three essays on the economics of education miguel e. martinez

124
Three Essays on the Economics of Education Miguel E. Martinez Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy under the Executive Committee of the Graduate School of Arts and Sciences COLUMBIA UNIVERSITY 2017

Upload: others

Post on 30-Apr-2022

2 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Three Essays on the Economics of Education Miguel E. Martinez

Three Essays on the Economics of Education

Miguel E. Martinez

Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy

under the Executive Committee of the Graduate School of Arts and Sciences

COLUMBIA UNIVERSITY

2017

Page 2: Three Essays on the Economics of Education Miguel E. Martinez

2

Copyright 2017

Miguel E. Martinez

All Rights Reserved

Page 3: Three Essays on the Economics of Education Miguel E. Martinez

3

ABSTRACT

Three Essays on the Economics of Education

Miguel E. Martinez

Essay 1: Determinants of NCLEX-RN Success Beyond the HESI Exit Exam: Performance in Nursing

Courses and Academic Readiness

Abstract: In this study, I examine whether demographics, pre-college academic readiness

measures, and performance in nursing courses improve the correct identification of students who

will pass/fail the NCLEX-RN above and beyond the HESI exit exam scores. I find that their inclusion

can improve the identification of those who will fail but not those who will pass.

Essay 2: The Impact of Remediation on NCLEX-RN Success: Positive, Neutral, or Negative?

Abstract: In this study, I use two nationally representative samples to explore the impact of

required remediation on passing the NCLEX-RN on the first attempt using regression discontinuity

design both as local randomization and as continuity at the cutoff using. As the former, I find

some evidence that remediation has a negative impact on passing the NCLEX-RN on the first

attempt. As the latter, I find limited evidence of positive treatment effects.

Essay 3: Testing a Rule of Thumb: For STEM degree attainment, More Selective is Better

Abstract: In this essay, I test the rule of thumb that, for STEM students, attending a highly

selective institution instead of a moderately selective institution improves the probability of

obtaining a STEM degree. Overall, I find that highly selective institutions have a comparative

advantage in producing STEM graduates among those already interested in STEM but not among

those initially not interested in STEM.

Page 4: Three Essays on the Economics of Education Miguel E. Martinez

4

This page intentionally left blank

Page 5: Three Essays on the Economics of Education Miguel E. Martinez

i

Contents List of Figures………………………………………………………………………………………………………………………………………… iii

Acknowledgements ...................................................................................................................................... iv

Chapter 1: Determinants of NCLEX-RN Success Beyond the HESI Exit Exam: Performance in Nursing

Courses and Academic Readiness ................................................................................................................. 1

1.1 Introduction ........................................................................................................................................ 1

1.2 Review of the Literature ..................................................................................................................... 4

1.3 Data ..................................................................................................................................................... 6

1.4 Organizational Context ....................................................................................................................... 7

1.5 Methodological Approaches ............................................................................................................. 11

1.6 Results ............................................................................................................................................... 14

1.6.1 Explanatory Models ................................................................................................................... 14

1.6.2 Out-of-Sample Predictive Models .............................................................................................. 18

1.6.3. Severe Error Rates ..................................................................................................................... 21

1.7 Conclusion and Further Research ..................................................................................................... 24

Chapter 2: The Impact of Remediation on NCLEX-RN Success: Positive, Neutral, or Negative? ............. 26

2.1 Introduction ...................................................................................................................................... 26

2.2 Review of the Literature ................................................................................................................... 30

2.2.1 HESI Exit Exam ............................................................................................................................ 33

2.3 Data ................................................................................................................................................... 34

2.4 Identification Strategy....................................................................................................................... 36

2.5 The Assumptions of Regression Discontinuity are Met .................................................................... 37

2.5.1 Manipulation of the Running Variable: Visual Evidence ........................................................... 38

2.5.2 Manipulation of the Running Variable: Statistical Evidence ..................................................... 39

2.5.3 Continuity of Pre-treatment Variables: Statistical Evidence ..................................................... 40

2.5.4 Covariate Balance: Statistical Evidence ..................................................................................... 40

2.6 Results ............................................................................................................................................... 42

2.6.1 Estimation of Treatment Effects: RD as Discontinuity at the Cutoff ......................................... 42

2.6.2 RD Estimates Conclusion and Discussion ................................................................................... 43

2.7 Robustness Checks ............................................................................................................................ 43

2.7.1 Logistic Regression ..................................................................................................................... 44

Page 6: Three Essays on the Economics of Education Miguel E. Martinez

ii

2.7.2 Propensity Score Matching Analysis .......................................................................................... 46

2.8 Looking for Heterogenous Treatment Effects................................................................................... 47

2.8.1 Logistic Regression ..................................................................................................................... 48

2.8.2 Propensity Score Matching ........................................................................................................ 48

2.9 Conclusion and Policy Implications ................................................................................................... 49

Chapter 3: Testing a Rule of Thumb: For STEM Degree Attainment, More Selective is Better ................ 52

3.1 Introduction ...................................................................................................................................... 52

3.2 Competing Definitions of STEM ........................................................................................................ 55

3.2.1 STEM as a Career........................................................................................................................ 55

3.2.2 STEM as a College Major ............................................................................................................ 57

3.3 Review of the Literature ................................................................................................................... 58

3.3.1 Interest in STEM Careers/Majors ............................................................................................... 59

3.3.2 College Choice ............................................................................................................................ 61

3.3.3 Matching in the College Choice Process .................................................................................... 62

3.3.4 STEM Attrition and Institutional Selectivity ............................................................................... 66

3.4 Data ................................................................................................................................................... 67

3.4.1 Defining Interest in STEM Interest in ELS of 2002/06/12 .......................................................... 68

3.4.2 Defining a Highly Selective Post-Secondary Institution in ELS of 2002/06/12 .......................... 69

3.4.3 Defining a STEM Degree in ELS of 2002/06/12 .......................................................................... 69

3.5 STEM Students .................................................................................................................................. 70

3.6 Methodology ..................................................................................................................................... 72

3.7 Results ............................................................................................................................................... 75

3.7.1 Logistic Regression Analysis ....................................................................................................... 76

3.7.2 Matching Analysis ...................................................................................................................... 77

3.7.3 Summary of Results ................................................................................................................... 79

3.8 Conclusion ......................................................................................................................................... 80

Works Cited/Consulted ............................................................................................................................... 81

Appendices .................................................................................................................................................. 92

Page 7: Three Essays on the Economics of Education Miguel E. Martinez

iii

List of Figures

Figure 1. Predicted Average Marginal Probabilities of Passing the NCLEX-RN by Foundations I Course

Grade........................................................................................................................................................... 17

Figure 2. Predicted Average Marginal Probabilities of Passing the NCLEX-RN by Health & Illness ICA

Course Grade. ............................................................................................................................................. 17

Figure 3. Predicted Average Marginal Probabilities of Passing the NCLEX-RN by Psych/Mental Health

Course Grade .............................................................................................................................................. 18

Figure 4. Predictive Validity of Models (1) and (3) and Current Progression Rule based on HESI Exit Exam.

.................................................................................................................................................................... 19

Figure 5. Possible Impacts of HESI exit exams scores on ability to sit for the NCLEX-RN. ......................... 33

Figure 6. Distribution of HESI Exit Exam Scores by Required Remediation Status ..................................... 39

Figure 7. Remediation Marginal Effects over the HESI Score Response Surface ....................................... 45

Page 8: Three Essays on the Economics of Education Miguel E. Martinez

iv

Acknowledgements To my dad.

I would like to express my sincere gratitude to Thomas Bailey and Peter Bergman who

oversaw my work during the research proposal stage and writing of the dissertation. The

insightful comments of the defense committee greatly improved the quality of my work. I am

very thankful to Luis Huerta, Aaron Pallas, and Clyde Belfield.

The dissertation would not have been possible without the help of Rutgers School of

Nursing. I would like to thank Bill Holzemer, Jeannie Cimiotti, Mary Johansen, Edna Cadmus,

Pam de Cordoba, and John Runfeldt for their encouragement and support during the dissertation

process.

To the teachers and instructors who inspired and influenced me along the way: Mr.

Reeves, Mrs. Williams, Mrs. Schaeffer, Dr. Dorce, Dr. Haynes, Dr. Rumiano and Dr. Rivera-Batiz.

To my parents for the love and affection they showered upon me since my birth and for

fostering the desire to understand. To my siblings for their example of hard work and

perseverance in their professional careers and for their unwavering support in all my endeavors.

Most of all, to my wife who supported my academic studies every step of the way. To

my children who generously gave up their time with me so I could push the dissertation

forward.

Page 9: Three Essays on the Economics of Education Miguel E. Martinez

1

Chapter 1: Determinants of NCLEX-RN Success

Beyond the HESI Exit Exam: Performance in

Nursing Courses and Academic Readiness

1.1 Introduction

First time NCLEX-RN (National Council Licensure Examination for Registered Nurses) pass

rates play a significant role in nursing programs. National accrediting bodies like the Commission

on Collegiate Nursing Education (CCNE) and the National League for Nursing Commission

for Nursing Education Accreditation (CNEA) require minimum first-time NCLEX-RN pass rates of

80 percent to maintain accreditation. These standards must be met by all pre-licensure

programs (Generic Baccalaureate, 2nd Degree Accelerated, pre-licensure Masters in Nursing

Science) and campuses (home and satellite) within an institution. Failure to meet the standard

risks the loss of accreditation and/or placement on probationary status (CCNE, 2014) (CNEA,

2015). Nursing programs must also meet the NCLEX-RN pass rate standards established by their

state boards of nursing which may be higher than those established by national accrediting

bodies.

First time NCLEX-RN pass rates have a marked signaling function. For prospective

students, high pass rates suggest a high-quality nursing program (Norton et al, 2006). Thus,

improving first time pass rates can increase the quality of applicants and enrollees which can

further improve NCLEX-RN outcomes while prolonged decreased performance can have the

opposite effect (Beeson and Kissling, 2001). For administrators within and outside of nursing

Page 10: Three Essays on the Economics of Education Miguel E. Martinez

2

decanal units, first time pass rates are often perceived as an indication of the quality of

instruction (Giddens et al, 2009; Taylor et al, 2014).

Thus, nursing programs have a strong incentive to identify students who are likely to fail

the NCLEX-RN. Once identified, administrators may assign students to remediation or may

prevent them from taking the licensure exam (progression). Typically, nursing programs identify

at risk-students based solely on the results of exit exams like the HESI (Health Education Systems,

Inc.), ATI (Assessment Technologies Institute), ERI (Educational Resources Inc), or NLN (National

League of Nursing).

Eleven validation studies have found that the HESI exit exam (E2) is an accurate predictor

of passing NCLEX-RN - National Council Licensure Examination for Registered Nurses on the first

attempt (Adamson and Britt, 2009; Launcher et al, 1999; Langford and Young, 2013, Lewis, 2005;

Newman et al; 2000, Nibert and Young, 2001; Nibert et al, 2002; Young and Willson, 2012;

Zweighaft, 2013; Schreiner at el, 2014; Zweighaft, forthcoming). E2 results have been used to

gauge student readiness for successful entry into the nursing profession (Pennington et al, 2010).

Yet the sole use of the examination for progression or remediation may not be justified since it

has predictive power for NCLEX-RN success but for not NCLEX-RN failure (Spurlock, 2006).

Students scoring above 850 on the HESI exit exam may be more likely to pass the NCLEX-RN than

those who do not but the HESI exit exam is not able to identify those who score above 850 but

do not pass the NCLEX-RN.

Nationally, approximately 97 percent of students who have scored 850 and above on the

HESI have passed the NCLEX (Young & Willson, 2012; Lewis, 2006). Not meeting the HESI cut

score, however, does not entail a precipitous drop in the predicted likelihood of NCLEX success.

Page 11: Three Essays on the Economics of Education Miguel E. Martinez

3

Those who achieve scores between 800 and 849 have more than a 93 percent probability of

passing the NCLEX. Students with HESI scores between 700 and 799 and those with HESI scores

between 600 and 699 have predicted probabilities of 85 percent and 69 percent, respectively

(Lewis, 2006). Within each of the ranges, HESI is not able to identify which students are most

likely to fail the NCLEX. The current use of HESI exit exam scores raises questions of equity and

efficiency. To require students, who have at least an 85% probability of passing to undergo

remediation may be a suboptimal allocation of resources. Attaching strong penalties

(withholding degree or eligibility to sit for the NCLEX) for not meeting the HESI exam cutoff may

also introduce a market inefficiency. As such, progression rules should not be based on a single

metric but on a broader assessment of student readiness to pass the NCLEX (Spurlock, 2006).

Using a rich administrative database of more than 700 students from a single nursing

school in the northeast, the first paper on the economics of education examines whether nursing

course grades, demographics, and SAT scores can improve the correct identification of students

who will pass/fail above and beyond the HESI exit exam. I do this by comparing the predictive

validity of the most common progression rule of scoring 850 on the HESI exit exam to two

statistical models: one that contains demographics, SAT scores, and nursing course performance

and one that, in addition, also contains an indicator of scoring 850 or above on the HESI exit exam

score. I find the HESI-based progression rule is best in identifying students who pass the NCLEX-

RN on the first attempt. In contrast, the model-based predictions are better at identifying

students who do not to pass it on the first attempt. Temporal variance in the predictive accuracy

of models exists. The model that includes HESI scores is better for the period after the more

Page 12: Three Essays on the Economics of Education Miguel E. Martinez

4

rigorous NCLEX-RN standard was implemented (in March of 2014) and the other for the previous

period.

I also assess the degree to which students are severely misplaced as a result of the HESI-

based remediation protocol and investigate if model-based protocols could potentially reduce

such misplacement using Community Health course outcomes as a proxy for NCLEX-RN

outcomes. In comparison to HESI-based progression rule, I find that the models result in

modestly lower under-placement and over-placement rates. The different models are not,

however, equally effective at reducing the severe error rate for all student subgroups. Yet the

variability in predictions across the models can be leveraged to more accurately place students.

Establishing inter-model agreement for placement decisions eliminates over-placement rates

and lowers under-placement rates by 30 percent.

The structure of the essay is as follows. In Section 2, I review the current literature on

determinants of NCLEX-RN success. The next section describes the context in which the

outcomes were generated and profiles the students. In Section 4, I briefly detail the analytical

sample. The methodological approaches are articulated in Section 5. Results are discussed in

the following section. Finally, I summarize results and advance ideas for further research in

Section 7.

1.2 Review of the Literature The literature suggests that demographic variables are associated with NCLEX-RN

outcomes. Males and African-American students fail the NCLEX-RN at significantly higher rates

Page 13: Three Essays on the Economics of Education Miguel E. Martinez

5

(Daley et al, 2003; Has et al, 2003). No difference in pass rates exist between the students based

in the main campuses and those students based in the satellite campuses (Has et al, 2003). Older

students were more likely to pass the NCLEX-RN on the first attempt (Briscoe and Anema 1999;

Beeson and Kissling, 2001; Daley et al, 2003; Fortier, 2010). English as primary language is

positively correlated with NCLEX-RN success (Arathuzik and Aber, 1998).

Studies also indicate that academic readiness and academic performance correlate with

positive NCLEX-RN outcomes. Students who passed had statistically higher verbal SAT scores

(Has et al, 2003). The number of grades below B in nursing courses is negatively correlated with

NCLEX-RN results (Beeson and Kissling, 2001). Standardized nursing end-of-course exam scores

have a statistically significant and positive correlation with passing the NCLEX-RN (Briscoe and

Anema, 1999; Daley et al, 2003). Grades in anatomy, pathophysiology, and medical surgical

courses were positively associated to passing the NCLEX-RN on the first attempt (Daley et al,

2003). Pre-nursing GPA in college, performance in college science courses, and final college GPA

are correlated with positive NCLEX-RN outcomes (Arathuzik and Aber, 1998; Sayles et al, 2003;

Tipton et al, 2007; Fortier, 2010; Bosch et al, 2011). Confidence in test taking and critical thinking

further increase the likelihood of NCLEX-RN success (Arathuzik and Aber, 1998; Giddens and

Gloeckner 2004; Santiago, 2013). In addition, family demands and negative emotions (anxiety,

anger and guilt) negatively correlate with performance on the NCLEX while an oral dependent

learning style have a positive association (Arathuzik and Aber, 1998; Sayles et al, 2003)

Overall, the literature examines the association among demographic, academic, and

contextual factors and NCLEX-RN outcomes. The analyses suggest that pre-admission measures

of academic readiness and performance are correlated with passing the NCLEX-RN on the first

Page 14: Three Essays on the Economics of Education Miguel E. Martinez

6

attempt (Has et al, 2003; Sayles et al, 2003; Bosch et al 2011). Non-minority and older students

tend to have better NCLEX outcomes (Daily, 2003; Has et al, 2003; Fortier, 2010; Bosch, 2011).

Performance on nursing standardized end-of-course exams, exit exams, and nursing courses are

associated with passing the NCLEX-RN (Beeson et al, 2011; Briscoe and Anema, 1999; Fortier,

2010; Tipton et al, 2007; Santiago, 2013). Critical thinking and test-taking skills, learning style,

and non-academic time demands are also predictive of NCLEX-RN success. Although the

methodologies leveraged by the above-mentioned studies are not rigorous they examine

theoretically relevant determinants of NCLEX-RN success. My determinants of NCLEX-RN model

incorporates the various aspects taken up by the studies – demographics, pre-admissions

measures of academic performance, and nursing course and standardized test performance – to

build the most comprehensive determinants model in the literature and to test its predictive

power. In addition, my essay will be the first study on determinants that uses NCLEX-RN

outcomes from 2014 and 2015 – the first two full calendar in which the new NCLEX-RN cut-score

(0.00 logits) was implemented.

1.3 Data I use a student level database of almost 750 students who attended a four-year nursing

program who took a HESI exit exam at a university in the northeast. The database draws data

on each student from three distinct sources: a university-managed student records information

system, a nursing school controlled HESI management system, and NCLEX-RN reports generated

by a state Board of Nursing. The university-based data source contains demographic

information: race, gender, status in the Educational Opportunity Fund program (a measure of

Page 15: Three Essays on the Economics of Education Miguel E. Martinez

7

poverty), SAT scores, and performance in nursing courses. The HESI management system

contains HESI exit exam scores while the Board of Nursing NCLEX-RN reports provide the

outcomes of NCLEX-RN test takers.

1.4 Organizational Context During the period covered by the analysis, the organizational context in which students

progressed through the nursing program experienced structural changes. In 2009, a new dean

took leadership of the nursing school after two years of having an interim dean. As the new

dean took the helm, a satellite campus opted to break off and become its own nursing school. In

2013, the main campus compensated for its previous loss by establishing a new satellite campus.

That same year, the university was mandated to merge with another university. Since both

institutions of higher education had nursing schools they were merged the following year. In

anticipation of the merger, both schools dedicated considerable resources to aligning curricula

and business processes. To this end, 22 joint-committees were convened. Program outcomes,

measures, and performance criteria were agreed upon for each pre-licensure and post-licensure

programs by the end of the 2015-2016 school year (Martinez et al, 2016).

At the end of calendar year 2015, the newly created nursing school was not fully

integrated. After the merger, faculty and staff of the two institutions still operated under

different statutes and unions. During the period covered by the data, the pre-licensure four-year

program remained the same since it was not a redundancy between the two institutions. There

were no changes to admissions criteria, progression rules, and/or graduation requirements.

Faculty attrition for that program was minimal (Martinez et al, 2016).

Page 16: Three Essays on the Economics of Education Miguel E. Martinez

8

Before the merger, the institution that absorbed the other institution had two pre-

licensure nursing programs: a second degree program and traditional four-year program. The

former was geared for students who already had a bachelor’s degree and wanted to receive a

nursing degree in a compressed time period. Cohorts of the accelerated program increased from

15 in 2009 to 38 in 2015. Historically, these students fared better on the NCLEX-RN than those

of the four-year nursing program. As such, they were not required to take the HESI exit exam

and, consequently, were not subjected to a remediation policy. The traditional program had

much larger enrollment and had a less seasoned student body.

Students apply to the four-year program pre-licensure program during high school. For

the cohorts covered by the period of analysis, yearly applications ranged from approximately

2,500 to 3,100 for 105-125 seats. Admissions were granted to 15 percent of applicants. About

one in three took up the admissions offer. Those who opted to attend school elsewhere had,

on average, higher SAT scores than those who did not (Martinez, 2014). The vast majority of

these students attended research-intensive universities in the northeast (Martinez et al, 2016).

The nursing program is highly structured. During the first two years of study, students

take their general education courses. Unlike their counterparts in the social sciences and

humanities, they take a considerable amount of entry-level science courses: Anatomy and

Physiology I and II, Organic Biochemistry, Microbiology, and World of Chemistry Lab. They are

also required to complete Statistics and Nutrition courses. Year three of the program is

exclusively dedicated to nursing courses. During the last year, students are required to take a

general elective and a literature course in addition to their nursing courses.

Page 17: Three Essays on the Economics of Education Miguel E. Martinez

9

Cohorts progress through the program synchronously. On average, about 10 percent of

students leave the program during the first year and an additional eight to nine percent during

the second year. Attrition during the third and fourth years is almost non-existent. Students

who receive a “C” or lower in a nursing course are placed on academic probation and required

to retake the course. Failure to successfully complete the course a second time triggers dismissal

from the program. The same is true of withdrawing from two nursing courses. During the last

semester, students take Community Health Nursing and Leadership and Management courses

and take the HESI exit exam. The former course focuses on the application of technical nursing

skills into a specific context while the latter emphasizes a softer set of competencies. The HESI

exit exam is taken early in the semester and students receive the scores almost instantaneously.

Those who score less than 850 on the HESI exit exam are mandated to go through remediation.

Remediation consists of reviewing nursing specific content on a one-on-one basis until they reach

an 80% score on an online HESI practice exam. The online practice exam is not

proctored/monitored and actual scores are not recorded on an official database. No students

have been reported to have not met the remediation requirements. Once the remediation

requirement is met the names of the students are sent to the state Board of Nursing so they may

be eligible to sit for the NCLEX-RN.

Differences exist between those who meet the HESI exit exam standard and those who

do not. As Table 1 points out, about 53 percent of the sample met the HESI standard and the

average difference in HESI exit exam scores between the two groups is almost 200 points. In

comparison to those who did not, those who met the standard are more likely to be white, have

higher SAT math scores, and have nursing course GPAs about 0.40 point higher (with the

Page 18: Three Essays on the Economics of Education Miguel E. Martinez

10

exception of the last two nursing courses). A twenty-four percentage point difference exists in

first-time NCLEX pass rates between those who met the HESI standard and those who did not.

Table 1. Summary Statistics for Analytical Sample

NCLEX

HESI>=850 HESI<850

n=392 n=351

Demographics

EOF Status 5% 16%

African American 8% 13%

American Indian 0% 1%

Asian 25% 28%

Hispanic 10% 14%

International 1% 1%

Multiracial 4% 4%

Native 1% 4%

Other 1% 0%

Unknown 3% 6%

White 49% 30%

Male 9% 13%

Campus 1 49% 43%

Campus 2 49% 54%

Campus 3 2% 2%

Academic Readiness

SAT Verbal Score 613 656

SAT Math Score 591 564

Academic Performance

Healthcare Delivery 3.65 3.28

Pathophysiology 3.63 3.22

Health Assessment 3.67 3.24

Foundations I 3.52 3.14

Childbearing Family 3.39 2.95

Health and Illness ICA 3.56 3.09

Health and Illness AOA I 2.89 2.36

Foundations II 3.61 3.30

Pharmacotherapeutics 3.58 3.14

Psych Mental Health 3.73 3.38

Health and Illness AOA II 3.54 3.27

Leadership & Management 3.67 3.55

Community Health 3.78 3.59

HESI Score 948 750

NCLEX Pass Rate 94% 70%

Page 19: Three Essays on the Economics of Education Miguel E. Martinez

11

1.5 Methodological Approaches Using an administrative dataset of more than 700 nursing students from a research

intensive university in the northeast covering calendar years 2009-2015, the first essay examines

the explanatory and predictive power of: a) pre-admission achievement, demographics, and

academic performance in nursing courses on passing the NCLEX-RN on the first attempt and

comparing them to the explanatory and predictive power of b) HESI exit exam scores on passing

the NCLEX-RN on the first attempt.

A simple logistic regression assesses the explanatory power of a). The model takes the

general form:

(1) Yi= β0 + βdΣXd + βpΣXp + βaΣXa + βyΣYeari + ei,

- where Y is passing the NCLEX-RN on the first attempt for student i, β0 is the intercept term, Xd

is a vector of student demographic variables (race, gender, and socio-economic status), Xp is a

vector of preadmission academic performance measures (verbal and quantitative SAT scores), Xa

is a vector of academic performance measures in nursing courses (Health Assessment;

Foundations of Nursing Practice I and II; the Childbearing Family; Pharmacotherapeutics; Health

and Illness of Infants, Children, and Adolescents; Health and Illness of Adults and Older Adults I

and II; Psychiatric Mental Health Nursing; Community Health Nursing; and Leadership and

Management), βy captures the cumulative impact of year fixed effects, and ei is an error term

with a mean of zero and variance of one. The betas capture the cumulative impact of their

respective vectors or variables.

A second model assesses the explanatory power of b). The model takes the general form:

Page 20: Three Essays on the Economics of Education Miguel E. Martinez

12

(2) Yi= β0 + βhXh +ei,

- where Y is passing the NCLEX-RN on the first attempt for student i, β0 is the intercept term, βh

captures the association between not meeting the HESI exam cutoff and the outcome, and ei is

an error term with the usual properties.

A third explanatory model will combine both models a) and b). The model will take the

general form:

(3) Yi= β0 + βdΣXd + βpΣXp + βaΣXa + βhXh + βyΣYeari + ei

- where the terms are as previously defined.

Since nursing programs find it useful to identify students most at risk of not passing the

NCLEX-RN on the first attempt, the predictive power of the three models will be tested as follows.

For models (1) and (3), three years of data (period p) will be used to predict NCLEX-RN outcomes

at p+1. Predicted outcomes will be compared to actual outcomes to identify the best performing

model (Hansen and Timmerman, 2012).

Finally, I explore a method to ascertain the degree to which students are severely

misplaced [assigned to remediation but passed the NCLEX-RN or was not assigned to remediation

but failed the NCLEX-RN] as a result of the current NCLEX-RN remediation protocol and to

investigate if other protocols could potentially reduce such misplacement. Inspired by Clayton

and Belfield (2014), I simulate what researchers/administrators could do if they had access to the

NCLEX-RN scores rather than NCLEX-RN pass/fail designations by calculating the severe error rate

for each of the three above-mentioned models for the Community Health Nursing course. The

Community Health Nursing course is ideal for this purpose since a) it is chronologically the last

Page 21: Three Essays on the Economics of Education Miguel E. Martinez

13

nursing-specific course that students take, b) it requires that all other nursing courses have been

successfully completed, c) the grade is granted after students take and received the HESI exit

exam scores and d) the distribution of grades is similar to NCLEX-RN pass/fail outcomes. About

11 percent of students in the sample earn a grade of B or lower in Community Health Nursing

which approximates the 10 to 11 percent of students who failed the NCLEX-RN on the first

attempt nationally during the period covered by the data (National Council Licensure

Examination for Registered Nurses, 2016).

Following Scott-Clayton and Belfield (2014), for individual students, the probability of

being severely misplaced is the sum of the probabilities of over-placement and under-placement.

For subgroups, it is average of the sum of those probabilities. The probability of being severely

under-placed is operationalized as:

Pr(Severally under-placed=1) = Pr(Grade of A=1)

if remediated, 0 else.

The grade of an “A” as the criterion to assess under-placement is not excessively high since the

average grade for nursing courses in the sample is barely below a 3.5 (B+). Moreover, more than

half of the students in the sample received an “A” in the course. The probability of the being

severely over-placed is:

Pr(Severally under-placed=1) = Pr(Grade of B or lower=1)

if not remediated, 0 else.

Approximately 11 percent of the sample earned a grade of B or below.

Page 22: Three Essays on the Economics of Education Miguel E. Martinez

14

Unlike the other analyses, severe error rates are derived by only using the observations

of students who were not subjected to remediation and extrapolating results to those students

who were remediated since the associations between the independent variables and the

outcomes have to be estimated net of remediation effects. Failure to do so would introduce

further bias into estimates. Those who have a probability higher than .50 will be considered to

have been under-placed or over-placed, respectively. The under(over)-placement rate is the

number of (non-)remediated students predicted to be under(over)-placed divided by the total

number of (non-)remediated students included in the model. The under-placement and over-

placement rates derived from two separate models – the aforementioned model (3) and a version

of model (3) without SAT scores. The latter approximates the data availability of typical two-

year nursing programs at community colleges where SAT/ACT scores are often lacking. The

misplacement rates of the two models are compared to those of the current progression

standard based on meeting the HESI exit exam cut-score. For the progression rule, the under-

placement rate is the number of remediated students who earned an “A” in the course divided

by the total number of remediated students. The over-placement rate is the number of non-

remediated students who earned a “B” or below divided by the number of non-remediated

students.

1.6 Results

1.6.1 Explanatory Models

Nursing programs need to know what student characteristics and/or academic pre-

admission measures are associated with passing the NCLEX-RN on the first attempt so they can

Page 23: Three Essays on the Economics of Education Miguel E. Martinez

15

admit students most likely to succeed and, once enrolled, can intervene if needed to. Table 2

below summarizes the point estimates for the full model for the entire period and for four-year

periods. The Foundations I course was the only nursing course found to have a statistically

significant relationship with the outcome in all the time periods covered. The Health and Illness

of Infants, Children, and Adolescents course also had a statistically significant association with

NCLEX-RN success for all but the period covering years 2012-2015. In periods 2011-2014 and

2012-2015, the Psych/Mental Health course was statistically significantly associated with the

outcome.

Table 2. Logistic regression results pooled and four-year periods

2009-2015

2009-2012

2010-2013

2011-2014

2012-2015

n=455 n=230 n=238 n=269 n=287

NCLEX Odds Ratio

Odds Ratio

Odds Ratio

Odds Ratio

Odds Ratio

Demographics EOF Status 1.15 0.99 1.30 1.03 2.55 American Indian 1.00 1.00 1.00 1.00 1.00 Asian 1.13 1.39 2.31 1.56 1.62 Hispanic 0.48 1.18 1.95 2.40 0.63 International 1.00 1.00 1.00 1.00 1.00 Multiracial 0.60 1.38 2.43 1.66 0.46 Native 1.67 0.73 3.98 2.99 4.23 Other 1.00 1.00 1.00 1.00 1.00 Unknown 0.40 1.00 1.00 0.81 0.30 White 1.08 1.63 1.62 0.98 0.90 Male 1.60 2.77 1.52 0.66 0.99 Campus 2 1.00 0.62 0.69 1.00 0.31 Campus 3 0.54 1.00 1.00 0.27 1.00

Academic Readiness SAT Verbal 1.00 1.01 1.01 1.00 1.00 SAT Math 1.00 0.99 1.00 1.00 1.00

Academic Performance HC Delivery 0.59 0.39 0.48 0.83 0.80

Page 24: Three Essays on the Economics of Education Miguel E. Martinez

16

Pathophysiology 1.40 1.77 1.29 1.29 1.01 Health Assessment 1.30 0.56 0.95 2.04 1.71 Foundations I 3.21†1 3.18† 4.36† 3.24† 4.94† Childbearing Family 1.54 1.62 2.46 1.07 1.35 Health and Illness ICA 3.03† 6.88† 5.02† 4.00† 2.45 Health and Illness AOA I 2.50 1.93 2.79 3.06 2.65 Foundations II 0.64 1.20 0.45 0.46 0.42 Pharmacotherapeutics 1.10 0.80 1.03 1.45 1.24 Psych Mental Health 2.55 1.23 1.45 4.12† 4.71† Health and Illness AOA II 1.15 1.60 0.89 0.32 0.80

Intercept 0.00 0.00 0.00 0.00 0.00

The figures below depict the average marginal probabilities for the three courses with

statistically significant relationships with the outcome for the 2009 to 2013 period and the 2014

to 2015. The latter period covers the years in which the NCLEX-RN exam cut-score was raised

from -0.16 logits to 0.00 logits. For the Foundations I course, the pattern is clear (Figure 1). In

comparison to the period 2014-2015, grades during the antecedent period have more

discriminating power. The average predicted probability ranges from 69 percent (C) to 93

percent (A) with the decreasing marginal rates. In contrast, average marginal probabilities for

the subsequent period only range from 73 percent to 77 percent with increases at a monotonic

rate. The difference in discriminating power between the two periods is similar for the Health

Assessment of Infants, Children, Adolescents (Figure 2) course but it is less pronounced. The

discriminating power for the Psych/Mental Health course, in contrast, significantly increases in

the latter period (Figure 3). For the 2009-2015 period, the average marginal probability of

1 † indicates statistical significance at 95% level. See Appendix 1.8 for greater detail.

Page 25: Three Essays on the Economics of Education Miguel E. Martinez

17

passing the NCLEX-RN on the first attempt is constant across the range of grades. The average

marginal probabilities in the adjacent period range from 36 percent (C) to 91 percent (A).

Figure 1. Predicted Average Marginal Probabilities of Passing the NCLEX-RN by Foundations I Course Grade.

Figure 2. Predicted Average Marginal Probabilities of Passing the NCLEX-RN by Health & Illness ICA Course Grade.

69%

77%

84%

89%

93%

73% 74% 75% 76% 77%

50%

55%

60%

65%

70%

75%

80%

85%

90%

95%

100%

C C + B B + A

Period 2009-2013 Period 2014-2015

65%

74%

81%

87%

92%

71%

74%

78%

81%85%

50%

55%

60%

65%

70%

75%

80%

85%

90%

95%

C C + B B + A

Period 2009-2013 Period 2014-2015

Page 26: Three Essays on the Economics of Education Miguel E. Martinez

18

Figure 3. Predicted Average Marginal Probabilities of Passing the NCLEX-RN by Psych/Mental Health Course Grade

Since the Foundations I course is taken during the first semester after completion of all

general education courses it the allows the program to intervene early in the program to optimize

the likelihood of passing the NCLEX-RN on the first attempt. Students take the Health and Illness

of Infants, Children, and Adolescents course during the second semester of the third year and the

Psych/Mental Health course during the first semester of the fourth year giving the program a

string of markers to assess if interventions are having the desired effects or if further

interventions are needed. The temporal instability of statistical relationships between given

performance on nursing courses presents a challenge only insofar as NCLEX-RN undergoes a

structural change such as a more rigorous standard as it was established beginning in 2014.

1.6.2 Out-of-Sample Predictive Models

In discerning the best progression rule for sitting for the NCLEX-RN exam and applying it,

nursing programs have to use previous results to extrapolate into the future. In this section, I

85% 85% 85% 85% 85%

36%

53%

69%

82%

91%

30%

40%

50%

60%

70%

80%

90%

100%

C C + B B + A

Period 2009-2013 Period 2014-2015

Page 27: Three Essays on the Economics of Education Miguel E. Martinez

19

use three years of data (period p) to predict NCLEX-RN outcomes at p+1 for models (1) and (3). I

then compare the predictive outcomes of the models and the current progression rule to actual

outcomes. Under the HESI-based progression rule, all students who do not score above a score

of 849 on the HESI exam exit are predicted to fail the NCLEX-RN while those who score above it

are predicted to pass it. Individuals with model-based probabilities (MBP) at or above .50 are

predicted to pass the NCLEX-RN.

Figure 4 below compares the predictive validity of models (1) and (3) and the current

progression rule that only uses HESI scores. With the exception of 2012, the current progression

rule (HESI Only) best identifies students who pass the NCLEX-RN. The model with HESI scores is

best for identifying students who will not pass the NCLEX-RN in years 2012 and 2013 while the

model without HESI scores is best for years 2014 and 2015. Overall, the results suggest a

predictive validity trade-off between identifying those students who fail and those who pass.

Figure 4. Predictive Validity of Models (1) and (3) and Current Progression Rule based on HESI Exit Exam.

Page 28: Three Essays on the Economics of Education Miguel E. Martinez

20

The optimal predictive method may be the one that minimizes that tradeoff. Overall,

HESI Only is the predictive method where the gap between correctly predicted NCLEX success

and failure is biggest for all years. MBP without HESI score is the only method that accurately

predicts failures for at least half of those who fail for all four years. MBP with HESI score, on the

other hand, rightly predicts the highest percentage of failures in years 2014 and 2015. The

change in NCLEX-RN cut-score starting in March of 2014 may, in part, explain the better

performance of MBP with HESI score in the last two years. The data suggest that, going forward,

MBP with HESI score may be the most appropriate method to identify students likely to fail and

HESI only is the best way to identify students likely to pass. Using both methods may be optimal

and, when model predictions contradict each other, side on the side of caution and assume that

the student is more likely to fail.

86%

97%100%

94%100%

94%

78%

88%

80%84%

93%87%

50%

20%24%

50%

24%

45%

69%

46%

71%

50%

28%

57%

0%

20%

40%

60%

80%

100%

120%

2012 2012 2012 2013 2013 2013 2014 2014 2014 2015 2015 2015

MBP woHESI

HESI Only MBP wHESI

MBP woHESI

HESI Only MBP wHESI

MBP woHESI

HESI Only MBP wHESI

MBP woHESI

HESI Only MBP wHESI

% of predicted to pass who actually passed % of predicted to fail who actually failed

Page 29: Three Essays on the Economics of Education Miguel E. Martinez

21

1.6.3. Severe Error Rates

Subjecting students to remediation who have a high probability of success without

remediation may be a sub-optimal allocation of resources and may result in negative unintended

consequences like poorer performance through moral discouragement and/or opportunity costs.

In this section, I assess the degree to which students are severely misplaced as a result of the

current NCLEX-RN remediation protocol and investigate if other protocols could potentially

reduce such misplacement using Community Health course outcomes as a proxy for NCLEX-RN

outcomes.

Model results converge. Students are much more likely to be under-placed rather than

over-placed. As Table 3 indicates, for both models, students are more likely to under-place than

to over-place by a factor of approximately five.

Table 3. Predicted Average Probabilities of Under-Placement or Over-Placement

Demographics Variables

Over-placed model (3)

Over-placed wo SAT Scores

Under-placed model (3)

Under-placed wo SAT Scores

n=224 n=300 n=156 n=245

Female 8.70% 7.62% 51.19% 43.62%

Male 11.11% 13.64% 23.10% 17.29%

EOF 40.00% 25.00% 45.77% 37.76%

non-EOF 7.95% 7.59% 48.29% 40.69%

Minority 14.63% 11.86% 47.30% 35.95%

White 2.70% 4.72% 49.01% 47.36%

All Students 8.97% 8.16% 47.80% 40.20%

Predicted probabilities mirror predicted misplacement rates. As Table 4 indicates, under-

placements are more likely by a factor of six or more. Both model-based predicted over-

placement rates are lower than the rate of the progression rule and model without SAT scores

has the lowest rate for “all students” and the majority of student subgroups. This pattern is

Page 30: Three Essays on the Economics of Education Miguel E. Martinez

22

slightly different for under-placement rates. The HESI progression rule has lower predicted

under-placement rates than model (3). The model without SAT Scores also has the lowest overall

under-placement rates.

As Tables 4 and 5 show, the model without SAT scores performs best for students in

general but not for each student subgroup. The model without SAT scores minimizes under-

placement and under-placement rates for three subgroups: male, non-EOF, and minority

students. This result may be driven by the differential predictive validity of SAT scores across

student subgroups (Mattern et al, 2008; Santelices et al, 2010). Model (3) performs best for white

students for both misplacement rates. The current HESI progression rule does not outperform

the models for any specific subgroup.

The convergence in predictions across the models can be leveraged to more accurately

place students. A simple business rule where students are identified as misplaced only when the

two models agree with the progression rule decreases severe error rates. The second column

of Table 6 captures the under-placement and over-placement rates under the above-mentioned

business rule. As can be seen in the second and third columns of Table 6, this approach of

including three distinct data sources eliminates over-placements and produces a modest

improvement in under-placement rates. The final column captures the results of adding a third

model to the business rule. The third regression model acting as the fourth data source

contains only an intercept and the nursing courses found in model (3). This model specification

may prove useful in combating perceptions that personal characteristics may be used to identify

students who could potentially benefit from remediation. The expansion of the business rule to

Page 31: Three Essays on the Economics of Education Miguel E. Martinez

23

include this model has marked effects on under-placement rates. It reduces the under-

placement rate by slightly more than 30 percent – from 39 percent to 27 percent. This inter-

model agreement approach proves beneficial in reducing severe misplacements.

Page 32: Three Essays on the Economics of Education Miguel E. Martinez

24

Table 6. Predicted Under/Over-Placement Rates By Number of Data Sources.

Demographics Variables

Under-placed Over-placed Under-placed

# of sources = 3 # of sources = 3 # of sources = 4

n=223 n=222 n=223

Female 43.88% 0.00% 30.10% Male 7.41% 0.00% 7.41% EOF 39.53% 0.00% 25.58% non-EOF 39.44% 0.00% 27.78% Minority 38.4%* 0.00% 24.00% White 40.00% 0.00% 30.00%

All Students 39.46% 0.00% 27.35%

1.7 Conclusion and Further Research I find that demographics, SAT scores, and performance in nursing courses provide

predictive power beyond that of the HESI but only for those who fail. Using them in statistical

models increases the percentage of students correctly predicted to fail by at least 50 percent

(2014) and, as much, as 100 percent (2012, 2013, and 2015). Nursing programs may improve

their NCLEX-RN pass rates by dedicating more resources to help students predicted to fail. Using

Community Health course outcomes as a proxy for NCLEX-RN outcomes, I also demonstrate that

model-based protocols could potentially reduce the number of students assigned to remediation

who will pass it with a high score and eliminate the number of students not assigned to

remediation who will fail it.

The analyses found in this paper could be further extended to include other four-year

nursing programs, community college nursing programs, and accelerated programs as well as

additional periods in which NCLEX-RN standard underwent a change to ascertain the

generalizability of findings and to investigate the possibility of non-HESI exit exam “universal”

Page 33: Three Essays on the Economics of Education Miguel E. Martinez

25

predictors of NCLEX-RN success. In addition, severe placement error rates should be calculated

using actual NCLEX-RN scores across various periods in which the passing standards changed to

investigate potential adverse impact of such changes on various student subgroups.

Page 34: Three Essays on the Economics of Education Miguel E. Martinez

26

Chapter 2: The Impact of Remediation on

NCLEX-RN Success: Positive, Neutral, or

Negative?

2.1 Introduction “That makes me nervous now. Our school drops us out of the program, if we do not pass

the HESI. They only give us 1 chance. There were plenty of people last year that failed the HESI and could not proceed.…The grades that you make and your average do not matter here when you take the HESI. You can make all A's through the nursing program, but if you fail that test, you are out” by Ivana (retrieved from http://allnurses.com/general-nursing-student/failed-the-hesi-137577-page2.html).

“In my case, our school required us to pass the Hesi Exit Exam with a 900 in order to officially graduate. Fellow students who failed the test had to take 5 weeks of remediation classes... it did not help the students at all. Most of them also failed the second, even the third time around.” by Claire (retrieved from http://www.yourbestgrade.com/hesi/letter/).

The experiences described by Ivana and Claire in their respective online fora, in some way,

capture key commonalities and divergences in the undergraduate nursing program market. Both

identify high stake testing as part of their programs. One offers remediation to assist their

students achieve desired performance levels while the other offers no such help. Despite their

apparent differences, both performance management approaches have the same aim – to

maximize the likelihood that their students will pass the NCLEX-RN (national licensure exam for

nurses) on the first attempt.

First time NCLEX-RN pass rates shape nursing programs. State Boards of Nursing make

first time NCLEX-RN pass rates publicly available as part of their charge to inform consumer

choice. Ceteris paribus, programs with relatively higher pass rates attract more applicants and/or

more qualified applicants, enroll and maintain them (Beeson and Kissling, 2001; Norton et al,

Page 35: Three Essays on the Economics of Education Miguel E. Martinez

27

2006). The first-time pass rate is also a discrete metric which can assist administrators

overseeing nursing programs to gauge their overall effectiveness and competitiveness (Giddens

et al, 2009; Taylor et al, 2014).

Regulations incentivize nursing programs to make remediation as effective as possible.

The Commission on Collegiate Nursing Education (CCNE) and the National

League for Nursing Commission for Nursing Education Accreditation (CNEA) require minimum

first-time NCLEX-RN pass rates to maintain accreditation. In addition, nursing programs must

also meet the NCLEX-RN pass rate standards established by their state boards of nursing which

may be higher than those established by national accrediting bodies. Prolonged inability to meet

standards could result in loss of accreditation and/or placement on probationary status (CCNE,

2014) (CNEA, 2015).

Nevertheless, nursing programs that offer remediation – like the one attended by Claire

- may not provide an advantage to its students. If the remediation is not effective then outcomes

may not be any different than those who do not offer them. The cycle of taking the HESI exit

exam (standardized exams used to determine eligibility to receive a nursing degree or sit for the

NCLEX-RN) may be extended through repeated testing but their probabilities of passing the

NCLEX-RN will not improve.

Remediation may not be effective for a variety of different reasons. The length of

remediation may be inadequate to compensate for knowledge not acquired over the course of

several years. If the established HESI exit exam performance is not met because of instructional

deficiencies then remediation conducted by the same instructors may not yield improved results.

Page 36: Three Essays on the Economics of Education Miguel E. Martinez

28

If students are not motivated to acquire the knowledge they did not learn during their regular

course of study then remediation may not help increase their probabilities of NCLEX-RN success.

Benchmarking practices may render remediation ineffective. Nursing programs set a

minimum HESI exam score to identify students who may benefit from remediation before taking

the NCLEX-RN (Langford and Young, 2013; Young and Willson, 2012; Zweighaft, 2013; Schreiner

at el, 2014). Most institutions set the cutoff score at 850. About 97 percent of students who

met or surpass the 850 benchmark have passed the NCLEX (Young & Willson, 2012; Lewis, 2006).

Not reaching a score of 850 on the HESI, however, does not indicate a sharp decrease in the

predicted probabilities of passing the NCLEX. More than 93 percent of students with HESI scores

between 800 and 849 pass the NCLEX pass the NCLEX-RN on their first attempt (Lewis, 2006).

The predicted probabilities at the lower end of the HESI score distribution are considerably lower.

Eighty-five percent of students with HESI scores between 700 and 799 pass the NCLEX-RN on the

first attempt while approximately seven out of ten of those with HESI scores between 600 and

699 do so (Lewis, 2006). Hypothetically, a remediation intervention that moves students 100

points along the HESI score distribution from 699 to 799 would increase the probability of passing

the NCLEX-RN on the first attempt by somewhere in the range of 15 percentage points while the

same hypothetical intervention for students with scores of 849 would increase their probability

of passing the NCLEX-RN around three percentage points. Thus, all things being equal, the

precise point at which programs establish the benchmark may affect the magnitude of the impact

of remediation, if one does indeed exist.

Using two national samples from the seventh and eight HESI exit exam validity studies

augmented by the sample from Chapter 1, the second economics of education essay aims to be

Page 37: Three Essays on the Economics of Education Miguel E. Martinez

29

assess the impact of required remediation on passing the NCLEX-RN on the first attempt through

regression discontinuity (RD). First, I find that the assumptions of RD as met. Next, I explore the

impact of required remediation using regression discontinuity design both as local randomization

and as continuity at the cutoff. As the former, I find some evidence that remediation has a

negative impact on passing the NCLEX-RN on the first attempt. As the latter, I find very limited

evidence that remediation may improve NCLEX-RN outcomes. Both sets of statistically significant

findings, however, are sensitive to bandwidth, functional assumptions and/or sample trimming.

Most of the RD evidence suggests that remediation has no impact on NCLEX-RN outcomes.

Logistic regression and propensity score matching (PSM) results do not contradict the overall

findings from the RD analysis. They suggest no association between remediation and success on

the NCLEX-RN. After running the robustness checks, I look for possible heterogeneous treatment

effects of remediation. Logistic regression results suggest that students who score below 600 on

the HESI score are more likely to pass the NCLEX-RN on the first attempt than those who do not

receive remediation. The results of PSM suggest otherwise.

Overall, I conclude that remediation most likely does not have an impact on students’

probability of passing the NCLEX-RN on the first attempt. This may be due to a ceiling effect

(Scott-Clayton and Rodriguez, 2012). Prior to remediation, remediated students just below the

modal cutoff of 850 have about a 93 percent chance of passing the NCLEX-RN – a few percentage

points above the national first-time pass rate. Measurably increasing the probability of passing

the NCLEX-RN of those students may not be possible. If remediation does have an impact, the

impact is likely different along the HESI score distribution. Students at the low end of the

distribution are most likely to benefit from it. I recommend that nursing programs lower their

Page 38: Three Essays on the Economics of Education Miguel E. Martinez

30

HESI benchmarks to increase per student allocation of remediation resources to students who

are most likely to benefit from it.

The structure of the essay is as follows. In Section 2, I review the current literature on

interventions to improve the likelihood of NCLEX-RN success. I then describe the analytical

sample. The identification strategy is articulated in Section 4. Section 5 demonstrates that the

assumptions of RD are met. RD estimates are discussed in the following section. In Section 6,

RD estimates are compared to logistic regression and propensity score matching (PSM)

estimates. Section 7 explores the possibility of heterogeneous treatment effects using logistic

regression and PSM. Finally, I summarize results and advance ideas for further research in

Section 8.

2.2 Review of the Literature In recent years, strategies to help students pass the NCLEX-RN on the first attempt have

received considerable attention in the literature. Overall, most of these studies are descriptive

or correlational and have modest sample sizes. Some even lack a control group. Seldom are the

interventions covered by the studies “single treatments”. Most often, the interventions involve

multiple simultaneous or staggered treatments. Like the larger remediation literature, the

studies reviewed in this section do not leverage adequate methodologies to capture causation

between interventions and outcomes (Bailey and Alonso, 2005).

The literature suggests that remediation is associated with improved NCLEX-RN

outcomes. Bondmass, Moonie, and Kowalski (2008) find an almost 9 percent point increase in

first-time NCLEX-RN pass rates after implementation of a remediation program. Heroff (2008)

Page 39: Three Essays on the Economics of Education Miguel E. Martinez

31

reports a 17 percent point increase in first-time pass rates after the introduction of an ATI exit

exam with an accompanying progression rule and remediation policy. Morton (2006) evaluates

the effect of structured remediation throughout the nursing program of study on NCLEX-RN pass

rates. Students who did not meet the benchmark score on HESI end-of-course exams and were

required to attend structured learning assistance workshops to review course specific content

had higher first-time NCLEX-RN pass rates than those who did not. Norton, Relf, Cox, Farley,

Lachat, Tucker and Murray (2006) find that the introduction of end-of-course and exit

standardized testing, a remediation course, and higher progression policy with NCLEX outcomes

increased the first-time NCLEX-RN pass rates by 11 percentage points. Schroeder finds an

eight-percentage point difference in first time pass rates between students under a new

testing/remediation policy and their predecessors (2013).

Studies on curricular interventions yield mixed results. Morris and Hancock investigate

the results of the introduction of new curriculum on HESI exit exam and NCLEX-RN results (2008).

The authors find no statistically significant difference in results between last cohort of graduates

under the old curriculum to the first cohort of graduates of the new curriculum. Frith, Sewell,

and Clark (2005) report that introduction of a capstone course and remediation in the form of

developing a study plan based on performance on the HESI exit exam increased first-time NCLEX-

RN pass rates. Lyons (2008) investigates the relationship between a problem-based NCLEX-RN

review course and NCLEX-RN outcomes. The four-month review course took place during

students last semester of study and was designed to increase their critical thinking as measured

by the ATI Critical Thinking Test. After students completed a baseline critical thinking test, the

author randomly assigned the 54 students to either the problem-based review course

Page 40: Three Essays on the Economics of Education Miguel E. Martinez

32

(treatment) or the traditional lecture based (control) to assess the impact of the problem-based

methodology. Groups were comparable in the baseline critical thinking test, overall GPA, nursing

GPA, ACT scores, and age. No statistically significant difference was found between treatment

and control groups in the post critical thinking test. The treatment group, however, had higher

first-time NCLEX-RN pass rates.

The literature associates interventions aimed at improving outcomes of specific student

subgroups with positive outcomes. Sifford and McDaniel investigate the impact of remediation

and test-taking course among at-risk students identified through HESI exit exam results (2007).

The authors find a statistically significant difference between pre and post exit exam results

among course participants. Parrone, Sredl, Miller, Phillips, and Donaubaur (2008) find that a

program that identified at risk students, provided students with faculty mentoring, individualized

remediation services, and instituted progression/graduation based on HESI exit exam, improved

NCLEX-RN pass rates from 70 percent to 100 percent. Sutherland, Hamilton, and Goodman

(2005) find that the Affirming At-Risk Minorities for Success (ARMS) program improved retention

rates and performance on a capstone course and mitigated the association of minority status

with passing the NCLEX-RN.

Overall, the literature suggests that the probability of students to pass the NCLEX-RN can

be improved through various interventions. Remediation has been consistently found to

improve the probability of NCLEX-RN while curricular interventions have proved less so. The

remediation literature, however, relies on methodologies that cannot establish causation. This

study makes a contribution to the NCLEX-RN remediation literature by using, for the first time, a

Page 41: Three Essays on the Economics of Education Miguel E. Martinez

33

quasi-experimental technique (regression discontinuity) to establish a causal link between

remediation and NCLEX-RN outcomes.

2.2.1 HESI Exit Exam

In practice, the HESI exit exam is used as a formative or as a summative assessment.

Figure 5 depicts the various ways nursing programs use the HESI exit exam to determine if

students will be allowed to take their licensure exam.

Figure 5. Possible Impacts of HESI exit exams scores on ability to sit for the NCLEX-RN.

Nursing programs administer the HESI exit exam during students’ last semester. Programs

with remediation may ask students who did not meet the cutoff to remediate until they are able

to reach the HESI exam exit cutoff. The number of attempts allowed to reach the HESI exit exam

cutoff may be limited. Those who reach it become eligible to sit for the NCLEX-RN while those

who do not will not get their nursing licenses. Programs with required remediation may not

require retaking of the HESI exit exam. Instead, they may ask students to correctly answer a

certain number or percentage of review questions or to meet with faculty members a certain

number of times to go over content-specific materials. Programs that offer remediation but do

not require it view the HESI exit exam as a formative assessment. In these cases, students review

their respective HESI exit exam scores by subject areas and concentrate their review efforts on

Completion of

General

Education and

Nursing Courses

HESI Exit Exam

(Cutoff)

Remediation

a) Required

b) Non-Required

redN No Remediation

c) Removal

d) Review

redN

Page 42: Three Essays on the Economics of Education Miguel E. Martinez

34

the areas they deem most in need of improvement. Some programs that do not offer

remediation may also use the HESI exit exam as a way to identify what content to review while

others use it as the definitive measure of NCLEX-RN readiness. For the latter, those who do not

meet the HESI exit exam benchmark are automatically kicked out of the program. No publication

exists that reports the number of students who are not allowed to sit for the NCLEX-RN because

of failure to meet the HESI exit exam cutoff or to meet remediation requirements.

2.3 Data Not all students in the full sample are in programs that require remediation. Thirty seven

of the 141 schools do not have compulsory remediation. As such, approximately one third of

the 10,022 students in the combined sample are not required to remediate if they do not meet

their respective HESI benchmark or do not have a benchmark at all. As Table 7 points out, among

those that require remediation, the distribution of students and institutions mirror each other.

850 is the modal HESI benchmark option that covers about two thirds of institutions and students

while 900 is the second most popular choice. This is not coincidence since the maker of the HESI

recommends “that students seriously remediate any subject area category in which they

obtained a score of less than 850 …” and also maintains that the “recommended level is 900”

(2014).2

Table 7. Distribution of Schools and Students by HESI Benchmark: Required Remediation

HESI Benchmark Schools # Schools % Students Students %

700

2 RD analyses are restricted to programs with required remediation with benchmarks of 850 and 900 to minimize the possibility of cutoff levels not being exogenously determined.

Page 43: Three Essays on the Economics of Education Miguel E. Martinez

35

725

750 1 1% 57 1%

800 1 1% 30 0%

825 1 1% 25 0%

850 69 66% 4,464 67%

875 1 1% 166 2%

900 15 14% 777 12%

950 1 1% 146 2%

Not Applicable 15 14% 981 15%

Total 104 100% 6,646 100%

All students in the full sample have HESI exit exam scores and NCLEX-RN outcomes, the

type of program attended (Bachelor of Nursing Science [BSN], Associate Degree in Nursing [ADN],

Practical Nursing and Diploma), remediation policy of their program, and HESI benchmark.

Approximately 725 students of those students also have demographics, SAT scores and nursing

course performance data.

Theoretically, for the subsample of almost 725 students, regression discontinuity (RD) as

a “local randomized experiment” can be leveraged to obtain unbiased estimates of treatment

effects of required remediation based on a HESI benchmark since: 1. HESI scores are anterior to

the treatment (i.e. HESI exam is taken prior to the assignment to treatment [remediation] or

control groups), 2. the HESI benchmark is exogenously determined (i.e. established prior to

assignment to treatment [remediation] or control groups, and 3. students are not able to

manipulate their score on the HESI. When the above conditions are met then the groups of

students just above it and just below should be comparable if not identical, with the only

difference between them being remediation (Trochim, 1984; Lee and Lemieux, 2010). This

situation approximates a randomized control trial (RCT) such that researchers can obtain

unbiased estimates of the local average treatment effect (LATE) of remediation. The relatively

Page 44: Three Essays on the Economics of Education Miguel E. Martinez

36

small sample size, however, would only allow for the detection of a very large treatment effect.

Instead, the maximum bandwidth at which covariate balance between treatment and control

groups in the subsample will be used as the maximum bandwidth at which the entire sample may

have covariate balance. Using the full sample, treatment effects will be estimated using

bandwidths of 10 points and less.

2.4 Identification Strategy The identification strategy restricts itself to local estimates. The econometric RD model

to identify the LATE takes the following general form (Jacob et al, 2012):

Yi = β0 + β1Remediationi* HESIi + β2HESIi + εi

- where i indexes students, Y is passing the NCLEX-RN on the first attempt, β0 is the average value

of the outcome for those in the treatment group conditional on the HESI score, β1 captures the

local average treatment effect [where Remediation=1 if x<HESI exit exam cutoff score and

Remediation=0 otherwise] and allows for different slopes on either side of the cutoff, β2 captures

the relationship between the outcome and residual differences of HESI scores between the

treatment and control group, and ε is an error term with the usual properties. School fixed

effects and school type are not included in the model because of collinearity. Their inclusion

reduces the sample by almost one third. 3

Sensitivity analysis will test the stability of estimates using the full and trimmed samples.

Although it is not likely that the assignment variable was manipulated by students the McCrary

3 Although not discussed in the essay, the local treatment estimates were nevertheless calculated with school fixed effects and school type in the models. None of the estimates were statistically significant.

Page 45: Three Essays on the Economics of Education Miguel E. Martinez

37

and Cattaneo, Janssen, and Ma (CJM) test will be performed to assess the presence of a

discontinuity in the density of observations at the cutoff (McCrary, 2008; Cattaneo, Janssen, and

Ma, 2016). Estimates of treatment effects will be derived by employing both local linear

regression and polynomials. Optimal bandwidth will be determined by three bandwidth

selectors: Ludwig and Miller (2007), Imbens and Kalyamanaraman (2012), and Calinco, Cattaneo,

and Titiunik (2014). Polynomials are capped at the third order to avoid noisy estimates (Gelman

and Imbens, 2014; Gelman and Zelizer, 2015).

Although not presented in the essay global effects were estimated with models that

included school fixed effects and school type using first, second, third, fourth, fifth and sixth order

polynomials with their respective interactions with the treatment term (Jacob et al, 2012). The

results are not presented because a) none of the treatment estimates were statistically

significant and b) the cross-validation band selectors for local second and third order polynomials

cover almost the entirety of the sample space – making them virtually global treatment

estimates.

2.5 The Assumptions of Regression Discontinuity are Met Heaping on either side of the cutoff may bias causal estimates of the treatment effect of

remediation based on HESI exit exam scores. If remediation is effective students who are able

to manipulate their scores just enough to meet the cutoff may upwardly bias estimates of

remediation effects since those are who manipulated the HESI score would be less ready for pass

the NCLEX-RN score than those just below the threshold who were helped by the remediation.

Conversely, if the remediation has a negative impact on NCLEX-RN performance then the

Page 46: Three Essays on the Economics of Education Miguel E. Martinez

38

treatment effect estimates would be downwardly biased. Students may also have an incentive

to manipulate HESI exam scores just below the established benchmark. If not meeting the

benchmark does not carry a significant penalty (i.e. lower course grade, losing eligibility to

graduate and/or sit for the NCLEX-RN) and remediation grants access to NCLEX-RN readiness

resources not available otherwise (i.e. one-on-one tutoring and mentoring) then students may

also have an incentive to manipulate HESI exam exit scores just below the benchmark. In this

case, if remediation has a negative impact the treatment effect estimates may be upwardly

biased. If the impact of remediation is negative, treatment estimates would be biased in the

opposite direction. If remediation has no impact, the direction of the bias would upward if

manipulation lands students above the cutoff and downward if they land below it. Table 8

captures the direction of bias based on the direction of manipulation and treatment

effectiveness. Under the assumption that not meeting the benchmark carries more disincentives

than incentives, if present, the modal manipulation would be above the benchmark.

Table 8. Direction of Bias Based on Remediation Effectiveness and Direction of Manipulation Remediation Positive Remediation Negative Remediation Neutral

Manipulation Above Upward Bias Downward Bias Upward Bias Manipulation Below Downward Bias Upward Bias Downward Bias

2.5.1 Manipulation of the Running Variable: Visual Evidence

I find no visual evidence of manipulation of the running variable induced by requiring

remediation for not meeting a HESI exit exam benchmark. Figure 6 displays the distribution of

centered HESI exit exam scores in 30 point bins – one fourth of a standard deviation- by required

Page 47: Three Essays on the Economics of Education Miguel E. Martinez

39

remediation policy status. 4 Both distributions are normal and have their peaks to the right of

their benchmarks (indicated by the red reference line). Most of the observations in both

distributions are also found to the right of the benchmarks. The median centered HESI score is

20.5 for the students in schools with required remediation and 40.5 for those without. This

suggests that whatever heaping exists around the cutoff may not be triggered by something other

than the required remediation policy.

Figure 6. Distribution of HESI Exit Exam Scores by Required Remediation Status

2.5.2 Manipulation of the Running Variable: Statistical Evidence

I find that required remediation does not induce manipulation of the running variable. I

use the McCrary test (2008) using a variety of a bandwidths. I start with a 5-point bandwidth and

increase it by 5 point increments until they reach the maximum of 250 points (slightly less than

two standards deviations). None of the 47 bandwidths tested yield statistically significant results.

4 Bin widths of 15 and 7.5 display the same pattern. See Appendix L.

0

.002

.004

.006

-1000 -500 0 500-1000 -500 0 500

Remediation Not Required Remediation Required

Den

sity

Centered HESI Exit Exam Score

Page 48: Three Essays on the Economics of Education Miguel E. Martinez

40

See results in Appendix B. Results of the heaping using test proposed by Cattaneo, Janssen,

and Ma (CJM) also suggest no manipulation. See Appendix C for results.

2.5.3 Continuity of Pre-treatment Variables: Statistical Evidence

The full sample of schools with required remediation and a benchmark of 850 or 900 does

not have demographics, academic readiness, and nursing course performance data. More than

720 students at one of these schools with a benchmark of 850 have such data. To assess the

continuity assumption needed to derive unbiased treatment effects of remediation under the

characterization of regression discontinuity design as discontinuity at the cutoff, each

demographic and academic variable is treated as a dependent variable in a regression

discontinuity model under three functional specifications (Hahn, Todd, and van der Klaauw,

1999). If any of the variables display at discontinuity at the cutoff point then the effect of

remediation could be confounded with that of those variables. Some evidence exist that the

continuity assumption is not met for a single nursing course but the finding is highly sensitive to

functional form and does not hold under a falsification exercise. I, thus, conclude that the

continuity assumption holds. See Appendix D for results and further discussion.

2.5.4 Covariate Balance: Statistical Evidence

Covariate balance in pre-treatment variables to assess the unconfoundedness

assumption of regression discontinuity design as a localized experiment is also tested using, once

again, the subset of the full sample with pre-treatment measures (Lee and Lemieux, 2010). Table

9 captures the differences in the student characteristics, measures of pre-admissions academic

aptitude and performance in nursing courses 10 points above and below the cutoff of 850. There

are no statistically significant differences at that bandwidth and at any smaller bandwidth.

Page 49: Three Essays on the Economics of Education Miguel E. Martinez

41

Statistical significant differences in performance in nursing courses begin to appear when the

bandwidth is extended beyond 10 points. For student characteristics and pre-admissions

academic aptitude measures, the bandwidth can be extended to 50 points before statistically

significant differences are detected between students subjected to remediation and those who

were not. See Appendix E for details. The findings of balance in all pre-treatment covariates

within 10 points of the benchmark are extrapolated to the larger samples to find the local average

treatment effect (LATE).

Table 9. Summary of Pre-Treatment Variables by Remediation Status: 10 Point Bandwidth

Variables No Remediation Remediation Difference

Student Characteristics

Non-White 63% 62% 1%

White 37% 38% -1%

Male 10% 5% 4%

Main Campus 48% 57% -13%

EOF 4% 5% -1%

Pre-admission Academics

Composite SAT 1110 1145 -35

SAT Verbal 549 566 -17

SAT Math 561 579 -18

Nursing Courses

Pathophysiology 3.4 3.5 -0.1

Health Assessment 3.5 3.3 0.2

Foundations I 3.4 3.1 0.3

Child Bearing Family 3.0 3.1 -0.1

Health and Illness Children 3.3 3.5 -0.2

Health and Illness Adults I 2.5 2.7 -0.2

Foundations II 3.2 3.3 -0.1

Research Process 3.2 3.3 -0.1

Pharmacotherapeutics 3.2 3.4 -0.2

Psych/Mental Health 3.7 3.4 0.2

Health and Illness Adults II 3.3 3.4 -0.1

Observations 22 21

% of treatment/control 6% 7%

**statistically significant at 95% level *** statistically significant at 99% level

Page 50: Three Essays on the Economics of Education Miguel E. Martinez

42

2.6 Results In this section, RD estimates are discussed and then followed by an overall conclusion

2.6.1 Estimation of Treatment Effects: RD as Discontinuity at the Cutoff

Treatment effects are estimated using three different bandwidth estimators under first,

second, and third order polynomial specifications. All specifications used a triangular distribution

that gives more weight to observations near the cutoff (Pereillon, 2013). As Table 10 points

out, all cross-validation (CV) estimates are not statistical significant. It is the same for estimates

derived using the CCT bandwidth selector. The second order specifications of the models using

the Imbens and Kalyamanaraman (IK) selector, however, yielded negative statistically significant

results.

Table 10: Treatment Effects Coefficients By Bandwidth Selector

Bandwidth Selector CV IK CCT

Polynomial Order 1st 2nd 3rd 1st 2nd 3rd 1st 2nd 3rd Bandwidth (b) 184 462 438 108 74 91 75 111 134 Bandwidth (h) 184 462 438 79 134 77 119 162 189

Methodology β β β β β β β β β

Conventional (b) 0.01 0.01 0.02 0.03 0.03 (0.02) 0.03 0.02 0.01 Bias-corrected (h) 0.02 0.02 0.03 0.02 (0.26) †† (0.05) 0.03 0.02 0.00 Robust (h) 0.02 0.02 0.03 0.02 (0.26) † (0.05) 0.03 0.02 0.00

†statistically significant at 95% level ††statistically significant at 99% level

Statistically significant results are sensitive to trimming of the sample by five percent of

the observations at either end of the distribution. Trimming renders the bias-corrected estimate

for the second order polynomials statistically insignificant.5 Robust estimates lose their

5 Conversely, trimming the sample by 5 percent for all the model specifications covered

in Table 4 does not generate any additional statistically significant results.

Page 51: Three Essays on the Economics of Education Miguel E. Martinez

43

significance when the bandwidths are decreased by 10 or 20 percent. See Appendix E for results

and discussion. “RD as a local randomized experiment” estimates also suggest no treatment

effects. They can be can be found in Appendix F.

2.6.2 RD Estimates Conclusion and Discussion

Overall, most RD estimates suggest that remediation has no impact on the probability of

passing the NCLEX-RN on the first attempt. This may be the result of a ceiling effect (Scott-

Clayton and Rodriguez, 2012). RD results that suggest an impact are sensitive to kernel functional

assumptions, sample trimming, and/or bandwidth. Resources allocated to remediation by both

nursing programs and students alike do not seem to be having their intended impact. The lack

of effectiveness may also be in part to their large percentage of students who are required to

remediate. From the nursing programs with HESI exam exit with required remediation and

benchmarks of 850 or 900, about two out of five students partake in remediation. Nursing

programs may not have adequate resources to cover such a large proportion of students who are

finishing their degrees. Also, since so many students are asked to remediate remediation may

also be regarded as a “common and normal experience”. The expectation/high probability of

having to remediate may be a disincentive to students to dedicate the necessary time to study

for the HESI exit exam.

2.7 Robustness Checks To assess the stability of the potential impact of the remediation on the likelihood of

passing the NCLEX-RN on the first attempt in the previous sections, outcomes of students subject

to remediation in institutions where remediation is required are compared to students in

Page 52: Three Essays on the Economics of Education Miguel E. Martinez

44

institutions where remediation is not required over the same range of HESI exit exam scores.

The comparison is done in two ways: logistic regression analysis and propensity score matching.

For both approaches, the analytical sample is restricted to institutions with required

remediation and that have benchmarks of 850 and those institutions that do not require a

remediation and do not have a benchmark. This ensures that all the observations from the

institutions with required remediation receive the intervention and all the observations from

institutions that do not require remediation do not receive the intervention.

2.7.1 Logistic Regression

The logistic regression model takes the following form:

Yi = β0 + β1Remediationi + β2HESIi + βjInstitutioni + βdInstitutionTypei + εi

- where i indexes students, Y is the outcome, β0 is the average value of the outcome for those in

the treatment group conditional on the HESI score, β1 captures the association between

remediation and the outcome [where Remediation=1 if x<850 HESI exit exam score, attendance

at institution where remediation is required with a benchmark of 850, Remediation=0

otherwise], β2 captures the relationship between the outcome and HESI scores for both the

treatment and control group, each βj captures institutional fixed effects and each βd captures the

association between institutional type and the outcome.

The results of the logistics regression analysis converge with those of the regression

discontinuity analysis. As Table 11 points out, holding constant school fixed effects and school

type, no statistically significant association was detected between remediation and passing the

NCLEX-RN on the first attempt. The average marginal effect of treatment was also statistically

Page 53: Three Essays on the Economics of Education Miguel E. Martinez

45

not significant and had a coefficient of .13. Including observations with HESI exit exam scores

above 850 and adding a dummy variable indicating attendance at an institution with required

remediation policy does not change the direction or the significance of the treatment variable.

See Appendix F.

Table 11. Logistic Regression Results (controlling for fixed school effects and school type effects)

Variable Odds Ratio Std. Err. z P>z UB LB

HESI Score 1.01 0.00 12.50 0.00 1.01 1.01

Treatment 2.34 1.90 1.05 0.30 0.48 11.47

Marginal effects over the HESI score response surface suggest the possibility of positive

treatment effects at the lower end of the HESI score distribution. Although statistically

insignificant at each 25-point interval, Chart 4 illustrates that the marginal effects of remediation

increase at a decreasing rate up to the 600-point mark. Starting at the 700-point mark, marginal

effects decrease at an almost monotonic rate. This suggests the possible existence of

heterogeneous treatment effects which will be explored in Section 2.8.

Figure 7. Remediation Marginal Effects over the HESI Score Response Surface

Page 54: Three Essays on the Economics of Education Miguel E. Martinez

46

2.7.2 Propensity Score Matching Analysis

To potentially reduce bias induced by imbalance in the pre-treatment variables,

propensity score matching is leveraged to estimate treatment effects. The matching model takes

the form:

Yi = β0 + β1HESI + βdInstitutionTypei + εi

- where Yi is remediation and the rest of the terms are as previously defined. Since, in the

restricted sample, in any given school all students are either subject to remediation or not subject

to remediation, institutional effects cannot be included in the model. The matching model used

logistic regression with no replacement. Covariate balance was achieved. See Appendix E.

Results were not sensitive to symmetrical trimming of the sample by five percent. In the

outcome model, observations were restricted to the area of common support. See Appendix F

for graph of area of common support.

0%

2%

4%

6%

8%

10%

12%

14%

16%

18%

20%

400 425 450 475 500 525 550 575 600 625 650 675 700 725 750 775 800 825 850

HESI Score

Page 55: Three Essays on the Economics of Education Miguel E. Martinez

47

The PSM results suggest no treatment effect. Table 12 summarizes pre and post

matching differences between the two groups. Prior to matching, students who received

remediation in schools with required remediation were statistically less likely to pass the NCLEX-

RN on the first attempt. Once matched, the difference disappears. The finding holds with

covariate adjusted PSM6 7. Overall, the PSM results converge with the logistic regression analysis

results.

Table 12: Pre and Post-Matching Differences Outcomes between Treated and non-Treated Variable Sample Treated Controls Difference S.E. T statistic

NCLEX-RN Pass Unmatched 0.80 0.83 -0.03 0.02 2.00 NCLEX-RN Pass ATT 0.79 0.83 -0.03 0.04 0.82

2.8 Looking for Heterogenous Treatment Effects8 Regression discontinuity cannot detect treatment effects far from the cutoff. In this

section, logistic regression and PSM are used to ascertain the possible existence and magnitude

of such effects.

6 Running the matched sample (in the area of common support) through an outcome model containing the same independent variables as the matching model. 7 The covariate adjusted PSM results are not the result of the violation of one or more of the three conditions needed for unbiased estimates under the usual PSM assumptions (Rubin, 2001). Roughly, the three conditions are: a) the difference in average propensity scores must be less than half a standard deviation, b) the ratio of the variances of the propensity score in the two groups must be more than one half and less than two and c) the ratio of the variances of the residuals of the explanatory variables after adjusting for the propensity score must be must more than one half and less than two (Rubin, 2001). All three conditions are met. Specifically, the results approximate the ideal for all the conditions – parity for conditions b and c and .01 for condition a. 8 Using the same sample as in Section 2.7

Page 56: Three Essays on the Economics of Education Miguel E. Martinez

48

2.8.1 Logistic Regression

To test for the presence of heterogeneous treatment effects of remediation, the following

model is formulated:

Yi = β0 + β1Remediationi + β2HESIi + βjInstitutioni + βdInstitutionTypei +βbBelow600i + βiBelow600 x Remediation + εi

- where the terms are as previously defined, βb quantifies the association between having

a HESI score below 600 and the outcome and βi captures the difference between the

observations below 600 who received remediation and those that did not.

The results from Table 13 suggest that students who score below 600 on the HESI exit

exam are less likely to pass the NCLEX-RN on the first attempt (marginal effect of -.12). At the

same time, those students with below 600 on the HESI exit exam who receive remediation are

more likely to pass the NCLEX-RN than those who did not receive remediation (marginal effect of

.14).

Table 13. Logistic Regression Results (controlling for fixed school effects and school type) Variable Odds Ratio Std. Err. z P>z UB LB

Remediation 2.02 1.70 0.83 0.41 0.39 10.52 HESI Score 1.01 0.00 9.36 0.00 1.01 1.01 Below 600 x Remediation 2.49 1.18 1.93 0.05 0.98 6.29 Below 600 0.45 0.19 -1.87 0.06 0.19 1.04

2.8.2 Propensity Score Matching

Once the analytical sample is further reduced to observations with HESI scores of 600

points and below, non-covariate adjusted PSM results do not support the logistic regression

findings of heterogeneous treatment effects among those students with HESI score below 600.

Table 14 summarizes the PSM results for the subgroup at the low end of the distribution.

Page 57: Three Essays on the Economics of Education Miguel E. Martinez

49

Matching achieves covariate balance. See Appendix I. Prior to matching, students who received

remediation were 20 percent points more likely to pass the NCLEX-RN on the first attempt. Once

matched, the statistically significant difference disappears.

Table 14: Pre and Post-Matching Differences Outcomes between Treated and non-Treated

Variable Sample Treated Controls Difference S.E. T statistic

NCLEX-RN Pass Unmatched 0.48 0.28 0.20 0.1 1.98

NCLEX-RN Pass ATT 0.50 0.56 -0.06 0.2 0.37

The covariate adjusted PSM result also provides evidence that does not support the regression

analysis findings of heterogeneous treatment effects among those students with HESI score

below 600.

2.9 Conclusion and Policy Implications Overall, remediation most likely does not have an impact on students’ probability of

passing the NCLEX-RN on the first attempt for most students. Evidence suggests a possible

positive treatment effect but restricted to the low end of HESI score distribution. Lowering HESI

benchmarks to increase per student allocation of remediation resources to students who are

most likely to benefit from it may improve NCLEX-RN outcomes.

State Boards of Nursing are charged with informing consumer choice. Historically, they

have fulfilled this obligation by making publicly available lists of accredited nursing programs

along with their first-time NCLEX-RN pass rates. For prospective consumers, first-time pass rates

are the primary indicator of program quality.

Page 58: Three Essays on the Economics of Education Miguel E. Martinez

50

Nursing programs, however, can artificially inflate their first-time pass rates by restricting

the number of their students who will be allowed to sit for the NCLEX-RN. Exit exams have been

the primary tool to restrict access to the NCLEX-RN. In some instances, students who do not

meet exit exam cutoff are automatically removed from the program. In other cases, students

who do not reach the exit exam cutoff are required to remediate and retake the exit exam until

they meet the cutoff. If they do not meet the cutoff they cannot take the NCLEX-RN. Still in

other cases, those who do not meet the exit exam threshold are required to remediate but not

to retake the test. If students refuse to remediate or do not meet the remediation standard then

they are not allowed to sit for the NCLEX-RN.

To address the potential for manipulation of first-time NCLEX-RN pass rates,

National Council of State Boards of Nursing (NCSBN) should encourage State Boards of Nursing

to develop and make publicly available new metrics of nursing programs productive efficiency.

The first-time pass rate is easily calculated and understood. For a given program, it is the number

of students who passed the NCEX-RN on the first attempt divided by the number of students who

took it for the first time. In addition to first-time pass rates, State Boards of Nursing should

employ two simple cohort-based metrics. The first would be the number of students who passed

the NCLEX-RN on the first attempt within a year of expected graduation divided by the number

of students in a given cohort. In practice, this would treat the students who left the program or

did not take the NCLEX-RN [either because they were not allowed to take it or because they chose

not to] as having failed it. Thus, it would “punish” programs for attrition. The metric would give

the public, the average probability of passing the NCLEX-RN on the first time of a student upon

entering the program. The second metric would be the number of students who were not

Page 59: Three Essays on the Economics of Education Miguel E. Martinez

51

allowed to sit for the NCLEX-RN after successfully completing their nursing courses divided by the

number of students in their cohort. This “sequester rate” would help state boards of nursing

and state nursing workforce centers identify how many students are potentially kept from the

nursing labor force, and, as such, may be contributing to nursing shortages. This metric would

be of particular interest to the public since it would indicate the percentage of students who after

having allocated considerable resources to their nursing course of study and having performed

well on nursing courses were not allowed even to sit for the NCLEX-RN.

Since exit exams have robust predictive validity for those who will pass the NCLEX-RN but

not for those who will fail, state Boards of Nursing should prohibit nursing programs from not

allowing their students to sit for the NCLEX-RN solely based on exit exam scores and/or

subsequent remediation outcomes. If exit exams are used in the decision to prohibit students

from taking the NCLEX-RN the cutoff score should not be above the score associated the

minimum pass rates established for programs to stay out of probation by state boards of nursing.

For example, in a state where the state board of nursing sets the minimum first-time pass rates

for nursing program to maintain accreditation at 80 percent, the HESI exit exam cutoff could not

exceed 650. In the analytical sample, the implementation of such a cutoff would reduce the

number of students required to undergo remediation by approximately 90 percent.

The current essay is limited by its data. The similarity of curricula across nursing programs

and increasing use of HESI end-of-course exams could facilitate the possibility of having

standardized nursing course outcome data to include in analyses. The inclusion of such data

could improve the precision of treatment estimates and provide the necessary inputs to test

covariate balance as localized random experiment.

Page 60: Three Essays on the Economics of Education Miguel E. Martinez

52

Chapter 3: Testing a Rule of Thumb: For STEM

Degree Attainment, More Selective is Better

3.1 Introduction The rise of the internet and other digital technologies in the mid to late 1990’s increased

demand for employees with STEM backgrounds (Coble and Allen, 2005). During that time, the

wage gap between the college educated and the rest of the population continued to widen at an

accelerated rate as markets became increasingly liberalized (Autor and Dorm, 2009). Thus, post-

secondary education in technical fields became increasingly perceived not only as a way to stay

globally competitive but to achieve equitable social outcomes domestically (Russell and Atwater,

2005) (Ma, 2009). By 2005, more than three fourths of the fastest growing fields were in science

and technology (Coble and Allen, 2005). From 2004 to 2008, the demand for science and

engineering workers increased at double the rate of the non-STEM sectors (Bureau of Labor

Statistics, 2008).

The supply of STEM workers depends to some degree on the ability of post-secondary

institutions to keep those students already interested in STEM engaged and to generate interest

among those initially not interested. High attrition rates among college students in STEM are

well recorded and give pause to high school students considering STEM college majors and their

guidance counselors and parents (Grandy, 1998; Bonous-Hammarth, 2000; Benbow et al, 2010).

About half of STEM majors do not get their bachelor’s degree within six years (Chen, 2013). The

institutional attributes which may exert positive or negative influences on STEM degree

attainment are many and students and parents may not be able to assess the status of each factor

Page 61: Three Essays on the Economics of Education Miguel E. Martinez

53

or a bundle of factors for specific institutions in their college choice sets. In this regard, a rule of

thumb may prove useful to help overcome the “dark side of choice,” (Scott-Clayton, 2012).

In this essay, I test the rule of thumb that, for STEM students, attending a highly selective

institution instead of a moderately selective institution improves the probability of obtaining a

STEM degree at the first attended institution among those interested in STEM among and among

those who are not initially interested. The potential formal mechanisms for the hypothesized

advantage of highly selective institutions cover a wide range. Frequency of faculty-student

interactions, participation in STEM peer groups, the size of the graduate school population,

participation in STEM research activities or co-op programs, peer effects and membership in

student learning communities have been found to have exerted a positive influence on STEM

persistence (Leslie et al; 1998; Springer et al, 1999; Gloria et al, 2005; Hernandez, 2005; Cole and

Espinoza, 2008; Hurtado et al, 2009; Ost, 2010; Griffith, 2010; Light and Micari; 2013). On the

other hand, attending a highly selective institution may hinder STEM degree attainment since

average institutional achievement has a negative relationship to academic self-concept in science

(Nagengast and Marsh, 2012). The challenge of isolating the impact of selective institutional

status on STEM degree attainment is controlling for self-selection. Students who opt to attend

highly selective institutions are systematically different among observable and non-observable

characteristics from those who attend other types of institutions.

Inspired by Krueger and Dale (1998), I estimate of the impact of attending a highly

selective post-secondary institution (vs. attending a moderately selective institution) on STEM

degree attainment among students interested in STEM at the beginning of their collegiate careers

through matching. To address potential omitted variable bias, I include total number of

Page 62: Three Essays on the Economics of Education Miguel E. Martinez

54

applications and number of admissions to highly selective and moderately selective institutions

for each student in matching model which confers information on typically unobserved student

characteristics such as motivation. I use regression on the matched sample to more precisely

estimate treatment estimates (covariate adjusted PSM estimates). To assess the impact of the

possibility that theoretically relevant variables were not included in the models, I subject

unadjusted propensity score matching (PSM) treatment estimates for the general population to

the Mantel and Haenzel (MH) test.

Overall, I find that highly selective institutions have a comparative advantage in producing

STEM graduates among those already interested in STEM but not among those initially not

interested in STEM. Attending a highly selective post-secondary institution has a positive impact

on STEM degree attainment among those interested in STEM fields. They are 16 to 30 percentage

points more likely to earn their degree at their first attended institution than their peers at

moderately selective institutions. On the other hand, among those not interested in STEM fields,

attending a highly selective institution makes them two to three percentage points less likely to

graduate with such a degree. For females, the negative association is one percent or non-

existent.

The rest of the essay is structured as follows. First, competing definitions of STEM are

discussed. Review of the literature follows. The next section provides an overview of the data

source. The analytical sample is described in the subsequent section. The next section explains

the methodology. The penultimate section is dedicated to results and the final section provides

an overall conclusion.

Page 63: Three Essays on the Economics of Education Miguel E. Martinez

55

3.2 Competing Definitions of STEM

3.2.1 STEM as a Career

What constitutes a STEM – formerly known as “SMET” - career has varied over time and

across stakeholders (Sanders, 2009) (Zinth, 2006). The most narrow STEM definition restricts

STEM employees to individuals directly involved in the physical and biological sciences,

mathematics, information and computer sciences, and electrical, chemical, civil, and mechanical

engineering. The majority of studies on the subject and the most recent legislative efforts to

influence STEM education have used this highly focused definition (Chen and Weko, 2009). For

example, in the report STEM Education: Preparing Jobs for the Future by the U.S. Congress Joint

Economic Committee, STEM careers only extend to “life sciences (except medical sciences),

physical sciences, mathematics and statistics, computing and engineering” (1). The U.S.

Department of Commerce definition of a STEM career has slightly broader parameters and

includes individuals in managerial positions in STEM fields (2012).

More inclusive definitions of STEM careers also exist. In some cases, they incorporate the

behavioral sciences in addition to the afore-mentioned fields. The National Science Foundation,

for example, groups economics, psychology, sociology and political science under its STEM

umbrella (Chen and Weko, 2009). The Organization for Economic Development and Cooperation

(OECD) goes even farther by including “manufacturing, processing, architecture and building” in

its STEM definition (van Langen and Dekkers, 2005).

Unable to reach consensus on a linguistic definition of a STEM career, researchers have

attempted to bring more clarification on the topic by applying quantitative techniques. Koonce

et al (2011) looked at Standard Occupational Classification (SOC) Codes of the Bureau of Labor

Page 64: Three Essays on the Economics of Education Miguel E. Martinez

56

Statistics from years 2000 and 2010 to ascertain the likelihood that a certain code would be

categorized as a STEM discipline by “papers, conference reports, websites for programs,

government documents, and other statistical data” (4). Formally,

p= (nj-ne)/N

where the term ni is the number of definitions where the code is included, ne is the number of

definitions where the code is explicitly excluded, and N is the total number of definitions. Using

eleven sources for STEM definitions, the authors find that the disciplines most likely to be

included in those STEM definitions were as follows:

P Score Program (SOC Codes 2000)

The vast majority (684 out of 860) of SOC codes, however, were never included in any of the 11

STEM definitions. Although the authors’ results do not give a definitive answer to what

Page 65: Three Essays on the Economics of Education Miguel E. Martinez

57

constitutes a STEM career, they help identify which disciplines are never or almost never

considered to fall under the STEM career umbrella. The competing definitions of a STEM career

are not merely a question of semantics. They represent a challenge to the interpretation and

comparison of empirical findings on the subject.

3.2.2 STEM as a College Major

The ambiguity about the definition of STEM also extends to college majors. The National

Center for Education Statistics (NCES) of the U.S. Department of Education produces more than

2,200 Classifications of Instructional Program (CIP) codes. NCES does not categorize any of the

CIP codes as STEM-related or non-STEM related. Federal agencies have selected different CIPs

as indicating a STEM college major. U.S. Immigration and Customs Enforcement places more

than 100 CIPs under its definition of a STEM major. These are further divided into six different

subcategories: computer and information sciences, engineering and engineering technologies,

biological and biomedical sciences, mathematics and statistics, physical sciences, and science

technologies. This demarcation of what is considered a STEM major allows the U.S. Immigration

and Customs Enforcement to offer work visas to foreign students majoring in those fields (U.S.

Department of Education, 2011). In contrast, research sponsored by the National Science

Foundation has defined STEM majors based on the amount of math and science courses required

(Hartwell, 2012). Under this definition, nursing is a STEM major.

In the same way that they attempted to refine the meaning of a STEM occupation, Koonce et

al (2011) used the afore-mentioned CIP codes to decipher which college majors were most likely

to be designated as STEM. Like their SOC counterparts, a sizable fraction of CIP codes were not

consistently included in the 40 separate STEM definitions. More than one third of them were

Page 66: Three Essays on the Economics of Education Miguel E. Martinez

58

considered STEM in less than five percent of the definitions (two out of 40). A sensible approach

to starting the process of standardizing a definition of a STEM major might be to first consider

the removal of these CIP codes from the definition. Mathematics, physical and computer sciences

were the most likely to be considered STEM in this case.

P Score Program (CIP 2000 Code)

0.875 Mathematics General

0.800 Chemistry, General

0.775 Computer Science, Biology/Biological Sciences, General

0.750 Physics, General

0.725 Mathematics, Other

The authors’ intuitive analyses suggest an overall lack of consensus on a working understanding

of what should be designated as a STEM major or as a STEM career. This is further exacerbated

by an even more ambiguous link between STEM majors and STEM occupations. This is

problematic for ascertaining the effectiveness of policy interventions directed at increasing the

number of STEM workers in the U.S. via increases in the number of students in STEM disciplines.

The essay, in its review of the literature and its analyses, restricts STEM students to those directly

involved in the physical and biological sciences, mathematics, information and computer

sciences, and electrical, chemical, civil, and mechanical engineering (Chen and Weko, 2009).

3.3 Review of the Literature

Page 67: Three Essays on the Economics of Education Miguel E. Martinez

59

3.3.1 Interest in STEM Careers/Majors

The STEM interest related literature is extensive. Overall, the literature reaches relative

consensus on personal traits associated with interest, commitment, pathways to and persistence

in STEM. The roles of family and schools are more nuanced and admit to a greater variety of

findings. The heterogeneity of conclusions is concomitant with the analytical samples and

modeling choices used to explore their respective hypotheses.

The literature suggests that interest in STEM is shaped and solidified after elementary

school. Eighth graders who reported interest in future science careers at age 30 had almost

double the odds of getting a life science degree and more than three times the odds of getting a

degree in a physical science (Tai et al, 2006). STEM students take more academically rigorous

courses in HS, perform better academically, come from more affluent families and are more likely

to study full-time while in college (Chen and Weko, 2009). Students who enter IT majors at

unconventional ages do not have this profile (Chen and Weko, 2009). Math achievement,

number of math courses taken and math efficacy have a positive impact on a student’s decision

to select a college major in a technical field (Ma, 2009). Perception of better-than-average

quality of instruction in math and science during high school is positively associated with the

decision to pursue a physical science college major (Leslie, McClure, and Oaxaca, 1998). Harris

Interactive (2011) finds that parents and guidance counselors were influential in students’

decision to pursue STEM majors, females were primarily motivated by intellectual challenge

while males were motivated by monetary compensation. More than 75% of students decided to

pursue STEM while in HS (Harris Interactive, 2011). STEM summer program for gifted high school

students does not increase STEM identity and STEM salience among participants (Lee, 2002).

Page 68: Three Essays on the Economics of Education Miguel E. Martinez

60

Participation of women in tertiary STEM education increases as the achievement gap between

females and males in secondary school decreases (van Langen and Dekkers, 2005). Female

students attending single-sex schools are not more likely to enter male dominated college majors

(Jennifer Thompson, 2003). The same holds true for male-only schools (James and Richards,

2003). Both boys and girls in single sex schools are more likely to enter more “gender neutral”

fields but not fields historically dominated by the opposite sex (Karpiak, Buchanan, Hosey, and

Smith, 2007).

A considerable number of studies associate personal and family characteristics with

developing and maintaining STEM interest. Mathematical ability is not the only determinant of

vocational choice, even for those who opt for careers in STEM (Benbow et al, 2010). Verbal and

spatial ability also play a role. In addition, males are more likely to report STEM interest (Benbow

et al, 2010). An increase in socio-economic status is associated with a decreased likelihood of

choosing a technical field (Ma, 2009). Female students from families with lower levels of socio-

economic status are as likely to choose a technical field as their male counterparts (Ma, 2009).

Mothers with college degrees increase the likelihood of their daughters getting an advanced

degree in physical science/engineering while the opposite was true for their sons (Leslie,

McClure, and Oaxaca, 1998). Expectation of graduate school attendance also increased the

likelihood of choosing a major in physical sciences, engineering or math (Leslie, McClure, and

Oaxaca, 1998). Plans of marriage shortly after college graduation increased the likelihood of

choosing a physical science/engineering major for white females and Hispanic males (Leslie,

McClure, and Oaxaca, 1998).

Page 69: Three Essays on the Economics of Education Miguel E. Martinez

61

Overall, students interested in STEM are systematically different from those who are not.

They are more academically ready than those who are not. Interest in STEM is not stable in junior

high school, high school, or college. Together this suggests that students in college could develop

interest in STEM in college and that the environment in which this interest is maintained and

fomented among these students could be different from the environment in which students

already interested find it easier to maintain theirs.

3.3.2 College Choice

College choice process has typically been conceptualized as having three distinct stages

with each stage having a set of hypothesized inputs and resulting outcomes (Hossler, 1989;

Cabrera and LaNasa, 2000). These stages are not strictly compartmentalized. They overlap and

interact with each other (Cabrera and LaNasa, 2000). The initial phase is typically referred to as

the predisposition phase which may begin as early as seventh grade. Factors external to the

student (i.e. family and school resources) dominate the input side while academic skills and

educational and career aspirations constitute the primary outputs. The outputs of the

predisposition stage, in turn, become part of the inputs of the subsequent stage: search. Taking

place during the last three years of high school, the perceived attraction of potential post-

secondary institutions of attendance is the characteristic input of the penultimate phase. The

resulting outcomes are formulating a list of potential institutions, gathering information about

them and finalizing a choice set for decision-making. (Cabrera and LaNasa, 2000). The search for

appealing post-secondary alternatives is a dynamic process in which newly discovered knowledge

about a potential institution may lead to further research into previously unconsidered

Page 70: Three Essays on the Economics of Education Miguel E. Martinez

62

institutions or to the rethinking of search criteria itself (Hossler, Schmit and Vesper, 1998 ; Dawes

and Brown, 2002; Moogan and Baron, 2003). In the choice stage during the junior and senior

years of high school, perceptions of affordability and institutional attributes become the most

salient inputs while post-secondary applications, registration, and attendance are the results of

interest (Cabrera and LaNasa, 2000).

Three types of models have been primarily used to explore college choice (Hossler and

Palmer, 2007). Economic models assume that agents (i.e. students and parents) are rational and

evaluate the relevant information for the decision at hand, including personal preferences, to

arrive at a utility-maximizing decision (Hossler, Braxton, & Coopersmith, 1985; Hossler, Schmit,

& Vesper, 1999; Gemici et al, 2014; Wiswall & Zafar, 2015; Reuben et al, 2015). Studies using

economic models often explore the role of college costs and financial constraints on college

choice. Sociological models, in contrast, focus on how personal and contextual factors contribute

to the various decisions in the college choice process. Finally, mixed models split the difference,

acknowledging that decision-making takes place within textured contexts (Perna, 2006). The

matching and outcome models in this essay are mixed. They include measures of income which

may not only influence the decision to select a highly selective institution over a moderately

selective one but also STEM degree attainment. The hypothesis tested in this essay is implicitly

sociological in that it seeks to ascertain if the academic environment in which students carry out

their studies may influence STEM degree attainment.

3.3.3 Matching in the College Choice Process

Page 71: Three Essays on the Economics of Education Miguel E. Martinez

63

The ideal outcome of the college choice process is fit. Fit may be best characterized as

the degree to which applicants’ overall expectations are met by an institution (Cabrera et al,

2011). The better the fit the more likely that student goals will be achieved. The expectation may

include proximity to family and friends, academic reputation/rigor, offerings of majors and/or

programs, characteristics of the student body, time-to-graduation and post-graduation market

outcomes. Match, for its part, is only concerned with a single dimension of fit - the

correspondence of a student’s academic ability to the modal or average academic profile of an

institution (Cabrera et al, 2011). In this regard, students may construct choice sets which include

“safety” (i.e. under-matched), “par” (i.e. matched) and “reach” (i.e. over-matched) schools

(Hoxby and Avery, 2012).

The literature suggests that the majority of low-income high achieving students do not

apply to selective institutions. Not applying may stem from not having a network of achievement

peers from whom to learn and emulate optimal behavior in the college choice process nor

teachers/counselors to guide and/or encourage them in that process (Hoxby and Avery, 2012).

Expanding College Opportunities (ECO) project which provided tailored information related to

the college choice process and “paper-free” application fee waivers to approximately 175

selective post-secondary institutions to high-achieving (i.e. 1300 combined Math and Verbal SAT

or 28 ACT or higher) low-income (i.e. bottom third of the income distribution) students increased

the number of applications by 19 percent and the probability of college match by slightly more

than 40 percent (Hoxby and Turner, 2013). The application patterns of low-income students

present a challenge to any matching methodology that uses income and/or number of

applications/admissions to highly selective institutions to assess the impact of attending highly

Page 72: Three Essays on the Economics of Education Miguel E. Martinez

64

selective institutions. The low number of low income students who apply to these institutions

diminishes the pool of students from which matches can be drawn.

Under-matching is found over the entire academic performance distribution, but is more

pronounced in rural areas and among those from families with lower levels of income and

educational attainment. Despite decreases in under-matching over time, more than 40 percent

of students of the graduating class of 2004 were still under-matched (Smith, Pender and Howell,

2013).

The literature links institutional selectivity to better student outcomes. The gap in

graduation rates between the most selective institutions and the least selective exceeds 30

percentage points (Bowen, Chingos, and McPherson, 2009). The increase in the average time-

to-degree is driven by institutions at the lower end of the selectivity distribution that experienced

declines in overall funding eroding student support services and increased the direct cost of

attendance (Bound, Lovenheim, and Turner, 2012). To cover costs, students increased their time

dedicated to work, which in turn, diminished the availability to engage in academics. Hoekstra

(2009) finds a 20 percent wage premium for having attended the flagship university.

The cost of attendance may also contribute to lack of overall fit. Four and six-year

graduation rates increase as net price increases conditional on student SAT scores (Smith et al,

2013). Thus, in choosing a post-secondary institution, students face an inherent trade-off

between cost and quality. The cost of attendance may also be related to favorable labor market

outcomes. Students who attend institutions with higher tuitions- not those who attended more

selective institutions - earn more than their peers (Krueger and Dale, 1999). In a follow-up study,

long-term elite school earnings premium, however, was found among Hispanic and Black

Page 73: Three Essays on the Economics of Education Miguel E. Martinez

65

students and those from disadvantaged backgrounds (Krueger and Dale, 2011). At the same

time, cost may also hinder access to post-secondary education. The literature converges on the

overall negative relationship between tuition cost and enrollment but the magnitude of the

relationship varies by type of admitting institution and student characteristics (Heller, 1997).

Conditional on financial aid received and student characteristics, non-Asian minority students are

more sensitive to grants and loans than their white counterparts (Heller, 1997). Some of the

differences in reactions to the cost of attendance among student subgroups may be explained by

imprecise perceptions of the cost of college attendance, late acquisition of financial aid

information and difficulty in navigating the financial aid process among lower income and

minority students (Gronsky and Jones, 2004: Heller, 2006; Dynarsky and Scott-Clayton, 2006 ).

Over-matching may not hinder student outcomes. It does not increase time-to-degree or

decrease course loads (Kurlaender and Grodsky, 2013). Others find evidence that increased

selectivity is not always better or neutral. Aggregate school performance exerts a negative

influence on student academic self-concept while individual academic achievement makes a

positive contribution to academic self-concept (Marsh and Hau, 2003). Something similar

happens with students interested in STEM careers, average institutional achievement had a

negative relationship to academic self-concept in science while individual achievement was

positively related (Nagengast and Marsh, 2012). Affirmative action-induced overmatching

among students leads to lower academic outcomes for those students who benefitted from it

(Sander, 2004).

Page 74: Three Essays on the Economics of Education Miguel E. Martinez

66

3.3.4 STEM Attrition and Institutional Selectivity

Student and post-secondary institutional characteristics have been associated with STEM

degree completion. Academic achievement in high school has been linked to persistence in

STEM during college. Students with higher high school GPAs and SAT scores have lower rates of

attrition and higher achievement (Grandy, 1998; Bonous-Hammarth, 2000; Benbow et al, 2010).

The academic rigor of high school math and science courses and performance in those courses

also predict success in STEM majors (Ellington, 2006; Anderson and Kim, 2006; Tyson et al, 2007).

More selective post-secondary institutions have been found to be more efficient at

producing STEM graduates (Eagen, 2009; Chen, 2013). Yet non-Asian minority STEM students at

these institutions do not have the same rates of success (Elliot et al, 1996; Chang et al, 2008;

Chang et al, 2010). The same is true for women of color (Espinosa, 2011). The different levels of

institutional selectivity have their own comparative advantage where more selective institutions

are better at graduating more academically prepared STEM majors and less selective institutions

are more efficient at graduating less academically prepared students (Arcidiacono, 2013). In

contrast to non-Historically Black College and Universities (HBCUs), as selectivity increases

among HBCUs the persistence rate of non-Asian minorities in science increase (Chang et al, 2008).

In addition, the organizational climate around STEM education has been linked to student

outcomes. High Faculty-student interactions, inclusion in STEM peer groups, the size of the

graduate school population and participation in formal research or co-op programs and student

learning communities have been found to have serve as a positive influence on STEM persistence

(Leslie et al; 1998; Springer et al, 1999; Gloria et al, 2005; Hernandez, 2005; Cole and Espinoza,

2008; Hurtado et al, 2009; Ost, 2010; Griffith, 2010; Light and Micari; 2013). Grades in first

Page 75: Three Essays on the Economics of Education Miguel E. Martinez

67

college STEM courses and their relative standing to non-STEM course grades have been positively

linked to STEM degree attainment (Crist et al, 2009; Rask, 2010; Ost, 2010). In addition, STEM

courses in which students are more actively engaged in learning and/or delivery of content are

associated with higher rates of retention for STEM majors (Watkins and Mazur, 2013; Freeman

et al, 2014).

Three fourths of STEM majors make their decision to pursue STEM majors in high school

and the college choice facing them is not easy (Harris Interactive, 2011). As discussed above, the

factors associated with successful completion of a STEM degree are many and slightly less than

half of STEM majors do not earn their bachelor’s degree within six years (Chen, 2013). Even if

students, parents, and counselors are aware of the various institutional attributes which may

exert positive or negative influences on desired academic outcomes they may not be able to

assess the status of each factor or a bundle of factors for specific institutions in their college

choice sets. In this regard, a rule of thumb may prove useful to help overcome the “dark side of

choice,” (Scott-Clayton, 2012). In my third essay, I propose to test the rule of thumb that, if in

doubt, “more selective is better.”

3.4 Data The Education Longitudinal Study (ELS) of 2002 tracks a representative sample of high

school sophomores in 2002 through high school and into college and/or the labor market. The

sample was drawn from 752 schools out of a sampling frame of approximately 1,200 public and

private schools from a total population of about 27,000 schools with 10th grades. From each

school, a sample of 26 to 30 sophomores was randomly selected. The slightly more than 16,000

Page 76: Three Essays on the Economics of Education Miguel E. Martinez

68

students participated in the study represented more than three million sophomores in 2002

(Ingels et al, 2014).

The baseline survey and its first follow-up contained information on the high schools they

attended, their teachers, peers, and families. The first two follow-ups were conducted at two-

year intervals in 2004 and 2006. The former captured most students at the end of their high

school careers. In addition to high school outcomes, it gathered information on long-term career

plans and the college choice process. The latter follow-up encapsulated their first years in post-

secondary education or the workforce. The third wave conducted in 2012 concentrated on

gathering information on post-secondary and labor market outcomes (Ingels et al, 2014).

3.4.1 Defining Interest in STEM Interest in ELS of 2002/06/12

Student interest in STEM is defined by student responses captured in variable F2B15. The

second follow-up to the Education Longitudinal Study of 2002 asks post-secondary attendees

which field of study they were most likely to pursue upon entering (F2B15)9. Specifically,

students are asked “When you began at [name of institution], what field of study did you think

you would most likely pursue?”. Response choices are categorized into 14 broad fields. For this

essay, interest in STEM is identified when respondents opt for either of three categories:

engineering or engineering technologies, computer or information sciences, and natural sciences

or mathematics.

9

Page 77: Three Essays on the Economics of Education Miguel E. Martinez

69

3.4.2 Defining a Highly Selective Post-Secondary Institution in ELS of 2002/06/12

The variable F2PS1SLC captures the highest selectivity category of the first post-secondary

institution for each student based on 2005 Carnegie classifications from IPEDS (US Department

of Education, 2008). The variable contains six valid values: highly selective (F2PS1SLC=1),

moderately selective (F2PS1SLC=2), 4-year inclusive (F2PS1SLC=3), 4-year not classified

(F2PS1SLC=4), 2-year not classified (F2PS1SLC=5), and less than 2-years (F2PS1SLC=6). Roughly,

the highly selective category identifies institutions whose student bodies have average SAT/ACT

scores in the highest quintile of the distribution while the moderately selective category houses

colleges and universities in the adjacent quintile of the distribution. The rest of the institutions

in the remaining categories either do not report average ACT/SAT scores or do not require them

for admissions (US Department of Education, 2008).

3.4.3 Defining a STEM Degree in ELS of 2002/06/12

The variable F3TZBCH1CP2 captures the first known bachelor’s degree for each student.

Students were identified as having earned a STEM degree under the most narrowly defined

definition of STEM if they had one of the following values: 10=Communication Technology and

Support, 11=Computer/Information Science Support, 14=Engineering, 15=Engineering

technologies/technicians, 27=Mathematics and Statistics, 40=Physical Sciences and 41=Science

technologies/technicians. To identify that degree was earned from the first-attended post-

secondary institution, the F3TZBCH1CP2 variable was coupled with the variable F3PS1RETAIN

which summarizes the status relative to first-attended postsecondary institution at the third

Page 78: Three Essays on the Economics of Education Miguel E. Martinez

70

follow-up. If they had obtained a STEM degree under the above-criterion, only those with values

of 1 or 2 (1=Earned a credential from first post-secondary institution attended; still attending

post-secondary institution as of the third follow up, 2=Earned a credential from first post-

secondary institution attended; no longer attending post-secondary institution as of the third

follow up, 3= Did not earn a credential from first post-secondary institution attended; still

attending first post-secondary institution as of the third follow up; 4= Did not earn a credential

from first post-secondary institution attended; not attending first post-secondary institution as

of the third follow up; did attend another post-secondary institution; 5= Did not earn a credential

from first post-secondary institution attended; not attending first post-secondary institution as

of the third follow up; did not attend another post-secondary institution) were identified as

having received them from their first post-secondary institution.

3.5 STEM Students Interest in STEM careers among high school students fluctuates during their high school

experience. In the sample, approximately 13 percent of the 55 percent of tenth graders who

responded to the question in the baseline survey regarding occupational expectations at age 30

reported plans of having a STEM career.10 By the first follow-up two years later, the 10 percent

10 The Education Longitudinal Study of 2002 probes high school sophomores about their long-term career plans (US Department of Education, 2004). Question 64 of the baseline survey asks, “Write in the name of the job or occupation that you expect or plan to have at age 30.” Below the blank line designated for the answer, two other options are found: 1. I don’t plan to work when I am 30 and 2. I don’t know. The verbatim responses provided by student may be interpreted in different ways. A response of “engineer” may mean a “train conductor” or a “chemist.” In all such cases, I assumed that the responses to indicate a STEM career. In the cases in which multiple career aspirations are expressed by a single student such student is considered to be a STEM career aspirant as long as one of them is STEM. The same question was posed to students in the first follow-up in 2004.

Page 79: Three Essays on the Economics of Education Miguel E. Martinez

71

of the 68 percent of students who replied to the same question reported STEM career

expectations.11

As Table 15 points out about 17 percent of the students in the sample entered post-

secondary institutions with the idea of becoming a STEM major. In contrast to their non-STEM

peers, these students opted to attend a) highly selective institutions at higher rates and b) two-

year colleges at lower rates.

Table 1512. First post-secondary institution (PSI) by STEM interest as freshmen

Selectivity of First PSI Attended

Freshmen: STEM Freshmen: non-STEM

n % n %

Highly - 4 year 96,820 30% 322,830 17%

Moderately - 4 year 94,020 29% 498,810 27%

Inclusive - 4 year 24,730 8% 131,100 7%

Selectivity not classified - 4 year 18,680 6% 90,680 5%

Selectivity not classified - 2 year 81,190 25% 744,870 40%

Selectivity not classified - less than 2 years 6,110 2% 66,010 4%

Total 321,550 100% 1,854,300 100%

Table 16 suggests that, in comparison to their peers at moderately selective institutions,

students interested in STEM as freshmen at highly selective institutions earned a degree at higher

rates by the third follow-up survey in 2012. They also transferred to other institutions or

dropped out at lower rates. Overall, this is suggestive of the potential advantage of highly

selective institutions over moderately selective institutions in the production of STEM degrees.

11 The pattern of missing data in STEM career aspirations does not allow for inclusion of STEM career aspirations in the matching/outcome models since they would severely restrict the number of observations. 12 See Appendix L for un-weighted counts.

Page 80: Three Essays on the Economics of Education Miguel E. Martinez

72

Namely, students at highly selective institutions may be more likely to obtain the degree to which

they initially aspired because they are less likely to drop out or transfer out. The differences in

graduation and retention rates between the two types of institutions, however, may be the

results of systematic differences in academic readiness and motivation between the students

who attended the institutions.

Table 1613. Earning of credentials by first post-secondary institutions selectivity among STEM freshmen

Enrollment Status of STEM Freshmen at First PSI Attended at Third Follow-up

Selectivity

Highly Moderately

n % n %

Earned a credential from PSI; still attending PSI as of 2012 3,010 3% 2,500 3%

Earned a credential from PSI; no longer attending PS1 as of 2012 58,920 64% 34,210 41%

No cred from PSI; still attending PSI as of 2012 560 1% 3,380 4%

No cred from PSI; no longer attending PSI; did attend another PS institution 28,240 30% 36,990 44%

No cred from PSI; no longer attending PSI; did not attend another PS institution 1,960 2% 6,380 8%

Total 92,691 100% 83,461 100%

3.6 Methodology After conducting logistic regression, propensity score matching will also be used. The

complex survey design of ELS of 2002/12 will be accommodated in the propensity score matching

approach (Zanutto, Lu, & Hornik, 2005). Sampling weights will be applied in the matching and

outcome models (Zanutto, 2006). The matching model will be a binary logistic regression instead

of a multinomial logistic regression for two reasons. First, moderately selective institutions are,

13 See Appendix M for un-weighted counts.

Page 81: Three Essays on the Economics of Education Miguel E. Martinez

73

theoretically, the closest substitute for highly selective institutions. Second, institutions “below”

the moderately selective line are less likely to require SAT/ACT scores for admissions, thus,

reducing the possibility of adequate matching. Thus, the analysis will compare STEM degree

attainment between comparable students attending highly selective institutions and students

attending moderately selective institutions. The binary matching model will take the following

general form:

(1) Yi= β0 + βdΣXd + βaΣXa+ βcΣXc + βoΣXo + ei

-where Y is attendance at a highly selective institution (1=highly selective institution

0=moderately selective institution), Xd is a vector of student demographic variables, Xa is a vector

of academic performance and exposure measures, Xc is a vector of college choice behaviors and

outcomes, Xo is a vector of contextual variables-and ei is an error term with the usual properties,

and the betas capture the cumulative association of each vector on the outcome. The

demographics variables include gender, race with four values -Asian, Multi-racial, Under-

represented Minorities (URM) and White – and income with three values: low income (low

[below $35,000], middle [$35,001 to $75,000], and high income [above $75,001]. The academic

readiness variables are Math and Verbal SAT scores and high school GPA. The exposure variables

capture the highest level of science and math courses completed in high school. The seven

categories of exposure for each subject are found in Appendix N. The college choice behavior

and outcome variables are total number of applications and number of admissions to a) highly

selective and b) moderately selective institutions. The contextual variables include school

control (public, Catholic and private), urban status (suburban, urban, and rural) and region

(Northeast, Midwest, South, and West).

Page 82: Three Essays on the Economics of Education Miguel E. Martinez

74

Meeting the conditions of common support, the outcomes model takes the form:

(2) Yi= β0 + β1T + βrΣXr + ei

- where Y is the attainment of a STEM degree for student i at the first post-secondary institution

attended (where 1=STEM degree 0=else), T identifies attendance at a highly selective post-

secondary institution, β1 captures the average treatment on the treated (ATT) of T on the

outcome, Xr is vector of all the variables included in the matching model and βr captures its

cumulative impact of any residual differences between “treatment” and “control” groups on the

outcome in the matched sample (DuGoff, Schuler, & Stuart, 2012). The Mantel and Haenzel (MH)

test will be used to assess the sensitivity of treatment estimates (Becker and Caliendo, 2007;

Caliendo and Kopeinig, 2008).

The number of total applications and number of admissions to highly selective institutions

and the number of admissions to moderately selective institutions play an important role in the

identification strategy. Post-secondary institutions make their admissions decision on a variety

of factors. Some of those factors – like academic achievement and readiness - may be observed

by researchers but others like motivation, maturity, and cultural fit may only be observed by

admissions personnel through admissions essay, interviews, and campus visits (Dale and Krueger,

1998). Thus, admissions decisions signal/confer valuable information about applicants not

directly observed by the researchers. They act as proxies for non-observables. Failure to include

them would lead to biased estimates. By definition, highly selective institutions have higher

admission thresholds for academics and, and most likely also for, non-academic factors than

moderately selective institutions. If non-academic factors or their proxies are not included in the

Page 83: Three Essays on the Economics of Education Miguel E. Martinez

75

models and are positively correlated with academic factors then the estimates on the impact of

attending a highly selective institution on STEM degree attainment would be upwardly biased

since it would be confounded with academic factors despite matching on academic factors.

Formally following Dale and Krueger (1998), for every student i applying to school j the

admissions committees grant admission or do not grant admission according the following rule:

(3) Zij= γ1X1i + γ2X2i + eij >=Tj then admit student i , else do not admit.

- where Z is the quality of the applicant, X1 are student traits observable to the researcher, X2

are student attributes not known to the researcher, their accompanying gammas are the relative

weights given by the admission committees to each type of characteristic, ei represents the

unique perspectives of personnel making admissions decisions at college j (and orthogonal to

STEM degree attainment) and T is the minimum threshold at which students gain admissions. If

X2 is not included in (1) as Xc then β1 in (2) will be upwardly biased if X2 is positively correlated

with X1.

3.7 Results The results section is structured as follows. It first identifies differences in academic

readiness/exposure measures and admission outcomes between the two types of attendees and

then presents the results of logistic regression analysis first for those interested in STEM (all

students and female students) and then for those who are not interested in STEM (all students

and female students). For the sake of brevity, the comparisons in the body of the paper are

restricted to SAT math scores, percentage of students exposed to the highest levels in the math

and science pipelines, and the number of acceptances to highly selective institutions.

Page 84: Three Essays on the Economics of Education Miguel E. Martinez

76

Comparison of all the variables in the matching/outcome models can be found in the appendices.

Next, reduction of average differences in the afore-mentioned variables between students

attending the two types of institutions as result of propensity score matching is quantified and is

followed by unadjusted and covariate adjusted PSM “treatment” estimates for the those

interested in STEM (all students and females) and for those not interested in STEM (all students

and females). Finally, the stability of unadjusted PSM estimates are tested.

The analyses will be restricted to those students who were admitted to at least one highly

selective institution. The control groups of students who opted to attend moderately selective

institutions will be composed of students who had the chance to attend at least one highly

selective institution. In this way, the control and treatment groups will be comparable, at least,

in applying and getting admissions to at least one highly selective institution.

3.7.1 Logistic Regression Analysis

STEM covers students interested in physical and biological sciences, mathematics,

information and computer sciences, and engineering. Approximately 322,000 students reported

interest in such fields at the beginning of their collegiate careers. The differences in academic

readiness and exposure measures between the approximately 190,000 students attending highly

selective and moderately selectively institutions, in some cases, are considerable. Students who

opted to attend highly selective institutions score 50 points higher on the math section of the

SAT, have two percentage point advantage in exposure to the highest level in the math course

pipeline and an eight percentage point advantage in the science pipeline and have been accepted

by highly selective institutions by .6 more of an admission. See Appendix O for a more detailed

view of differences.

Page 85: Three Essays on the Economics of Education Miguel E. Martinez

77

Logistic regression analysis suggests a positive association of attending a highly selective

institution with STEM degree attainment among those interested in STEM. The odds ratio is for

the highly selective status is 1.51. The corresponding marginal effect estimate suggest that for

every 100 students who attend highly selective institutions and every 100 students who attend

moderately selective institution highly selective institutions will graduate 22 more students with

STEM degrees. See Appendix P for a more detailed view of regression results. Marginal effects

for females interested in STEM were estimated at 30 percentage points. 14

Among students not interested in STEM fields upon entering post- secondary education,

attending a highly selective institution decreases the probability of graduating with a STEM

degree by three percentage points (marginal effect). Attending a highly selective institution

decreases the probability of obtaining a

STEM degree by one percentage point among females not interested in STEM fields (marginal

effect).

3.7.2 Matching Analysis

The matching models yield more comparable “treatment” and “control” groups.

Differences in SAT math scores are reduced by 56 percent, highest exposure in the science course

pipeline and number of acceptances to highly selective institutions are both reduced by 87

percent but differences in the highest exposure to math courses are increased by almost fourfold,

from two percent to almost eight percent. See Appendix Q. Most differences remain statistically

significant. About forty percent of the 12,276 students in the treatment group were not

14 Based on a linear probability model.

Page 86: Three Essays on the Economics of Education Miguel E. Martinez

78

matched. The unmatched units were all at high end of the propensity score distribution. See

Appendix R.

The unadjusted PSM results suggest positive average treatment on the treated (ATT)

effects for attending a highly selective institution. As Table 17 points out, on average, the

probability of graduating with a STEM degree is 16 percentage points higher if one attends a

highly selective institution. Manzel-Haenszel bound results suggest that the unadjusted PSM

estimates are stable. This is a lower estimate than the marginal effect of the logistic regression

analysis. Despite the differences in the magnitude of the estimate both coincide on the positive

effect of attending a highly selective institution. Ideally, when there is less than perfect parity

between the synthetic control and treatment groups as in the case at hand, more precise

estimates can be calculated by running the matched sample through the matching regression

model. Covariate adjusted PSM estimates suggest a positive association of 30 percentage points

between attending a highly selective institution and STEM degree attainment. PSM estimates

for female students could not converge on an answer.

Table 17. Pre and Post Matching Differences in STEM Degree Attainment Variable Sample Treated Controls Difference S.E. t-stat

STEM Degree Unmatched 0.52 0.42 0.10 0.01 15.03

STEM Degree Matched 0.58 0.42 0.16 0.01 20.71

Finally, it should be noted that highly selective institutional status has the opposite effect

when the analytical sample is reduced to those who did not have STEM interest upon arriving at

their post-secondary institutions. Highly selective institutions are not as effective at generating

and maintaining STEM interest among those initially not interested in STEM. Unadjusted PSM

Page 87: Three Essays on the Economics of Education Miguel E. Martinez

79

results suggest that students not interested in STEM at moderately selective institutions earn

STEM degrees at a rate of six percent while their counterparts at highly selective institutions earn

degrees at half the rate. The covariate adjusted PSM results put the difference between slightly

lower at one and half percent. PSM results suggest that females who attend moderately selective

institutions do not have an advantage in obtaining STEM degrees. The covariate adjusted

estimate suggests a one percentage point advantage. Although the possible positive association

between moderately selective attendance and STEM degree attainment is relatively small the

sheer number of students not interested in STEM at those institutions make it a non-negligible

source of STEM graduates.

3.7.3 Summary of Results

Table 18 summarizes results from logistic regression (Reg), propensity score matching

(PSM) and covariate adjusted PSM (CA PSM). Among all students initially interested upon

entering college, attending a highly selective institution is associated with a higher probability of

attaining a STEM degree from the first attended institution. Attending a moderately selective

college, on the other hand, is associated with higher probabilities of STEM degree attainment

among students not initially interested in STEM. This advantage may not extend to female

students or is considerably less for them.

Table 18. Summary of Results

Group

Initially Interested in STEM Initially Not Interested in STEM

Reg PSM CA PSM Reg PSM CA PSM

All Students 0.22 0.16 0.30 -0.02 -0.03 -0.02

Female Students 0.3015 - - 0.00 0.00 -0.01

15 Linear regression

Page 88: Three Essays on the Economics of Education Miguel E. Martinez

80

3.8 Conclusion The rule-of-thumb that “more selective is better” overall holds true for students

interested in STEM. This suggest that parents and counselors should encourage students to

attend highly selective institutions when students have the chance of doing so. Scholarships

directed at students interested in STEM should also take this into account and provide incentives

for students to opt for more selective institutions. On the other hand, highly selective

institutions are not as effective at generating and maintaining interest in STEM among those

initially not interested. Parties involved in the college choice process of students who are

undecided about their major at the end of high school who believe these students could

potentially succeed in STEM should encourage them to attend moderately selective institutions.

Page 89: Three Essays on the Economics of Education Miguel E. Martinez

81

Works Cited/Consulted Acemoglu, D. (2012). What Does Human Capital Do? A Review of Goldin and Katz's The Race between

Education and Technology (No. w17820). National Bureau of Economic Research.

Adamson, C., & Britt, R. (2009). Repeat testing with the HESI Exit Exam-sixth validity study. Computers

Informatics Nursing, 27(6), 393-397.

Aghion, P. A., Howitt, P. A., & Peñalosa, C. G. (1998). Endogenous growth theory. MIT press.

Altonji, J. G., Elder, T. E., & Taber, C. R. (2005). An evaluation of instrumental variable strategies for

estimating the effects of catholic schooling. Journal of Human Resources, 40(4), 791-821.

Ansalone, G. (2010). Tracking: Educational Differentiation or Defective Strategy.Educational Research

Quarterly, 34(2), 3-17.

Arathuzik, D., & Aber, C. (1998). Factors associated with national council licensure examination-

registered nurse success. Journal of Professional Nursing, 14(2), 119-126.

Austin, P. C., Grootendorst, P., & Anderson, G. M. (2007). A comparison of the ability of different

propensity score models to balance measured variables between treated and untreated subjects: a

Monte Carlo study. Statistics in medicine, 26(4), 734-753.

Autor, David & Dorn, David (2009). The Growth of Low Skill Service Jobs and the Polarization of the U.S.

Labor Market, NBER Working Papers 15150, National Bureau of Economic Research, Inc.

Ballantine, J.H., & Hammack, F.M. (2009). The sociology of education: A systematic analysis. Upper

Saddle River, NJ: Pearson.

Becker, S. O., & Caliendo, M. (2007). Mhbounds-sensitivity analysis for average treatment effects (No.

2542). IZA Discussion Papers.

Beeson, S. A., & Kissling, G. (2001). Predicting success for baccalaureate graduates on the NCLEX-

RN. Journal of Professional Nursing, 17(3), 121-127.

Benbow, C. P., & Minor, L. L. (1986). Mathematically talented males and females and achievement in the

high school sciences. American Educational Research Journal, 23(3), 425-436.

Betts, J. R., & Shkolnik, J. L. (2000). The effects of ability grouping on student achievement and resource

allocation in secondary schools. Economics of Education Review, 19(1), 1-15.

Bifulco, R. (2002). Addressing Self-Selection Bias in Quasi-Experimental Evaluations of Whole-School

Reform A Comparison of Methods. Evaluation Review, 26(5), 545-572.

Bondmass, M. D., Moonie, S., & Kowalski, S. (2008). Comparing NET and ERI Standardized Exam Scores

between Baccalaureate Graduates Who Pass or Fail the NCLEX-RN©. International Journal of Nursing

Education Scholarship,5(1), 1-15.

Bosch, P. C., Doshier, S. A., & Gess-Newsome, J. (2012). Bilingual nurse education program: Applicant

characteristics that predict success. Nursing education perspectives, 33(2), 90-95.

Page 90: Three Essays on the Economics of Education Miguel E. Martinez

82

Bound, J., Lovenheim, M., & Turner, S. (2009). Why have college completion rates declined? An analysis

of changing student preparation and collegiate resources (No. w15566). National Bureau of Economic

Research.

Bowen, W. G., Chingos, M. M., & McPherson, M. S. (2009). Crossing the finish line: Completing college at

America's public universities. Princeton University Press.

Bray, M., & Thomas, R. M. (1995). Levels of comparison in educational studies: Different insights from

different literatures and the value of multilevel analyses.Harvard Educational Review, 65(3), 472-491.

Briscoe, V. J., & Anema, M. G. (1999). The relationship of academic variables as predictors of success on

the National Council Licensure Examination for Registered Nurses (NCLEX-RN) in a selected associate

degree program. ABNF journal, 10(4), 80.

Burdett, K. (2013). How students choose a college: Understanding the role of internet based resources in

the college choice process. Unpublished dissertation.

Bureau of Labor Statistics (2008). Occupational Employment Series Survey, (May 2004 to May 2008).

Washington, D.C.

Business Higher Education Forum (2010). Increasing the Number of STEM Graduates: Insights from the

U.S. STEM Education & Modeling Project.

Cabrera, A. F., & La Nasa, S. M. (2000). Understanding the College-Choice Process. New Directions for

Institutional Research, 2000(107), 5-22.

Cabrera, A. F., & La Nasa, S. M. (2001). On the path to college: Three critical tasks facing America's

disadvantaged. Research in Higher Education, 42(2), 119-149.

Caliendo, M., & Kopeinig, S. (2008). Some practical guidance for the implementation of propensity score

matching. Journal of economic surveys, 22(1), 31-72.

Calonico, S., Cattaneo, M. D., & Titiunik, R. (2014). Robust Nonparametric Confidence Intervals for

Regression-Discontinuity Designs. Econometrica, 82(6), 2295-2326.

Calvó-Armengol, A., Patacchini, E., & Zenou, Y. (2009). Peer effects and social networks in education. The

Review of Economic Studies, 76(4), 1239-1267.

Card, D., & Payne, A. A. (2002). School finance reform, the distribution of school spending, and the

distribution of student test scores. Journal of Public Economics, 83(1), 49-82.

Carrell, S., Sacerdote, B., & West, J. (2009). Beware of economists bearing reduced forms? an

experiment in how not to improve student outcomes.

Carrell, Scott, Page, Marianne, and West, James (2010). Sex and Science: How Professor Gender

Perpetuates The Gender Gap: The Quarterly Journal of Economics: 1101-1144.

Catsambis, Sophia (1994). The Path to Math: Gender and Racial-Ethnic Differences in Mathematics

Participation from Middle School to High School. Sociology of Education, Vol. 67 (3), 199-215.

Page 91: Three Essays on the Economics of Education Miguel E. Martinez

83

Cattaneo, M. D., Frandsen, B. R., & Titiunik, R (2015) . Randomization Inference in the Regression

Discontinuity Design: An Application to Party Advantages in the US Senate.

Cattaneo, M. D., Frandsen, B. R., & Titiunik, R. Randomization Inference in the Regression Discontinuity

Design: An Application to Party Advantages in the US Senate.

Cattaneo, M. D., Jansson, M., & Ma, X. (2015). Simple Local Regression Distribution Estimators with an

Application to Manipulation Testing. working paper, University of Michigan.

Cavanagh, S. E., & Fomby, P. (2012). Family Instability, School Context, and the Academic Careers of

Adolescents. Sociology of Education, 85(1), 81-97.

CCNE (2013). Standards for Accreditation of Baccalaureate and Graduate Nursing Programs.

Chen, X. and Weko, T (2009). Students Who Study Science, Technology, Engineering, And Mathematics

(STEM) In Postsecondary Education. Stats in Brief (NCES 2009-161). National Center for Education

Statistics, U.S. Department of Education.

CNEA (2016). Accreditation Standards for Nursing Education Programs.

Coble and Allen (2005). Keeping America Competitive. Denver, CO, ECS.

Comparative Education, Vol. 41 (3), 329-350.

Crosnoe, Robert. 2009. “Low-Income Students and the Socioeconomic Composition of Public High

Schools.” American Sociological Review, Vol. 74: 709-730

Dale, S. B., & Krueger, A. B. (1999). Estimating the payoff to attending a more selective college: An

application of selection on observables and unobservables (No. w7322). National bureau of economic

research.

Dale, S., & Krueger, A. B. (2011). Estimating the return to college selectivity over the career using

administrative earnings data (No. w17159). National Bureau of Economic Research.

Daley, L. K., Kirkpatrick, B. L., Frazier, S. K., Chung, M. L., & Moser, D. K. (2003). Predictors of NCLEX-RN

success in a baccalaureate nursing program as a foundation for remediation. Journal of Nursing

Education, 42(9), 390-398.

Dawes, P. L., & Brown, J. (2002). Determinants of awareness, consideration, and choice set size in

university choice. Journal of Marketing for Higher Education, 12(1), 49-75.

DuGoff, E. H., Schuler, M., & Stuart, E. A. (2014). Generalizing observational study results: Applying

propensity score methods to complex surveys. Health services research, 49(1), 284-303.

Dynarski, Susan M. and Judith E. Scott-Clayton. 2006. The Cost of Complexity in Federal Student Aid:

Lessons from Optimal Tax Theory and Behavioral Economics. Cambridge, MA: National Bureau of

Economic Research.

Eder, D. (1981). Ability grouping as a self-fulfilling prophecy: A micro-analysis of teacher-student

interaction. Sociology of Education, 151-162.

Page 92: Three Essays on the Economics of Education Miguel E. Martinez

84

Enders, C. K., & Tofighi, D. (2007). Centering predictor variables in cross-sectional multilevel models: a

new look at an old issue. Psychological methods,12(2), 121.

Engberg, Mark and Wolniak, Gregory C., "College Student Pathways to the STEM Disciplines"

(2013). School of Education: Faculty Publications & Other Works.Paper 10.

Flanagin, A. J., & Metzger, M. J. (2000). Perceptions of Internet information credibility. Journalism &

Mass Communication Quarterly, 77(3), 515-540.

Flanagin, A. J., & Metzger, M. J. (2007). The role of site features, user attributes, and information

verification behaviors on the perceived credibility of web-based information. New Media & Society, 9(2),

319-342.

Fleming, J., and Morning, C. (1995). Correlates of the SAT in Minority Engineering Students: An

Exploratory Study. Journal of Higher Education, Vol. 69(1), 89-108.

Fortier, M. E. (2012, July). Predictors of Success on the NCLEX-RN Among Transfer BSN Students.

In Sigma Theta Tau International's 23rd International Nursing Research Congress.

Frith, K. H., Sewell, J. P., & Clark, D. J. (2005). Best practices in NCLEX-RN readiness preparation for

baccalaureate student success. Computers Informatics Nursing, 23(6), 322-329.

Gardiner, J. C., Luo, Z., & Roman, L. A. (2009). Fixed effects, random effects and GEE: What are the

differences?. Statistics in medicine, 28(2), 221-239.

Gelman, A., & Imbens, G. (2014). Why high-order polynomials should not be used in regression

discontinuity designs (No. w20405). National Bureau of Economic Research.

Gelman, A., & Zelizer, A. (2015). Evidence on the deleterious impact of sustained use of polynomial

regression on causal inference. Research & Politics, 2(1).

Gemici, A., & Wiswall, M. (2014). Evolution of gender differences in post-secondary human capital

investments: College Majors. International Economic Review, 55(1), 23-56.

Giddens, J. and Groecker (2005). The relationship of critical thinking to performance on the NCLEX-

RN®. Journal of Nursing Education, 44(2), 85.

Grandy, Jerilee (1998). Persistence in Science of High-Ability Minority Students: Results of a Longitudinal

Study. The Journal of Higher Education, Vol. 69 (6), 589-620.

Grodsky, Eric and Melanie T. Jones (2004). Real and Imagined Barriers to College Entry: Perceptions of

Cost. Social Science Research 36:745–766.

Hanushek, E. A., Kain, J. F., & Rivkin, S. G. (2004). Why public schools lose teachers. Journal of human

resources, 39(2), 326-354.

Harris Interactive (2011). STEM Perceptions: Student & Parent Study Parents and Students Weigh in on

How to Inspire the Next Generation of Doctors, Scientists, Software Developers and Engineers.

Commissioned by Microsoft Corporation Corp. Seattle, WA.

Page 93: Three Essays on the Economics of Education Miguel E. Martinez

85

Harwell, E. (2012). An Analysis of Parent Occupation and Student Choice in STEM Major. Project STEP-

UP. University of Illinois at Urbana-Champaign.

Hearn, J. C. (1984). The relative roles of academic, ascribed, and socioeconomic characteristics in college

destinations. Sociology of Education, 22-30.

Heller, D. E. (1997). Student price response in higher education: An update to Leslie and

Brinkman. Journal of Higher Education, 624-659.

Heller, D.E. (2006). Early Commitment of Financial Aid Eligibility. American Behavioral Scientist 49(12):

1719–1738.

Heroff, K. (2009). Guidelines for a progression and remediation policy using standardized tests to

prepare associate degree nursing students for the NCLEX-RN at a rural community college. Teaching and

Learning in Nursing, 4(3), 79-86..

Hoekstra, M. (2009). The Effect of Attending the Flagship State University on Earnings: A Discontinuity-

Based Approach. The Review of Economics and Statistics, 91(4), 717-724.

Hofmann, D. A., & Gavin, M. B. (1998). Centering decisions in hierarchical linear models: Implications for

research in organizations. Journal of Management, 24(5), 623-641.

Hong, G., & Raudenbush, S. W. (2005). Effects of kindergarten retention policy on children’s cognitive

growth in reading and mathematics. Educational Evaluation and Policy Analysis, 27(3), 205-224.

Hossler, D., & Palmer, M. (2008). Why understand research on college choice. National Association for

College Admission Counseling (NACAC). Fundamentals of college admission counseling: A textbook for

graduate students and practicing counselors, 42-53.

Hossler, D., Braxton, J., & Coopersmith, G. (1989). Understanding student college choice. Higher

education: Handbook of theory and research, 5, 231-288.

Hossler, D., Schmit, J., & Vesper, N. (1999). Going to college: How social, economic, and educational

factors influence the decisions students make. JHU Press.

Hoxby, C. M., & Avery, C. (2012). The missing" one-offs": The hidden supply of high-achieving, low

income students (No. w18586). National Bureau of Economic Research.

Hoxby, C., & Turner, S. (2013). Expanding college opportunities for high-achieving, low income

students. Stanford Institute for Economic Policy Research Discussion Paper, (12-014).

Hubbard, A. E., Ahern, J., Fleischer, N. L., Van der Laan, M., Lippman, S. A., Jewell, N., ... & Satariano, W.

A. (2010). To GEE or not to GEE: comparing population average and mixed models for estimating the

associations between neighborhood risk factors and health. Epidemiology, 21(4), 467-474.

Hughes, J. N., Chen, Q., Thoemmes, F., & Kwok, O. M. (2010). An investigation of the relationship

between retention in first grade and performance on high stakes tests in third grade. Educational

evaluation and policy analysis,32(2), 166-182.

Ingels, S.J., Pratt, D.J, Alexander, C.P., Jewell, D.M., Lauff, E. Mattox, T.L., and Wilson, D. (2014).

Education Longitudinal Study of 2002 Third Follow-up Data File Documentation (NCES 2014-364).

Page 94: Three Essays on the Economics of Education Miguel E. Martinez

86

National Center for Education Statistics, Institute of Education Sciences, U.S. Department of Education.

Washington, DC. Retrieved [date] from http://nces.ed.gov/pubsearch.

Jean Giddens PhD, R. N. (2005). The relationship of critical thinking to performance on the NCLEX-

RN®. Journal of Nursing Education, 44(2), 85.

Karpiak, C. P., Buchanan, J. P., Hosey, M., & Smith, A. (2007). University students from single-sex and

coeducational schools: Differences in majors and attitudes at a Catholic university. Psychology of

Women Quarterly, Vol. 31: 282–289

Kelcey, B. M. (2009). Improving and assessing propensity score based causal inferences in

Koonce, D. , Anderson, D., Zhou, J. , Hening, D., & Conley, V. (2011). What is STEM? Paper presented to

the American Society for Engineering Education’s 2011 Annual Conference & Exposition.

Kurlaender, M., & Grodsky, E. (2013). Mismatch and the paternalistic justification for selective college

admissions. Sociology of Education, 0038040713500772.

Langdon, D. McKittrick, G., Beede, D., Khan, B., & Doms, M. (July, 2011). STEM: Good jobs now and for

the future. Washington D.C.: U.S. Department of Commerce, Economics and Statistics Administration.

Retrieved from www.esa.doc.gov/sites/default/files/reports/.../stemfinalyjuly14_1.pdf

Langford, R. and Young, A (2013). Predicting NCLEX-RN success with the HESI Exit Exam: Eighth validity

study. Journal of Professional Nursing. 29: S5–S9.

Lankford, H., Loeb, S., & Wyckoff, J. (2002). Teacher sorting and the plight of urban schools: A

descriptive analysis. Educational Evaluation and Policy Analysis, 24(1), 37-62.

Lau, S., & Roeser, R. W. (2002). Cognitive abilities and motivational processes in high school students'

situational engagement and achievement in science.Educational Assessment, 8(2), 139-162.

Lauder, H., & Hughes, D. (1990). Social inequalities and differences in school outcomes. New Zealand

Journal of Educational Studies, 25(1), 37-60.

Lee, A. (2013). Determining the Effects of Pre-College STEM Contexts on STEM Major Choices in 4-year

Post-Secondary Institutions Using Multilevel Structural Equation Modeling. Journal of Pre-College

Engineering Education,3(2).

Lee, James (2002). More Than Ability: Gender and Personal Relationships Influence Science and

Technology Involvement. Sociology of Education, Vol. 75 (4), 349-373.

Lee, V. E., & Ready, D. D. (2009). US high school curriculum: Three phases of contemporary research and

reform. The Future of Children, 19(1), 135-156.

Lent, R. W., & Brown, S. D. (2006). On conceptualizing and assessing social cognitive constructs in career

research: A measurement guide. Journal of Career Assessment, 14(1), 12-35.

Leslie, L. Larry, McClure, Gregory T., and Oaxaca, Ronald L. (1998). Women and Minorities in Science and

Engineering: A Life Sequence Analysis. The Journal of Higher Education, Vol. 69 (3), 239-276

Page 95: Three Essays on the Economics of Education Miguel E. Martinez

87

Levin, H. M. (1985). Solving the shortage of mathematics and science teachers.Educational Evaluation

and Policy Analysis, 7(4), 371-382.

Lewis, C. C. (2005). Predictive accuracy of the HESI Exit Exam on NCLEX-RN pass rates and effects of

progression policies on nursing student Exit Exam scores (Doctoral dissertation).

Lim, N., Haddad, A., Butler, D. M., & Giglio, K. (2013). First Steps Toward Improving DoD STEM

Workforce Diversity. RAND Corporations

Lüdtke, O., Marsh, H. W., Robitzsch, A., Trautwein, U., Asparouhov, T., & Muthén, B. (2008). The

multilevel latent covariate model: a new, more reliable approach to group-level effects in contextual

studies. Psychological Methods,13(3), 203.

Lyons, E. M. (2008). Examining the effects of problem-based learning and NCLEX-RN scores on the

critical thinking skills of associate degree nursing students in a southeastern community

college. International Journal of Nursing Education Scholarship, 5(1), 1-17.

Ma, X., & Klinger, D. A. (2000). Hierarchical linear modelling of student and school effects on academic

achievement. Canadian Journal of Education, Vol. 25: 41-55.

Ma, X., Ma, L., & Bradley, K. D. (2008). Using multilevel modeling to investigate school effects. In A. A.

O’Connell & D. B. McCoach (Eds.), Multilevel modeling of educational data (pp. 59-110). Charlotte, NC:

Information Age.

Manski, C. F. (1993). Identification of endogenous social effects: The reflection problem. The review of

economic studies, 60(3), 531-542.

Marcus, S. M., & Gibbons, R. D. (2012). Caution should be used in applying propensity scores estimated

in a full cohort to adjust for confounding in subgroup analyses: commentary on “Applying propensity

scores estimated in a full cohort to adjust for confounding in subgroup analyses”.

Pharmacoepidemiology and drug safety, 21(7), 710-712.

Marsh, H. W., Lüdtke, O., Robitzsch, A., Trautwein, U., Asparouhov, T., Muthén, B., & Nagengast, B.

(2009). Doubly-latent models of school contextual effects: Integrating multilevel and structural equation

approaches to control measurement and sampling error. Multivariate Behavioral Research, 44(6), 764-

802.

Marsh, J. L., Hutton, J. L., & Binks, K. (2002). Removal of radiation dose response effects: an example of

over-matching. BMJ: British Medical Journal, 325(7359), 327.

Martinez, Miguel (2014). Assessment Report: Center for Teaching Advancement and Assessment

Research.

Martinez, Miguel (2016). Climate Survey Report.

Mattern, K. D., Patterson, B. F., Shaw, E. J., Kobrin, J. L., & Barbuti, S. M. (2008). Differential Validity and

Prediction of the SAT®. Research Report No. 2008-4. College Board

McGann, E., & Thompson, J. M. (2008). Factors related to academic success in at-risk senior nursing

students. International Journal of Nursing Education Scholarship, 5(1), 1-15.

Page 96: Three Essays on the Economics of Education Miguel E. Martinez

88

Michigan.

Miller, L. C., & Mittleman, J. (2012). High Schools That Work and college preparedness: Measuring the

model's impact on mathematics and science pipeline progression. Economics of Education Review.

31(6), 1116–1135.

Moogan, Y. J., & Baron, S. (2003). An analysis of student characteristics within the student decision

making process. Journal of Further and Higher Education,27(3), 271-287.

Morgan, S. L., Gelbgiser, D., & Weeden, K. A. (2013). Feeding the Pipeline: Gender, Occupational Plans,

and College Major Selection∗. Social science research.

Morris, T., & Hancock, D. (2008). Program exit examinations in nursing education: Using a value added

assessment as a measure of the impact of a new curriculum. Educational Research Quarterly, 32(2), 19.

Morton, A. M. (2006). Improving NCLEX scores with structured learning assistance. Nurse

Educator, 31(4), 163-165.

Muller, C. Farkas, G., and Riegle-Crumb, C. (2006). The role of gender and friendship in course-taking.

Sociology of Education, Vol. 79(3), 208-228.

multilevel and nonlinear settings. Unpublished doctoral dissertation. University of

Nagengast, B., & Marsh, H.W. (2012). Big fish in little ponds aspire more: Mediation and cross-cultural

generalizability of school-average ability effects on self-concept and career aspirations in

science. Journal of Educational Psychology, 104, 1033-1053.

Nibert, A. and Young, A. (2001). A third study on predicting NCLEX success with the HESI Exit Exam.

Computers in Nursing. 19: 172–178.

Nibert, A. T., & Young, A. (2006). A Third Study on Predicting NCLEX Success With the HESI Exit

Exam. Computers Informatics Nursing, 24, 21S-27S.

Nibert, A., Young, A., and Adamson, C (2002). Predicting NCLEX success with the HESI Exit Exam: Fourth

annual validity study. Computers, Informatics, Nursing. 20: 261–267.

Norfleet James, A., & Richards, H. C. (2003). Escaping stereotypes: Educational attitudes of male alumni

of single-sex and coed schools.Psychology of Men & Masculinity, 4(2), 136.

Norton, C. K., Relf, M. V., Cox, C. W., Farley, J., Lachat, M., Tucker, M., & Murray, J. (2006). Ensuring

NCLEX-RN success for first-time test-takers. Journal of Professional Nursing, 22(5), 322-326.

Parker, C.E. (2010). Project RISE Pilot Study: Methods in Longitudinal Studies of Youth in Informal STEM

Education. Presented at the American Educational Research Association Annual Conference, Denver,

Colorado, May 2010.

Parrone, J., Sredl, D., Miller, M., Phillips, M., & Donaubauer, C. (2008). An evidence-based

teaching/learning strategy for foreign nurses involving the Health Education Systems Incorporated

examination as a predictor for National Council Licensure Examination for Registered Nurses

success. Teaching and Learning in Nursing, 3(1), 35-40.

Page 97: Three Essays on the Economics of Education Miguel E. Martinez

89

Pennington, T. D., & Spurlock Jr, D. (2010). A systematic review of the effectiveness of remediation

interventions to improve NCLEX-RN pass rates. Journal of Nursing Education, 49(9), 485.

Rabe-Hesketh, S., & Skrondal, A. (2006). Multilevel modelling of complex survey data. Journal of the

Royal Statistical Society: Series A (Statistics in Society), 169(4), 805-827.

Raudenbush, S. W. (2002). Hierarchical linear models: Applications and data analysis methods (Vol. 1).

Sage.

Raudenbush, S. W., & Willms, J. (1995). The estimation of school effects.Journal of educational and

behavioral statistics, 20(4), 307-335.

Robertson, K. F., Smeets, S., Lubinski, D., & Benbow, C. P. (2010). Beyond the Threshold Hypothesis Even

Among the Gifted and Top Math/Science Graduate Students, Cognitive Abilities, Vocational Interests,

and Lifestyle Preferences Matter for Career Choice, Performance, and Persistence. Current Directions in

Psychological Science, 19(6), 346-351.

Romer, P. M. (1990). Endogenous technological change. Journal of Political Economy, S71-S102.

Rosenbaum, J. E. (1980). Track misperceptions and frustrated college plans: An analysis of the effects of

tracks and track perceptions in the National Longitudinal Survey. Sociology of Education, 74-88.

Rothwell, J. (2013). The Hidden STEM Economy: The Surprising Diversity of Jobs Requiring Science,

Technology, Engineering, and Math Knowledge.

Rothwell, Jonathan and Ruiz, Neil (2012). H-1B Visas and the STEM Shortage. Brookings Institute.

Retrieved from http://www.brookings.edu/research/papers/2013/05/10-h1b-visas-stem-rothwell-ruiz

Ruggles, S., J. T. Alexander, K. Genadek, R. Goeken, M. B. Schroeder, and M. Sobek. (2010). 2010

American Community Survey data. Edited by University of Minnesota, Integrated Public Use Microdata

Series, Minneapolis, MN.

Rule, R. , Has, and Nugent (2004). The use of discriminant function analysis to predict student success on

the NCLEX-RN. Journal of Nursing Education, 43(10), 440.

Russell, M. L., & Atwater, M. M. (2005). Traveling the road to success: A discourse on persistence

throughout the science pipeline with African American students at a predominantly white

institution. Journal of Research in Science Teaching, 42(6), 691-715.

Salzman, H., Kuehn, D., & Lowell, B. L. (2013). Guestworkers in the High-Skill US Labor Market. Economic

Policy Institute Briefing Paper, 359.

Sander, R. H. (2004). A systemic analysis of affirmative action in American law schools. Stanford Law

Review, 367-483.

Sanders, Mark (2009). STEM, STEM Education, STEMmania. The Technology Teacher,

December/January: 20-26.

Santelices, M. V., & Wilson, M. (2010). Unfair treatment? The case of Freedle, the SAT, and the

standardization approach to differential item functioning. Harvard Educational Review, 80(1), 106-134.

Page 98: Three Essays on the Economics of Education Miguel E. Martinez

90

Santiago, L. A. (2013). Forecasting nursing student success and failure on the NCLEX-RN using predictor

tests (Doctoral dissertation).

Sayles, S., Shelton, D., & Powell, H. (2003). Predictors of success in nursing education. ABNF

Journal, 14(6), 116.

Schreiner, B. Barton, L., Willson, P., & Langford, R (2014). Standardized Predictive Testing: Practices,

Policies and Outcomes. Administrative Issues Journal 4(2), 68-76.

Schroeder, J. (2013). Improving NCLEX-RN pass rates by implementing a testing policy. Journal of

Professional Nursing, 29(2), S43-S47.

Sifford, S., & McDANIEL, D. M. (2007). Results of a remediation program for students at risk for failure

on the NCLEX exam. Nursing Education Perspectives, 28(1), 34-36.

Smith, J., Pender, M., & Howell, J. (2013). The full extent of student-college academic

undermatch. Economics of Education Review, 32, 247-261.

Southworth, S. and Mickelson, R. (2007). The Interactive Effects of Race, Gender, and School

Composition on College Track Placement. Social Forces, Vol. 86(2), 497-523.

Spurlock Jr, D. R. (2008). A study of the usefulness of the HESI exit exam in predicting NCLEX-RN

failure. Journal of Nursing Education, 47(4), 157.

Stuenkel, D. L. (2006). At-risk students: do theory, grades, standardized examinations, promote

success?. Nurse Educator, 31(5), 207-212.

Sutherland, J. A., Hamilton, M. J., & Goodman, N. (2007). Affirming At-Risk Minorities for Success

(ARMS): Retention, graduation, and success on the NCLEX-RN®. Journal of Nursing Education, 46(8), 347-

353.

Tai, R. H., Liu, C., Maltese, A, and Fan, X. (2006). CAREER CHOICE: Enhanced: Planning Early for Careers in

Science. Life sci, 1, 1443-1444.

Taylor, H., Loftin, C., & Reyes, H. (2014). First-Time NCLEX-RN Pass Rate: Measure of Program Quality or

Something Else?. Journal of Nursing Education,53(6).

Thompson, Jennifer (2003). The Effect of Single-Sex Secondary Schooling on Women’s Choice of College

Major. Sociological Perspectives, Volume 46 (2), 257–278.

Tipton, P., Pulliam, M., Beckworth, C., Illich, P., Griffin, R., & Tibbitt, A. (2008). Predictors of associate

degree nursing students’ success students. Southern Online Journal of Nursing Research, 8(1), 8.

Toma, J. D., & Cross, M. E. (1998). Intercollegiate athletics and student college choice: Exploring the

impact of championship seasons on undergraduate applications. Research in Higher Education, 39(6),

633-661.

Tracey, T. J., Robbins, S. B., & Hofsess, C. D. (2005). Stability and change in interests: A longitudinal study

of adolescents from grades 8 through 12. Journal of Vocational Behavior, 66(1), 1-25.

Page 99: Three Essays on the Economics of Education Miguel E. Martinez

91

Uyehara, J., Magnussen, L., Itano, J., & Zhang, S. (2007, January). Facilitating Program and NCLEX-RN

Success in a Generic BSN Program. In Nursing Forum (Vol. 42, No. 1, pp. 31-38).

Van Langen, Annemarie and Dekkers, Hetty, (2005). Cross-National Differences in Participating in

Tertiary Science, Technology, Engineering and Mathematics Education.

Wang, X. (2013). Why Students Choose STEM Majors Motivation, High School Learning, and

Postsecondary Context of Support. American Educational Research Journal.

Ware, N. C. & Lee, V. E. (1988). Sex differences in choice of college science majors. American Educational

Research Journal, Vol. 25: 593-614.

White House Office of Science and Technology Policy (2013). Preparing a 21st Century Workforce

Science, Technology, Engineering, and Math

Wilkinson, Ian, Judy M. Parr, Irene Y.Y. Fung, John A.C. Hattie, Michael A.R. Townsend (2002).

Discussion: modeling and maximizing peer effects in school. International Journal of Educational

Research, Vol. 37(5), 521-535.

Wiswall, M., & Zafar, B. (2015). Determinants of college major choice: Identification using an

information experiment. The Review of Economic Studies, 82(2), 791-824.

Wiswall, M., & Zafar, B. (2015). How do college students respond to public information about

earnings?. Journal of Human Capital, 9(2), 117-169.

Wray, K., Whitehead, T., Setter, R., & Treas, L. (2006). Use of NCLEX preparation strategies in a hospital

orientation program for graduate nurses. Nursing Administration Quarterly, 30(2), 162-177.

Young, A., & Willson, P. (2012). Predicting NCLEX-RN success: The seventh validity study HESI exit

exam. Computers Informatics Nursing, 30(1), 55-60.

Yu, Bing (2012). Variable Selection and Adjustment in Relation to Propensity Scores and Prognostic

Scores: From Single Level to Multi-level Data. Unpublished doctoral dissertation. University of Toronto.

Zanutto, E. L. (2006). A comparison of propensity score and linear regression analysis of complex survey

data. Journal of Data Science, 4(1), 67-91.

Zanutto, E., Lu, B., & Hornik, R. (2005). Using propensity score subclassification for multiple treatment

doses to evaluate a national antidrug media campaign. Journal of Educational and Behavioral Statistics,

30(1), 59-73.

Zinth, Kyle (2006). Recent State STEM Initiatives. Denver: Education Commission of the States.

Zweighaft, E. L. (2013). Impact of HESI specialty exams: The ninth HESI Exit Exam validity study. Journal

of Professional Nursing, 29(2), S10-S16.

Page 100: Three Essays on the Economics of Education Miguel E. Martinez

92

Appendices Appendix A

Years 2009-2015

NCLEX Odds Ratio SE z P>|z|

Lower Limit

Upper Limit

Demographics

EOF Status 1.15 0.58 0.28 0.78 0.43 3.07

American Indian 1.00 (omitted)

Asian 1.13 0.64 0.21 0.83 0.37 3.46

Hispanic 0.48 0.29 -1.21 0.23 0.15 1.57

International 1.00 (omitted)

Multiracial 0.60 0.48 -0.64 0.53 0.13 2.87

Native 1.67 1.60 0.53 0.60 0.25 10.97

Other 1.00 (omitted)

Unknown 0.40 0.43 -0.86 0.39 0.05 3.28

White 1.08 0.64 0.13 0.90 0.34 3.44

Male 1.60 0.80 0.94 0.35 0.60 4.25

Newark Campus 0.54 0.24 -1.40 0.16 0.23 1.28

Blackwood Campus 1.00 (omitted)

Academic Readiness

SAT Verbal Score 1.00 0.00 0.85 0.39 1.00 1.01

SAT Math Score 1.00 0.00 -0.02 0.98 0.99 1.01

Academic Performance

HC Delivery 0.59 0.19 -1.60 0.11 0.31 1.13

Pathophysiology 1.40 0.39 1.21 0.23 0.81 2.40

Health Assessment 1.30 0.47 0.73 0.47 0.64 2.64

Foundations I 3.21† 1.24 3.04 0.00 1.51 6.83

Childbearing Family 1.54 0.63 1.05 0.30 0.69 3.45

Health and Illness ICA 3.03† 1.27 2.63 0.01 1.33 6.91

Health and Illness AOA I 2.50 1.11 2.06 0.04 1.05 5.95

Foundations II 0.64 0.27 -1.05 0.30 0.28 1.47

Pharmacotherapeutics 1.10 0.33 0.31 0.76 0.61 1.97

Psych Mental Health 2.55 0.95 2.51 0.01 1.23 5.30

Health and Illness AOA II 1.15 0.54 0.29 0.77 0.45 2.90

Year Fixed Effects

Year 2010 0.15 0.11 -2.56 0.01 0.04 0.64

Year 2011 0.30 0.26 -1.41 0.16 0.06 1.61

Year 2012 0.22 0.20 -1.70 0.09 0.04 1.25

Page 101: Three Essays on the Economics of Education Miguel E. Martinez

93

Year 2013 0.51 0.50 -0.69 0.49 0.08 3.46

Year 2014 0.13 0.11 -2.43 0.02 0.03 0.68

Year 2015 1.00 (omitted)

Intercept 0.00 0.00 -4.82 0.00 0.00 0.00

Years 2009-2012

NCLEX Odds Ratio SE z P>|z|

Lower Limit

Upper Limit

Demographics

EOF Status 1.30 1.05 0.32 0.75 0.27 6.31

American Indian 1.00 (omitted)

Asian 2.31 1.88 1.03 0.30 0.47 11.36

Hispanic 1.95 1.69 0.77 0.44 0.36 10.62

International 1.00 (omitted)

Multiracial 2.43 3.40 0.64 0.53 0.16 37.66

Native 3.98 5.98 0.92 0.36 0.21 75.77

Other 1.00 (omitted)

Unknown 1.00 (omitted)

White 1.62 1.27 0.61 0.54 0.35 7.53

Male 1.52 1.24 0.51 0.61 0.31 7.56

Newark Campus 0.69 0.42 -0.60 0.55 0.21 2.30

Blackwood Campus 1.00 (omitted)

Academic Readiness

SAT Verbal Score 1.01 0.00 1.87 0.06 1.00 1.02

SAT Math Score 1.00 0.00 -0.39 0.70 0.99 1.01

Academic Performance

HC Delivery 0.48 0.23 -1.51 0.13 0.18 1.25

Pathophysiology 1.29 0.55 0.59 0.56 0.56 2.98

Health Assessment 0.95 0.59 -0.08 0.94 0.28 3.20

Foundations I 4.36† 2.41 2.66 0.01 1.47 12.90

Childbearing Family 2.46 1.58 1.40 0.16 0.70 8.69

Health and Illness ICA 5.02† 3.20 2.53 0.01 1.44 17.51

Health and Illness AOA I 2.79 1.79 1.60 0.11 0.79 9.81

Foundations II 0.45 0.30 -1.19 0.24 0.12 1.68

Pharmacotherapeutics 1.03 0.46 0.07 0.95 0.43 2.46

Psych Mental Health 1.45 0.85 0.63 0.53 0.46 4.60

Health and Illness AOA II 0.89 0.64 -0.16 0.87 0.22 3.60

Year Fixed Effects

Year 2010 0.47 0.48 -0.73 0.46 0.06 3.51

Year 2011 0.76 0.74 -0.28 0.78 0.11 5.15

Year 2012 0.49 0.45 -0.78 0.44 0.08 3.00

Page 102: Three Essays on the Economics of Education Miguel E. Martinez

94

Intercept 0.00 0.00 -3.77 0.00 0.00 0.00

Years 2010-2013

NCLEX Odds Ratio SE z P>|z|

Lower Limit

Upper Limit

Demographics

EOF Status 1.30 1.05 0.32 0.75 0.27 6.31

American Indian 1.00 (omitted)

Asian 2.31 1.88 1.03 0.30 0.47 11.36

Hispanic 1.95 1.69 0.77 0.44 0.36 10.62

International 1.00 (omitted)

Multiracial 2.43 3.40 0.64 0.53 0.16 37.66

Native 3.98 5.98 0.92 0.36 0.21 75.77

Other 1.00 (omitted)

Unknown 1.00 (omitted)

White 1.62 1.27 0.61 0.54 0.35 7.53

Male 1.52 1.24 0.51 0.61 0.31 7.56

Newark Campus 0.69 0.42 -0.60 0.55 0.21 2.30

Blackwood Campus 1.00 (omitted)

Academic Readiness

SAT Verbal Score 1.01 0.00 1.87 0.06 1.00 1.02

SAT Math Score 1.00 0.00 -0.39 0.70 0.99 1.01

Academic Performance

HC Delivery 0.48 0.23 -1.51 0.13 0.18 1.25

Pathophysiology 1.29 0.55 0.59 0.56 0.56 2.98

Health Assessment 0.95 0.59 -0.08 0.94 0.28 3.20

Foundations I† 4.36 2.41 2.66 0.01 1.47 12.90

Childbearing Family 2.46 1.58 1.40 0.16 0.70 8.69

Health and Illness ICA† 5.02 3.20 2.53 0.01 1.44 17.51

Health and Illness AOA I 2.79 1.79 1.60 0.11 0.79 9.81

Foundations II 0.45 0.30 -1.19 0.24 0.12 1.68

Pharmacotherapeutics 1.03 0.46 0.07 0.95 0.43 2.46

Psych Mental Health 1.45 0.85 0.63 0.53 0.46 4.60

Health and Illness AOA II 0.89 0.64 -0.16 0.87 0.22 3.60

Year Fixed Effects

Year 2010 0.47 0.48 -0.73 0.46 0.06 3.51

Year 2011 0.76 0.74 -0.28 0.78 0.11 5.15

Page 103: Three Essays on the Economics of Education Miguel E. Martinez

95

Year 2012 0.49 0.45 -0.78 0.44 0.08 3.00

Intercept 0.00 0.00 -3.77 0.00 0.00 0.00

Years 2011-2014

NCLEX Odds Ratio SE z P>|z|

Lower Limit

Upper Limit

Demographics

EOF Status 1.03 0.73 0.04 0.97 0.25 4.16

American Indian 1.00 (omitted)

Asian 1.56 1.15 0.60 0.55 0.37 6.61

Hispanic 2.40 2.11 0.99 0.32 0.42 13.51

International 1.00 (omitted)

Multiracial 1.66 1.69 0.50 0.62 0.23 12.21

Native 2.99 4.17 0.78 0.43 0.19 46.07

Other 1.00 (omitted)

Unknown 0.81 0.94 -0.18 0.86 0.08 7.93

White 0.98 0.74 -0.02 0.98 0.23 4.29

Male 0.66 0.46 -0.59 0.55 0.17 2.62

Blackwood Campus 1.00 (omitted)

Newark Campus† 0.27 0.16 -2.18 0.03 0.08 0.87

Academic Readiness

SAT Verbal Score 1.00 0.00 1.14 0.25 1.00 1.01

SAT Math Score 1.00 0.00 -0.64 0.52 0.99 1.01

Academic Performance

HC Delivery 0.83 0.37 -0.43 0.67 0.35 1.97

Pathophysiology 1.29 0.47 0.71 0.48 0.63 2.64

Health Assessment 2.04 1.03 1.42 0.16 0.76 5.50

Foundations I† 3.24 1.62 2.35 0.02 1.22 8.64

Childbearing Family 1.07 0.60 0.12 0.91 0.35 3.23

Health and Illness ICA† 4.00 2.33 2.38 0.02 1.28 12.52

Health and Illness AOA I 3.06 1.88 1.82 0.07 0.92 10.23

Foundations II 0.46 0.28 -1.29 0.20 0.14 1.49

Pharmacotherapeutics 1.45 0.56 0.95 0.34 0.67 3.10

Psych Mental Health† 4.12 2.12 2.76 0.01 1.51 11.28

Health and Illness AOA II 0.32 0.23 -1.58 0.12 0.08 1.32

Year Fixed Effects

Year 2011 1.40 1.22 0.38 0.70 0.25 7.75

Year 2012 1.44 1.27 0.41 0.68 0.26 8.09

Year 2013 3.88 3.02 1.74 0.08 0.84 17.83

Intercept 0.00 0.00 -3.70 0.00 0.00 0.00

Page 104: Three Essays on the Economics of Education Miguel E. Martinez

96

Years 2012-2015

NCLEX Odds Ratio SE z P>|z|

Lower Limit

Upper Limit

Demographics

EOF Status 2.55 1.72 1.38 0.17 0.67 9.60

American Indian 1.00 (omitted)

Asian 1.62 1.21 0.65 0.52 0.38 7.00

Hispanic 0.63 0.50 -0.58 0.56 0.13 3.04

International 1.00 (omitted)

Multiracial 0.46 0.48 -0.75 0.45 0.06 3.50

Native 4.23 5.30 1.15 0.25 0.36 49.24

Other 1.00 (omitted)

Unknown 0.30 0.37 -0.96 0.34 0.03 3.45

White 0.90 0.69 -0.14 0.89 0.20 4.08

Male 0.99 0.65 -0.02 0.98 0.27 3.61

Newark Campus 0.31 0.18 -2.04 0.04 0.10 0.96

Blackwood 1.00 (omitted)

Academic Readiness

SAT Verbal Score 1.00 0.00 0.47 0.64 1.00 1.01

SAT Math Score 1.00 0.00 -0.14 0.89 0.99 1.01

Academic Performance

HC Delivery 0.80 0.36 -0.50 0.62 0.34 1.91

Pathophysiology 1.01 0.37 0.04 0.97 0.50 2.06

Health Assessment 1.71 0.80 1.14 0.25 0.68 4.30

Foundations I† 4.94 2.61 3.02 0.00 1.75 13.91

Childbearing Family 1.35 0.79 0.51 0.61 0.43 4.25

Health and Illness ICA 2.45 1.30 1.69 0.09 0.87 6.92

Health and Illness AOA I 2.65 1.55 1.67 0.10 0.84 8.33

Foundations II 0.42 0.24 -1.49 0.14 0.13 1.32

Pharmacotherapeutics 1.24 0.50 0.53 0.60 0.56 2.71

Psych Mental Health† 4.71 2.41 3.02 0.00 1.72 12.86

Health and Illness AOA II 0.80 0.53 -0.34 0.73 0.22 2.90

Year Fixed Effects

Year 2012 0.53 1.13 -0.30 0.77 0.01 33.81

Year 2013 1.00 1.99 0.00 1.00 0.02 50.03

Year 2014 0.38 0.75 -0.49 0.62 0.01 18.62

Page 105: Three Essays on the Economics of Education Miguel E. Martinez

97

Intercept 0.00 0.00 -3.21 0.00 0.00 0.01

Appendix B

15 Point Bandwidth

7.5 Point Bandwidth

0

.00

2.0

04

.00

6

-1000 -500 0 500-1000 -500 0 500

Remediaton Not Required Remediation Required

Den

sity

Centered HESI Exit Exam Score

Page 106: Three Essays on the Economics of Education Miguel E. Martinez

98

Appendix C

McCrary Test Results by Bandwidth.

Bandwidth for Wald Test Statistic Coef. SE z P>z LB UB

0 0.10 0.12 0.84 0.40 -0.13 0.33

25 0.00 (omitted)

30 0.00 (omitted)

35 0.00 (omitted)

40 0.00 (omitted)

45 0.00 (omitted)

50 0.00 (omitted)

55 0.00 (omitted)

60 0.10 0.12 0.84 0.40 -0.13 0.33

65 0.00 (omitted)

70 0.10 0.12 0.84 0.40 -0.13 0.33

75 0.10 0.12 0.84 0.40 -0.13 0.33

80 0.10 0.12 0.84 0.40 -0.13 0.33

85 0.10 0.12 0.84 0.40 -0.13 0.33

90 0.10 0.12 0.84 0.40 -0.13 0.33

95 0.10 0.12 0.84 0.40 -0.13 0.33

0

.00

2.0

04

.00

6.0

08

-1000 -500 0 500-1000 -500 0 500

Remediation Not Required Remediation RequiredD

en

sity

Centered HESI Exit Exam Score

Page 107: Three Essays on the Economics of Education Miguel E. Martinez

99

100 0.00 (omitted)

105 0.10 0.12 0.84 0.40 -0.13 0.33

110 0.00 (omitted)

115 0.27 0.22 1.20 0.23 -0.17 0.71

120 0.27 0.22 1.21 0.23 -0.17 0.71

125 0.27 0.22 1.22 0.22 -0.17 0.71

130 0.27 0.22 1.22 0.22 -0.17 0.71

135 0.27 0.22 1.23 0.22 -0.16 0.71

140 0.27 0.22 1.23 0.22 -0.16 0.71

145 0.28 0.22 1.23 0.22 -0.16 0.71

150 0.28 0.22 1.23 0.22 -0.16 0.71

155 0.28 0.22 1.24 0.22 -0.16 0.71

160 0.28 0.22 1.24 0.22 -0.16 0.71

165 0.28 0.22 1.24 0.22 -0.16 0.72

170 0.27 0.22 1.24 0.22 -0.16 0.71

175 0.22 0.21 1.05 0.29 -0.19 0.63

180 0.19 0.20 0.92 0.36 -0.21 0.59

185 0.16 0.20 0.82 0.42 -0.23 0.56

190 0.15 0.20 0.74 0.46 -0.24 0.53

195 0.13 0.20 0.68 0.50 -0.25 0.52

200 0.12 0.19 0.63 0.53 -0.26 0.50

205 0.11 0.19 0.58 0.56 -0.27 0.49

210 0.11 0.19 0.55 0.59 -0.27 0.48

215 0.10 0.19 0.51 0.61 -0.28 0.48

220 0.09 0.19 0.49 0.63 -0.28 0.47

225 0.09 0.19 0.46 0.65 -0.29 0.46

230 0.08 0.18 0.46 0.64 -0.27 0.43

235 0.08 0.17 0.47 0.64 -0.25 0.40

240 0.07 0.16 0.47 0.64 -0.24 0.39

245 0.07 0.15 0.47 0.64 -0.23 0.37

250 0.07 0.15 0.46 0.64 -0.22 0.36

Appendix D

The CJM heaping test under triangular, uniform, and Epinochev kernel distributional

assumptions do not yield statistical evidence of manipulation of the running variable at the 95%

confidence level. The Table below summarizes the parameters associated with the CJM heaping

test under the triangular distribution.

Page 108: Three Essays on the Economics of Education Miguel E. Martinez

100

Results of CJM density test

Cutoff c = 0.000 Left of c Right of c

Number of obs 3,404 2,105

Effective Number of obs 1,551 840

Order local polynomial 2 2

Order BC 3 3

Bandwidth 63 70

T Statistic -1.66

Appendix E

The table below summarizes the continuity of the demographics, performance in nursing

courses, and academic readiness measures for the 18 percent of the sample that has those data

available. Seventeen out of the eighteen variables do not yield any evidence of violating the

continuity assumption. The first order polynomial specification, however, picks up a negative

statistically significant (s) discontinuity for Psych/Mental Health course at the cutoff of 850. This

holds true even when the cut-off is moved to 825 (negative discontinuity) and 800 (positive

discontinuity). The discontinuities present at 825 and 800 suggest that the discontinuity at 850

may be a statistical artifact. None of the other 17 variables pick up significance when the cut-off

is artificially moved to 800 and to 825. In addition, second and third order polynomial

specifications detect only non-statistically significant (ns) discontinuities for Psych/Mental Health

at 850. Thus, the finding for the negative discontinuity for Psych/Mental Health is relatively

unstable. This leads to the conclusion that there is no discontinuity in pre-treatment measures

Page 109: Three Essays on the Economics of Education Miguel E. Martinez

101

and extrapolating this finding from the subsample to the full sample of students in institutions

with required remediation. Nevertheless, it should be noted that if the finding were to be valid

then the discontinuity would downwardly bias the effect of remediation.

Continuity of Demographics and Academic Performance and Readiness at Cutoff

Order of Polynomials

1 2 3

Demographics

EOF Status ns ns ns

Minority ns ns ns

Male ns ns ns

Main Campus ns ns ns

Academic Readiness

SAT Verbal Score ns ns ns

SAT Math Score ns ns ns

Academic Performance

HC Delivery ns ns ns

Pathophysiology ns ns ns

Health Assessment ns ns ns

Foundations I ns ns ns

Childbearing Family ns ns ns

Health and Illness ICA ns ns ns

Health and Illness AOA I ns ns ns

Foundations II ns ns ns

Pharmacotherapeutics ns ns ns

Psych Mental Health s ns ns

Health and Illness AOA II ns ns ns

Page 110: Three Essays on the Economics of Education Miguel E. Martinez

102

Appendix F

As the table below points out, second order polynomial bias-corrected estimates are not

sensitive to increases/decreases of bandwidth of 10 and 20 percent.

Falsification Exercise for Table 7 Statistically Significant Results

Polynomial Order Baseline 2nd: b=137 & h=74

Difference from Baseline -0.2 -0.1 0.1 0.2

Bandwidth (h) 60 67 81 89

Bandwidth (b) 110 123 150 164

Methodology β β β β

Conventional (b) 0.02 0.03 0.03 0.03

Bias-corrected (h) (0.22) †† (0.25) †† (0.29) †† (0.25) ††

Robust (h) (0.22) (0.25) (0.29) †† (0.25) ††

†statistically significant at 95% level ††statistically significant at 99% level

Statistically insignificant results are not sensitive to changes in bandwidth. The

falsification exercise was structured as follows. For each order polynomial n for bandwidth

selector type i on Table 4, the difference between the maximum and the minimum values is

Page 111: Three Essays on the Economics of Education Miguel E. Martinez

103

calculated and then divided by four to get x. Three bandwidths for the falsification/sensitivity

exercises for each order polynomial n for bandwidth selector type i were calculated as:

Minimum value + x Minimum value + 2x Minimum value + 3x Thus, for each order polynomial, treatment effects are estimated at three additional bandwidths

within the values of the largest and smallest bandwidths produced by the various bandwidth

selectors.

Falsification Exercise for All Table 4 Results

Polynomial Order 1st 2nd 3rd

Bandwidth (b) 102 129 156 171 268 365 178 265 352

Bandwidth (h) 105 131 157 128 240 322 167 258 339

Methodology β β β β β β β Β β

Conventional (b) 0.02 0.02 0.02 0.03 0.02 0.02 0.02 0.03 0.03

Bias-corrected (h) 0.02 0.03 0.03 0.02 0.03 0.03 0.01 0.02 0.03

Robust (h) 0.02 0.03 0.03 0.02 0.03 0.03 0.01 0.02 0.03

†statistically significant at 95% level ††statistically significant at 99% level

All the estimates are statistically insignificant.

Page 112: Three Essays on the Economics of Education Miguel E. Martinez

104

Appendix G

The observations of the sample that have demographics, SAT scores, and performance in

nursing courses suggest that covariate balance may exist within 10 points of the cutoff. Table 5

summarizes tests of proportions bandwidths of ten points and lower. At all bandwidths, the

first-time pass rates for students subject to remediation are lower than those who met or

surpassed the cutoff. None of the differences in pass rates are, however, statistically significant.

Overall, the results based on zero order polynomials do not suggest either a negative nor positive

impact.

Table 5. Results of Test of Proportions by Bandwidth

Treatment Control Bandwidth Pass Rate n Pass Rate n Difference

10 89.4% 104 93.1% 250 -3.7%

9 88.7% 97 92.2% 226 -3.6%

8 88.5% 87 93.0% 206 -4.5%

7 88.3% 77 92.9% 186 -4.6%

6 88.6% 70 93.2% 169 -4.7%

Page 113: Three Essays on the Economics of Education Miguel E. Martinez

105

5 87.9% 58 92.6% 148 -4.6%

4 89.4% 47 92.1% 121 -2.7%

3 87.9% 33 90.6% 101 -2.7%

2 92.0% 25 93.5% 85 -1.5%

1 83.3% 12 93.3% 62 -10.0%

†statistically significant at 95% level ††statistically significant at 99% level

Localized Randomized Experiment Estimates: First, Second, and Third Order Polynomials

Results of Table 5 are sensitive to functional form and distributional assumptions. Table

6 summarizes the statistical significance of treatment estimates derived using bandwidths of 10

points or less. All estimates using bandwidths under each kernel function - uniform (Uni),

triangular(Tri), and Epanechnikov (Epa) – are statistically insignificant (ns). The only statistically

significant (s) result is positive is at the four-point bandwidth but it is sensitive to functional

assumptions.

Table 6. Results of Treatment Effect Estimates by Polynomial Order/Functional Assumption

1st Order Polynomial 2nd Order Polynomial 3rd Order Polynomial

Bandwidth Uni Tri Epa Uni Tri Epa Uni Tri Epa

10 ns ns ns ns ns ns ns ns ns

9 ns ns ns ns ns ns ns ns ns

8 ns ns ns ns ns ns ns ns ns

7 ns ns ns ns ns ns ns ns ns

6 ns ns ns ns ns ns ns ns ns

5 ns ns ns ns ns ns ns ns ns

4 ns ns ns ns ns ns s ns ns

3 ns ns ns ns ns ns ns ns ns

2 ns ns ns ns ns ns ns ns ns

1 -16 - - - - - - - -

16 - indicates that estimates could not be calculated

Page 114: Three Essays on the Economics of Education Miguel E. Martinez

106

Appendix H.

Variable Odds Ratio Std. Err. z P>z UB LB

HESI Score 1.01 0.00 16.05 0.00 1.01 1.01

Treatment 1.14 0.18 0.82 0.41 0.83 1.56

Required Remediaton Policy 2.28 1.86 1.01 0.31 0.46 11.31

Appendix I

Variables No Remediation Remediation Difference

Page 115: Three Essays on the Economics of Education Miguel E. Martinez

107

Student Characteristics

Non-White 66% 62% 5%

White 34% 38% -5%

Male 13% 9% 3%

Main Campus 51% 47% 4%

EOF 6% 10% -4%

Pre-admission Academics

Composite SAT 1117 1152 -37

SAT Verbal 543 564 -22

SAT Math 574 588 -15

Observations 98 93

% of treatment/control 23% 33%

**statistically significant at 95% level *** statistically significant at 99% level

Appendix J

Area of Common Support for All Observations with Scores Lower than 850

Page 116: Three Essays on the Economics of Education Miguel E. Martinez

108

Variable Treated Control %bias T-Statistic p>t

HESI Score 742.13 737.65 5.7 1.35 0.178

BSN School 51% 51% 0 0 1

PN School 10% 11% -0.5 -0.13 0.895

P Score 0.59 0.59 0 0.01 0.993

Appendix K

Area of Common Support for All Observations with Scores Lower than 600

.4 .5 .6 .7 .8Propensity Score

Untreated Treated: On support

Treated: Off support

Page 117: Three Essays on the Economics of Education Miguel E. Martinez

109

Variable Treated Control %bias T-Statistic p>t

HESI_score 553.67 557.91 -7.9 -0.73 0.469

BSN School 82% 81% 2.3 0.19 0.852

P Score 0.78 0.79 -1.5 -0.13 0.898

.4 .5 .6 .7 .8 .9Propensity Score

Untreated Treated: On support

Treated: Off support

Page 118: Three Essays on the Economics of Education Miguel E. Martinez

110

Appendix L

Selectivity of First PSI Attended Freshmen: STEM

Freshmen: non-STEM

n % n %

Highly - 4 year 1,799 20% 569 36%

Moderately - 4 year 2,418 27% 449 28%

Inclusive - 4 year 610 7% 113 7%

Selectivity not classified - 4 year 445 5% 82 5%

Selectivity not classified - 2 year 3,318 37% 364 23%

Selectivity not classified - less than 2 years 288 3% 25 2%

Total 8,878 100% 1,602 100%

Appendix M

Selectivity

Page 119: Three Essays on the Economics of Education Miguel E. Martinez

111

Enrollment Status of STEM Freshmen at First PSI Attended at Third Follow-up

Highly Moderately

n % n %

Earned a credential from PSI; still attending PSI as of 2012

15 3% 11 3%

Earned a credential from PSI; no longer attending PS1 as of 2012

336 64% 181 45%

No cred from PSI; still attending PSI as of 2012

4 1% 14 4%

No cred from PSI; no longer attending PSI; did attend another PS institution

161 30% 170 43%

No cred from PSI; no longer attending PSI; did not attend another PS institution

12 2% 24 6%

Total 528 100% 400 100%

Appendix N

Math Course Pipeline

Page 120: Three Essays on the Economics of Education Miguel E. Martinez

112

No Math

Non-Academic

Low Academic

Middle Academic I

Middle Academic II

Advanced Academic I

Advanced Academic II/Pre-Calc

Advanced Academic III / Calculus

Science Course Pipeline

No Science

Primary Physical Science

Secondary Physical Science and Basic Bio

General Biology

Chemistry 1 or Physics 1

Chemistry 1 and Physics 1

Chemistry 2 or Physics 2

Chemistry and Physics and Level 7

Appendix O. Pre-matching Differences between Treatment and Control

Variables Moderately

Selective Highly

Selective Difference

Page 121: Three Essays on the Economics of Education Miguel E. Martinez

113

SAT Math 580.06 629.70 -49.633***

SAT Verbal 554.42 589.73 -35.317***

Math Pipeline 2 0.00 0.00 -0.003***

Math Pipeline 3 0.00 0.00 0.00

Math Pipeline 4 0.01 0.01 0.002**

Math Pipeline 5 0.08 0.03 0.052***

Math Pipeline 6 0.10 0.10 -0.005**

Math Pipeline 7 0.19 0.24 -0.043***

Math Pipeline 8 0.61 0.63 -0.018***

Science Pipeline 2 0.01 0.00 0.008***

Science Pipeline 3 0.00 0.01 -0.006***

Science Pipeline 4 0.06 0.02 0.042***

Science Pipeline 5 0.16 0.13 0.030***

Science Pipeline 6 0.24 0.26 -0.019***

Science Pipeline 7 0.15 0.13 0.024***

Science Pipeline 8 0.37 0.45 -0.079***

Black 0.02 0.04 -0.019***

URM 0.13 0.06 0.071***

White 0.75 0.83 -0.077***

Female 0.28 0.30 -0.024***

Middle Income 0.52 0.37 0.151***

High Income 0.31 0.52 -0.205***

Moderately Selective Admissions 1.53 1.36 0.175***

High Selective Admissions 1.58 2.17 -0.597***

Total Applications 4.16 3.93 0.232***

Catholic 0.07 0.10 -0.026***

Private 0.04 0.09 -0.045***

Suburban 0.47 0.47 -0.01

Rural 0.27 0.18 0.089***

Observations

14,145

69,577

* Significant at 90% Confidence

** Significant at 95% Confidence

*** Significant at 99% Confidence

Appendix P: Unmatched Sample Logistic Regression Results

Variables Odds Ratio

Std. Error

z P>z Lower Bound

Upper Bound

Page 122: Three Essays on the Economics of Education Miguel E. Martinez

114

Highly Selective 2.50 0.11 20.66 0.00 2.29 2.73

SAT Math 1.00 0.00 12.21 0.00 1.00 1.01

SAT Verbal 1.00 0.00 -5.83 0.00 1.00 1.00

Math Pipeline 2 1.00 (omitted) Math Pipeline 3 1.00 (omitted) Math Pipeline 4 1.00 (omitted) Math Pipeline 5 1.00 (omitted) Math Pipeline 6 0.87 0.08 -1.54 0.12 0.73 1.04

Math Pipeline 7 0.50 0.03 -13.37 0.00 0.45 0.55

Math Pipeline 8 1.00 (omitted) Science Pipeline 2 1.00 (omitted) Science Pipeline 3 1.00 (omitted) Science Pipeline 4 594.15 103.36 36.72 0.00 422.50 835.55

Science Pipeline 5 7.15 0.53 26.30 0.00 6.18 8.28

Science Pipeline 6 0.94 0.04 -1.40 0.16 0.85 1.03

Science Pipeline 7 2.67 0.18 14.81 0.00 2.35 3.04

Science Pipeline 8 1.00 (omitted) Black 2.92 0.35 8.96 0.00 2.31 3.69

URM 12.48 1.21 26.00 0.00 10.31 15.09

White 16.76 1.44 32.70 0.00 14.15 19.84

Female 0.04 0.00 -53.84 0.00 0.04 0.05

Middle Income 0.49 0.03 -13.00 0.00 0.44 0.54

High Income 0.16 0.01 -30.11 0.00 0.14 0.18

Moderately Selective Admissions

1.19 0.04 4.95 0.00 1.11 1.27

High Selective Admissions 1.77 0.05 19.40 0.00 1.67 1.87

Total Applications 0.81 0.02 -10.22 0.00 0.78 0.84

Catholic 0.27 0.02 -22.21 0.00 0.24 0.30

Private 0.09 0.01 -29.90 0.00 0.08 0.10

Suburban 0.80 0.04 -4.91 0.00 0.73 0.87

Rural 0.22 0.02 -20.14 0.00 0.19 0.25

Intercept 0.04 0.01 -15.28 0.00 0.03 0.06

Observations 20,166 Pseudo R^2 0.4848

Page 123: Three Essays on the Economics of Education Miguel E. Martinez

115

Appendix Q. Post-matching Differences between Treatment and Control

Variables Highly

Selective Moderately

Selective Difference

SAT Math 579.63 599.40 -19.77

SAT Verbal 555.76 554.72 1.04

Math Pipeline 2 0% 0% 0%

Math Pipeline 3 0% 0% 0%

Math Pipeline 4 0% 0% 0%

Math Pipeline 5 12% 2% 11%

Math Pipeline 6 0% 0% 0%

Math Pipeline 7 23% 26% -3%

Math Pipeline 8 65% 73% -7%

Science Pipeline 2 0% 0% 0%

Science Pipeline 3 0% 0% 0%

Science Pipeline 4 7% 1% 6%

Science Pipeline 5 7% 7% -1%

Science Pipeline 6 34% 37% -4%

Science Pipeline 7 5% 11% -6%

Science Pipeline 8 48% 44% 4%

Black 5% 4% 0%

URM 15% 11% 4%

White 74% 82% -8%

Female 22% 22% 0%

Middle Income 42% 43% 0%

High Income 37% 40% -3% Moderately Selective Admissions

159% 140% 19%

High Selective Admissions

160% 147% 13%

Total Applications 457% 381% 76%

Catholic 11% 13% -2%

Private 4% 3% 0%

Suburban 41% 60% -19%

Rural 17% 13% 4%

Observations 8,447 8,447

Page 124: Three Essays on the Economics of Education Miguel E. Martinez

116

Appendix R. Area of Common Support

0 .2 .4 .6 .8 1Propensity Score

Untreated Treated: On support

Treated: Off support