Research Article   Embracing the Generalized Propensity Score Method: Measuring the Effect of Library Usage on First-Time-In-College Student Academic Success   Jingying Mao Department of Statistics Florida State University Tallahassee, Florida, United States of America Email: mjy_jean@hotmail.com    Kirsten Kinsley Assessment Librarian Florida State University Libraries Tallahassee, Florida, United States of America Email: kkinsley@fsu.edu    Received: 2 Aug. 2017    Accepted: 9 Nov. 2017        2017 Mao and Kinsley. This is an Open Access article distributed under the terms of the Creative Commons‐Attribution‐Noncommercial‐Share Alike License 4.0 International (http://creativecommons.org/licenses/by-nc-sa/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly attributed, not used for commercial purposes, and, if transformed, the resulting work is redistributed under the same or similar license to this one.     Abstract   Objective – This research focuses on First-Time-in-College (FTIC) student library usage during the first academic year as number of visits (frequency) and length of stay (duration) and how that might affect first-term grade point average (GPA) and first-year retention using the generalized propensity score (GPS). We also want to demonstrate that GPS is a proper tool that researchers in libraries can use to make causal inferences about the effects of library usage on student academic success outcomes in observation studies.   Methods – The sample for this study includes 6,380 FTIC students who matriculated in the fall 2014 and fall 2015 semesters at a large southeastern university. Students’ library usage (frequency and duration), background characteristics, and academic records were collected. The Generalized Propensity Score method was used to estimate the effects of frequency and duration of FTIC library visits. This method minimizes self-selection bias and allows researchers to control for   demographic, pre-college, and collegiate variables. Four dose-response functions were estimated for each treatment (frequency and duration) and outcome variable (GPA and retention).   Results – The estimated dose-response function plots for first-term GPA and first-year retention rate have similar shapes, which initially decrease to the minimum values then gradually increase as the treatment level increases. Specifically, the estimated average first-term GPA is minimized when the FTIC student only visits the library three times or spends one hour in the library during his/her first semester. The threshold for first-year retention occurs when students visit the library 15 times or spend 21 hours in the library during their first semester. After those thresholds, an increase in students’ library usage is related to an increase in their academic success.   Conclusions – The generalized propensity score method gives the library researcher a scientifically rigorous methodological means to make causal inferences in an observational study (Imai & van Dyk, 2004). Using this methodological approach demonstrates that increasing library usage is likely to increase FTIC students’ first-term GPA and first-year retention rates past a certain threshold of frequency and duration.     Introduction   The collegiate experience often includes a diversity of opportunities and experiences to foster student development and engagement affecting the retention and academic success of the first-time-in-college (FTIC) student. According to Astin’s Input-Environment-Output (I-E-O) Model of Student Involvement, student inputs—such as high school grade point average (GPA), ACT scores, and gender—are often associated predictors of first-year student success outputs (or outcomes), such as grades and retention (Strauss, 2014; Astin, 1997). The collegiate environment, including a student’s major, enrolled credit hours, involvement in athletics, living in learning communities, and employment is also an important influence on student outputs. Another potential environmental factor that may affect student success outputs is time spent in the library.   The research study presented in this article attempts to isolate the treatment variables of number of library visits (frequency) and total hours of stay (duration) during the first year of college while controlling for other potential predictors of college success, such as student input and other collegiate environmental variables, by measuring the effects of frequency and duration of library visits on retention and GPA. Since randomizing a control group of students who do not use the library and those who do is ethically impossible, how do we measure FTIC students’ success and the effects of library usage while also controlling for student inputs and other non-library environmental impacts?    We decided to apply the generalized propensity score (GPS) method for a number of reasons. Using GPS in addition to the I-E-O design gives a more rigorous approach to measuring library impact on student academic success because we attempt to control for as many inputs and other environmental collegiate variables as possible. In addition, it allows us to “make causal inferences from correlational data” and to “minimize the chances that our inferences are wrong” (Astin & Antonio, 2012, p. 31). As Astin & Antonio (2012) emphatically state, “Although we can never be sure that we have controlled all such variables, the more we control, the greater confidence we can have in our causal inferences” (p. 31). Furthermore, using the GPS method reduces the effects of self-selection bias (Astin & Antonio, 2012, p. 31). The bias may be caused because students who have certain characteristics, such as higher ACT scores and higher high school GPA, may self-select to use the library frequently and for long durations. This may cause an overestimation of the treatment effect of library usage. GPS also allows us to measure the effect of continuous library usage variables over time by frequency and duration. Moreover, we can predict that with each treatment or dose of library time, retention and GPA for FTIC students will increase. If more library visits and duration of stay are related to increasing retention rates and higher grades, we will have more confidence to say that as library visits increase so do the student success variables of first-year retention and GPA.   Literature Review   According to Astin’s Input-Environment-Output (I-E-O) Model of Student Involvement (1970, 1990, 1993), both student inputs and the college environment influence student outputs (arrows B and C on Figure 1). (Please note: The terms output and outcomes will be used interchangeably throughout this paper as they relate to Astin’s theory, even though outputs are typically defined differently than outcomes.) At the same time, student inputs (arrow A on Figure 1) affect how students experience the college environment.   According to the model, input variables such as pre-college high school grades and college entrance exam scores (e.g., SAT scores) collectively impact whether a student succeeds in college. Higher education research has been exploring the environmental and engagement variables that contribute to student academic success or outputs. These variables may includes student engagement, investment in “educationally purposeful activities” (Kuh, 2001, p. 12), involvement in student organizations, social interactions, and engagement with faculty (Braxton, Hirschy, & McClendon, 2004; Kuh, Cruce, Shoup, Kinzie, & Gonyea, 2008; Roksa & Whitley, 2017). “Without knowing how students spend their time, it’s almost impossible to link student learning outcomes to the educational activities and processes associated with them” (Kuh, 2001, p. 15).   Librarians who research what factors the library contributes to student success would benefit from applying Astin’s Model since it offers a practical, holistic theoretical approach to looking at the interaction between student attributes and their environment and can easily incorporate library activities as part of the environmental variables. It acknowledges what academic librarians already know—that “many other factors besides the library contribute to students’ academic success . . .” (Jantii & Cox, 2012, p. 4). Even so, libraries provide many services and resources that help to engage students in “educationally purposeful activities” that contribute to student success. “Students engage in a wider variety of interactions with their libraries and it is important to examine the differences those interactions can have on student outcomes” (Soria, Fransen, & Nackerud, 2013, p. 149).     Figure 1 Astin’s Input-Environment-Output Model.     In 2003, Kuh and Gonyea stated that “relatively little is known about what and how students’ academic library experiences contribute to desired outcomes of college . . .” (p. 258). Over 15 years later, Soria et al. (2017a, 2017b) report a similar dearth of research in this area, though more and more research is rapidly being published on this topic. Almost 50 years ago, Kramer and Kramer (1968) looked at the retention rates of freshman who used the library and found that borrowing library books was associated with retention. Mezick (2007) found a significant positive association between library expenditures and student persistence for all Carnegie Classifications and between retention and the “number of library professional staff . . . at doctoral granting institutions” (p. 564).   Although other studies have looked at student outcomes and library use, it was not until the Value of Academic Libraries’ initiative of the Association of College & Research Libraries (ACRL) that a collective, concentrated effort was made to create a body of research demonstrating academic library value and impact related to student success measures (Oakleaf, 2010). Following the commencement of the Value of Academic Libraries initiative, current library research demonstrates connections between FTIC student library usage and its impact on GPA and retention outcomes. Emmons & Wilkinson (2011) found that library input variables (e.g., wages, library volumes, and expenditures) had an effect on student retention. Using a linear regression model while controlling for socioeconomic status, race, and ethnicity, they discovered that an increase in the ratio of professional library staff to students had a positive effect on both student retention (measured by students returning for their second year) and six-year graduation rates. Interestingly, Stemmer and Mahan (2016) found that the ways that freshman used the library (outputs) were associated with GPA and retention. Using the library for academic purposes like checking out books or using online resources were associated with GPA and retention, but using the library computers for personal use and the late-night study rooms for cramming sessions was negatively associated with success outcomes.   Nine recent studies examined by the authors found that a combination of library space, instruction, and resource usage by FTIC students was positively associated with retention, GPA, or both (Kot & Jones, 2015; Soria et al., 2013, 2014, 2017a, 2017b; Haddow, 2013; Murray, Ireland, & Hackathorn, 2016; Stemmer & Mahan, 2016; Stone & Ramsden, 2012). Note that of the studies examined, most focused on library space and resource usage effects on student outcomes which included workstation logins, study room usage, e-resources and print books usage, interactions with library personnel, use of ILL and reference, and other similar resources. Kot & Jones (2015), Soria et al. (2017b), and Murray et al. (2016) also included library instruction in their list of environmental variables. Some of the studies controlled for other input and environmental variables that may impact student success (Kot & Jones, 2015; Soria et al., 2013, 2014, 2017a, 2017b). Some used the propensity score matching methodology (Kot & Jones, 2015; Soria et al. (2017b) and some studies applied Astin’s I-E-O model as their conceptual framework (Kot & Jones, 2015; Soria et al., 2014, 2017a, 2017b; Stemmer & Mahan, 2016).   Another study, conducted by masters of economics students at Florida State University using our local library turnstile data, found that students who had low GPAs showed “larger academic gains from additional library usage than their high-GPA library user counterparts” (Holcombe, Lukashevich, & Alvarez (2016, p. 14). Note that though this study examined undergraduate student library usage and GPA, it was not limited to the FTIC population. The use of the GPS methodology is unique to this library study since we were predicting outcomes based on continuous variables of library usage over time from actual turnstile data. It is interesting to note that the two outcomes measured in this study, GPA and retention, have been correlated: higher individual GPAs “may well be the single best predictors of student persistence . . .” (Pascarella & Terenzini 2005, p. 396). In addition, scholarship that focuses exclusively on the critical role of library instruction and its effect on first-year retention and GPA is not reviewed here.   Aims   This study aims to evaluate the effect of library usage (frequency of visits and duration of stay) over the course of a semester on FTIC student academic success measured in first-term GPA and first-year retention rate. In our study, student outputs or dependent variables are first-term GPA and first-year retention rate. The independent variables include the environmental variables of library usage (library visit frequency and duration) while controlling for other non-library related college environment variables. Other controlled variables include student inputs, such as demographic characteristics and other pre-college academic variables. By studying first-year students we by default control for the effects “of later collegiate experiences that may also influence students’ outcomes . . .” (Soria et al., 2017a, p. 10).   This is an observational study where we could not randomly assign students to different amounts of library visit treatment during their first year. As a result, students have self-selected themselves into different levels of treatment because of their different input variables, such as gender, class, major etc. So we also tried to find a statistical method to minimize the self-selection bias in our sample.   Specifically, the research questions for this study are:   1)       Does library usage measured in frequency (visits per semester) and duration (length of stay per semester) impact student academic success in terms of first-term GPA and first-year retention rate? 2)       Are these impacts still observed after controlling for other input and environmental variables? and 3)       Does embracing generalized propensity scoring give librarians more rigorous research results?   Methods   Data   The sample for this study includes 6,380 FTIC students who matriculated in the fall 2014 and fall 2015 semesters at a large southeastern university. Here FTIC refers to an entering freshman or a first-year student attending college for the first time at the undergraduate level. This includes students who attended college for the first time in the prior summer term and are also enrolled in the fall term. Also included are students who entered with advanced standing (having earned college credits before graduation from high school). For the purposes of this paper, retention is measured for FTIC students by their “persistence between the first and second year at college” (Kuh, et al., 2008, p. 555).   Data in the study comes from two sources: the C-Cure System (card swipe system) and the Office of Institutional Research. The campus has two major libraries and these were chosen sites for the study because they have turnstiles that could provide primary data for our study. Each library has six turnstiles, including two entrances, two exits, and a handicap entrance and exit. Both libraries require students to swipe student IDs at the turnstiles to enter or exit libraries. The C-Cure System collects card-swipe data that includes student identification information, time that students enter or exit the library, direction (in or out), and which turnstile they use. By matching swipe-in and swipe-out records, we extracted frequency and duration of individual library usage for each semester.   At our request, the Office of Institutional Research provided all other student background characteristics and academic records for all FTIC students. By merging card-swipe data and student information data, the final data set was ready for analysis. This data was coded to keep student information anonymous. The output (dependent) variables of interest were first-term GPA and first-year retention rate.   The environment (treatment) variables of interest were library usage measures, defined as first-term library visit frequency and duration (measured in hours). Other environment variables that we controlled for include major (college), class (freshman, sophomore, junior, senior or non-degree), military status, participation in athletics or sports, current load (credit hours enrolled in the first term), matriculation year (2014 or 2015), housing status (whether living on or off campus), and participation in the Center for Academic Retention and Enhancement program (provides transition support for minority students).   The input variables for the study included students’ demographic characteristics and pre-college academic variables. Demographic characteristics included the student’s gender, race, citizenship, age at matriculation, parent income level, and education levels of students’ mothers and fathers. Pre-college academic variables included the student’s high school GPA, ACT scores, and transfer credits. Some of students were admitted with SAT or ACT scores only. To compare those two measures, we transferred SAT scores into corresponding ACT scores using an SAT/ACT concordance/comparison chart. For those students who had both test scores, only the ACT scores were used. Table A1 in the Appendix presents summary statistics for all variables.   Generalized Propensity Score Method   To adjust for self-selection bias and control for the inputs and other environmental variables in a scientifically rigorous way, we use the GPS method developed by Hirano and Imbens (2004). This method is a generalization of the binary treatment propensity score matching method (Rosenbaum & Rubin, 1983) and is used to make causal inference in the observational studies (Imai & Dyk, 2004).   In this study, the treatment variables (library visit frequency and duration per student) are continuous measurements that can take the value of all positive integers. So, we decided to use the GPS method instead of the binary propensity score matching method to estimate the effects of continuous treatments—that is, the number of library visits and the number of hours spent in the library over time on student grades and retention.   Following Hirano and Imbens (2004), we have random samples of FTIC students indexed by . For each sample , there is a set of potential outcomes,  (i.e. first-term GPA, first-year retention rate) with a given level of treatment , referred to as the unit-level dose-response function. In our study, treatment  is the first-term library visit frequency and duration and  is an interval . For each sample , we observed a vector of covariates, , its actual treatment received, , and actual outcome corresponding to the actual treatment received, . Our goal was to estimate the average dose-response function: . Hereafter, we will omit  to simplify the notation.   The key assumption for the GPS method is weak unconfoundedness introduced by Hirano and Imbens (2004):   .   We assumed that the level of treatment received is independent of the potential outcome given observed covariates. This assumption requires us to get a rich set of covariates including all possible variables that may influence selection into different levels of treatment.   Based on this assumption, we were able to estimate the GPS. If we write the conditional density of the treatment given the covariates as , then the GPS is defined as:   .   If the GPS is correctly estimated, then it has a balance property as the binary propensity score:   .   Hirano and Imbens (2004) mentioned that this property does not require unconfoundedness. In combination with weak unconfoundedness, it implies that the level of treatment received is unconfounded given the GPS as well.   Given this result, GPS can be used to remove bias caused by difference in covariates in the following two steps. First, we estimated the conditional expectation of potential outcome as a function of the treatment level and estimated GPS:   .   Second, we estimated the dose-response function at each treatment level by taking the average of this conditional expectation over the GPS evaluated at that particular treatment level:   .   Implementation   The first step is to estimate the GPS. Since our treatment variables (frequency and duration) are counts and highly skewed with a large amount of zero values, a negative binomial generalized linear model with log link function is used to model the conditional distribution:   .   Then the GPS is estimated via the following:   .   There are many other ways to specify the distribution and estimate the GPS. As long as the balance of covariates is achieved after adjusting for the GPS, the model specification is not the key point here.   The second step is to specify the conditional expectation of potential outcome given the treatment level and estimated GPS using OLS. In our study, a quadratic approximation including the interaction term was used when the outcome variable is first-year GPA:   .   When the outcome is first-year retention rate, we used a logistic regression model to estimate the conditional expectation of potential outcome because retention is a binary outcome with value 0 as not being retained and 1 as being retained:   .   However, there is no direct causal interpretation of those estimated coefficients (Hirano & Imbens, 2004).   The final step was to estimate the average dose-response function at treatment levels of interest given the estimated parameters in the last step. In the case of first-term GPA, the dose-response function was estimated as the following:   .   And in the case of first-year retention, the dose-response function is estimated as the following:   .   We also computed the 95% confidence bands for the dose-response function based on 1,000 bootstrap replications, considering all estimation steps including GPS and -parameters.   Common Support Condition and Balancing of Covariates   As in the standard propensity score matching method, we needed to check the common support condition. We adapted the approach from Kluve, Schneider, Uhlendorff, & Zhao (2012). First, we divided the sample into three groups by the 30th and 70th quartiles of the treatment. For each group, we evaluated the GPS for the whole sample at the group mean of the treatment. Then we plotted the distribution of the evaluated GPS for that group against the distribution of the evaluated GPS for the rest of the sample. The overlap of those two distributions is the common support. We repeated the above procedures for all three groups. Finally, we restricted our final sample to individuals who are comparable across all three groups simultaneously. In other words, we deleted individuals whose GPS fell out of any common support of the three groups.   Besides assessing the common support condition, balancing of covariates is also very important to the GPS method. We regressed each covariate on the treatment with and without conditioning on the predicted level of treatment  (Imai & van Dyk, 2004).  If there was no correlation between treatment and any covariate after conditioning on the predicted treatment, then we concluded that the covariate balance is achieved after adjusting for the GPS.   Results   First-Term GPA   All tables and figures regarding the process of implementing the GPS method are included in the Appendix. As previously noted, Table A1 provides summary statistics. Table A2 provides the estimated coefficients from the negative binomial generalized linear models using the first-term GPA as the outcome variable. Both models showed that age, participation in athletics, ACT scores, college attended, current academic load, matriculation year, and race had influence on student library usage.   We assessed the common support condition using the method we described in the methodology section. Figures A1 and A2 in the Appendix illustrate the distribution of the evaluated GPS before and after deleting the non-overlap for the treatment variables of frequency and duration, respectively. After imposition of common support for the frequency treatment, we deleted only 0.4% of our original sample. For the duration treatment, we deleted 0.3% of our original sample.   Then we checked the balancing properties of the GPS using the method proposed by Imai & van Dyk (2004). Table A3 presents the coefficient and its standard error for each covariate with and without conditioning on . Table A3 clearly demonstrates that before we conditioned on  multiple covariates were significant. After we conditioned on, no significant covariate was observed. For example, participation in athletics had a high positive correlation with both treatments (frequency and duration). However, once we conditioned on the predicted level of treatment, athletic participation was not significant in either case. So, we concluded that the balancing properties of the GPS were achieved in both treatment cases.     Figure 2 The dose-response function of first-term library usage frequency vs. first-term GPA.     Figure 3 The dose-response function of first-term library usage duration vs. first-term GPA.     The final step of our study was to estimate the dose-response function. We regressed the outcome: first-term GPA on the treatment variable and the GPS. The estimated coefficients are listed in Table A4. As was mentioned before, the estimated coefficients did not have any direct causal interpretation.   The dose-response function was estimated for each treatment level of interest by averaging the estimated regression function over the GPS evaluated at the desired treatment level. Figures 2 and 3 present the dose-response function of first-term GPA for the treatment variables of frequency and duration, respectively. The dotted lines were 95% confidence bands based on 1,000 bootstrap replications that accounted for all estimation steps.   Figures 2 and 3 show the dose-response functions for frequency and duration have similar shapes. First-term GPA first decreased and reached its minimum value, then gradually increased when the library usage frequency and duration increased.   For frequency, first-term GPA was minimized at 3.19066 when the FTIC student only visited the library three times in their first semester. Once the student visited the library over three times, library usage had a continued positive relationship with their first-term GPA.   Similarly, for duration, first-term GPA was minimized at 3.177407 when the FTIC student only spent one hour in the library during their first semester. When the student spent an hour or longer in the library there were gains in first-term GPA. The longer the time spent in the library, the larger the increase in first-term GPA.   First-Year Retention Rate Analysis procedures for first-year retention rate were almost the same as the procedures for first-term GPA, except that we included first-term GPA as a covariate when the outcome variable was retention rate. We then used a logistic regression model in order to estimate the conditional expectation of outcome.   In the Appendix, Table A5 presents the estimated coefficients from the GPS estimation step.  Figures A3 and A4 and Table A6 (see the Appendix) verified the common support condition and the balancing properties. The estimated coefficients from the logistic regression model are presented in Table A7.   The dose-response functions were finally estimated at each treatment level of interest. Figures 4 and 5 present the dose-response function of first-year retention rate for the treatment variables of frequency and duration, respectively. The dotted lines are 95% confidence bands based on 1,000 bootstrap replications that accounted for all estimation steps.     Figure 4 The dose-response function of first-term library usage frequency vs. first-year retention.     Figure 5 The dose-response function of first-term library usage duration vs. first-year retention.     Both dose-response functions have a shape similar to Figures 2 and 3. Both plots indicate that first-year retention rate first declined to its minimum value within the lower value of the treatment and then gradually increased as the treatment increased.   For frequency, when students visited the library only fifteen times in their first semester, they had the lowest first-term retention rate at 93.89%. For duration, the minimum retention rate was achieved at 93.84% when FTIC students spent only twenty-one hours in the library during their first semester. After that, further increases in first-term library usage frequency and duration both resulted in higher first-year retention rate.   The estimated dose-response function plots for first-term GPA and first-year retention rate have similar shapes, which initially decrease to minimum values and then gradually increase as the treatment levels increase. In other words, there was a threshold of frequency and duration of library visits where an increase of students’ library usage had a negative effect on their first-term GPA and retention rates. Specifically, the estimated average first-term GPA was minimized when FTIC students visited the library only three times or spent only one hour in the library during their first semester. The threshold for measurable increases in first-year retention occurred when students visited the library fifteen times or spent twenty-one hours in the library during their first semester.   As the estimated dose-response functions reveal, increasing library usage was likely to increase FTIC students’ first-term GPA and first-year retention rates past a certain threshold of frequency and duration. When FTIC students visited more than three times or spent more than two hours in the library during their first semester, library usage positively affected students’ first-term GPAs. After FTIC students crossed the threshold of visiting the library more than fifteen times or spending more than twenty-one hours there in their first semester, students with higher library usage had higher first-year retention rates.   Discussion   The small drop of both first-term GPA and retention rate before reaching the thresholds for frequency and duration may be explained in several possible ways. First, we did not account for those FTIC students who may go to other libraries on campus other than the two major libraries included in this study. For example, engineering majors may not choose to come to the two on-campus libraries because their department and library are located off-campus. Some students may only come to the libraries at the beginning of the semester or during finals. Holcombe et al. (2016), using the same cohort and data set, found that those students who come to the library only to cram during finals week do not seem to benefit from low frequency, high duration library usage per semester.     The study has several limitations. The definition of library usage used here (total frequency and duration in one semester) may be too broad. We consider only when and how long the students entered the building, ignoring what they might be doing while in the building such as using other   library services, collections, and spaces (such use of study rooms)  (Soria et al. 2017a; 2017b). Furthermore, we cannot presume that students are studying when they visit the library. We can only assume they are doing some form of “educationally purposeful activities” that include using databases to conduct research and studying (Kuh, 2001, p. 12; Kuh & Gonyea, 2003). In one recent survey by Cengage, results showed that student library users spend their time studying alone, using the databases and reference materials, and meeting study groups (Strang, 2015). In a fall 2016 survey, the activities our students reported coming to the library for were to 1) work on a paper, project, or homework; 2) study for an exam; 3) print something; or 4) wait between classes (Dawson, 2016). Another limitation of this study is that it is not possible to control or account for all possible covariates that may influence the student success outcomes of GPA and first-year retention rates. Especially difficult to measure are intangible, intrinsic, and individual student inputs. For example, one study found that a student’s “grit” or “mindset,” which is the “willingness to work hard for an extended period in search of a long-term goal,” was a key factor in college student success (Barton, 2015, para. 9).   Conclusion   Our results indicate that increasing library usage contributes to higher FTIC students’ first-term GPAs and first-year retention rates past a certain threshold of frequency and duration. In addition, GPS is a valid methodology to use because it minimizes self-selection bias and estimates the potential outcome, GPA and retention rate, at every possible value of library usage (frequency and duration).   Using the GPS method, future studies could build on the findings of this study by looking at library usage and the relative impact on student four-to-six-year graduation rates, library usage across different academic disciplines, and other populations of library users, such as faculty and graduate students. Furthermore, future analyses could triangulate these results by analyzing the effects of library e-resource and equipment usage, instruction, and participation in library outreach and engagement activities to gain a more comprehensive understanding of how the academic library services, spaces, and resources collectively impact student success.   References   Astin, A. W. (1970). The methodology of research on college impact, part one. Sociology of Education, 43(3), 223–254. https://dx.doi.org/10.2307/2112065   Astin, A. W. (1990). Assessment for excellence: The philosophy and practice of assessment and evaluation in higher education. New York: Maxwell Macmillan International.   Astin, A. W. (1993). What matters in college?: Four critical years revisited (1st ed.). San Francisco: Jossey-Bass.   Astin, A. W. (1997). How “good” is your institution’s retention rate? Research in Higher Education, 38(6), pp. 647-658. https://dx.doi.org/10.1023/A:1024903702810     Astin, A. W., & Antonio, A. L. (2012). Assessment for excellence: The philosophy and practice of assessment and evaluation in higher education. Lanham, MD: Rowman & Littlefield Publishers.   Barton, D. (2015, September 16). The most important factor in a college student’s success [Blog post]. Retrieved from https://blogs.wsj.com/experts/2015/09/16/the-most-important-factor-in-a-college-students-success/   Braxton, J. M., Hirschy, A. S., & McClendon, S. A. (2004). Understanding and reducing college student departure. ASHE-ERIC higher education report, volume 30, issue 3. Indianapolis: Jossey-Bass, An Imprint of Wiley.   Cox, B. L., & Jantti, M. (2012, July 17). Discovering the impact of library use and student performance. EDUCAUSE Review. Retrieved from http://er.educause.edu/articles/2012/7/discovering-the-impact-of-library-use-and-student-performance   Dawson, A. (2016). Strozier daytime visit feedback fall 2016 results. Retrieved from https://docs.google.com/document/d/1y8RJdKgnVUvvwmS6dTNHwSOFkqM1IUonq2qvb-enRgo/edit?usp=drive_web&usp=embed_facebook   Donnelly, P. J. (2010). Examining pre-college academic variables: Investigating future college success. Retrieved from ProQuest Dissertations & Theses Global. (3398089).   Haddow, G. (2013). Academic library use and student retention: A qualitative analysis. Library & Information Science Research, 35(2), 127-136. https://dx.doi.org/10.1016/j.lisr.2012.12.002   Hirano, K., & Imbens, G. W. (2004). The propensity score with continuous treatments. In A. Gelman & X. L. Meng (Eds.), Applied Bayesian modeling and causal inference from incomplete-data perspectives (pp. 73–84). John Wiley & Sons, Ltd.   Holcombe, C., Lukashevich, I., & Alvarez, J. (2016, July 29). Measuring the effects of increased library use on GPA outcomes of FSU undergraduates. Symposium on Applied Economics 2016: The Final Presentations for the M.S. in Applied Economics. Florida State University. Retrieved from https://www.lib.fsu.edu/sites/default/files/sites/default/files/upload/executive_summary.pdf   Imai, K., & van Dyk, D. A. (2004). Causal inference with general treatment regimes. Journal of the American Statistical Association, 99(467), 854–866. https://dx.doi.org/10.1198/016214504000001187   Kot, F. C., & Jones, J. L. (2014). The impact of library resource utilization on undergraduate students’ academic performance: A propensity score matching design. College & Research Libraries, 76(5), 566-586. https://dx.doi.org/10.5860/crl.76.5.566   Kramer, L. A., & Kramer, M. B. (1968). The college library and the drop-out. College & Research Libraries, 29(4), 310–312.     Kuh, G. D. (2001). Assessing what really matters to student learning: Inside the national survey of student engagement. Change: The Magazine of Higher Learning, 33(3), 10–17. https://dx.doi.org/10.1080/00091380109601795   Kuh, G. D., & Gonyea, R. M. (2003). The role of the academic library in promoting student engagement in learning. College & Research Libraries, 64(4), 256–282. https://dx.doi.org/10.5860/crl.64.4.256   Kuh, G. D., Cruce, T. M., Shoup, R., Kinzie, J., & Gonyea, R. M. (2008). Unmasking the effects of student engagement on first-year college grades and persistence. The Journal of Higher Education, 79(5), 540–563. https://doi.org/10.1080/00221546.2008.11772116   Kluve, J., Schneider, H., Uhlendorff, A., & Zhao, Z. (2012). Evaluating continuous training programmes by using the generalized propensity score. Journal of the Royal Statistical Society: Series A (Statistics in Society), 175(2), 587–617. https://dx.doi.org/10.1111/j.1467-985X.2011.01000.x   Mezick, E. M. (2007). Return on investment: Libraries and student retention. The Journal of Academic Librarianship, 33(5), 561–566. https://dx.doi.org/10.1016/j.acalib.2007.05.002   Murray, A., Ireland, A., & Hackathorn, J. (2016). The value of academic libraries: Library services as a predictor of student retention. College & Research Libraries, 77(5), 631–642. https://dx.doi.org/10.5860/crl.77.5.631   Oakleaf, M. (2010). The value of academic libraries: A comprehensive research review and report. Chicago: Association of College and Research Libraries. http://www.ala.org/acrl/sites/ala.org.acrl/files/content/issues/value/val_report.pdf   Pascarella, E. T., & Terenzini, P. T.  (1991). How college affects students: Findings and insights from twenty years of research. San Francisco: Jossey-Bass Publishers.   Roksa, J., & Whitley, S. E. (2017). Fostering academic success of first-year students: Exploring the roles of motivation, race, and faculty. Journal of College Student Development, 58(3), 333–348. https://dx.doi.org/10.1353/csd.2017.0026   Rosenbaum, P. R., & Rubin, D. B. (1983). The central role of the propensity score in observational studies for causal effects. Biometrika, 70(1), 41–55. https://dx.doi.org/10.2307/2335942                 Soria, K. M., Fransen, J., & Nackerud, S. (2013). Library use and undergraduate student outcomes: New evidence for students’ retention and academic success. Portal: Libraries and the Academy, 13(2), 147–164. https://dx.doi.org/10.1353/pla.2013.0010   Soria, K. M., Fransen, J., & Nackerud, S. (2014). Stacks, serials, search engines, and students’ success: First-year undergraduate students’ library use, academic achievement, and retention. The Journal of Academic Librarianship, 40(1), 84–91. https://dx.doi.org/10.1016/j.acalib.2013.12.002   Soria, K. M., Fransen, J., & Nackerud, S. (2017a). Beyond books: The extended academic benefits of library use for first-year college students. College & Research Libraries, 78(1), 8–22. https://dx.doi.org/10.5860/crl.v78i1.16564   Soria, K. M., Fransen, J, & Nackerud, S. (2017b). The impact of academic library resources on undergraduates’ degree completion. College & Research Libraries, 78(6), 812-823.   https://dx.doi.org/10.5860/crl.78.6.812   Stemmer, J. K., & Mahan, D. M. (2016). Investigating the relationship of library usage to student outcomes. College & Research Libraries, 77(3), 359-375. https://dx.doi.org/10.5860/crl.77.3.359    Stone, G., & Ramsden, B. (2013). Library impact data project: Looking for the link between library usage and student attainment. College and Research Libraries, 74(6), 546-559. https://dx.doi.org/10.5860/crl12-406   Strang, T. (2015, July 2). Top four reasons students use their college library [Blog post]. Retrieved from https://blog.cengage.com/top-four-reasons-students-use-their-college-library     Appendix   Table A1 Summary Statistics Variables Mean Standard deviation Output Variables GPA 3.278 0.690 Retention 0.957 0.204 Environment (treatment) Variables Frequency 35.066 39.705 Duration 56.019 74.300 Other Environment Variables Military 0.026 0.160 Athlete 0.018 0.134 Housing 0.821 0.384 CARE 0.000 0.022 Current Load 12.869 1.842 Class Freshman 0.711 0.453 Sophomore 0.253 0.435 Junior 0.036 0.186 Senior 0.001 0.025 Non-Degree 0.000 0.013 College Applied Studies 0.000 0.018 Arts & Sciences 0.301 0.459 Business 0.150 0.357 Communication & Information 0.046 0.210 Criminology 0.029 0.167 Education 0.021 0.143 Engineering 0.070 0.255 Film School 0.005 0.071 Fine Arts 0.006 0.075 Human Sciences 0.072 0.259 Music 0.027 0.163 Nursing 0.025 0.157 Registrar 0.000 0.013 Social Sciences 0.071 0.257 Social Work 0.006 0.078 Undergraduate Studies 0.146 0.353 Visual Arts, Theatre, & Dance 0.024 0.153 Matriculation Year 2014 0.453 0.498 2015 0.547 0.498 Input Variables Age 20.749 0.776 US Citizen 0.978 0.146 HS GPA 4.045 0.340 ACT 27.145 2.740 Transfer or Exam Credit 21.679 16.793 Race White 0.683 0.465 Hispanic/Latino 0.177 0.382 Black/African American 0.046 0.210 Asian 0.031 0.174 American Indian/Alaska Native 0.002 0.041 Native Hawaiian/Other Pacific Islands 0.002 0.040 Two or More Races 0.041 0.199 Not Specified 0.018 0.131 Gender Female 0.593 0.491 Male 0.407 0.491 Father's Education Level College 0.057 0.231 High School 0.028 0.165 Middle School 0.001 0.028 Unknown 0.914 0.280 Mother's Education Level College 0.058 0.235 High School 0.024 0.153 Middle School 0.002 0.040 Unknown 0.916 0.277 Parent Income Level < $1000 0.008 0.091 $1000-$40000 0.018 0.132 $40000-$75000 0.017 0.130 $75000-$100000 0.013 0.114 $100000+ 0.036 0.187 Unknown 0.907 0.290     Table A2 Estimated Coefficients from the GPS Estimation   Treatment: Frequency Treatment: Duration Covariates Estimate Std. Error Estimate Std. Error   military -0.0713 0.1049 -0.0987 0.1201   athlete -0.5749a 0.1277 -0.6429a 0.1459   housing 0.0723 0.0444 0.1100c 0.0509   CARE 0.2593 0.7719 -0.2987 0.8878   current load 0.0319a 0.0096 0.0239c 0.0110   class.Freshman 2.3445 1.5296 1.6594 1.6231   class.Sophomore 2.3328 1.5314 1.6293 1.6253   class.Junior 2.3246 1.5368 1.7034 1.6319   class.Senior 2.5368 1.6819 1.5132 1.8105   college.Applied.Studies -2.1336c 1.0327 -0.9900 1.1066   college.Arts & Sciences 0.2616c 0.1132 0.5256a 0.1298   college.Business 0.0681 0.1180 0.3541b 0.1352   college.Communication & Information 0.0556 0.1338 0.3242c 0.1533   college.Criminology 0.0034 0.1471 0.1712 0.1686   college.Education -0.1176 0.1593 -0.0022 0.1824   college.Engineering 0.3619b 0.1272 0.6368a 0.1459   college.Film.School -0.0923 0.2603 -0.2208 0.2986   college.Fine.Arts -0.1341 0.2564 -0.1448 0.2934   college.Human.Sciences 0.2856c 0.1257 0.6087a 0.1440   college.Music -0.2808d 0.1488 -0.5593b 0.1707   college.Nursing 0.2225 0.1511 0.5199b 0.1731   college.Social.Sciences 0.2755c 0.1260 0.5308a 0.1444   college.Social.Work 0.2448 0.2405 0.3673 0.2756   college.Undergraduate.Studies 0.0885 0.1178 0.3144c 0.1350   MatriculationYearTer.20149 -0.1387b 0.0427 -0.1155c 0.0489   age 0.0755b 0.0276 0.0652c 0.0317   US citizen -0.1027 0.1189 -0.0344 0.1363   HS GPA 0.0655 0.0591 0.0191 0.0677   ACT -0.0139c 0.0070 -0.0236b 0.0080   Transfer Or Exam Credit -0.0009 0.0019 -0.0014 0.0022   Race.White -0.1113 0.1281 -0.0075 0.1468   Race.Hispanic.Latino -0.0377 0.1327 0.0865 0.1522   Race.Black.African.American 0.0161 0.1490 0.0765 0.1709   Race.Asian 0.2804d 0.1585 0.3924c 0.1817   Race.American.Indian.Alaska 0.1095 0.4228 0.1406 0.4849   Race.Native.Hawaiian.Oth.Pa 0.2246 0.4402 0.0388 0.5055   Race.Two.or.More.Races -0.0897 0.1509 0.0016 0.1730   Gender.Male 0.1047b 0.0368 -0.0265 0.0422   EducationFather.College -0.2234 0.2676 -0.3814 0.3063   EducationFather.High.School -0.1018 0.2706 -0.3427 0.3098   EducationFather.Middle.School -0.6790 0.5771 -1.3476c 0.6611   EducationMother.College 0.1792 0.2459 0.1591 0.2815   EducationMother.High.School 0.0914 0.2560 0.0494 0.2930   EducationMother.Middle.School -0.0591 0.4932 -0.2111 0.5648   ParentIncome....1000 0.0774 0.2275 0.2417 0.2605   ParentIncome..1000..40000 -0.2691 0.2604 -0.1351 0.2981   ParentIncome..40000..75000 -0.0937 0.2729 0.0781 0.3124   ParentIncome..75000.100000 -0.3024 0.2875 -0.0555 0.3290   ParentIncome..100000 -0.2199 0.2579 -0.0599 0.2952   aSignificant at the 0.1% level bSignificant at the 1% level cSignificant at the 5% level dSignificant at the 10% level       Figure A1 Common support condition for frequency.     Figure A2 Common support condition for duration.     Table A3 Covariate Balance With and Without Conditioning on Treatment: Frequency Treatment: Duration   Without Condition Condition Without Condition Condition Covariates Est Std. Error Est Std. Error Est Std. Error Est Std. Error military -1.936 3.108 0.526 3.057 -2.772 5.784 1.440 5.688 athlete -14.716a 3.801 -0.301 3.864 -25.067a 6.985 -0.703 7.060 housing 1.317 1.301 -0.655 1.285 4.120d 2.422 -0.504 2.399 CARE 11.519 22.951 -2.977 22.564 -17.749 42.721 -1.992 41.973 currentload 1.118a 0.270 0.075 0.275 0.892d 0.503 0.028 0.498 class.Freshman -0.359 1.100 -0.267 1.080 1.943 2.046 -0.724 2.017 class.Sophomore 0.657 1.147 0.267 1.127 -1.968 2.134 0.912 2.105 class.Junior -1.365 2.692 0.260 2.646 -0.424 5.000 -0.542 4.911 class.Senior -5.406 19.878 -7.218 19.525 -20.587 37.000 -7.299 36.352 college.Arts & Sciences 4.786a 1.085 0.087 1.115 8.660a 2.018 0.145 2.068 college.Business -3.644b 1.394 -0.120 1.390 -4.658d 2.594 -0.519 2.563 college.Communication & Infor -4.302d 2.369 0.145 2.346 -4.382 4.410 -0.331 4.340 college.Criminology -6.015c 2.996 -0.233 2.969 -11.474c 5.562 0.612 5.523 college.Education -9.548b 3.481 0.706 3.489 -19.332b 6.455 0.973 6.487 college.Engineering 8.641a 1.951 -0.073 2.010 11.458b 3.637 -0.691 3.666 college.Film.School -6.655 7.043 -0.487 6.930 -25.044d 13.107 -0.126 12.981 college.Fine.Arts -15.940c 7.153 0.055 7.108 -33.430b 12.358 -1.315 12.331 college.Human.Sciences 3.163 1.926 -0.034 1.903 13.329a 3.585 1.253 3.617 college.Music -11.205a 3.087 0.574 3.137 -33.874a 5.820 1.200 6.247 college.Nursing 0.914 3.163 -0.293 3.108 6.410 5.887 -1.413 5.806 college.Social.Sciences 3.658d 1.939 0.037 1.920 5.109 3.606 -0.348 3.560 college.Social.Work -1.366 6.384 -0.835 6.270 -6.215 11.882 -0.654 11.676 college.Undergraduate.Studies -3.875b 1.410 0.001 1.409 -6.168c 2.622 0.457 2.613 MatriculationYearTer.20149 -1.728d 1.002 0.089 0.991 -2.460 1.863 -0.307 1.836 age 1.165 d 0.649 0.013 0.643 1.267 1.200 -0.483 1.184 UScitizen -6.396 d 3.444 0.910 3.418 -9.033 6.389 -1.147 6.297 HSGPA 1.525 1.468 -0.031 1.445 -1.398 2.734 -0.209 2.686 ACT -0.168 0.182 -0.045 0.179 -1.129a 0.339 -0.029 0.341 Transfer Or Exam Credit 0.005 0.030 0.006 0.029 -0.054 0.055 0.016 0.054 Race.White -5.577a 1.069 -0.653 1.107 -9.178a 1.990 -0.937 2.038 Race.Hispanic.Latino 2.554d 1.305 0.497 1.289 5.730c 2.427 0.340 2.412 Race.Black.African.American 4.788c 2.377 0.488 2.352 7.144 4.417 1.146 4.357 Race.Asian 16.637a 2.869 0.482 3.049 26.383a 5.344 0.283 5.561 Race.American.Indian.Alaska 8.863 11.993 0.214 11.794 7.120 22.325 -0.085 21.932 Race.Native.Hawaiian.Oth.Pa 8.661 12.578 -2.597 12.377 -6.284 23.412 -3.421 22.996 Race.Two.or.More.Races 0.594 2.503 0.275 2.459 0.310 4.651 0.560 4.568 Gender.M 3.301b 1.01 -0.371 1.027 -1.505 1.888 -0.548 1.856 EducationFather.College -9.825a 2.176 -0.668 2.231 -16.136a 4.041 -1.312 4.099 EducationFather.High.School -4.732 3.063 1.153 3.034 -11.430c 5.653 1.782 5.622 EducationFather.Middle.School -18.915 19.877 -5.248 19.546 -31.344 36.999 3.270 36.412 EducationMother.College -8.686a 2.148 -0.640 2.183 -14.549a 3.989 -1.151 4.025 EducationMother.High.School -7.704c 3.304 0.271 3.290 -14.884c 6.090 0.268 6.068 EducationMother.Middle.Scho -7.497 13.257 1.780 13.036 -11.964 26.172 7.415 25.736 ParentIncome....1000 -0.193 5.534 -2.504 5.438 5.859 10.301 -0.247 10.125 ParentIncome..1000..40000 -8.768c 3.821 -0.093 3.799 -14.877c 7.113 1.502 7.072 ParentIncome..40000..75000 -4.059 3.822 0.459 3.766 -7.888 7.147 0.072 7.039 ParentIncome..75000.100000 -10.535 4.470c 0.485 4.453 -17.606c 8.220 -1.511 8.146 ParentIncome..100000. -9.337a 2.719 -0.525 2.738 -16.642a 5.028 -1.055 5.052 aSignificant at the 0.1% level bSignificant at the 1% level cSignificant at the 5% level dSignificant at the 10% level     Table A4 Estimated Coefficients of Conditional Distribution of GPA Given Treatment and GPS Treatment: Frequency Treatment: Duration Estimate Std. Error Estimate Std. Error Intercept 3.0990a 0.1311 Intercept 3.2390a 0.0880 Frequency 0.0039a 0.0010 Duration 0.0008c 0.0003 Frequency^2 0.0000b 0.0000 Duration^2 0.0000 0.0000 GPS 1.9740 3.2350 GPS -2.1390 1.7650 GPS^2 -7.1340 20.2600 GPS^2 15.7100d 9.0180 Frequency*GPS 0.1875 0.3512 Duration*GPS 0.1173 0.3676 aSignificant at the 0.1% level bSignificant at the 1% level cSignificant at the 5% level dSignificant at the 10% level     Table A5 Estimated Coefficients from the GPS Estimation Treatment: Frequency                            Treatment: Duration Covariates Estimate Std. Error Estimate Std. Error GPA 0.2140a 0.0283 0.2084a 0.0324 military -0.0586 0.1045 -0.0911 0.1199 athlete -0.6102a 0.1273 -0.6777a 0.1457 housing 0.0482 0.0442 0.0838 0.0507 CARE 0.2962 0.7687 -0.2671 0.8852 current load 0.0069 0.0099 -0.0008 0.0113 class.Freshman 2.3856 1.5245 1.6851 1.6187 class.Sophomore 2.3646 1.5263 1.6456 1.6209 class.Junior 2.3548 1.5316 1.7060 1.6276 class.Senior 2.7508 1.6760 1.6585 1.8056 college.Applied.Studies -2.1790c 1.0298 -1.0332 1.1037 college.Arts & Sciences 0.3227b 0.1131 0.5905a 0.1298 college.Business 0.1013 0.1176 0.3939b 0.1350 college.Communication & Information 0.0601 0.1333 0.3383c 0.1529 college.Criminology 0.0301 0.1466 0.2083 0.1682 college.Education -0.1111 0.1586 0.0170 0.1819 college.Engineering 0.4516a 0.1274 0.7308a 0.1463 college.Film.School -0.0267 0.2593 -0.1574 0.2978 college.Fine.Arts -0.2152 0.2556 -0.2062 0.2927 college.Human.Sciences 0.3389b 0.1254 0.6642a 0.1439 college.Music -0.2335 0.1483 -0.4992b 0.1703 college.Nursing 0.2451 0.1506 0.5535b 0.1727 college.Social.Sciences 0.2956 0.1255 0.5550a 0.1440 college.Social.Work 0.3005 0.2394c 0.4239 0.2748 college.Undergraduate.Studies 0.1287 0.1175 0.3610b 0.1348 MatriculationYearTer.20149 -0.1213 0.0425 -0.1020c 0.0487 age 0.0595c 0.0276 0.0515 0.0316 US citizen -0.0978 0.1184 -0.0327 0.1359 HS GPA -0.0619 0.0616b -0.1012 0.0707 ACT -0.0163c 0.0070 -0.0264a 0.0080 Transfer Or Exam Credit -0.0003 0.0019 -0.0007 0.0022 Race.White -0.1151 0.1275 -0.0031 0.1464 Race.Hispanic.Latino -0.0488 0.1322 0.0879d 0.1517 Race.Black.African.American 0.0074d 0.1484 0.0770 0.1704 Race.Asian 0.2748 0.1578 0.3960c 0.1812 Race.American.Indian.Alaska 0.1277 0.4212 0.1675 0.4837 Race.Native.Hawaiian.Oth.Pa 0.2161 0.4384 0.0228 0.5042 Race.Two.or.More.Races -0.0783 0.1503 0.0170 0.1725 Gender.Male 0.1198 0.0368 -0.0140 0.0422 EducationFather.College -0.2083 0.2665 -0.3705 0.3054 EducationFather.High.School -0.0887 0.2695 -0.3393 0.3089 EducationFather.Middle.School -0.5736 0.5743 -1.2376d 0.6588 EducationMother.College 0.1774 0.2449 0.1567 0.2807 EducationMother.High.School 0.0653 0.2549 0.0233 0.2922 EducationMother.Middle.School -0.0353 0.4912 -0.1828 0.5631 ParentIncome....1000 0.1116 0.2267 0.2785 0.2598 ParentIncome..1000..40000 -0.2009 0.2596b -0.0745 0.2975 ParentIncome..40000..75000 -0.0625 0.2719 0.1074 0.3116 ParentIncome..75000.100000 -0.3075 0.2864 -0.0456 0.3281 ParentIncome..100000. -0.1718 0.2570 -0.0019 0.2945 aSignificant at the 0.1% level bSignificant at the 1% level cSignificant at the 5% level dSignificant at the 10% level       Figure A3 Common support condition for frequency.     Figure A4 Common support condition for duration.     Table A6 Covariate Balance With and Without Conditioning on Treatment: Frequency Treatment: Duration Without Condition Condition Without Condition Condition Covariates Est Std. Error Est Std. Error Est Std. Error Est Std. Error   GPA 5.674a 0.727 0.327 0.781 7.768a 1.350 0.429 1.395   military -1.898 3.102 0.923 3.026 -2.804 5.793 2.214 5.666   athlete -15.467a 3.713 -0.152 3.721 -25.607a 6.935 -0.445 6.940   housing 1.262 1.299 -0.087 1.268 4.324d 2.423 0.579 2.378   CARE 11.556 22.913 -3.566 22.338 -17.780 42.785 -0.885 41.806   currentload 1.053a 0.275 0.284 0.271 0.887d 0.506 0.443 0.495   class.Freshman -0.397 1.098 -0.143 1.069 1.966 2.050 -0.482 2.007   class.Sophomore 0.649 1.144 0.075 1.115 -1.988 2.138 0.522 2.094   class.Junior -1.110 2.699 0.708 2.631 -0.456 5.008 0.236 4.892   class.Senior -5.369 19.845 -14.297 19.339 -20.618 37.056 -12.298 36.201   college.Arts...Sciences 4.909a 1.082 -0.303 1.096 8.608a 2.022 -0.461 2.049   college.Business -3.680b 1.391 0.252 1.373 -4.695d 2.598 -0.191 2.552   college.Communication...Infor -4.263d 2.365 0.465 2.319 -4.415 4.417 -0.438 4.321   college.Criminology -5.976c 2.991 0.432 2.936 -11.501c 5.570 1.372 5.493   college.Education -9.678b 3.462 1.356 3.429 -19.364b 6.464 1.681 6.436   college.Engineering 8.373a 1.952 -0.750 1.971 11.757b 3.647 -0.627 3.637   college.Film.School -6.618 7.031 -0.666 6.858 -25.075d 13.127 0.707 12.912   college.Fine.Arts -17.297b 6.628 0.784 6.536 -33.462b 12.376 0.447 12.255   college.Human.Sciences 3.219d 1.922 -0.502 1.884 12.929a 3.587 0.253 3.585   college.Music -11.371a 3.064 0.921 3.064 -33.435a 5.918 2.781 6.196   college.Nursing 0.952 3.158 0.182 3.077 6.378 5.896 -1.268 5.777   college.Social.Sciences 3.795c 1.934 -0.128 1.897 5.076 3.612 -0.439 3.543   college.Social.Work -1.329 6.373 -2.496 6.209 -6.246 11.900 -2.610 11.627   college.Undergraduate.Studies -3.782b 1.409 0.352 1.392 -6.205c 2.626 0.820 2.598   college.Visual.Arts..Theatre. -5.690d 3.246 -0.013 3.178 -20.066a 6.058 -0.168 6.035   MatriculationYearTer.20149 -1.807d 0.999 0.134 0.979 -1.994 1.867 0.121 1.828   Age 1.308c 0.644 0.053 0.631 1.649 1.201 -0.155 1.178   UScitizen -6.704c 3.414 2.121 3.361 -8.449 6.398 0.951 6.274   HSGPA 1.396 1.465 -0.311 1.430 -1.224 2.737 -0.355 2.674   ACT -0.183 0.182 -0.015 0.177 -1.161a 0.339 0.058 0.339   TransferOrExamCredit 0.004 0.030 -0.002 0.029 -0.057 0.055 0.002 0.054   Race.White -5.593a 1.067 -0.339 1.083 -9.155a 1.994 -0.598 2.016   Race.Hispanic.Latino 2.671c 1.303 0.303 1.276 5.693c 2.431 -0.392 2.402   Race.Black.African.American 4.739c 2.369 0.602 2.319 6.910 4.417 1.316 4.327   Race.Asian 16.431a 2.871 -0.979 2.975 27.114a 5.392 0.059 5.520   Race.American.Indian.Alaska 8.900 11.973 1.262 11.673 7.089 22.358 1.818 21.843   Race.Native.Hawaiian.Oth.Pa 8.698 12.556 -3.485 12.251 -6.315 23.448 -2.985 22.905   Race.Two.or.More.Races 0.508 2.494 0.468 2.430 0.277 4.658 1.287 4.550   Gender.M 3.468a 1.012 -0.257 1.008 -1.556 1.892 0.040 1.851   EducationFather.College -9.672a 2.170 -0.011 2.185 -16.042a 4.053 -0.649 4.066   EducationFather.High.School -5.033d 3.032 1.256 2.975 -11.238c 5.678 2.428 5.603   EducationFather.Middle.School -18.878 19.843 -5.654 19.347 -31.375 37.055 3.601 36.254   EducationMother.College -8.656a 2.136 -0.218 2.136 -14.566a 3.995 -0.715 3.990   EducationMother.High.School -7.534c 3.299 0.959 3.248 -14.423c 6.139 1.349 6.069   EducationMother.Middle.Scho -7.460 13.235 1.873 12.904 -11.995 26.211 7.913 25.629   ParentIncome....1000 0.682 5.473 -2.517 5.335 4.650 10.220 -2.945 9.992   ParentIncome..1000..40000 -8.851c 3.815 0.328 3.752 -14.909c 7.124 2.560 7.033   ParentIncome..40000..75000 -4.022 3.816 0.834 3.727 -7.920 7.157 0.953 7.011   ParentIncome..75000.100000 -10.584c 4.435 1.780 4.375 -17.638c 8.232 -0.373 8.106   ParentIncome..100000. -9.479a 2.702 -0.412 2.682 -16.324b 5.058 -1.122 5.023   aSignificant at the 0.1% level bSignificant at the 1% level cSignificant at the 5% level dSignificant at the 10% level     Table A7 Estimated Coefficients of Conditional Distribution of GPA Given Treatment and GPS Treatment: Frequency Treatment: Duration     Estimate Std. Error Estimate Std. Error Intercept 5.2350a 0.9633 Intercept 3.0600a 0.6626 Frequency 0.0127 0.0083 Duration 0.0148a 0.0028 Frequency^2 0.0000 0.0000 Duration^2 0.0000b 0.0000 GPS -53.1700c 24.3100 GPS 17.2900 12.1200 GPS^2 366.7000c 161.7000 GPS^2 -123.1000c 55.9100 Frequency*GPS -8.7860a 2.6010 Duration*GPS -4.1490 2.9280 aSignificant at the 0.1% level bSignificant at the 1% level cSignificant at the 5% level dSignificant at the 10% level