.02 -43178.04 Clustered-robust standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1% Control Variables include: Log Income, Married, Divorced/Separated, Widowed, Age, Education, Number of Children

114

Table 4.9: HRS Pooled Negative Binomial Results for Frequency of Attendance and Unemployment Men Women Labor Market Variables Unemployed (at time of interview) -0.15 0.14 (0.12) (0.09) Months Unemployed (since last job worked) 0.01 -0.00 (0.01) (0.00) Religion Variables Catholic 0.07 -0.06 (0.06) (0.04) Jewish -0.59*** -0.76*** (0.22) (0.15) Other -0.32* -0.17 (0.17) (0.18) None -1.33*** -1.27*** (0.11) (0.12) Constant 2.33*** 1.74*** (0.49) (0.35) Control Variables Yes Yes Year Effects Yes Yes Census Division Effects Yes Yes Observations 6713 9547 Individuals 2509 3484 Wald Chi-Squared 348.01 494.41 Prob. Chi-Squared 0.00 0.00 Log Pseudolikelihood -1.98e+08 -2.39e+08 Alpha (dispersion parameter) 2.59 1.90 Clustered-robust standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1% Control Variables include: Log Income, Black, Hispanic, Married, Divorced/Separated, Widowed, Age, Education, and Number of Children

115

Table 4.10: HRS Pooled Negative Binomial Results for Frequency of Attendance and Labor Force Status Men Men Women Women Labor Market Variables Out of Labor Force -0.17*** -0.11* -0.08* -0.04 (at time of interview) (0.06) (0.06) (0.04) (0.04) Months Out of Labor Force -0.00 -0.00 0.00 0.00 (since last job worked) (0.00) (0.00) (0.00) (0.00) Health -0.08*** -0.08*** (0.02) (0.02) Religion Variables Catholic 0.08 0.08 -0.06 -0.07* (0.06) (0.06) (0.04) (0.04) Jewish -0.59*** -0.62*** -0.77*** -0.78*** (0.22) (0.22) (0.15) (0.15) Other -0.30* -0.31* -0.19 -0.20 (0.17) (0.17) (0.18) (0.17) None -1.33*** -1.32*** -1.27*** -1.26*** (0.11) (0.11) (0.12) (0.12) Constant 2.20*** 2.53*** 1.68*** 2.04*** (0.49) (0.50) (0.35) (0.35) Control Variables Yes Yes Yes Yes Year Effects Yes Yes Yes Yes Census Division Effects Yes Yes Yes Yes Observations 6713 6713 9547 9547 Individuals 2509 2509 3484 3484 Wald Chi-Squared 360.92 380.26 496.73 495.18 Prob. Chi-Squared 0.00 0.00 0.00 0.00 Log Pseudolikelihood -1.98e+08 -1.98e+08 -2.39e+08 -2.38e+08 Alpha (dispersion parameter) 2.58 2.58 1.90 1.89 Clustered-robust standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1% Control Variables include: Log Income, Black, Hispanic, Married, Divorced/Separated, Widowed, Age, Education, and Number of Children

116

Table 4.11: HRS Poisson Fixed Effects Results for Frequency of Attendance and Unemployment Men Women Key Variables Unemployed (at time of interview) 0.07 0.02 (0.07) (0.05) Months Unemployed (since last job) -0.00 -0.00 (0.00) (0.00) Control Variables Yes Yes Year Effects Yes Yes Census Division Effects Yes Yes Observations 5240 8180 Individuals 1840 2851 Wald Chi-Squared 26.52 24.67 Prob. Chi-Squared 0.09 0.13 Log Likelihood -20757.72 -32391.04 Clustered-robust standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1% Control Variables include: Log Income, Married, Divorced/Separated, Widowed, Age, and Number of Children

117

Table 4.12: HRS Poisson Fixed Effects Results for Frequency of Attendance and Labor Force Status Men Men Women Women Key Variables Out of Labor Force -0.00 -0.00 0.05** 0.06** (at time of interview) (0.04) (0.04) (0.02) (0.02) Months Out of Labor Force 0.00 0.00 0.00 0.00 (since last job) (0.00) (0.00) (0.00) (0.00) Health -0.00 -0.01 (0.01) (0.01) Control Variables Yes Yes Yes Yes Year Effects Yes Yes Yes Yes Census Division Effects Yes Yes Yes Yes Observations 5240 5240 8180 8180 Individuals 1840 1840 2851 2851 Wald Chi-Squared 24.73 24.74 32.38 34.50 Prob. Chi-Squared 0.13 0.17 0.02 0.02 Log Likelihood -20760.95 -20760.80 -32369.10 -32361.91 Clustered-robust standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1% Control Variables include: Log Income, Married, Divorced/Separated, Widowed, Age, and Number of Children

Table 4.13: HRS First Difference of Attendance Variable Obs Mean Std. Dev. Min Max First Difference Attendance 10267 -0.41 17.31 -78 78

118

Table 4.14: Distribution of Frequency of Attendance Per Year in the NLSY79 Restricted Sample Frequency of Attendance Year Obs Total Responses Percent of Total Responses Per Year in the Year in the Year 0 1979 548 2684 20.42 6 1979 742 2684 27.65 12 1979 257 2684 9.58 30 1979 311 2684 11.59 52 1979 578 2684 21.54 78 1979 248 2684 9.24 0 1982 972 4099 23.71 6 1982 1257 4099 30.67 12 1982 403 4099 9.83 30 1982 468 4099 11.42 52 1982 719 4099 17.54 78 1982 280 4099 6.83 0 2000 375 3728 10.06 6 2000 686 3728 18.40 12 2000 774 3728 20.76 30 2000 2 3728 0.05 52 2000 1041 3728 27.92 78 2000 850 3728 22.80

119

Table 4.15: NLSY79 Descriptive Statistics Restricted and Weighted 1979 Variable Obs Mean Std. Dev. Min Max Time Varying Variables Frequency of Attendance Per Year 2684 24.16 25.56 0 78 Unemployed 2684 0.40 0.49 0 1 (in past calendar year) Months Unemployed 2684 0.91 2.00 0 12 (in past calendar year) Out of Labor Force 2684 0.65 0.48 0 1 (in past calendar year) Months Out of Labor Force 2684 4.47 4.52 0 12 (in past calendar year) Student 2684 0.58 0.49 0 1 Secondary Student 2684 0.37 0.48 0 1 Postsecondary Student 2684 0.21 0.41 0 1 Full-Time Postsecondary Student 2684 0.18 0.38 0 1 Part-Time Postsecondary Student 2684 0.03 0.16 0 1 Protestant 2684 0.55 0.50 0 1 Catholic 2684 0.31 0.46 0 1 Jewish 2684 0.01 0.11 0 1 Other 2684 0.01 0.10 0 1 None 2684 0.12 0.32 0 1 Log Income (thousands of 2008 dollars) 2684 10.72 1.01 0 12.31 Married 2684 0.11 0.32 0 1 Never Married 2684 0.87 0.33 0 1 Divorced/Separated 2684 0.01 0.12 0 1 Widowed 2684 0 0 0 0 Age 2684 18.61 1.82 16 22 Education (highest grade completed) 2684 11.25 1.65 4 16 Number of Children 2684 0.11 0.39 0 4 Time Invariant Variables Black 2684 0.14 0.35 0 1 Hispanic 2684 0.07 0.25 0 1 Men 2684 0.50 0.50 0 1 Women 2684 0.50 0.50 0 1

120

Table 4.16: NLSY79 Descriptive Statistics Restricted and Weighted 1982 Variable Obs Mean Std. Dev. Min Max Time Varying Variables Frequency of Attendance Per Year 4099 20.73 24.09 0 78 Unemployed 4099 0.39 0.49 0 1 (in past calendar year) Months Unemployed 4099 1.04 2.18 0 12 (in past calendar year) Out of Labor Force 4099 0.59 0.49 0 1 (in past calendar year) Months Out of Labor Force 4099 3.40 4.24 0 12 (in past calendar year) Student 4099 0.39 0.49 0 1 Secondary Student 4099 0.15 0.35 0 1 Postsecondary Student 4099 0.25 0.43 0 1 Full-Time Postsecondary Student 4099 0.20 0.40 0 1 Part-Time Postsecondary Student 4099 0.05 0.22 0 1 Protestant 4099 0.56 0.50 0 1 Catholic 4099 0.32 0.46 0 1 Jewish 4099 0.01 0.10 0 1 Other 4099 0.02 0.14 0 1 None 4099 0.10 0.30 0 1 Log Income (thousands of 2008 dollars) 4099 10.53 1.18 0 12.03 Married 4099 0.24 0.43 0 1 Never Married 4099 0.72 0.45 0 1 Divorced/Separated 4099 0.03 0.18 0 1 Widowed 4099 0.00 0.04 0 1 Age 4099 20.72 2.33 17 25 Education (highest grade completed) 4099 12.12 1.85 3 18 Number of Children 4099 0.26 0.62 0 5 Time Invariant Variables Black 4099 0.13 0.33 0 1 Hispanic 4099 0.06 0.24 0 1 Men 4099 0.49 0.50 0 1 Women 4099 0.51 0.50 0 1

121

Table 4.17: NLSY79 Descriptive Statistics Restricted and Weighted 2000 Variable Obs Mean Std. Dev. Min Max Time Varying Variables Frequency of Attendance Per Year 3728 35.71 29.78 0 78 Unemployed 3728 0.06 0.25 0 1 (in past calendar year) Months Unemployed 3728 0.22 1.23 0 12 (in past calendar year) Out of Labor Force 3728 0.22 0.41 0 1 (in past calendar year) Months Out of Labor Force 3728 1.69 3.86 0 12 (in past calendar year) Student 3728 0.06 0.24 0 1 Secondary Student 3728 0.00 0.04 0 1 Postsecondary Student 3728 0.06 0.24 0 1 Full-Time Postsecondary Student 3728 0.03 0.16 0 1 Part-Time Postsecondary Student 3728 0.03 0.19 0 1 Protestant 3728 0.53 0.50 0 1 Catholic 3728 0.25 0.44 0 1 Jewish 3728 0.01 0.10 0 1 Other 3728 0.08 0.26 0 1 None 3728 0.13 0.34 0 1 Log Income (thousands of 2008 dollars) 3728 10.70 1.87 0 12.94 Married 3728 0.65 0.48 0 1 Never Married 3728 0.15 0.36 0 1 Divorced/Separated 3728 0.20 0.40 0 1 Widowed 3728 0.01 0.08 0 1 Age 3728 39.02 2.32 35 43 Education (highest grade completed) 3728 13.35 2.62 0 20 Number of Children 3728 1.43 1.27 0 9 Time Invariant Variables Black 3728 0.14 0.34 0 1 Hispanic 3728 0.06 0.24 0 1 Men 3728 0.49 0.50 0 1 Women 3728 0.51 0.50 0 1

122

Table 4.18: HRS Descriptive Statistics Restricted and Weighted 2004 Variable Obs Mean Std. Dev. Min Max Time Varying Variables Frequency of Attendance Per Year 5902 26.67 26.87 0 78 Unemployed (at time of interview) 5902 0.03 0.17 0 1 Months Unemployed (since last job) 5902 0.47 10.19 0 473 Out of Labor Force (at time of interview) 5902 0.26 0.44 0 1 Months Out of Labor Force (since last job) 5902 19.05 77.40 0 558 Log Income (thousands of 2008 dollars) 5902 8.48 4.09 0 14.64 Married 5902 0.70 0.46 0 1 Never Married 5902 0.06 0.23 0 1 Divorced/Separated 5902 0.20 0.40 0 1 Widowed 5902 0.05 0.22 0 1 Age 5902 55.64 3.39 50 62 Number of Children 5902 2.73 1.82 0 17 Health (1=excellent…5=poor) 5902 2.64 1.14 1 5 Time Invariant Variables Protestant 5902 0.58 0.49 0 1 Catholic 5902 0.27 0.44 0 1 Jewish 5902 0.02 0.15 0 1 Other 5902 0.01 0.12 0 1 None 5902 0.11 0.32 0 1 Education (highest year) 5902 13.35 2.89 0 17 Black 5902 0.11 0.31 0 1 Hispanic 5902 0.09 0.28 0 1 Men 5902 0.48 0.50 0 1 Women 5902 0.52 0.50 0 1

123

Table 4.19: HRS Descriptive Statistics Restricted and Weighted 2006 Variable Obs Mean Std. Dev. Min Max Time Varying Variables Frequency of Attendance Per Year 5404 26.80 26.94 0 78 Unemployed (at time of interview) 5404 0.02 0.15 0 1 Months Unemployed (since last job) 5404 0.66 14.14 0 494 Out of Labor Force (at time of interview) 5404 0.31 0.46 0 1 Months Out of Labor Force (since last job) 5404 22.53 79.38 0 583 Log Income (thousands of 2008 dollars) 5404 8.46 4.02 0 15.83 Married 5404 0.70 0.46 0 1 Never Married 5404 0.05 0.22 0 1 Divorced/Separated 5404 0.19 0.40 0 1 Widowed 5404 0.06 0.23 0 1 Age 5404 57.70 3.41 52 65 Number of Children 5404 2.75 1.82 0 15 Health (1=excellent…5=poor) 5404 2.63 1.13 1 5 Time Invariant Variables Protestant 5404 0.59 0.49 0 1 Catholic 5404 0.26 0.44 0 1 Jewish 5404 0.02 0.15 0 1 Other 5404 0.01 0.12 0 1 None 5404 0.11 0.32 0 1 Education (highest year) 5404 13.37 2.87 0 17 Black 5404 0.11 0.31 0 1 Hispanic 5404 0.08 0.27 0 1 Men 5404 0.47 0.50 0 1 Women 5404 0.53 0.50 0 1

124

Table 4.20: HRS Descriptive Statistics Restricted and Weighted 2008 Variable Obs Mean Std. Dev. Min Max Time Varying Variables Frequency of Attendance Per Year 4954 26.09 27.01 0 78 Unemployed (at time of interview) 4954 0.02 0.15 0 1 Months Unemployed (since last job) 4954 0.38 8.29 0 482 Out of Labor Force (at time of interview) 4954 0.35 0.48 0 1 Months Out of Labor Force (since last job) 4954 25.90 81.50 0 599 Log Income (thousands of 2008 dollars) 4954 8.53 3.91 0 14.17 Married 4954 0.68 0.46 0 1 Never Married 4954 0.05 0.22 0 1 Divorced/Separated 4954 0.20 0.40 0 1 Widowed 4954 0.07 0.25 0 1 Age 4954 59.39 3.26 54 65 Number of Children 4954 2.76 1.83 0 16 Health (1=excellent…5=poor) 4954 2.74 1.11 1 5 Time Invariant Variables Protestant 4954 0.58 0.49 0 1 Catholic 4954 0.27 0.44 0 1 Jewish 4954 0.02 0.15 0 1 Other 4954 0.01 0.12 0 1 None 4954 0.11 0.32 0 1 Education (highest year) 4954 13.39 2.89 0 17 Black 4954 0.11 0.31 0 1 Hispanic 4954 0.08 0.28 0 1 Men 4954 0.47 0.50 0 1 Women 4954 0.53 0.50 0 1

125

BIBLIOGRAPHY

Allison, Paul D. and Richard P. Waterman (2002). “Fixed-Effects Negative Binomial Regression

Models.” In Stolzenberg, Ross M. (Ed.), Sociological Methodology: 247-265.

Washington, D.C.: Blackwell.

Altonji, Joseph G. and Rebecca Blank (1999). “Race and Gender in the Labor Market.” In

Ashenfelter, Orley and David Card (Eds.), Handbook of Labor Economics, Volume 3C:

3-102. New York: Elsevier.

Angrist, Joshua D. and Alan B. Krueger (1991). “Does Compulsory School Attendance Affect

Schooling and Earnings?” The Quarterly Journal of Economics, 106(4): 979-1014.

Azzi, Corry and Ronald Ehrenberg (1975). “Household Allocation of Time and Church

Attendance.” Journal of Political Economy, 83(1): 27-56.

Baum, Christopher F., Mark E. Schaffer and Steven Stillman (2007). “Enhanced routines for

instrumental variables/generalize method of moments estimation and testing.” The Stata

Journal, 7(4): 465-506.

Becker, Gary S. (1965). “A Theory of the Allocation of Time.” Economic Journal, 75(299): 493-

517.

Becker, Penny Edgell and Heather Hofmeister (2001). “Work, Family, and Religious

Involvement for Men and Women.” Journal for the Scientific Study of Religion, 40(4):

707-722.

Benham, Lee (1974). “Benefits of Women’s Education within Marriage.” The Journal of

Political Economy, 82(2 part 2): S57-S71.

126

Berger, Mark C. (1985). “The Effect of Cohort Size on Earnings Growth: A Reexamination of

the Evidence.” Journal of Political Economy, 93(3): 561-573.

Bettendorf, Leon and Elbert Dijkgraaf (2005). “The bicausal relation between religion and

income.” Tinbergen Institute. Tinbergen Institute Discussion Papers: TI 2005-105/2.

Binder, Dennis (2007). “The Disinvestment in Public Higher Education.” Chapman University

School of Law, Chapman University Law Research Paper No. 2007-06.

Blanchflower, David G. and Andrew J. Oswald (2004). “Well-being over time in Britain and the

USA.” Journal of Public Economics, 88(7-8): 1359-1386.

Blanchflower, David G. and Andrew J. Oswald (2008). “Is well-being U-shaped over the life

cycle?” Social Science & Medicine, 66(8): 1733-1749.

Bound, John, David A. Jaeger, and Regina M. Baker (1995). “Problems with Instrumental

Variables Estimation When the Correlation Between the Instruments and the Endogenous

Explanatory Variable is Weak.” Journal of the American Statistical Association, 90(430):

443-450.

Bound, John, Michael Schoenbaum, Todd R. Stinebrickner and Timothy Waidmann (1998).

“The Dynamic Effects of Health on the Labor Force Transitions of Older Workers.”

National Bureau of Economic Research, NBER Working Paper No. 6777.

Boyte, Layne (2005). “Female Education and Religiosity: Their Institutional Impacts on

Economic Growth.” Atlantic Economic Journal, 33: 361-62.

Bradshaw, Matt and Christopher G. Ellison (2009). “The Nature-Nurture Debate Is over, and

Both Sides Lost! Implications for Understanding Gender Differences in Religiosity.”

Journal for the Scientific Study of Religion, 48(2): 241-51.

Branas-Garza, Pablo and Shoshana Neuman (2004). “Analyzing Religiosity within an Economic

127

Framework: The Case of Spanish Catholics.” Review of Economics of the Household,

2(1): 5-22.

Brooks, Arthur C (2007). Who Really Cares: The Surprising Truth About Compassionate

Conservatism. Cambridge: Basic Books.

Brown, Timothy Tyler (2009). “Rational praying: The economics of prayer.” The Journal of

Socio-Economics, 38(1): 37-44.

Brusentsev, Vera (2006). “Evolution of Female Labor Force Participation in the United States:

1967 to 2003.” International Advances in Economic Research, 12(3): 358-373.

Cameron, A. Colin and Pravin K. Trivedi (2010). Microeconometrics Using Stata Revised

Edition. College Station: Stata Press.

Carroll, Nick (2007). “Unemployment and Psychological Well-being.” The Economic Record,

83(262): 287-302.

Child Trends Data Bank (2009). “Religiosity: Percentage of 8th, 10th, and 12th Grade Students

in the United States Who Report That Religion Plays a Very Important Role in Their

Lives, Selected Years 1976-2006.” Child Trends Data Bank.

Childs, Ellen (2010). “Religious Attendance and Happiness: Examining Gaps in the Current

Literature--A Research Note.” Journal for the Scientific Study of Religion, 49(3): 550-

560.

Chiteji, Ngina (2010). “Time Preference, Noncognitive Skills and Well Being across the Life

Course: Do Noncognitive Skills Encourage Healthy Behavior?” American Economic

Review, 100(2): 200-204.

Clark, Andrew E. (2003). “Unemployment as a Social Norm: Psychological Evidence from

Panel Data.” Journal of Labor Economics, 21(2): 323-351.

128

Clark, Andrew E. (2006). “A Note on Unhappiness and Unemployment Duration.” Institute for

the Study of Labor (IZA), Discussion Paper Series No. 2406.

Clark, Andrew E. and Andrew J. Oswald (1994). “Unhappiness and Unemployment.” The

Economic Journal, 104(424): 648-659.

Clotfelter, Charles T (1999). “The Familiar but Curious Economics of Higher Education;

Introduction to a Symposium.” Journal of Economic Perspectives, 13(1): 3-12.

Cohen-Zada, Danny and Moshe Justman (2005). “The Religious Factor in Private Education.”

Journal of Urban Economics, 57(3): 391-418.

Cohen-Zada, Danny and William Sander (2008). “Religion, Religiosity, and Private School

Choice: Implications for Estimating the Effectiveness of Private Schools.” Journal of

Urban Economics, 64: 85-100.

Collett, Jessica L. and Omar Lizardo (2009). “A Power-Control Theory of Gender and

Religiosity.” Journal for the Scientific Study of Religion, 2009, 48(2): 213-231.

Cornwall, Marie (2009). “Reifying Sex Differences Isn't the Answer: Gendering Processes, Risk,

and Religiosity.” Journal for the Scientific Study of Religion, 48(2): 252-55.

Crain, Mark W. and Katherine J. Lee (1999). “Economic Growth Regressions for the American

States: A Sensitivity Analysis.” Economic Inquiry, 37(2): 242-57.

Currie, Janet and Brigitte C. Madrian (1999). “Health, health insurance and the labor market.” In

Orley Ashenfelter and David Card (Eds.) Handbook of Labor Economics, 3(3): 3309-

3416. New York: Elsevier.

D'Antonio, William V., James D. Davidson, Dean R. Hoge, and Mary L. Gautier (2007).

American Catholics Today: New Realities of Their Faith and Their Church. Lanham,

MD: Rowman & Littlefield.

129

DeVaus, David A. (1984). “Workforce Participation and Sex Differences in Church

Attendance.” Review of Religious Research, 25(3): 247-256.

DeVaus, David A. and Ian McAllister (1987). “Gender Differences in Religion: A Test of the

Structural Location Theory.” American Sociological Review, 52(4): 472-481.

Deaton, Angus S. (2009). “Aging, Religion, and Health.” National Bureau of Economic

Research, NBER Working Paper No. 15271.

Di Tella, Rafael, Robert J. MacCulloch and Andrew J. Oswald (2001). “Preferences over

Inflation and Unemployment: Evidence from Surveys of Happiness.” American

Economic Review, 91(1): 335-341.

Economist (2009). “No rush for pews; Church attendance in recessions.” The Economist. June

20, 2009 U.S. Edition.

Ferraro, Kenneth F. and Jessica A. Kelley-Moore (2000). “Religious Consolation Among Men

and Women: Do Health Problems Spur Seeking?” Journal for the Scientific Study of

Religion, 39(2): 220-234.

Fiori, Katherine L., Judith C. Hays and Keith G. Meador (2004). “Spiritual Turning Points and

Perceived Control Over the Life Course.” International Journal of Aging and Human

Development, 59(4): 391-420.

Flatau, Paul, June Galea and Ray Petridis (2000). “Mental Health and Wellbeing and

Unemployment.” The Australian Economic Review, 33(2): 161-181.

Fortin, Nicole M. (2008). “The Gender Wage Gap among Young Adults in the United States:

The Importance of Money versus People.” The Journal of Human Resources, 43(4): 884-

918.

Francis, Leslie J. and Carolyn Wilcox (1998). “Religiosity and Femininity: Do Women Really

130

Hold a More Positive Attitude toward Christianity?” Journal for the Scientific Study of

Religion, 37(3): 462-469.

Fritjers, Paul and Tony Beatton (2008). “The mystery of the U-shaped relationship between

happiness and age.” National Centre for Econometric Research (NCER). NCER Working

Paper Series No. 26.

Gerber, Alan, Jonathan Gruber, and Daniel M. Hungerman (2008). “Does Church Attendance

Cause People to Vote? Using Blue Laws' Repeal to Estimate the Effect of Religiosity on

Voter Turnout.” National Bureau of Economic Research, Working Paper No. 14303.

Ghez, Gilbert R. and Gary S. Becker (1975). Allocation of Time and Goods over the Life Cycle

(Human Behavior and Social Institutions: No. 6). New York: Columbia University Press.

Gill, Sam D. (2005). “Prayer.” In Johnes, Lindsay (Ed.), Encyclopedia of Religion: 7367-72.

Macmillan Reference USA.

Glanville, Jennifer L, David Sikkink and Edwin I. Hernandez (2008). “Religious Involvement

and Educational Outcomes: The Role of Social Capital and Extracurricular

Participation.” The Sociological Quarterly, 49(1): 105-137.

Goldin, Claudia and Lawrence F Katz (1999). “The Shaping of Higher Education: The

Formative Years in the United States, 1890 to 1940.” Journal of Economic Perspectives,

13(1): 37-62.

Greene, William (2007). “Fixed and Random Effects Models for Count Data.” NYU Working

Paper No. EC-07-16. Available at SSRN: http://ssrn.com/abstract=1281928.

Groen, Jeffrey A. and Michelle J. White (2004). “In-State Versus out-of-State Students: The

Divergence of Interest between Public Universities and State Governments.” Journal of

Public Economics, 88(9-10): 1793-1814.

131

Groothuis, Peter and Paul E. Gabriel (2010). “Positive assortative mating and spouses as

complementary factors of production: a theory of labour augmentation.” Applied

Economics, 42(7-9): 1101-1111.

Groves, Melissa Osborne (2005). “How important is your personality? Labor market returns to

personality for women in the U.S. and U.K.” Journal of Economic Psychology, 26(6):

827-841.

Gruber, Jonathan (2004). “Pay or pray? The impact of charitable subsidies on religious

attendance.” The Journal of Public Economics, 88(12): 2635-2655.

Gruber, Jonathan and Daniel M. Hungerman (2008). “The Church versus the Mall: What

Happens When Religion Faces Increased Secular Competition?” Quarterly Journal of

Economics, 123(2): 831-862.

Grun, Carola, Wolfgang Hauser and Thomas Rhein (2010). “Is Any Job Better than No Job? Life

Satisfaction and Re-employment.” Journal of Labor Research, 31(3): 285-306.

Gunnoe, Marjorie Linder and Kristin A. Moore (2002). “Predictors of Religiosity among Youth

Aged 17-22: A Longitudinal Study of the National Survey of Children.” Journal for the

Scientific Study of Religion, 41(4): 613-622.

Hayashi, Fumio (2000). Econometrics. Princeton: Princeton University Press.

Heath, W. C., M.S. Waters and J. K. Watson (1995). “Religion and economic welfare: An

empirical analysis of state per capita income.” Journal of Economic Behavior and

Organization, 27(1): 129-42.

Heckman, James J., Jora Stixrud and Sergio Urzua (2006). “The Effects of Cognitive and

Noncognitive Abilities on Labor Market Outcomes and Social Behavior.” Journal of

Labor Economics, 24(3): 411-82.

132

Huang, Chong, Honbin Li, Pak Wai Liu and Junsen Zhang (2009). “Why Does Spousal

Education Matter for Earnings? Assortative Mating and Cross-Productivity.” Journal of

Labor Economics, 27(4): 633-652.

Iannaccone, Laurence R. (1988). "Formal Model of Church and Sect." American Journal of

Sociology, 94(2): 241-268.

Iannaccone, Laurence R. (1990). “Religious Practice: A Human Capital Approach.” Journal for

the Scientific Study of Religion, 29(3): 297-314.

Iannaccone, Laurence R. (1998). “Introduction to Economics of Religion.” Journal of Economic

Literature, 36(3): 1456-1496.

Iannaccone, Laurence R. (2006). “Economy.” In Ebaugh, Helen Rose (Ed.), Handbook of

Religion and Social Institutions: 21-39. New York: Springer Science Business Media.

Johnson, O. David, Marc Matre, and Gigi Armbrecht (1991). “Race and Religiosity: An

Empirical Evaluation of a Causal Model.” Review of Religious Research, 32(3): 252-266.

Kahoe, Richard D. (1974). “Personality and achievement correlates of intrinsic and extrinsic

religious orientations.” Journal of Personality and Social Psychology, 29(6): 812-818.

Kelley-Moore, Jessica A. and Kenneth F. Ferraro (2001). “Functional Limitations and Religious

Service Attendance in Later Life: Barrier and/or Benefit Mechanism?” Journal of

Gerontology, 56B(6): S365-S373.

Killingsworth, Mark R. and James J. Heckman (1986). “Female Labor Supply: A Survey.” In

Ashenfelter, Orley and Richard Layard (Eds.), Handbook of Labor Economics, Volume 1,

103-204. New York: North-Holland.

133

Kraft, Kornelius and Stefanie Neimann (2009). “Impact of Educational and Religious

Homogamy on Marital Stability.” Institute for the Study of Labor (IZA), Working Paper

No. 4491.

Lefgren, Lars and Frank McIntyre (2006). “The Relationship between Women’s Education and

Marriage.” Journal of Labor Economics, 24(4): 787-830.

Lehrer, Evelyn (2004). “Religiosity as a Determinant of Educational Attainment: The Case of

Conservative Protestant Women in the United States.” Review of Economics of the

Household, 2(2): 203-219.

Lipford, Jody W. and Robert D. Tollison (2003). “Religious participation and income.” Journal

of Economic Behavior and Organization, 51(2): 249-260.

Loh, Eng Sing (1996). “Productivity Differences and the Marriage Wage Premium for White

Males.” The Journal of Human Resources, 31(3): 566-589.

Loury, Linda D. (2004). “Does Church Attendance Really Increase Schooling?” Journal for the

Scientific Study of Religion, 43(1): 119-127.

Mare, Robert D. (1991). “Five Decades of Educational Assortative Mating.” American

Sociological Review, 56(1): 15-32.

Marks, Gary N. and Nicole Fleming (1998). “Influences and Consequences of Well-Being

Among Australian Young People: 1980-1995.” Social Indicators Research, 46(3): 301-

323.

McIntosh, Daniel N., Roxanne Cohen Silver and Camille B. Wortman (1993). “Religion’s role in

adjustment to a negative life event: Coping with the loss of a child.” Journal of

Personality and Social Psychology, 65(4): 812-821.

Miller, Alan S. and John P. Hoffman (1995). “Risk and Religion: An Explanation of Gender

134

Differences in Religiosity.” Journal for the Scientific Study of Religion, 34(1): 63-75.

Miller, Lisa (2010). “No Atheists in Foxholes; And other myths of the recession.” Newsweek,

156(5): 29. August 2, 2010.

Muller, Chandra and Christopher G. Ellison (2001). “Religious Involvement, Social Capital, and

Adolescents’ Academic Progress: Evidence from the National Education Longitudinal

Study of 1988.” Sociological Focus 34(2): 155-183.

Murphy, Gregory C. and James A. Athanasou (1999). “The effect of unemployment on mental

health.” Journal of Occupational and Organizational Psychology, 72(1): 83-99.

Musick, Marc A., James S. House and David R. Williams (2004). “Attendance at Religious

Services and Mortality in a National Sample.” Journal of Health and Social Behavior,

45(2): 198-213.

Neuman, Shoshana (1986). “Religious Observance Within a Human Capital Framework: Theory

and Application.” Applied Economics, 18(11): 1193-1202.

Neuman, Shoshana and Adrian Ziderman (1992). “Benefit’s of Women’s Education within

Marriage: Results for Israel in a Dual Labor Market Context.” Economic Development

and Cultural Change, 40(2): 413-424.

O’Connor, Thomas P., Dean R. Hoge and Estrelda Alexander (2002). “The Relative Influence of

Youth and Adult Experiences on Personal Spirituality and Church Involvement.” Journal

for the Scientific Study of Religion, 41(4): 723-732.

Oreopoulos, Philip, Till von Wachter and Andrew Heisz (2006). “The Short-and Long-Term

Career Effects of Graduating in a Recession: Hysteresis and Heterogeneity in the Market

for College Graduates.” National Bureau of Economic Research, Working Paper No.

12159.

135

Park, Jerry Z. and Christian Smith (2000). “‘To Whom Much Has Been Given...’: Religious

Capital and Community Volunatrism among Churchgoing Protestants.” Journal for the

Scientific Study of Religion, 39(3): 272-286.

Pencavel, John (1986). “Labor Supply of Men: A Survey.” In Ashenfelter, Orley and Richard

Layard (Eds.), Handbook of Labor Economics, Volume 1, 3-102. New York: North-

Holland.

Pencavel, John (1998). “Assortative Mating by Schooling and the Work Behavior of Wives and

Husbands.” American Economic Review, 88(2): 326-329.

Petersen, Larry R. (1994). “Education, Homogamy, and Religious Commitment.” Journal for the

Scientific Study of Religion, 33(2): 122-135.

Pew Forum (2009). “Is a Bad Economy Good for Church Attendance?” Online at

http://pewforum.org/Religious-Attendance/Is-a-Bad-Economy-Good-for-Church-

Attendance.aspx. March 12, 2009. Accessed May 16, 2011.

Powdthavee, Nattavudh (2005). “Unhappiness and Crime: Evidence from South Africa.”

Economica, 72(3): 531-547.

Presser, Stanley and Mark Chaves (2007). “Is Religious Service Attendance Declining?” Journal

for the Scientific Study of Religion, 46(3): 417-423.

Putnam, Robert D. and David E. Campbell (2010). American Grace: How Religion Divides and

Unites Us. New York: Simon & Scheuster.

Regnerus, Mark D., Christian Smith, and Brad Smith (2004). “Social Context in the

Development of Adult Religiosity.” Applied Developmental Science, 8(1): 27-38.

Riley, Naomi Schaefer (2004). God on the Quad: How Religious Colleges and the Missionary

Generation Are Changing America. New York: St. Martin's Press.

136

Rosenberg, Morris (1965). Society and the adolescent self-image. Princeton, NJ: Princeton

University Press.

Rotter, Julian B. (1996). Generalized expectancies for internal versus external control of

reinforcement. Washington, DC: American Psychological Association.

Sacerdote, Bruce and Edward L. Glaeser (2001). “Education and Religion.” National Bureau of

Economic Research, Working Paper No. 8080.

Sander, William (2005). “Catholics and Catholic Schooling.” Education Economics, 13(3): 257-

268.

Sander, William (2002). “Religion and Human Capital.” Economics Letters, 75: 303-307.

Senik, Claudia (2004). “When information dominates comparison: Learning from Russian

subjective panel data.” Journal of Public Economics, 88(9-10): 2099-2124.

Smith, Christian (2009). Souls in Transition: The Religious and Spiritual Lives of Emerging

Adults. New York: Oxford University Press.

Smith, Christian, Melinda Lundquist Denton, Robert Faris and Mark Regnerus (2002). “Mapping

American Adolescent Religious Participation.” Journal for the Scientific Study of

Religion, 41(4): 597-612.

Smith, Ian and John W. Sawkins (2003). “The economics of regional variation in religious

attendance.” Applied Economics, 35(14): 1577-1588.

Smith-Lovin, Lynn and Ann R. Tickamyer (1979). “Nonrecursive Models of Labor Force

Participation, Fertility Behavior and Sex Role Attitudes.” American Sociological Review,

43(4): 541-557.

Soydemir, Gokce A., Elena Bastida and Genaro Gonzalez (2004). “The Impact of Religiosity on

Self-Assessments of Health and Happiness: Evidence from the US Southwest.” Applied

137

Economics, 36(7): 665-672.

Spilka, Bernard, Philip Shaver and Lee A.Kirkpatrick (1985). “A General Attribution Theory for

the Psychology of Religion.” Journal for the Scientific Study of Religion, 24(1): 1-20.

Staiger, Douglas and James H. Stock (1997). “Instrumental Variables Regression with Weak

Instruments.” Econometrica, 65(3): 557-586.

Stark, Rodney (2002). “Physiology and Faith: Addressing the ‘Universal’ Gender Differences in

Religious Commitment.” Journal for the Scientific Study of Religion, 41(3): 495-507.

Sullivan, Dennis H. (1985). “Simultaneous Determination of Church Contributions and Church

Attendance.” Economic Inquiry, 23(2): 309-320.

Thompson Jr., Edward H. (1991). “Beneath the Status Characteristic: Gender Variations in

Religiousness.” Journal for the Scientific Study of Religion, 30(4): 381-394.

Tiefenhalter, Jill (1997). “The Productivity Gains of Marriage: Effects of Spousal Education on

Own Productivity across Market Sectors in Brazil.” Economic Development and Cultural

Change, 45(3): 633-650.

Toosi, Mitra (2010). “Labor force projections to 2018: older workers staying more active.”

Monthly Labor Review Online. Available at

http://www.bls.gov/opub/mlr/2009/11/art3full.pdf.

Torgler, Benno (2005). “The Importance of Faith: Tax Morale and Religiosity.” Journal of

Economic Behavior and Organization, 61: 81-109.

Uecker, Jeremy E. (2008). “Alternative Schooling Strategies and the Religious Lives of

American Adolescents.” Journal for the Scientific Study of Religion, 47(4): 563-584.

Ulbrich, Holley and Myles Wallace (1984). “Women's Work Force Status and Church

Attendance.” Journal for the Scientific Study of Religion, 23(4): 341-350.

138

Weber, Max (2001)[1905]. Kalberg, Stephen (Trans.). The Protestant ethic and the spirit of

capitalism. Chicago: Fitzroy Dearborn.

Welch, Finish (1979). “Effects of Cohort Size on Earnings: The Baby Boom Babies’ Financial

Bust.” Journal of Political Economy, 87(5 pt.2): S65-S97.

Winkelmann, Liliana and Rainer Winkelmann (1998). “Why Are the Unemployed So Unhappy?

Evidence from Panel Data.” Economica, 65(257): 1-15.

Wooldridge, Jeffrey M. (2000). Introductory Econometrics: A Modern Approach. New York:

South-Western College Publishing.

Wooldridge, Jeffrey M. (2002). Econometric Analysis of Cross Section and Panel Data.

Cambridge: MIT Press.

139

APPENDIX A

APPENDICES TO CHAPTER 3

A.1 EXCLUSION OF OBSERVATIONS

A.1.1 POOLED CROSS SECTION EXCLUSIONS

For the pooled cross section, I lose 1,990 observations by only including respondents with a religious identity, 6,575 observations due to missing or non-response data on frequency of prayer, 717 observations due to missing or non-response data on family income and respondent income, 6 observations due to missing or non-response data on age, 60 observations due to missing or non-response data on frequency of attendance, 10 observations due to missing or non- response data on education degree completed, 1 observation due to missing or non-response data on marital status, 5 observations due to missing or non-response data on religious identity at age

16, 123 observations due to missing or non-response data on how fundamentalist was the religion in which the respondent was raised, 15 observations due to missing or non-response data on remaining in the same state, 371 observations due to missing or non-response data involved with construction of the wages measure, and 17 observations due to missing or non-response data for highest year of education completed by the respondent’s spouse.

A.1.2 PANEL DATA EXCLUSIONS

For the panel, I begin with 3,072 observations for 1,536 respondents. Some observations are eliminated due to inconsistent data on measures that should be fixed over time including gender, religion raised in, how fundamentalist religion raised in was, race, and ethnicity. I treat

140

these inconsistent individuals as missing data. When restricting I lose 491 observations by only

including respondents with a religious identity, 16 observations due to missing or non-response

data on frequency of prayer, 445 observations due to missing or non-response data on family

income and respondent income, 14 observations due to missing or non-response data on age, 1

observation due to missing or non-response data on frequency of attendance, 368 observations due to missing, inconsistent, or non-response data on how fundamentalist was the religion in which the respondent was raised, 1 due to missing or non-response data on religion in which the respondent was raised, 6 observations due to missing or non-response data on remaining in the same state, 11 observations due to missing or non-response data involved with construction of the wages measure, 112 observations due to missing, inconsistent, or non-response data on race and ethnicity, and 16 observations due to missing, inconsistent, or non-response data on gender and 294 observations from balancing the panel.

A.2 TREATMENT OF LABOR AND NON-LABOR INCOME

A.2.1 LABOR INCOME

GSS income data are collected categorically but are transformed into semi-continuous data by a three step process. The three steps include: using midpoints of the categories as a measure of central tendency, calculation of the mean income in the top unbounded category through use of the Pareto distribution, and scaling income data across years into constant year

2000 dollars. For example, a respondent reports that he or she is in income category 2, which means the respondent earned between $1000 and $5000 in the previous year. The transformation uses the midpoint of $3000 and adjusts for inflation to report an income in year 2000 dollars for the respondent. The top income category, meanwhile, uses a process involving the Pareto

141

distribution since it is unbounded from above. For more details, please see GSS Methodological

Reports No. 64 and 101. I obtain data for labor income from the GSS measure CONRINC. It is

based on the RINCOME measure which is built solely on asking each respondent what their

income was in the previous year from working. The GSS does not report any income data that

indicate a respondent has 0 labor income. To help keep observations where a respondent’s labor

income must be 0 in the sample, I replace missing or non-response labor income data with 0 in

cases where the respondent reports working 0 weeks in the previous year. This replacement is

justified since the GSS measure RINCOME is built solely on asking each respondent what their

income was in the previous year from working. A respondent must have no labor income in the

previous year if they did not work any weeks at all in the previous year. This replacement results

in saving 105 observations for the final cross section sample.

A.2.2 NON-LABOR INCOME

Non-labor income is computed by subtracting a respondent’s real labor income from real family

income using the GSS measures CONRINC and CONINC, respectively. In cases where a

respondent’s real labor income exceeds real family income, I set non-labor income to 0. This adjustment is only necessary for 75 respondents or less than 2 percent of the sample. In each case where the adjustment is made respondents report being in the top category for both real labor income and real family income. The transformation process of the top category of income data, therefore, is the reason the adjustment is made. For more details, please see Appendix A.2.1 and

GSS Methodological Reports No. 64 and 101.

142

A.3 INSTRUMENTAL VARIABLES

A.3.1 SPOUSE EDUCATION AND SPOUSE PRESTIGE

I begin by restricting the sample to all relevant variables and all six instrument possibilities. The resulting sample is 3496 observations. Using bivariate correlations, prestige of occupation and highest year of school completed are highly correlated. Bivariate correlations are run on spouse prestige and spouse education, mother prestige and mother education, and father prestige and father education. On this basis, I estimate first stage regressions with only the three prestige measures included and repeat the first stage with only the three education measures included.

For the first stage regressions using prestige, spouse prestige is statistically significant most frequently with 7 of 8 cases statistically significant at the ten percent level or lower. For the first stage regressions using education, spouse education is statistically significant most frequently with 7 of 8 cases statistically significant at the ten percent level or lower. On this basis, I investigate spouse prestige and spouse education further.

I begin the last phase of determining which instrument to use by restricting the sample

two different ways. In the first instance I restrict to all relevant variables and spouse prestige to

get a sample size of 4229, which I label sample A. In the second instance I restrict to all relevant

variables and spouse education to get a sample size of 4240, which I label sample B.

I run first stage regressions in sample A using spouse prestige and interactions of spouse prestige with nonlabor income and a not employed dummy. Spouse prestige is statistically significant at the ten percent level or lower in 7 of 8 cases. I also run F-tests on the three instruments for all 8 cases. Instrumental variables estimation is also carried out using the three instruments to test for overidentification using Hansen’s J-statistics.

143

I run first stage regressions in sample B using spouse education and interactions of

spouse education with nonlabor income and a not employed dummy. Spouse education is

statistically significant at the ten percent level or lower in 6 of 8 cases, where 1 of the 2

remaining cases is significant at the eleven percent level. I also run F-tests on the three instruments for all 8 cases. Instrumental variables estimation is also carried out using the three instruments to test for overidentification using Hansen’s J-statistics.

I compare the F-statistics for the F-tests run on the instruments in sample A and the instruments in sample B. The comparison reveals that the F-statistics are slightly higher in sample B in 5 of 8 cases. The three remaining cases display slightly higher F-statistics for sample

A. In other words, the F-statistics reveal that spouse education and its two interaction terms are slightly stronger instruments than spouse prestige and its two interactions.

Comparison of overidentification tests resulting from the two samples reveals that overidentification is a problem in 2 of 8 cases using sample A while it is a problem in 1 of 8 cases using sample B. In other words, spouse education and its interactions display fewer problems from overidentification than spouse prestige and its interactions. Overidentification is determined to be a problem when the Hansen’s J-statistic p-value is 0.05 or lower. With this evidence in mind, I use spouse education and its two interactions with nonlabor income and the not employed dummy as instruments. There are three reasons I use spouse education and its interactions instead of spouse prestige and its interactions. First, spouse education is a positive measure while spouse prestige is a normative measure. Second, spouse education and its two interactions are slightly stronger instruments. Third, spouse education and its interactions are less affected by overidentification.

144

A.3.2 ALTERNATIVE INSTRUMENT STRATEGIES

I investigate a number of alternative instrument choices. I use a sample size of 4,527

respondents in each case, which I obtain by restricting for all measures except spouse education,

to analyze each alternative. An idea stemming from Oreopoulos et al (2006), where the authors show that graduates entering the labor market in a recession year suffer damages to earnings for the first 8 to 10 years, I examine labor market entry in a recession in several different ways. I use data from the National Bureau of Economic Research on business cycle expansions and contractions available at http://www.nber.org/cycles.html. I create dummies for potential labor market entry in a recession year, fifteen recession cohorts to correspond with 15 recessionary periods from 1918-2001, potential labor market entry in a recession year conditional on being in the labor force less than or equal to 10 years, and potential entry in a recession year multiplied by the number of months in a recession during the year of entry. Running each dummy alone in

separate first stage regressions identical to those presented in Tables 3 and 4, evidence reveals

weak or no statistical significance for coefficients on the dummy measures in the regressions. On

this basis I rule out using labor market entry in a recession year as an instrument.

As another idea from Welch (1979) and Berger (1985), I try dummies for generational

cohorts as instruments. I specifically create birth cohort dummies for the Greatest Generation

(1901-1924), Silent Generation (1925-1945), Baby Boomers (1946-1964), Generation X (1965-

1981), and Generation Y (1982-2000). I omit the Greatest Generation dummy and run first stage

regressions as in Tables 3 and 4 with the remaining four cohort dummies as instruments. Results

reveal little or no statistical significance on the dummy coefficients in the 8 first stage

regressions run. This result remains regardless of which generation cohort is treated as omitted.

From these findings I rule out using generational birth cohort dummies as instruments.

145

As an alternative I create dummies for 10 year birth cohorts (1907-1916, 1917-1926,

1927-1936, 1937-1946, 1947-1956, 1957-1966, 1967-1976, and 1977-1986) in accordance with the range of birth years in the sample. As with the generation cohorts, little or no statistical significance appears on the dummy coefficients in the 8 first stage regressions regardless of which 10 year cohort is treated as the omitted category. Therefore, I dismiss using 10 year birth cohort dummies as instruments.

I also consider size of the birth cohort in the year a respondent is born by collecting data on the number of births per year and the birth rate per year as defined by number of births per

1,000 people in the population for use as possible instruments. Data are from Vital Statistics of the United States, 1995, Volume I, Natality, table 1-1, Live births, birth rates, and fertility rates, by race: United States, 1909-1995. I run 8 first stage regressions as in Tables 3 and 4 using number of births per year in the year the respondent is born as the instrumental variable and find statistical significance at the ten percent level or lower on the number of births per year coefficient in 3 of 8 regressions. I repeat this process separately for birth rate per year in the year the respondent is born as the instrumental variables and find statistical significance at the ten percent level or lower on the birth rate per year coefficient in 4 of 8 regressions. While there is some correlation for both measures, it is not as strong as spouse education, where 6 of 8 first stage regressions contain a statistically significant coefficient for spouse education. Thus, to avoid weak instrument problems, I eliminate number of births per year in the year the respondent is born and the birth rate per year in the year the respondent is born as instruments.

Exploring graduation cohorts as instruments, I examine size of a graduation cohort in the potential year a respondent received their highest terminal degree if their highest degree is either a high school diploma or a Bachelors degree. I collect U.S. data from 1925-2004 on the number

146

of high school graduates per year, the high school graduation rate per year as defined by the

number of high school graduates divided by the number of 17 year olds in the population in

October of the final school year, the number of bachelors degrees awarded per year, and the

number of bachelors degrees awarded per year by gender. The data are from the National Center

for Education Statistics Digest of Education Statistics for 1959-2004 and the Biennial Survey of

Education in the United States for 1925-1958. Each of the five graduate cohort measures is run separately as the only instrumental variable in 8 first stage regressions as in Tables 3 and 4. For the number of high school graduates, statistical significance at the ten percent level or lower on its coefficient is present in 1 of 8 regressions. The high school graduation rate, meanwhile, displays statistical significance at the ten percent level or lower on its coefficient in 3 of 8 cases.

The total number of bachelors degrees, number of bachelors degrees earned by men, and number of bachelors degrees earned by women, show statistical significance at the ten percent level or lower in 4 of 8 regressions, 4 of 8 regressions, and 5 of 8 regressions, respectively. None of these results for the five graduate cohort measures indicate a stronger instrument than spouse education. In essence, I rule out using any of the five graduate cohort measures as instruments in favor of using spouse education.

147

APPENDIX B

APPENDICES TO CHAPTER 4

B.1 NATIONAL LONGITUDINAL SURVEY OF YOUTH 1979 COHORT

B.1.1 CROSS-SECTIONAL SAMPLE

The military subsample was discontinued by the NLSY79 in1985 while the economically disadvantaged non Black/non Hispanic subsample was cut in 1991. Weights for the NLSY79 data, furthermore, only offer two options of weighting the entire dataset or weighting the cross- sectional dataset. Because of these two reasons, I only use the cross-sectional sample. I weight the data using 1979 cross sectional weights to adjust for sample design.

B.1.2 SAMPLE RESTRICTIONS

I eliminate 2,753 observations where a respondent is under age 16 because they are not age eligible for the labor force. Clark (2003), Clark (2006), and Winkelmann and Winkelmann

(1998) also make this restriction when studying life satisfaction and unemployment. I also eliminate 1,037 observations with a 1979 sample weight of 0. I lose observations from refusal, don’t know, invalid skip, and non-interview responses for the measures I include in this study. I specifically lose 47 observations from frequency of religious service attendance, 1,046 observations from number of months unemployed in the past calendar year, 94 observations from region of interview, 2,557 observations from income, 49 observations from highest grade completed, 7 observations from religious identity, 195 from residence in a metropolitan statistical area, and 37 from full-time or part-time status as a postsecondary student.

148

B.1.3 CHILDHOOD RELIGIOUS UPBRINGING

In earlier drafts I also controlled for childhood religious upbringing that I sorted into the five categories of Protestant, Catholic, Jewish, Other, and None. Bivariate correlations between childhood religious upbringing and present religious identity indicated significant positive correlation between the two types of religious controls. Moreover, estimation results do not differ significantly in any way when including childhood religious upbringing as an additional religious control. Therefore, I do not include childhood religious upbringing as a control.

Furthermore, I only use present religious identity rather than only childhood religious upbringing to keep NLSY79 and HRS estimating equations as similar as possible.

B.1.4 AVERAGE ATTENDANCE IN 2000

Analysis of raw sample data reveals a significant difference in the distribution of

responses to the NLSY79 survey question on frequency of religious service attendance. In 1979

and 1982 approximately 11-12% of all respondents interviewed indicate a response of attending

religious services “2-3 times a month”. In 2000 this percentage is only 0.1% or 10 people out

8020 respondents interviewed. Table 12 shows the distribution of frequency of religious service

attendance by year in the final sample unweighted. The table shows that the distribution of

frequency of attendance per year is significantly different in year 2000 when compared to years

1979 and 1982. In year 2000 there is a significant decline in the percent of respondents who

report “0” and “30” in and a significant increase in the percent of respondents who report “78”

and “52” compared to years 1979 and 1982. These distributions by year closely mirror those of

the unrestricted NLSY79 sample. Considering what the distributions indicate, the average

frequency of attendance per year I report in Tables 13-15 is correct

149

B.2 HEALTH AND RETIREMENT STUDY

B.2.1 DATASET VERSION

Because of the difficulty involved in using multiple waves of the Health and Retirement

Study (HRS), I use the user-friendly RAND Version J HRS dataset for all of the HRS measures in this paper except frequency of attendance to religious services. To add the frequency of attendance variable to the RAND dataset, I extract it from public use files and merge it with the

RAND dataset using identifier data from the HRS Tracker 2008 V 1.0.

B.2.2 HEALTH AND RETIREMENT STUDY DESCRIPTION

The Health and Retirement Study (HRS) is a national survey that originated in 1992 to collect data on individuals age 50 or over and their spouses regardless of the age of the spouse.

The survey oversamples for Blacks, Hispanics, and residents of Florida in an effort to provide panel data for analysis of health and retirement of the elderly in the United States. Since the study began, five cohorts have been interviewed: the initial HRS cohort (born 1931-1941), The

Study of Assets and Health Dynamics Amongst the Oldest Old (AHEAD; born before 1924),

The Children of the Depression (CODA; born 1924-1930), The War Babies (WB; born 1942-

1947), and The Early Baby Boomers (EBB; born 1948-1953). Respondents are assigned to these five cohorts by birth year and by the cohort the household was originally sampled in. Sampling is designed to take into account non-age eligible respondents already being interviewed as part of a previous household cohort. For example, a non age-eligible respondent may first be interviewed as part of the HRS household cohort but will later be an age-eligible respondent of the WB birth year cohort. Because sampling is designed to keep consistent representation by birth year, I use cohorts by birth year in this study. I use the WB and EBB birth year cohorts to follow Clark

(2003), Clark (2006), and Winkelmann and Winkelmann (1998) who limit their analyses to

150

individuals between ages 16 and 65 when studying life satisfaction and unemployment.

Frequency of religious service attendance data are available in 1992, 1994, 2004, 2006, and 2008 for the original HRS birth year cohort. With the age restriction imposed, however, the majority of the HRS cohort exceeds the age restriction in 2004 with the entire HRS cohort exceeding the age restriction by 2008. The WB and EBB cohorts, meanwhile, only lose observations in 2008 from respondents born in 1942 due to the age restriction. Therefore, I use the WB and EBB birth year cohorts to obtain a sample that maintains the representation of the original survey design while not being subject to attrition problems due to the age restriction.

B.2.3 YEARS OF DATA USED AND SAMPLE WEIGHTS

Since WB and EBB birth year cohort respondents could initially be interviewed as part of the HRS household cohort, there are a few respondents belonging to the WB and EBB birth year cohorts who have data on their frequency of religious service attendance in 1992 and 1994. I eliminate these 1992 and 1994 observations for two reasons. First, there are only 47 additional observations added to the final dataset of 16,268 observations spread over 5,993 individuals, which implies that the 47 observations are not meaningfully representative of the two cohorts’ birth year. Second, to use fixed effects panel estimation with sample weights, weights must be assigned to be constant over time. I use 2004 person-level sample weights to weight the data to be consistent and because some respondents leave the dataset over time and do not have a sample weight in 2006 or 2008. Furthermore, weights are not available in 1992 and 1994 for the vast majority of respondents in the WB and EBB birth year cohorts. The sample weights I use are designed to match the Current Population Survey, which includes living noninstitutionalized individuals.

151

B.2.4 SAMPLE RESTRICTIONS

I eliminate 3,826 observations where a respondent is over age 65 to follow Clark (2003),

Clark (2006), and Winkelmann and Winkelmann (1998) who also make this restriction when

studying life satisfaction and unemployment. I also drop 435 observations with missing sample

weights and 173 observations with a sample weight of 0. I lose observations from missing, don’t know, or unknown responses for the measures I include in this study. I specifically lose 28 observations from frequency of religious service attendance, 57 observations from religious identity, 6 observations from marital status, 4 observations from race and ethnicity, 96 observations from highest year of education, 329 observations from number of children, 30 observations from census division of interview, and 8 observations from self-reported overall health.

B.2.5 LABOR FORCE STATUS MEASURES

Respondents are counted as in the labor force if they are working full time, working part time, unemployed, or partly retired. They are otherwise not in the labor force and treated as such.

The unemployed are created primarily from the unemployed labor force category. Additional unemployed respondents are identified from the partly retired labor force category. The partly retired labor force category includes individuals who mention retirement and are working part time in addition to individuals who mention retirement and are looking for a part time job. Using data on the year the last job worked ended and months since last worked, as detailed in Appendix

B.2.5, I determine that individuals who are classified as partly retired must be unemployed if they have (1) reported working in their life time and (2) have a positive measure of months since last worked.

152

B.2.6 UNEMPLOYMENT MEASURES

To construct months unemployed since last job and months out of the labor force since last job, I use data on the month and year of interview, month and year last job worked ended, and month and year of birth. I begin by generating the number of months since last worked. I set the number of months since last worked equal to 0 if the respondent is currently working. If the respondent has never worked, I subtract 192 months (12 months in 16 years) from the respondent’s age in months, which I determine using month and year of interview and month and year of birth. This provides me with the number of months the respondent has been eligible to work since age 16. I keep this number as the number of months since last worked for never worked respondents. In the case of respondents who are not currently working and have worked,

I compute the number of months since last job using the month and year of interview and month and year last job worked ended. With the months since last worked measure complete, I generate months unemployed since last job by setting it equal to months since last worked in cases where the respondent is unemployed at the time of interview. I generate months out of the labor force since last job with a similar process by setting months out of the labor force equal to months since last worked in cases where the respondent is out of the labor force at time of interview.

B.3 COUNT DATA ESTIMATION METHODS

B.3.1 ADDRESSING OVERDISPERSION

According to Cameron and Trivedi (2010), overdispersion refers to when the conditional variance of the dependent variable is greater than the conditional mean. Overdispersion is a violation of the equidispersion property of the Poisson distribution where the conditional variance and conditional mean of the dependent variable are assumed to be equal. Pooled

153

negative binomial estimation adjusts for overdispersion and quantifies it with a dispersion parameter alpha, where values of alpha significantly greater than 0 indicate an overdispersion problem. Poission fixed effects is used instead of negative binomial fixed effects because software packages such as Stata use negative fixed effects estimates based on Hausman et al.

(1984) that addresses unobserved heterogeneity as it pertains to the overdispersion parameter rather than the regressors, according to Greene (2007) and Allison and Waterman (2002).

Pooled negative binomial estimation with individual dummies included is a feasible alternative, as Allison and Waterman (2002) highlight, provided the number of individual dummies to be created is not computationally onerous. Given the large number of individuals in both datasets, I use Poisson fixed effects with cluster-robust standard errors because it is computationally feasible and adequate for addressing overdispersion and unobserved heterogeneity. For a more technical discussion of the negative binomial fixed effects estimator and its pitfalls see Greene

(2007) and Allison and Waterman (2002).

154