Has the Predictive Validity of High School GPA and ACT Scores on Postsecondary Enrollment Changed Over Time?

Research Report

2024-01

Has the Predictive Validity of High

School GPA and ACT Scores on

Postsecondary Enrollment Changed

Over Time?

XIUHAN CHEN, M.ED., AND EDGAR I. SANCHEZ, PH.D.

ACT Research | Research Report | January 2024 2

Conclusions

Accounting for student characteristics and school effects, using both high school GPA (HSGPA)

and ACT Composite score was more predictive of postsecondary enrollment than using either

predictor alone. When used together, the predictive power of HSGPA increased from 2010 to

2021, while the predictive power of the ACT Composite score decreased during this period.

HSGPA had greater predictive power than did the ACT Composite score after 2014. We believe

the adoption of test-optional admissions policies before and especially after the COVID-19

pandemic helps explain the increase in predictive power for HSGPA.

So What?

Given the extensive evidence that HSGPA inflation has increased dramatically in recent years,

the predictive power of HSGPA alone is threatened. This can result in underprepared students

entering postsecondary education and experiencing less favorable outcomes such as higher

drop-out rates and higher student loan debt. Combining ACT scores with HSGPA helps provide

a more comprehensive picture of which students are prepared to enroll in postsecondary

education. This can help postsecondary institutions more accurately predict which students may

continue their education after high school.

ACT Research | Research Report | January 2024 3

Now What?

Given the potential threat to the predictive power of HSGPA posed by grade inflation, using ACT

scores can help offset the uncertainty of the interpretation of HSGPA and help differentiate

between students with the same HSGPA (e.g., students who all have a 4.0 HSGPA). HSGPA

and ACT scores measure different aspects of student achievement; therefore, they provide

complementary information. We recommend using both HSGPA and ACT scores to predict

postsecondary enrollment because using both yields the strongest predictive power.

About the Authors

Xiuhan Chen, M.Ed.

Xiuhan Chen is a Ph.D. student at the

University of Missouri studying Statistics,

Measurement, and Evaluation in Education.

Her research focuses on the combination of

causality and psychometrics. Throughout her

academic journey, Ms. Chen has aimed to

deepen the understanding of the causal

relationship between human behaviors and

personality traits.

Edgar I. Sanchez, Ph.D.

Edgar I. Sanchez is a lead research scientist

at ACT, where he studies postsecondary

admissions, national testing programs, test

preparation efficacy, and intervention

effectiveness. Throughout his career, Dr.

Sanchez has focused his research on the

transition between high school and college

and supporting the decision-making capacity

of college administrators, students, and their

families. His research has been widely cited

in academic literature and by the media,

including The Wall Street Journal, The

Washington Post, USA Today, and the

education trade press.

Acknowledgements

The authors wish to thank Joann Moore, Jill McVey, Jeff

Allen, and Jeffery Steedle for their comments on earlier

drafts of this report.

ACT Research | Research Report | January 2024 4

Introduction

Grades are among the most important indicators of students’ academic success, skill, and

knowledge. Not only do grades influence academic awards, academic intervention, and

advanced course placement (Feldman, 2018), they also affect athletic or extracurricular

eligibility, employment, driving permits, car insurance rates, college admission, scholarships,

and financial aid (Griffin & Townsley, 2021). In American high schools, high school grade point

average (HSGPA) is one quantitative measure of grades (ACT, 2005). HSGPA is often

compared with standardized test scores (such as ACT

test scores). Both ACT scores and

HSGPA are common measures of high school students’ academic achievement. These

measures are typically used to predict students’ academic performance or future success, such

as first-semester college GPA, first-year college GPA, and college completion. However,

HSGPA may not always be an accurate measure of a student’s true ability (ACT, 2005). The

presence of grade inflation raises questions about the extent to which we should rely on grades

to measure of academic achievement or predict future success.

Grade inflation refers to an increase in grades without a concurrent increase in students’ true

ability or other objective measures of academic performance (Bejar & Blew, 1981; Camara et

al., 2004; Gershenson, 2018; Godfrey, 2011). In other words, students’ grades may not reflect

their true level of content mastery (Chowdhury, 2018). The evidence of grade inflation has been

well documented, with research consistently demonstrating that HSGPA has steadily increased

over the past several decades, while standardized assessment scores have remained

unchanged or declined (ACT, 2005; Bejar & Blew, 1981; Camara et al., 2004; Gershenson,

2020; Godfrey, 2011; Ziomek & Svec, 1995). For example, Ziomek and Svec (1995) chose a

sample of 530,000 students from 5,136 public schools between the 1989–1990 and the 1993–

1994 school years to demonstrate that there was a steady increase in HSGPA while ACT

scores remained constant. A more recent study found that from 2009 to 2019 alone, HSGPA, as

reported on a 0.0 to 4.0 scale, increased from 3.00 to 3.11, while National Assessment of

Educational Progress (NAEP) scores remained constant (U.S. Department of Education,

Institute of Education Sciences,

National Center for Education Statistics [NCES], National

Assessment of Educational Progress [NAEP], 2022a). A second recent study found that from

2010 to 2021, HSGPA among ACT-tested students increased from 3.17 to 3.36 (Sanchez and

Moore, 2022).

Detecting grade inflation is challenging, as it requires both grades across time and a stable

measure against which to compare them (ACT, 2005; Bejar & Blew, 1981). Finefter-Rosenbluh

and Levinson (2015) summarized three approaches to assessing grade inflation. First, grade

inflation can be understood by comparing high school grades to standardized test scores using

longitudinal data. In the past, HSGPA has been compared to scores from several standardized

tests, such as ACT scores (ACT, 2005; Bellott, 1981; Woodruff & Ziomek, 2004; Zhang &

Sanchez, 2013), SAT scores (Bejar & Blew, 1981; Godfrey, 2011), NAEP data (U.S.

Department of Education, NCES, NAEP, 2020a), and end-of-course exams (Gershenson,

2018). Second, grade inflation can be understood as grade compression. For example, Hurwitz

& Lee (2018) used nationally representative data to estimate grade inflation among all high

ACT Research | Research Report | January 2024 5

school students who took the SAT. They found that the percentage of students with A averages

increased from 38.9% to 47.0% between 1998 and 2016. Meanwhile, there was a 4.2

percentage point decrease in the percentage of students with B averages, followed by a 3.8

percentage point decrease in the percentage of students with C averages. Similarly, Sanchez

and Moore (2022) found a decrease in the percentage of students reporting a B average

HSGPA from 2010 to 2021 (46.8% and 36.2%, respectively) and an increase in the percentage

of students reporting an A average HSGPA from 2010 to 2021 (40.3% and 54.9%,

respectively).

Lastly, grade inflation can be detected by comparing grades among high schools.

Nord (2011) recorded that grade inflation was concentrated in wealthier and white-majority high

schools. Similarly, Hurwitz and Lee (2018) observed greater grade inflation among white, Asian,

wealthy, and private school students than among students from low-income families and

students in public schools.

There are several possible reasons grade inflation occurs in high schools. Some researchers

argue that one reason is that high school students are taking advanced coursework that may

include “bonus” points that boost their HSGPA (Camara et al., 2004; Hurwitz & Lee, 2018).

Subjectivity in grade assignment among teachers also plays a role within and between schools.

For example, some teachers assign grades more strictly, while others are more lenient.

Gershenson (2018) also argued that teachers may be motivated to provide positive evaluations

of student performance not only to satisfy students and parents but also to enhance the

reputation of their schools or classrooms.

In recent years, grade inflation has become more pronounced, especially after the COVID-19

pandemic. Because of the COVID-19 outbreak, most American schools transitioned to remote

and distance learning in spring 2020 (U.S. Department of Education, NCES, NAEP, 2022b). In

response, some school districts moved away from the traditional A–F grading policy to more

flexible policies implemented by districts, schools, or classroom teachers (Arundel, 2020; Cano,

2020; Sawchuk, 2020). Grading policies included pass/fail grades, credit/no credit grades, and

“do no harm” grading policies (Castro et al., 2020). These changes in mode of instruction and

assignment of grades raise concerns about how grades were assigned during this time, and the

reliability and accuracy of grades from this period deserve careful consideration. Moreover,

grades given during the pandemic may be more a function of district or school policies than

accurate reflections of differences in academic achievement across students (Sanchez &

Moore, 2022).

With access to additional electronic devices while learning at home during the pandemic, some

students may have used multiple devices to check or change answers while completing

assignments or exams (Schramm et al., 2021). Consequently, these students may have

received higher grades on exams than they would have otherwise. According to Gonzalez et al.

(2020), higher grades during remote schooling may be correlated with cheating on online

exams. Moreover, students from low-income families may lack basic technology access, such

as access to high-speed internet and an adequate number of digital devices, thereby

perpetuating social inequities (Herold, 2020). This may have resulted in students from higher-

income families receiving artificially inflated grades because of their greater access to

technology that allows them to check answers.

ACT Research | Research Report | January 2024 6

These circumstances have increased the uncertainty in evaluating students’ achievement during

the COVID-19 pandemic. An early case study has revealed some evidence that high school

grades improved slightly during COVID-19 (Schramm et al., 2021), and additional research has

found that grade inflation became more apparent in 2020 and 2021 (Sanchez & Moore, 2022).

Grade inflation is potentially problematic for students, postsecondary institutions and employers,

and society at l

arge (Finefter-Rosenbluh & Levinson, 2015; Silva et al., 2023). For students,

inflated grades can create a false sense of proficiency, leading students to believe that they

have mastered core content skills (Chowdhury, 2018; Finefter-Rosenbluh & Levinson, 2015). As

a result, students may study less, and parents may not understand that their children may need

help to catch up (Gershenson, 2018). Grade inflation may prevent students from reaching their

full potential (Gershenson, 2018) and subsequently may reduce educational opportunities. For

example, some students may take less rigorous courses and subjects in which they can receive

higher grades in order to boost HSGPA (Chowdhury, 2018).

Additionally, grade inflation increases the rate at which underprepared high school students

enter college, increasing the risk that these students will later drop out of college (Gershenson,

2018). Inflated grades may also influence postsecondary institutions and employment decisions

(Finefter-Rosenbluh & Levinson, 2015). Grade inflation may provide misleading information to

colleges and employers who use grades as indices of ability and content mastery (Gershenson,

2018).

Grade inflation diminishes the ability of postsecondary institutions to distinguish among

prospective students (Finefter-Rosenbluh & Levinson, 2015; Godfrey, 2011; Silva et al, 2023).

Similarly, employers who use inflated grades as a criterion in making hiring decisions may not

be able to make meaningful distinctions among job applicants; likewise, they may not be able to

tell whether an applicant has truly acquired the skills or knowledge required for the job (Finefter-

Rosenbluh & Levinson, 2015).

Finally, grade inflation poses societal problems because it may increase social disparities and

inequities (Chowdhury, 2018; Finefter-Rosenbluh & Levinson, 2015). Students from families

with high socioeconomic status and those attending private or affluent high schools are more

likely to receive inflated grades (Gershenson, 2018; Nata et al., 2014; Neves et al., 2017). More

recent research has found that students at schools with higher percentages of traditionally

underserved students experienced higher rates of grade inflation and students at schools with

higher percentages of students eligible for free or reduced-price lunch experienced higher grade

inflation (Sanchez & Moore, 2022). These two studies examined affluence slightly differently

and found somewhat divergent results. That said, when privileged students receive inflated

grades, they may also receive additional unearned advantages in college and graduate

admissions, further entrenching their elite status. Although these students are not directly

responsible for this situation, this unearned advantage reinforces inequalities (Finefter-

Rosenbluh & Levinson, 2015).

Grade inflation also weakens the predictive validity of HSGPA (Zhang & Sanchez, 2013). The

ceiling effect of grades (when students score at or near the maximum for the HSGPA

distribution) makes it difficult to compare the academic performances of different students,

ACT Research | Research Report | January 2024 7

thereby reducing the practical range of HSGPA. As more students approach the “ceiling,” less

variability in grades is observed, which leads to less predictive power (Chan et al., 2007;

Hurwitz & Lee, 2018).

Much research has focused on the predictive power of HSGPA and standardized admissions

test scores (e.g., ACT and SAT). Although the findings consistently demonstrate that HSGPA

and test scores are highly correlated (Zhang & Sanchez, 2013), conclusions diverge when it

comes to the predictive power of HSGPA and test scores. Some research indicates that HSGPA

is a better predictor of college completion, college cumulative GPA, and first-year grades than

college admissions tests (Allensworth & Clark, 2020; Bowen et al., 2009; Galla et al., 2019;

Geiser & Santelices, 2007; Zwick, 2006). For example, a study that employed a sample of

80,000 first-year students in the University of California system found that HSGPA is

consistently the strongest predictor of four-year college outcomes, such as cumulative college

grades and graduation (Geiser & Santelices, 2007). A 1999 study of nearly 150,000 first-year

students illustrated that HSGPA was more predictive of college completion than admissions

tests (Bowen et al., 2009). More recently, Galla et al. (2019) replicated the 1999 predictive

validity study findings using a national sample of 47,303 students who applied to college in the

2009–2010 academic year. This was later confirmed by Allensworth and Clark (2020).

Other studies have found that test scores are more predictive than HSGPA. For example,

Ramist et al. (1994) examined data from 45 colleges to demonstrate that SAT scores were more

predictive of individual college course grades than HSGPA. More recently, Gershenson (2018)

used a sample of nearly all high school students in North Carolina from 2005 to 2016 to show

that Algebra 1 end-of-course exam scores predict ACT math scores much better than do high

school course grades.

HSGPA and ACT scores have different predictive utility depending on the postsecondary

outcome examined. For example, HSGPA was slightly more accurate than ACT scores in

predicting whether students earn a 2.00 or higher first-year college GPA (FYGPA; ACT, 2022;

Noble & Sawyer, 2002). Yet the ACT Composite score and HSGPA were equally accurate in

predicting whether students earn a 3.00 or higher FYGPA (ACT, 2022). Likewise, HSGPA is

better than the ACT Composite score in predicting minimal success (retention through the first

year and attaining a 2.0 or higher FYGPA in college), while the ACT score is better than HSGPA

at predicting high-level success (retention through first year and attaining a 3.5 or higher

FYGPA) and very high-level success (retention through first year and attaining a 3.7 or higher

FYGPA). In addition, ACT Composite scores had incremental validity beyond HSGPA alone

(Sawyer, 2010).

However, it is undeniable that together, HSGPA and test scores are more predictive of future

success in college than either predictor alone (ACT, 1997; ACT, 2022; Bridgeman et al., 2008;

Camara & Echternacht, 2000; Kobrin et al., 2008; Mattern & Patterson, 2011; Noble & Sawyer,

2002; Sawyer, 2010; Westrick et al., 2015; Westrick et al., 2019; Willingham et al., 1990; Zwick,

2006). Moreover, HSGPA and test scores demonstrate strong contributions in predicting post-

secondary success across all race/ethnicity and gender groups (Bridgeman et al., 2008).

ACT Research | Research Report | January 2024 8

Although the predictive validity of HSGPA and test scores on postsecondary outcomes after

enrollment has been extensively established, it is noteworthy that relatively little research has

been done on predicting college enrollment using HSGPA and test scores (ACT, 2010; ACT,

2022). Much research has indicated that college enrollment is positively correlated with high

school grades (Okpych & Courtney, 2017; Zhang & Sanchez, 2013). However, with the increase

in grade inflation, the longitudinal predictive power of HSGPA alone, test scores alone, and the

combination of HSGPA and test scores to predict college enrollment deserve to be explored

further. This study is an extension of Sanchez and Moore’s (2022) study of the predictive validity

of HSGPA and ACT scores. This study employs multilevel logistic regression to examine the

predictive power of HSGPA and ACT test scores to predict postsecondary enrollment between

2010 and 2021, including students who tested during the pandemic.

The study addresses the following research questions:

Has the predictive power of HSGPA on postseco

ndary enrollment changed over

time?

Has the predictive power of the ACT Composite scor

e on postsecondary enrollment

changed over time?

Has the predictive power of HSGPA and ACT Co

mposite scores combined on

postsecondary enrollment changed over time?

Does a model which incorporates only HSGPA

, only ACT Composite scores, or both

best predict postsecondary enrollment?

Methods

Analytical Sample

The present research incorporates student data from the 2010 to 2021 ACT-tested high school

graduating cohorts. Between 2010 and 2020, we included every other year due to

computational limitations. The 2021 cohort was included to explore the change in predictive

validity post-pandemic. This data includes the most recent ACT scores for students who took

the ACT more than once. The sample was constrained to public high schools that were matched

to 2- and 4-year college enrollment data from the National Student Clearinghouse.

Students in

the sample either took the ACT on a national test day or through a state or district contract. Only

schools with at least 30 tested students each year were included to ensure the stability of

multilevel modeling estimates. As such, the individual year statistics may not match those of the

official graduating class. The study sample consisted of 16,323,815 students.

Measures

Student demographics. As part of the ACT registration process, students were asked to report

certain demographic information. In this study, we used the following student characteristics:

race/ethnicity, gender, and family income.

Course grade information. Students were also asked to provide information related to courses

they took in high school and their grades for each course. Self-reported grades in up to 23

ACT Research | Research Report | January 2024 9

courses in English, mathematics, social studies, and natural science were averaged to calculate

students’ cumulative HSGPA. Evidence has shown that students’ self-reported HSGPA is highly

correlated with students’ transcript GPA (Sanchez & Buddin, 2015). Additionally, self-reported

data has been shown to be a good substitute for transcript-reported grades for research

purposes (Camara et al., 2004; Kuncel et al., 2005; Shaw & Mattern, 2009).

Enrollment information. Enrollment data for the fall after high school graduation was retrieved

from the National Student Clearinghouse (NSC). ACT-tested students in the sample were

matched to NSC records.

Data Analysis

We started by describing students’ characteristics and the number of schools. We then used

hierarchical logistic regression with the adaptive Gauss-Hermite quadrature approximation

method to estimate postsecondary enrollment for each year (see Appendix).

In the multilevel

model, Level 1 is students and Level 2 is schools. Students were nested in schools. Enrollment

was regressed on HSGPA, ACT Composite score, test type (i.e., national vs. state and district

testing), race/ethnicity, gender, and family income.

The Level 2 intercepts were allowed to vary across schools. Nakagawa and Shielzeth (2013)

proposed a simple method of obtaining pseudo 



in multilevel logistic regression, which

computes the variance of the predicted outcomes using their logit form. We calculated the

marginal and conditional 



to signify goodness-of-fit of the models. A marginal 



is based on

the fixed effects only, while a conditional 



is based on both the fixed and random effects

(Huang, 2022; Nakagawa & Shielzeth, 2013).

The conditional 



value represents the total

model variance explained at both Level 1 and Level 2. The marginal 



represents the

percentage of student-level variance explained by our models. HSGPA and ACT Composite

score were standardized across all years. Missing data on family income was imputed 5 times

using multiple imputation. Multiple imputation was conducted using the multivariate normal

distribution in SAS Proc Impute. The multiple imputation model included parental income,

race/ethnicity, gender, HSGPA, and ACT Composite score. One model was run per imputation

in R, and then the results were pooled to give final estimates. The data analysis for this study

was performed using R (R Core Team, 2022).

Results

Descriptive Statistics

Table 1 summarizes the number of schools and student characteristics in the analyses. From

2010 to 2021, the number of schools ranged from 19,283 to 22,700. Across years, there were

slightly more female students than male students. Among these students, the percentage of

Black students decreased from 13.5% to 10.7% from 2010 to 2021 and the number of White

students decreased from 63.2% to 61.1%. The proportion of Asian students and Hispanic

students increased from 2010 to 2020 but dropped slightly in 2021. From 2010 until 2021, the

proportion of students from low-income families (defined as students from a family with an

income less than $36,000 per year) decreased, while the proportion of students from high-

ACT Research | Research Report | January 2024 10

income families (those coming from a family with an income greater than $100,000 per year)

increased. College enrollment increased from 27.3% in 2010 to 32.7% in 2021. Students’ ACT

Composite scores increased by 0.6 scale score points between 2010 to 2021, while students’

HSGPA steadily increased from 2010 to 2021, with an especially pronounced increase in recent

years.

In the study sample, the percentages of students taking the ACT as part of the National testing

program (administered on Saturdays) decreased from 82.3% in 2010 to 62.4% in 2021, while

the percentage of students who participated in the State and District testing program (typically

during a school day) increased from 17.7% in 2010 to 37.6% in 2021. In 2016, about 30% of the

graduating class’s most recent record was from a State and District testing program. The State

and District testing program administers the ACT to nearly all students within a state or district.

The ACT National testing program is more likely to be used by college-bound students who plan

to use ACT scores as part of their college application package. The number of students who

use the National testing program has decreased; one possible contributor to this change (in

addition to the rise of State and District testing) is the growth of test-optional policies before and,

importantly, after the COVID-19 pandemic. On March 11, 2020, the World Health Organization

declared COVID-19 a global pandemic (WHO, 2020). In the United States, ACT canceled the

April, June, and July 2020 national test administrations and resumed limited testing in June

2020. As a result of the pandemic, many post-secondary institutions instituted test-optional

policies.

ACT Research | Research Report | January 2024 11

Table 1. Descriptive Statistics

Characteristic 2010 2012 2014 2016 2018 2020 2021

Number of students 1,353,918 1,424,173 1,522,907 1,644,802 1,367,430 1,014,662 748,592

Number of high schools 21,111 21,890 22,375 22,700 22,248 21,448 19,283

Testing program N (%)

National 82.3% 82.4% 77.3% 70.3% 73.1% 75.1% 62.4%

State and District 17.7% 17.6% 22.7% 29.7% 26.9% 24.9% 37.6%

Race/ethnicity N (%)

Asian 4.3% 4.2% 4.6% 4.8% 5.5% 5.8% 5.3%

Black 13.5% 13.3% 13.0% 12.8% 12.2% 11.6% 10.7%

Hispanic 9.9% 13.8% 15.1% 16.1% 16.2% 16.0% 12.8%

White 63.2% 59.7% 57.7% 55.7% 55.5% 56.4% 61.1%

Other 3.7% 4.4% 5.0% 5.3% 5.7% 5.9% 5.7%

Prefer not to respond 2.5% 3.6% 3.5% 3.6% 3.3% 3.2% 3.0%

Missing 2.9% 0.8% 1.1% 1.6% 1.7% 1.0% 1.4%

Student gender N (%)

Female 55.5% 55.3% 54.8% 53.9% 54.9% 56.4% 54.0%

Male 44.4% 44.6% 44.8% 44.9% 44.6% 43.2% 44.4%

Other/missing 0.1% 0.2% 0.4% 1.1% 0.4% 0.4% 1.6%

Enrolled in college N (%)

Yes 72.7% 71.6% 71.8% 67.8% 71.2% 68.8% 67.3%

No 27.3% 28.4% 28.2% 32.2% 28.8% 31.2% 32.7%

Family income N (%)

< $36K 27.5% 25.8% 25.1% 24.2% 22.1% 18.8% 15.9%

$36K–$60K 18.6% 17.2% 17.0% 16.9% 15.7% 13.8% 12.4%

$60K–$100K 19.7% 18.8% 18.7% 18.5% 18.1% 17.1% 16.7%

> $100K 14.2% 16.0% 18.1% 20.0% 23.9% 26.7% 30.0%

Missing 20.0% 22.2% 21.2% 20.4% 20.2% 23.6% 24.9%

ACT Composite mean (SD)

21.2

(5.20)

21.2

(5.23)

21.3

(5.35)

21.3

(5.48)

21.6

(5.65)

21.9

(5.83)

21.8

(5.97)

HSGPA mean (SD)

3.20

(0.64)

3.23

(0.63)

3.24

(0.63)

3.25

(0.63)

3.33

(0.61)

3.41

(0.58)

3.43

(0.59)

ACT Research | Research Report | January 2024 12

Q1: Has the Predictive Power of HSGPA on Postsecondary Enrollment

Changed Over Time?

Figure 1 shows the standardized coefficients of HSGPA on postsecondary enrollment from 2010

to 2021, controlling for test type, gender, race/ethnicity, and family income. From 2010 to 2016,

the predictive power of HSGPA changed little. In 2018, the predictive power of HSGPA

increased from 0.57 in 2016 to 0.63 in 2018. However, it decreased to previous levels in 2020

and increased again in 2021 to its peak. Figure 2 presents the odds ratios and 95% confidence

intervals of HSGPA on postsecondary enrollment from 2010 to 2021. It shows the same pattern,

in which the predictive power of HSGPA changed more dramatically starting in 2018 and 2021.

Figure 1.

Standardized Coefficients of HSGPA from 2010 to 2021

Note: For 2020, one of the five imputed data s

ets failed to meet one of two convergence criteria.

In that imputation, the optimizer convergence criterion was met but the gradient convergence

criterion was not. We calculated the pooled coefficient with the remaining four imputations, and

the coefficient was the same within three decimal places. Since there was no practical

difference, we report the results from all five imputations.

ACT Research | Research Report | January 2024 13

Figure 2. Odds Ratios of HSGPA from 2010 to 2021

Note: For 2020, one of the four imputed data sets failed to meet one of two convergence

criteria. In that imputation, the optimizer convergence criterion was met but the gradient

convergence criterion was not. We calculated the pooled coefficient with the remaining four

imputations, and the coefficient was the same within three decimal places. Since there was no

practical difference, we report the results from all five imputations.

Q2: Has the Predictive Power of the ACT Composite score on

Postsecondary Enrollment Changed Over Time?

Figure 3 shows the standardized coefficients of the ACT Composite score in predicting

postsecondary enrollment from 2010 to 2021, controlling for test type, gender, race/ethnicity,

and family income. The standardized coefficient decreased from 0.68 to 0.49 between 2010 and

2021. It reached its lowest coefficient in 2020 during the pandemic. Figure 4 shows the odds

ratios of the ACT Composite score in predicting postsecondary enrollment from 2010 to 2021. In

2010, the odds ratio was 1.97, indicating that with one standard deviation score increase on the

ACT Composite score, the odds of enrollment increased by 97%.

However, the odds ratio

dropped to 1.64 in 2021. The decrease in standardized coefficients and odds ratios indicates

that the predictive power of the ACT Composite score on postsecondary enrollment has

changed over time.

ACT Research | Research Report | January 2024 14

Figure 3. Standardized Coefficients of the ACT Composite Score from 2010 to 2021

ACT Research | Research Report | January 2024 15

Figure 4. Odds Ratios of ACT Composite Score from 2010 to 2021

Q3: Has the Predictive Power of Using Both HSGPA and ACT

Composite Scores on Postsecondary Enrollment Changed Over

Time?

Figure 5 and Figure 6 show the standardized coefficients and odds ratios of HSGPA and ACT

Composite scores in predicting postsecondary enrollment from 2010 to 2021, while controlling

for test type, gender, race/ethnicity, and family income. From 2010 to 2021, the HSGPA

coefficient increased from 0.40 to 0.51, and the odds ratios of HSGPA increased from 1.49 to

1.67. During the same period, the coefficients of ACT Composite scores decreased from 0.43 to

0.24, and the odds ratios of ACT Composite score coefficients decreased from 1.53 to 1.27.

Before 2014, the ACT Composite score was slightly more predictive of postsecondary

enrollment than HSGPA. After 2014, HSGPA was more predictive of postsecondary enrollment

than the ACT Composite score. In addition, the difference in the predictive power of HSGPA

and the ACT Composite score has widened since 2014. The gap between the predictive power

of the ACT Composite score and HSGPA became more pronounced after 2018. This may be

because of the growth of test-optional policies. As the use of the ACT has decreased in

importance during recent admissions cycles, it is logical that its predictive power would also

decrease. Similarly, as HSGPA has grown in importance in recent admissions cycles, its

predictive power should also increase.

ACT Research | Research Report | January 2024 16

Figure 5. Standardized Coefficients of HSGPA and ACT Composite Score from 2010 to 2021

ACT Research | Research Report | January 2024 17

Figure 6. Odds Ratios of HSGPA and ACT Composite score from 2010 to 2021

Note: The odds ratio for 2014 and 2016 for HSGPA are the same when rounded to two decimal

places.

In 2020, both the predictive power of HSGPA and the ACT Composite score in predicting

postsecondary enrollment declined; however, the predictive power of the ACT Composite score

declined more dramatically than did that of HSGPA. This decline may be related to the dramatic

changes required during the COVID-19 pandemic. Overall, HSGPA has more predictive power

than the ACT Composite score.

Q4: Does a Model Which Incorporates Only HSGPA, Only ACT

Composite, Or Both Best Predict Postsecondary Enrollment?

Figure 7 shows the conditional 



from hierarchical logistic regression models of HSGPA alone,

ACT Composite score alone, and HSGPA and ACT Composite score together predicting

postsecondary enrollment from 2010 to 2021 while controlling for test type, gender,

race/ethnicity, and family income. The conditional 



for the three models steadily increased

from 2010 to 2018, then dropped in 2020 and 2021. The model with both predictors always had

the largest 



value in all years examined. In other words, while HSGPA alone was more

predictive than ACT Composite score alone in predicting postsecondary enrollment, using both

measures of academic achievement together was more effective at predicting post-secondary

enrollment. This finding is consistent with previous research. The conditional 



for the ACT

Composite–alone model and HSGPA-alone model were similar between 2010 and 2016. In

years where they differed, the difference in conditional 



was very small, no more than 0.02.

ACT Research | Research Report | January 2024 18

Figure 7. Conditional 



of HSGPA and ACT Composite Score from 2010 to 2021

Figure 8 shows the marginal 



from hierarchical logistic regression models of HSGPA alone,

ACT Composite score alone, and HSGPA and ACT Composite score together predicting

postsecondary enrollment from 2010 to 2021 while controlling for test type, gender,

race/ethnicity, and family income. As found in the conditional 



models, ACT Composite score

combined with HSGPA was more predictive of postsecondary enrollment across the years

examined. Here again, while the marginal 



for each model differed slightly, there may not be

a practical difference in predictive validity.

ACT Research | Research Report | January 2024 19

Figure 8. Marginal 



of HSGPA and ACT Composite Score from 2010 to 2021

Discussion

This study provided evidence of changes in the predictive validity of HSGPA and ACT

Composite score before and during the first year of the pandemic. This study found little

practical difference in the predictive validity of the three models examined (HSGPA, ACT

Composite, and both HSGPA and ACT Composite). It is possible that we found these results

because we included student level covariates that had similar explanatory power across

models. In this study, we used data from 2010 to 2021 to explore the predictive power of

HSGPA and ACT Composite score on postsecondary enrollment using hierarchical logistic

regression.

We first explored the change in the predictive power of HSGPA from 2010 to 2021. From the

results of the hierarchical logistic regression, we saw evidence of the predictive power of

HSGPA on postsecondary enrollment changing over time. From 2010 to 2021, the predictive

power of HSGPA has steadily increased, increasing to its highest value in 2021. We then

explored the change in the predictive power of the ACT Composite score from 2010 to 2021. In

contrast to the performance of the predictive power of HSGPA, the predictive power of the ACT

Composite score has decreased over time, reaching its lowest value in 2021.

In this analysis, we also compared the predictive power of HSGPA alone, ACT Composite score

alone, and HSGPA and ACT Composite score together on postsecondary enrollment from 2010

to 2021. We found evidence that HSGPA currently has more predictive power for postsecondary

enrollment than the ACT Composite score after accounting for students’ characteristics and

school effects. In the years examined, HSGPA was not always the stronger predictor when

compared to the ACT Composite score in predicting postsecondary enrollment. Before 2014,

ACT Research | Research Report | January 2024 20

the ACT Composite score had stronger predictive power than HSGPA. After 2014, HSGPA had

more predictive power than the ACT Composite score, and the gap in predictive power between

HSGPA and the ACT Composite score in predicting postsecondary enrollment has increased

since 2014.

There could be several potential reasons for this trend. First, test-optional policies introduced by

postsecondary institutions meant that students were no longer required to submit their ACT

scores as part of their application for postsecondary institutions. In a test-optional

postsecondary institution application, HSGPA became the primary, and at times only,

quantitative measure of student achievement. Based on data from the Integrated Postsecondary

Education Data System (U.S. Department of Education, National Center for Education

Statistics, Integrated Postsecondary Education Data System, 2023), the number of

postsecondary institutions requiring ACT or SAT test scores steadily decreased from 2010

(1,183) through 2019 (985). There were more dramatic decreases in the number of

postsecondary institutions requiring test scores in 2020 (562) and 2021 (163). Therefore,

HSGPA received more attention than test scores.

Second, COVID-19 sped up the adoption of test-optional policies. During COVID-19, most high

schools transitioned to remote or distance learning. Most postsecondary institutions instituted

test-optional policies because of students’ limited access to ACT and SAT testing. Moreover,

there may have been changes in the ways in which HSGPA and test scores were interpreted

and used for decision making in the admissions process. In addition, students’ perceptions of

the value of a postsecondary education may be changing over time. Both of the latter points

deserve further study.

As the role of test scores in admissions has changed in recent years, the importance and high-

stakes nature of the use of HSGPA for admissions highlight the need for valid and reliable

grades. As the importance of HSGPA increases in postsecondary enrollment, there is also

evidence that grade inflation has been increasing at a significant rate in recent years. This

raises concerns about the way HSGPA is understood and used by postsecondary institutions.

We must make sure HSGPA remains a valid indicator in the absence of test scores. This is of

particular concern if underprepared students are entering postsecondary education and

experiencing less favorable outcomes.

As prior studies have demonstrated, there is greater predictive validity of test scores for higher

levels of success in college (e.g., successive level of first-year college GPA), but both measures

of student achievement have uses depending upon their application. In this study, the three

models examined were found to have similar predictive power on enrollment in any

postsecondary institution. We must bear in mind, however, that these models predict enrollment

at both 2- and 4-year institutions. This criterion of continuing education after high school is

useful for guiding high school students but should be coupled with what we know about the

predictive validity of both measures for other postsecondary outcomes such as college GPA,

retention, and degree completion.

This study clearly demonstrated that using HSGPA and ACT Composite score combined is

more effective for predicting enrollment than using HSGPA alone, as may be the case in test-

ACT Research | Research Report | January 2024 21

optional admissions policies. HSGPA and ACT scores show different aspects of students’

achievement and therefore provide complementary information. Being able to use both HSGPA

and a standardized test score such as an ACT score provides a better understanding of student

achievement and content mastery. For those reasons, we recommend the use of both HSGPA

and ACT Composite score in predicting postsecondary enrollment.

Limitations

As noted, this study was concerned with predicting students’ enrollment in either a 2- or 4-year

institution. Most two-year institutions have open-admissions policies, which means that ACT

scores might not be required or used in the admissions process. This study found similar

predictive validity for HSGPA and ACT Composite score when used alone and together. If we

were to change the scope of the question to include only 4-year institutions, the results would

likely change. To test this assertion, we reran the three models to predict 4-year enrollment in

2021. In that analysis, we found that the use of both measures simultaneously accounted for a

greater percentage of variance in enrollment (about 25%) than did using either HSGPA or ACT

Composite alone (21% and 22%, respectively). This suggests that when we focus on enrollment

in 4-year institutions, where admissions decisions are typically more selective than in 2-year

institutions, using both measures of achievement is a better option, consistent with the overall

recommendations of this study.

ACT Research | Research Report | January 2024 22

References

ACT. (1997). Prediction research services tables.

ACT. (2005). Are high school grades inflated? Issues in College Readiness [Research report].

ACT. (2010). Mind the gaps: How college readiness narrows achievement gaps in college

success. https://www.act.org/content/dam/act/unsecured/documents/MindTheGaps.pdf

ACT. (2022). The ACT technical manual.

https://www.act.org/content/dam/act/unsecured/documents/ACT_Technical_Manual.pdf

Allensworth, E. M., & Clark, K. (2020). High school GPAs and ACT scores as predictors of

college completion: Examining assumptions about consistency across high schools.

Educational Researcher, 49(3), 198–211.

Arundel, K. (2020, December). How educators are tweaking grading approaches in response to

the pandemic. K–12 Dive. Retrieved from

https://www.k12dive.com/news/how-

educators-are-tweaking-grading-approaches-in-response-to-the-pandemic/591729/

Bejar, I. I., & Blew, E. O. (1981). Grade inflation and the validity of the Scholastic Aptitude

Test. American Educational Research Journal, 18(2), 143–156.

Bellott, F. K. (1981). Relationships of declining test scores and grade inflation.

Bowen, W. G., Chingos, M. M., & McPherson, M. (2009). Crossing the finish line: Completing

College at America’s Public Universities. Princeton University Press.

Bridgeman, B., Pollack, J., & Burton, N. (2008). Predicting grades in different types of college

courses. ETS Research Report Series, 2008(1), i–27.

Camara, W. J., & Echternacht, G. (2000). The SAT

I and high school grades: Utility in

predicting success in college (Research Note No. 10). College Entrance Examination

Board, Office of Research and Development.

Camara, W., Kimmel, E., Scheuneman, J., & Sawtell, E. A. (2004). Whose grades are inflated?

Research Report No. 2003-4. College Entrance Examination Board.

Cano, R. (2020). How coronavirus has changed grading policies. Cal Matters.

https://calmatters.org/education/2020/05/how-coronavirus-has-changed-grading-policies/

Castro, M., Choi, L., Knudson, J., & O'Day, J. (2020). Grading policy in time of COVID-19:

Considerations and implications for equity. Policy and Practice Brief. California

Collaborative on District Reform.

Chan, W., Hao, L., & Suen, W. (2007). A signaling theory of grade inflation. International

Economic Review, 48(3), 1065–1090.

Chowdhury, F. (2018). Grade inflation: Causes, consequences and cure. Journal of Education

and Learning, 7(6), 86–92.

ACT Research | Research Report | January 2024 23

Feldman, J. (2018). Grading for equity: What it is, why it matters, and how it can transform

schools and classrooms. Corwin Press.

Finefter-Rosenbluh, I. & Levinson, M. (2015). What is wrong with grade inflation (if anything)?

Philosophical Inquiry in Education, 23(1), 3–21.

Galla, B. M., Shulman, E. P., Plummer, B. D., Gardner, M., Hutt, S. J., Goyer, J. P., D’Mello, S.

K., Finn, A. S., Duckworth, A. L. (2019). Why high school grades are better predictors of

on-time college graduation than are admissions test scores: The roles of self-regulation

and cognitive ability. American Educational Research Journal, 56(6), 2077–2115.

Geiser, S., & Santelices, M. V. (2007). Validity of high-school grades in predicting student

success beyond the freshman year: High-school record vs. standardized tests as

indicators of four-year college outcomes. Research & Occasional Paper Series: CSHE.

6.07. Center for Studies in Higher Education.

Gershenson, S. (2018). Grade inflation in high schools (2005–2016). Thomas B. Fordham

Institute.

Gershenson, S. (2020). Great expectations: The impact of rigorous grading practices on student

achievement. Thomas B. Fordham Institute.

Godfrey, K. E. (2011). Investigating Grade Inflation and Non-Equivalence. Research Report

2011-2. College Board.

Gonzalez, T., de la Rubia, M. A., Hincz, K. P., Comas-Lopez, M., Subirats, L., Fort, S., & Sacha,

G. M. (2020). Influence of COVID-19 confinement on students’ performance in higher

education. PLOS ONE, 15(10), Article e0239490.

Griffin, R., & Townsley, M. (2021). Points, points, and more points: High school grade inflation

and deflation when homework and employability scores are incorporated. Journal of

School Administration Research and Development, 6(1), 1–11.

Herold, B. (2020). The disparities in remote learning under coronavirus (in charts). Education

Week, 10.

Huang, F. L. (2022). Practical multilevel modeling using R. Sage Publications.

Hurwitz, M., & Lee, J. (2018). Grade inflation and the role of standardized testing. Measuring

success: Testing, grades, and the future of college admissions, 64–93.

Kobrin, J. L., Patterson, B. F., Shaw, E. J., Mattern, K. D., & Barbuti, S. M. (2008). Validity of the

SAT

for predicting first-year college grade point average. Research Report No. 2008-

5. College Board.

Kuncel, N. R., Credé, M., & Thomas, L. L. (2005). The validity of self-reported grade point

averages, class ranks, and test scores: A meta-analysis and review of the

literature. Review of Educational Research, 75(1), 63–82.

Mattern, K. D., & Patterson, B. F. (2006). Validity of the SAT® for predicting fourth-year grades:

2006 SAT validity sample. Statistical Report 2011-7. College Board.

http://files.eric.ed.gov/fulltext/ED563098.pdf

ACT Research | Research Report | January 2024 24

Nakagawa, S., & Schielzeth, H. (2013). A general and simple method for obtaining R

from

generalized linear mixed‐effects models. Methods in Ecology and Evolution, 4(2), 133–

142.

Nata, G., Pereira, M. J. and Neves, T. (2014). Unfairness in access to higher education: a 11

year comparison of grade inflation by private and public secondary schools in Portugal.

Higher Education, 68(6), 851–874.

Neves, T., Ferraz, H. and Nata, G. (2017). Social inequality in access to higher education:

Grade inflation in private schools and the ineffectiveness of compensatory education.

International Studies in Sociology of Education, 26(2), 190–210.

Noble, J., & Sawyer, R. (2002). Predicting different levels of academic success in college using

high school GPA and ACT Composite score. ACT Research Report Series.

Nord, C. (2011). America’s high school graduates: Results from the 2009 NAEP high school

transcript study. DIANE Publishing.

Okpych, N. J., & Courtney, M. E. (2017). Who goes to college? Social capital and other

predictors of college enrollment for foster-care youth. Journal of the Society for Social

Work and Research, 8(4), 563–593.

R Core Team (2022). R: A language and environment for statistical computing. R Foundation for

Statistical Computing. https://www.r-project.org/

Ramist, L., Lewis, C., & McCamley‐Jenkins, L. (1994). Student group differences in predicting

college grades: Sex, language, and ethnic groups. ETS Research Report

Series, 1994(1), i–41.

Sanchez, E., & Buddin, R. (2015). How accurate are self-reported high school courses, course

grades, and grade point average? ACT Working Paper Series No. WP-2015–03.

Sanchez, E. I., & Moore, R. (2022). Grade inflation continues to grow in the past decade.

Research Report. ACT.

Sawchuk, Stephen. (2020). Grading students during the coronavirus crisis: What’s the right call?

Education Week.

https://www.edweek.org/teaching-learning/grading-students-during-

the-coronavirus-crisis-whats-the-right-call/2020/04

Sawyer, R. (2010). Usefulness of high school average and ACT scores in making college

admission decisions. ACT Research Report Series 2010-2. ACT.

Schramm, H., Rubin, I., & Schramm, N. (2021). Covid‐19 and high school grades: An early case

study. Significance (Oxford, England), 18(2), 6.

Shaw, E. J., & Mattern, K. D. (2009). Examining the accuracy of self-reported high school grade

point average. Research Report No. 2009-5. College Board.

Silva, P. L., DesJardins, S., Biscaia, R., Sá, C., & Teixeira, P. (2023). Public and private school

grade inflations patterns in secondary education. IZA Discussion Paper No. 16016.

U.S. Department of Education, Institute of Education Sciences, National Center for Education

Statistics (NCES), National Assessment of Educational Progress (NAEP). (2020). NAEP

ACT Research | Research Report | January 2024 25

long-term trend assessment results: Reading and mathematics. Retrieved from

https://www.nationsreportcard.gov/ltt/?age=9

U.S. Department of Education, Institute of Education Sciences, National Center for Education

Statistics (NCES), National Assessment of Educational Progress (NAEP). (2022a).

2019 NAEP high school transcript study (HSTS) results. Retrieved from

https://www.nationsreportcard.gov/hstsreport/#home

U.S. Department of Education, Institute of Education Sciences, National Center for Education

Statistics (NCES), National Assessment of Educational Progress (NAEP).

(2022b). U.S. education in the time of COVID. Retrieved from

https://nces.ed.gov/surveys/annualreports/topical-studies/covid/

U.S. Department of Education, National Center for Education Statistics, Integrated

Postsecondary Education Data System (IPEDS), 2010-2020. Admissions survey.

Retrieved from https://nces.ed.gov/ipeds/survey-components/6 on August 16, 2023.

Westrick, P. A., Le, H., Robbins, S. B., Radunzel, J. M., & Schmidt, F. L. (2015). College

performance and retention: A meta-analysis of the predictive validities of ACT

scores,

high school grades, and SES. Educational Assessment, 20(1), 23-45.

Westrick, P., Marini, J., Young, L., Ng, H., & Shmueli, D. (2019). Validity of the SAT® for

predicting first-year grades and retention to the second year (College Board Research

Report 2019-5). The College Board.

World Health Organization (WHO). (2020). Timeline: WHO’s COVID-19 response. Retrieved

from https://www.who.int/emergencies/diseases/novel-coronavirus-2019/interactive-

timeline

Willingham, W. W., Lewis, C., Morgan, R., & Ramist, L. (1990). Predicting college grades: An

analysis of institutional trends over two decades. Educational Testing Service.

Woodruff, D. J., & Ziomek, R. L. (2004). High school grade inflation from 1991 to 2003.

Research Report Series 2004-04. ACT.

Zhang, Q., & Sanchez, E. I. (2013). High school grade inflation from 2004 to 2011. ACT

Research Report Series, 2013(3). ACT.

Ziomek, R. L., & Svec, J. C. (1995). High school grades and achievement: Evidence of grade

inflation. NASSP Bulletin, 81(587), 105–113.

Zwick, R. (2006). Higher education admission testing. In R. Brennan (Ed.), Educational

measurement (4th ed., pp. 647–679). American Council on Education, Praeger.

ACT Research | Research Report | January 2024 26

Appendix

Table A1. HLM Model Coefficients: HSGPA Only

Characteristic

Coefficients

2010 2012 2014 2016 2018 2020 2021

Intercept

1.405*** 1.222*** 1.486*** 1.322*** 1.498*** 1.227*** 1.141***

HSGPA

0.574*** 0.574*** 0.589*** 0.566*** 0.630*** 0.571*** 0.641***

Test type State and District

−1.234*** −1.218*** −1.262*** −1.169*** −1.272*** −1.204*** −0.794***

Gender

Male

−0.116*** −0.149*** −0.209*** −0.200*** −0.275*** −0.284*** −0.261***

Other/missing

−0.289*** −0.478*** −1.094*** −0.676*** −0.352*** −0.409*** −0.366***

Race

Black

0.234*** 0.286*** 0.168*** 0.129*** 0.166*** 0.153*** 0.191***

Hispanic

−0.146*** −0.029* −0.120*** −0.081*** −0.080*** −0.090*** −0.128***

Missing

−0.312*** 0.047 −0.120*** −0.113*** −0.104*** −0.059* −0.111**

Other

−0.033* 0.037* −0.073*** −0.092*** −0.106*** −0.102*** −0.119***

Prefer not to

respond

−0.147*** −0.014 −0.139*** −0.147*** −0.221*** −0.185*** −0.195***

White

0.128*** 0.206*** 0.037** 0.028** 0.000 −0.022 −0.063***

Family income

Linear

0.395*** 0.361*** 0.371*** 0.307*** 0.370*** 0.292*** 0.331***

Quadratic

−0.101*** −0.081*** −0.073*** −0.091*** −0.073*** −0.074*** −0.065***

Cubic

−0.001 −0.006 0.003 0.001 −0.002 −0.003 −0.002

Note: * p < 0.05, ** p < 0.01, ***p < 0.001; the reference categories in the model were National testing program, female, and Asian.

ACT Research | Research Report | January 2024 27

Table A2. HLM Model Odds Ratios: HSGPA Only

Characteristic

Odds Ratios

2010 2012 2014 2016 2018 2020 2021

Intercept

4.077*** 3.392*** 4.421*** 3.750*** 4.474*** 3.410*** 3.130***

HSGPA

1.776*** 1.776*** 1.802*** 1.761*** 1.877*** 1.770*** 1.898***

Test type State and District

0.291*** 0.296*** 0.283*** 0.311*** 0.280*** 0.300*** 0.452***

Gender Male

0.890*** 0.862*** 0.811*** 0.819*** 0.760*** 0.753*** 0.771***

Other/missing

0.749*** 0.620*** 0.335*** 0.509*** 0.703*** 0.664*** 0.694***

Race Black

1.264*** 1.332*** 1.183*** 1.137*** 1.181*** 1.165*** 1.211***

Hispanic

0.864*** 0.971* 0.887*** 0.922*** 0.923*** 0.914*** 0.880***

Missing

0.732*** 1.049 0.887*** 0.894*** 0.902*** 0.942* 0.895**

Other

0.967* 1.038* 0.930*** 0.912*** 0.899*** 0.903*** 0.888***

Prefer not to respond

0.864*** 0.987 0.870*** 0.863*** 0.801*** 0.831*** 0.823***

White

1.136*** 1.228*** 1.038** 1.029** 1.000 0.978 0.939***

Family income Linear

1.484*** 1.435*** 1.449*** 1.360*** 1.448*** 1.339*** 1.392***

Quadratic

0.904*** 0.922*** 0.929*** 0.913*** 0.929*** 0.928*** 0.937***

Cubic

0.999 0.994 1.003 1.001 0.998 0.997 0.998

Note: * p < 0.05, ** p < 0.01, ***p < 0.001; the reference categories in the model were National testing program, female, and Asian.

ACT Research | Research Report | January 2024 28

Table A3. HLM Model Coefficients: ACT Composite Only

Characteristic

Coefficients

2010 2012 2014 2016 2018 2020 2021

Intercept

1.572*** 1.436*** 1.730*** 1.532*** 1.742*** 1.446*** 1.425***

ACT Composite

0.677*** 0.658*** 0.647*** 0.580*** 0.615*** 0.469*** 0.492***

Test type State and District

−1.369*** −1.353*** −1.386*** −1.279*** −1.369*** −1.298*** −0.846***

Gender

Male

−0.253*** −0.291*** −0.360*** −0.347*** −0.442*** −0.430*** −0.439***

Other/missing

−0.462*** −0.577*** −1.148*** −0.748*** −0.411*** −0.561*** −0.614***

Race

Black

0.267*** 0.291*** 0.153*** 0.128*** 0.186*** 0.186*** 0.215***

Hispanic

−0.118*** −0.047*** −0.158*** −0.093*** −0.071*** −0.058*** −0.101***

Missing

−0.230*** 0.022 −0.183*** −0.136*** −0.136*** −0.028 −0.036

Other

−0.151*** −0.080*** −0.201*** −0.174*** −0.172*** −0.121*** −0.145***

Prefer not to respond

−0.265*** −0.115*** −0.261*** −0.227*** −0.281*** −0.207*** −0.220***

White

0.030* 0.089*** −0.090*** −0.054*** −0.063*** −0.025* −0.073***

Family income

Linear

0.340*** 0.319*** 0.335*** 0.283*** 0.338*** 0.282*** 0.345***

Quadratic

−0.114*** −0.089*** −0.083*** −0.100*** −0.093*** −0.095*** −0.096***

Cubic

−0.009 −0.015** −0.006 −0.008 −0.013* −0.008 −0.012

Note: * p < 0.05, ** p < 0.01, ***p < 0.001; the reference categories in the model were National testing program, female, and Asian.

ACT Research | Research Report | January 2024 29

Table A4. HLM Model Odds Ratios: ACT Composite Only

Characteristic

Odds Ratios

2010 2012 2014 2016 2018 2020 2021

Intercept

4.815*** 4.203*** 5.639*** 4.628*** 5.707*** 4.244*** 4.159***

ACT Composite

1.968*** 1.930*** 1.910*** 1.786*** 1.849*** 1.599*** 1.635***

Test type State and District

0.254*** 0.259*** 0.250*** 0.278*** 0.254*** 0.273*** 0.429***

Gender

Male

0.777*** 0.747*** 0.698*** 0.706*** 0.642*** 0.651*** 0.645***

Other/missing

0.630*** 0.562*** 0.317*** 0.473*** 0.663*** 0.571*** 0.541***

Race

Black

1.306*** 1.338*** 1.165*** 1.136*** 1.204*** 1.205*** 1.240***

Hispanic

0.889*** 0.955*** 0.854*** 0.911*** 0.931*** 0.944*** 0.904***

Missing

0.794*** 1.022 0.832*** 0.873*** 0.872*** 0.972 0.964

Other

0.860*** 0.923*** 0.818*** 0.840*** 0.842*** 0.886*** 0.865***

Prefer not to respond

0.767*** 0.891*** 0.770*** 0.797*** 0.755*** 0.813*** 0.803***

White

1.030 1.094*** 0.914*** 0.947*** 0.939*** 0.975* 0.930***

Family income

Linear

1.404*** 1.376*** 1.398*** 1.327*** 1.403*** 1.325*** 1.413***

Quadratic

0.892*** 0.915*** 0.920*** 0.905*** 0.911*** 0.910*** 0.908***

Cubic

0.991 0.985** 0.994 0.992 0.987* 0.992 0.988

Note: * p < 0.05, ** p < 0.01, ***p < 0.001; the reference categories in the model were National testing program, female, and Asian.

ACT Research | Research Report | January 2024 30

Table A5. HLM Model Coefficients: ACT Composite and HSGPA

Characteristic

Coefficients

2010 2012 2014 2016 2018 2020 2021

Intercept

1.434*** 1.263*** 1.533*** 1.339*** 1.506*** 1.214*** 1.130***

HSGPA

0.395*** 0.393*** 0.415*** 0.410*** 0.462*** 0.444*** 0.513***

ACT Composite

0.427*** 0.416*** 0.396*** 0.336*** 0.358*** 0.245*** 0.238***

Test type School and District

−1.207*** −1.193*** −1.227*** −1.141*** −1.232*** −1.176*** −0.765***

Gender

Male

−0.160*** −0.198*** −0.258*** −0.240*** −0.320*** −0.318*** −0.295***

Other/missing

−0.341*** −0.516*** −1.127*** −0.719*** −0.369*** −0.459*** −0.452***

Race

Black

0.388*** 0.431*** 0.298*** 0.253*** 0.318*** 0.284*** 0.324***

Hispanic

−0.045** 0.054*** −0.048* −0.001 0.019 0.000 −0.036*

Missing

−0.199*** 0.129*** −0.061*** −0.037 −0.025 0.030 0.009

Other

−0.024 0.050** −0.067*** −0.064*** −0.062*** −0.047** −0.062***

Prefer not to respond

−0.168*** −0.012 −0.148*** −0.133*** −0.189*** −0.149*** −0.160***

White

0.114*** 0.191*** 0.015 0.031** 0.02 0.019 −0.017

Family income

Linear

0.303*** 0.281*** 0.291*** 0.235*** 0.287*** 0.231*** 0.272***

Quadratic

−0.104*** −0.082*** −0.076*** −0.092*** −0.080*** −0.081*** −0.072***

Cubic

−0.007 −0.012* −0.003 −0.004 −0.008 −0.007 −0.008

Note: * p < 0.05, ** p < 0.01, ***p < 0.001; the reference categories in the model were National testing program, female, and Asian.

ACT Research | Research Report | January 2024 31

Table A6. HLM Model Odds Ratios: ACT Composite and HSGPA

Characteristic

Odds Ratios

2010 2012 2014 2016 2018 2020 2021

Intercept

4.195*** 3.536*** 4.633*** 3.815*** 4.510*** 3.367*** 3.097***

HSGPA

1.485*** 1.482*** 1.515*** 1.507*** 1.588*** 1.559*** 1.670***

ACT Composite

1.532*** 1.516*** 1.486*** 1.400*** 1.430*** 1.278*** 1.269***

Test type State and District

0.299*** 0.303*** 0.293*** 0.320*** 0.292*** 0.308*** 0.465***

Gender

Male

0.852*** 0.821*** 0.773*** 0.786*** 0.726*** 0.728*** 0.745***

Other/missing

0.711*** 0.597*** 0.324*** 0.487*** 0.691*** 0.632*** 0.637***

Race

Black

1.474*** 1.539*** 1.347*** 1.288*** 1.374*** 1.329*** 1.382***

Hispanic

0.956** 1.056*** 0.953*** 0.999 1.019 1.000 0.965***

Missing

0.819*** 1.138*** 0.941* 0.964 0.975 1.031 1.010

Other

0.976 1.051** 0.935*** 0.938*** 0.940*** 0.954** 0.940**

Prefer not to respond

0.845*** 0.988 0.863*** 0.876*** 0.828*** 0.862*** 0.853***

White

1.121*** 1.210*** 1.016 1.031** 1.020 1.020 0.983

Family income

Linear

1.354*** 1.324*** 1.337*** 1.265*** 1.332*** 1.260*** 1.313***

Quadratic

0.901*** 0.921*** 0.927*** 0.912*** 0.923*** 0.923*** 0.930***

Cubic

0.993 0.988* 0.997 0.996 0.992 0.993 0.992

Note: * p < 0.05, ** p < 0.01, ***p < 0.001; the reference categories in the model were National testing program, female, and Asian.

ACT Research | Research Report | January 2024 32

Notes

HSGPA is reported on a scale of 0.0 to 4.0. A letter grade of A was defined as an HSGPA of

3.5 or higher, a letter grade of B was defined as an HSGPA between 2.5 and 3.5, a letter

grade of C was defined as an HSGPA between 1.5 and 2.5, a letter grade of D was defined

as an HSGPA between 1.0 and 1.5, and an HSGPA below 1.0 was considered an F.

The National Student Clearinghouse is a nonprofit education service whose database contains

enrollment information for 97% of students at postsecondary institutions

(https://nscresearchcenter.org/current-term-enrollment-estimates/).

To deal with nonconvergence issues, we changed the default optimization method from

Nelder-Mead optimization to the quadratic approximation approach.

Family income was included in the models as an ordered factor. In R, ordered factors in

logistic regression are tested for higher order terms. As such, the models included a test for

the linear, quadratic, and cubic terms.

The simplified equations from Huang (2022) are used. The marginal 



formula: 







 





 















, where 



is the variance of fixed effects and 



is the variance of the

random intercept. The conditional 



formula: 











 













 















, where 



is the

variance of fixed effects and 



is the variance of the random intercept.

Across years, the standard deviation for ACT Composite score was about 1.0.

ABOUT ACT

ACT is a mission-driven, nonprofit organization

dedicated to helping people achieve education and

workplace success. Grounded in more than 60 years

of research, ACT is a trusted leader in college and

career readiness solutions. Each year, ACT serves

millions of students, job seekers, schools, government

agencies, and employers in the U.S. and around the

world with learning resources, assessments, research,

and credentials designed to help them succeed from

elementary school through career.

For more information, visit act.org